Beruflich Dokumente
Kultur Dokumente
paper industry
A production by COST E 36
Book title: Use of modeling and simulation in pulp and paper industry
2008
ISBN 978-91-977493-0-5
Editor: Erik Dahlquist
COST- the acronym for European COoperation in the field of Scientific and
Technical Research- is the oldest and widest European intergovernmental
network for cooperation in research. Established by the Ministerial
Conference in November 1971, COST is presently used by the scientific
communities of 35 European countries to cooperate in common research
projects supported by national funds.
The funds provided by COST - less than 1% of the total value of the projects
- support the COST cooperation networks (COST Actions) through which,
with EUR 30 million per year, more than 30.000 European scientists are
involved in research having a total value which exceeds EUR 2 billion per
year. This is the financial worth of the European added value which COST
achieves.
A bottom up approach (the initiative of launching a COST Action comes
from the European scientists themselves), la carte participation (only
countries interested in the Action participate), equality of access
(participation is open also to the scientific communities of countries not
belonging to the European Union) and flexible structure (easy
implementation and light management of the research initiatives) are the
main characteristics of COST.
As precursor of advanced multidisciplinary research COST has a very
important role for the realisation of the European Research Area (ERA)
anticipating and complementing the activities of the Framework
Programmes, constituting a bridge towards the scientific communities of
emerging countries, increasing the mobility of researchers across Europe
and fostering the establishment of Networks of Excellence in many
key scientific domains such as: Biomedicine and Molecular Biosciences;
Food and Agriculture; Forests, their Products and Services; Materials,
Physical and Nanosciences; Chemistry and Molecular Sciences and
Technologies; Earth System Science and Environmental Management;
Information and Communication Technologies; Transport and Urban
Development; Individuals, Societies, Cultures and Health. It covers basic
and more applied research and also addresses issues of pre-normative
nature or of societal importance.
Web: www.cost.esf.org
2.2.1.1. Physical models and their validation for pulp and paper applications. ..... 34
2.2.1.1.1 Abstract: .............................................................................................................................. 35
2.2.1.1.2 Introduction: ..................................................................................................................... 35
2.2.1.1.3 Simulator models ........................................................................................................... 36
2.2.1.1.3.1 Screens ............................................................................................................................. 36
2.2.1.1.3.2 Hydro cyclones, cleaners ........................................................................................ 39
2.2.1.1.4 Model validation and tuning with process data ............................................. 42
2.2.1.1.5 Simulators for other applications .......................................................................... 44
2.2.1.1.6 Conclusions ....................................................................................................................... 44
2.2.1.1.7 References: ........................................................................................................................ 45
2.2.2 Statistical models/data driven models ....................................................................................... 46
2.2.2.1 Finding dependencies and time lags between different signals ........................... 47
2.2.2.2 Correlation Measures .................................................................................................................. 47
2.2.2.3 Entropy and Mutual Information .......................................................................................... 47
2.2.2.4 The Mutual Information Rate ................................................................................................. 49
2.2.2.5 Test Results on Industrial Process Data ............................................................................ 50
2.2.2.6 Discussion and Practical Considerations .......................................................................... 51
2.2.2.7 References......................................................................................................................................... 51
2.2.3 Artificial neural networks (ANNs) ........................................................................................... 52
2.2.3.1 Types of ANNs ................................................................................................................................. 52
2.2.3.2 Way of working of ANNs ........................................................................................................... 54
2.2.3.3 Back propagation algorithm .................................................................................................... 54
2.2.3.4 Modifications of back propagation algorithm ................................................................ 58
2.2.3.5 Applications of ANNs in the pulp and paper industry ............................................... 59
2.2.3.6 Literature references .................................................................................................................. 60
2.2.4 Event driven models........................................................................................................................ 62
2.2.5 Hybrid models .................................................................................................................................... 62
Introduction .............................................................................................................. 74
2.3.2.3.1.2
2.3.2.3.1.3
Example ....................................................................................................................... 77
2.3.2.3.1.4
Conclusions................................................................................................................ 84
2.3.2.3.1.5
References .................................................................................................................. 84
9.1.3.1.3 2-D calculations with a simultaneous solver for both pressure flow
calculations and chemical reactions .................................................................................................141
9.1.3.2 Other applications.......................................................................................................................143
9.1.3.3 Modelling of delignification ...................................................................................................143
9.1.3.4 Conclusions ....................................................................................................................................147
9.1.3.5 References.......................................................................................................................................148
9.2 Economic Benefits of Advanced Digester Control. ..................................................................149
9.3 Application of softsensors for cooking ..........................................................................................155
9.3.1 Batch cooking ....................................................................................................................................155
9.3.2 Continuous cooking .......................................................................................................................156
9.4 Decision Support System for TMP Production..........................................................................157
9.4.1 Abstract ................................................................................................................................................157
9.4.2 Introduction .......................................................................................................................................157
9.4.3 Description of the Case.................................................................................................................157
9.4.4 Phases of the Optimization ........................................................................................................158
9.4.5 The Simulation Models.................................................................................................................159
9.4.6 Decision making and optimization ........................................................................................161
9.4.7Conclusions .........................................................................................................................................163
9.4.8 References ..........................................................................................................................................164
9.5 Optimisation of TMP production scheduling .............................................................................164
9.5.1 Abstract ................................................................................................................................................164
9.5.2 Introduction ............................................................................................................................................165
9.5.3 Process description ............................................................................................................................166
9.5.4 Application ...............................................................................................................................................167
9.5.4.1 Two Mill Model .............................................................................................................................168
9.5.4.2 Paper Machines Production Schedules and TMP Demands.................................169
9.5.4.3 Implementation............................................................................................................................169
9.5.4.4 Results...............................................................................................................................................169
9
10
11
14
PREFACE
The purpose of this book is to show how simulators can be utilized for many different
applications in Pulp and Paper industry. Theory is combined with many practical
applications. Both technical and economic benefits are covered. This publication is
supported by COST.
The intention is that many different cathegories can use the book. Process engineers can
learn about the actual methods, managers can find motivation for why the implementation
should be done, production managers can get ideas for process improvements and
automation engineers can gain information about a broad range of methods to use in the
future. For research and development staff the book will give an overview of the state of the
art both with respect to the methods as such, as well as with respect to implementations in
real life.
The contributors to the book are all members of an EU network under the COST action E36,
which has had the goal to show benefits with process simulation in Pulp and Paper industry.
The participants are from 14 EU-countries. The chairman of the action has been Johannes
Kappen from PTS in Munich, vice chairman prof Risto Ritala from Tampere Technical
university in Finland. The group leaders for the work groups have been Jussi Manninen
from VTT (off-line simulations), Espoo, Finland, Erik Dahlquist from Mlardalen University,
Vsters, Sweden (on-line simulations) and Carlos Negro University of Complutense,
Madrid (commercial softwares). Erik Dahlquist also has been the main editor of the book.
The book is divided into four major sections after the introduction. We will go through the
content briefly in each chapter.
Chapter 1 Introduction
Modelling and simulation
Here we want to introduce the basic methods used for mathematical modelling of industrial
processes. First we have the
Chapter 2 Modelling including pre-processing
where we go through how to handle the data before it is used to build or verify the models,
or later on used for predictions and control. In the next chapter
Chapter 3 Soft sensors
we show how a number of process measurements can be used together with lab data to
build new soft sensors. These are primarily for on-line use with respect to different quality
variables, that are difficult to measure directly due to a lack of good instruments, or due to
high price for the instruments.
In chapter four
Chapter 4 Transfer of process know how into models
15
the importance of understanding the processes either you want to make statistical or
physical models is discussed. It is sometimes stated that statistical models will fix that by
themselves, but that is normally not a good approach. The more information you have of the
process, the more efficient the work will be to get a good simulation model and a working
control system.
This takes us over to the next major part of the book,
Process control and decision making
This covers primarily how models can be used for in-line applications using different
methods. First we discuss
Chapter 5 Model based control (MPC Model predictive control,)
which is a strong tool for multivariable process control on-line. Theory as well as practical
examples is presented, although more experience is given in later chapters. Model based
control takes us over to the
Chapter 6 Production planning
which covers a longer time perspective, and how to schedule the production within a
certain time horizon. Methods for optimized production planning are presented. Decisions
support,
Chapter 7 Decision support
represents a number of methods helping operators as well as managers and process
engineers to make the right decision when different situations occur, and where the most
obvious action is not selfevident.
Now follows a larger number of real life experiences as
Applications and case studies
These cover different parts of the mills where we start with
Chapter 8 Applications in pulp mills (kraft, CTMP, recovery ) followed by
Chapter 9 Applications in paper mills (incl deinking) and
Chapter 10 Applications in utilities
The last major section more concerns tools, simulator environements and commercial
softwares.
Commercial HW/SW tools
In the chapter about design tools a number of tools have been tested for the same
applications, and comments on pros and cons are given.
Chapter 11 Design tools (off-line, steady state and dynamic)
16
18
From various sources we can learn that a rather large potential of operational income is
hidden in the way operations are performed today. This potential has to be unleashed by a
mill in order to compete successfully. Cost reduction alone is, of course, one step in the right
19
direction but will not enable a mill to win the war. Operations need to get smarter and
more intelligent. More needs to be done with less is a hollow statement, but describes the
way visionary managers think.
Changing papermaking operations is more and more the way to improve business results.
The proven concept of (strategic) investments is another, but no longer can be seen as the
only king of the castle. Globalization and mergers have made papermaking an economical
practice in which the companies want to increase value for the shareholders only.
Investments therefore have to follow very strict rules, generating cash within very short
timeframes. New ways of accomplishing things need to be explored and modeling and
simulation are surely among the possible candidates to become very important in the next
decades.
From the past
Operations have become complex to such an extent that the operators and engineers are not
able to follow the process anymore. The past decades have brought us more and more unit
operations in-line: from dilution head boxes to multiple wire sections, press section and
dryer section, size presses, infrared dryers, in-line coaters, calendars, and what-have-you.
All these are part of the same paper machine which used to consist of a simple head box,
forming section, press- and dryer section. This puts a lot of pressure on the material we try
to produce in the same process, namely paper. Specialty chemicals are added to enhance
properties during the production process, aside from all the chemicals that are added to
improve functional properties.
Currently we are at a point that adding another chemical gives unknown side effects
without us even knowing it. At the same time we see a downward trend in the quality of the
raw materials (stock), are it virgin fiber due to cost reasons, or be it recycled fibers due to
recycling itself.
Apart from the above, the knowledge level of operators and staff is decreasing every year.
The papermaking industry is not very attractive to newly skilled personnel and year-byyear we see elderly experienced people leave. This is a threat especially for existing (nonGreenfield) mills.
Towards the future
The above-mentioned trends force us to shift our efforts from ad-hoc solutions to more
structural changes in the way we work, or perform the operations. Implementing models in
the operations or in decision-making is one of the tools that enable us to do this. One
example can be a model of a certain quality parameter of the end-product, normally
measured hours or even days after the paper is being produced. At the exact time the paper
is produced, this model predicts this quality and enables us to adjust the process, if
necessary. This way, a lot of time and effort (money) is saved, improving the operational
result.
To be successful these models have to meet certain criteria, which actually do not comprise
of the type of model, but of the online applicability of the result. We have seen examples of
all types of successful models, from very simple univariate linear relations to very
20
complicated neural networks. The thing they all have in common is what can be defined as
completeness.
Completeness
Online models are pieces of software, running on a computer, generating output by
processing one or several inputs. First of all the inputs have to come from the running
process, preferably from the process information (PI) system in use. At the same time, the
output(s) should be redirected to the same PI-system, making the results fully transparent
to the user. In actual online application, these will most probably be the operators,
controlling the process. Any deviation from these basic rules will generate possible
impairment of the applicability, due to lack of completeness.
Examples: In an installation for a new green field mill in Malaysia MNI (Malaysian
Newsprint Industries) concluded that they got a 20% faster start up due to operator
training with simulator; compared to what they would expect from start ups with
experienced operators in Australia, although no one of the operators had any
previous experience from pulp and paper operations.
At one mill in Indonesia (APRIL group, RAPP) sheet brakes could be reduced by 29
% after using advanced tools for process diagnostics and loop analysis. The value of
this was 7 MUSD/y.
Within the Dutch Paper & Board industry a survey has been conducted under the 27 paper
& board mills and different suppliers (knowledge institutes / suppliers & consultants).
21
When the first contact person asked about models is the process engineer. Almost all mills
have been using models and simulations, by them selfs or by consultants for two main
areas.
-
Logistic challenges
These two areas have in common that the process is visible and can be made
understandable, although the complete process is still complex. Because setting up a good
working model/simulation an accurate plan of the process is necessary. To get this up-todate accurate process scheme the mill has to take a detailed look at the process. This work
already gives a lot of insight in the process at work, which also improves the understanding
of the results coming from the model and the simulation. When looking at other areas for
simulations and modelling like optimizing press section and dryers section. Models are
looked at with caution. First the processes inside these sections are complex and difficult to
master. Second these models and simulations arent always in line with practical
experience.
One mill had a visit of a supplier of press felts. The supplier used a simulation to show the
effectiveness of their felts. The mill asked to put in a higher moisture content of the felts
before entering the press nip. The simulation showed lower press efficiency, while practical
experience showed that the press efficiency was improved. These kinds of cases make
process engineers weary of applying advanced models and simulations.
The last but not the least remark made by process engineers is that a lot of models are much
too complex to be issued out in the mill. A simulation model may provide the operator /
manager / engineer with a traffic light on the performance of the machine. The
model/simulation gives a complete controlled traffic junction with all inputs / calculations
and outputs shown. Next to this the program asks continuously for different settings.
It is off course possible for mills to buy their own model, analysing and simulation
program(s), like KCL-Wedge, WinGems, Aspen and Mathlab. However these programs can
be used for a lot of modelling and simulation situations. That makes the program very
diverse but also because of a lot of options difficult to master. After mastering the program,
putting in the process in the right way, with the right parameters is another time consuming
activity. Most mills dont have the manpower to put in these kinds of projects. You have to
take into account also the possibility that the person in question finds another job after a lot
of special trainings.
Therefore most of these challenges are performed by consultants or knowledge institutes.
The bottleneck in this situation is that these parties know the modelling and simulation
program, but not the mill specific process at every detail. This often leads to modelling and
simulation projects taking more time than estimated.
When a model, analysis or simulation of a certain process has been set up in such a way that
it makes the right prediction, every change in the process has to be implemented also in the
model, analysis and simulation. This makes these programs very time consumable in order
to keep them accurate.
22
Often the analysis, model or simulation is made on a single data block. When loading the
solution on a new data block often the solution doesnt have a fit. Making the solution fit to
more data blocks and the process can only be done with a lot of manual analysis of the
solution and the physical process.
In general results from modelling and simulations were satisfactory and very useful.
Especially for processes which are hard to make visible or when theres too much risk for
empirical testing.
All interviewed people indicated they would like to make more use of models and
simulations. When asked for the criteria to use them, the following criteria where
mentioned:
-
Some models are not yet available (on-line stiffness prediction / mottle / internal
bond)
Models are available but do not yet seem to be in-line with empirical experience
Models are very complex in control, an easier understandable user interface would
be very welcome
o
A layered user interface for the different functions in a mill working with the
model/simulation.
outcome. However this screen full of data, pointers and flags is useful for the process
engineer.
The dilemma has it origin where the scientist, engineer makes the analysis, model or
simulation in very great detail in order to put all working variables and parameters in the
model and to find their mutual interaction. This makes the model too complex for the
manager to get a good insight in the risk and to make fast decisions.
These world will can come together if a manager had to time to learn the program and make
his/her own presentation layer or the scientist/engineer could produce this presentation
layer from the managers point of view. Another more practical solution is to equip the
model with different user interfaces, one for each level of operating the model.
However when the analysis, model or simulation has been set up in the right way and is
well presented the manager can benefit from it. Then it will be very easy for him what the
risk and benefit is on certain changes, investments or trials in the process. Before we get
there the mentioned programs still need to undergo development, since nowadays they are
still a helpful tool for specialists, which require not only adaptation by every change in the
physical process, but also a (self)critical well informed scientist/engineer who knows the
process at hand.
The overall conclusion is that mills see very much potential in modelling and simulation but
therefore the following criteria have to be met:
-
New models available for end product qualities which are now undetermined
Logical user interfaces which are understandable for different levels of operating
Run ability
24
Goals should be ambitious, but as soon as belief in success is lost, it is better to redefine
either the application-project or the goals. Commitment of all parties involved is crucial.
Top-management sets the vision and should live the message; otherwise lower levels of
the organization will not follow. Middle management must believe in the results, as they are
the key players in empowering the end-users. Application-projects in the next decade will
most certainly evolve around reshaping the way we work today and around evangelizing
the message of simulation results. Very often the end-users (operators) tend to believe
actual sensors together with their gut feeling, and will expose a deep-rooted disbelief to
calculated results, just because these results are calculated and not real or because the
calculated results sometimes do not match with their past experience.
Use the right tool for the job.
Part of the result comes from the tools. This is true in normal life, but is vital in application
of simulation results or modeling-projects, especially in the paper industry. Due to the large
number of modeling methods it is easy to pick the wrong one and make life more difficult
than it should be. The key to success is a combination of knowledge of available modeling
methods, tools in that area, and process-knowledge. Thorough knowledge about the process
being studied enables us to pick the suitable tools, improving the chances to success.
Innovate, dont imitate.
All efforts done in implementing simulations should generate outcome in terms of
operational result, otherwise no value is added and the application (actual software
program) is prone to be not used at all or the application-project itself will not succeed.
Therefore, implementing simulations should generate innovative results as opposed to
more of the same as seen in the past years.
Better is the enemy of good.
Applicable simulations should be implemented. There should be no time-lag between
development and application, due to polishing or trying to make it better. Very often the
end-users are in need of a simple traffic light with a red, yellow and green light, but a totally
controlled high-fashion junction is presented to them instead. A simple traffic light will do!
Simulations very often are done by highly educated staff with their own believes of what is
good and what is not. The truth is that the end-users will know best whats good for them or
whats fit for purpose.
It is better to shut-up than to lie.
Faith is easier lost than it is gained. New information will be trusted only after a rather long
time in which the information has been proven. False information in that period will reset
the faith-counter, as a matter of speaking. To be successful, it is very important for
simulation results to have some indication of the reliability, or trustfulness. This should be
an integral part of the modeling efforts or implementation project and will prove to be one
of the key factors for success.
Carrot or stick?
25
Involvement of the end-users in the development of any real-life object is crucial. Too often
we see things being marketed which obviously were developed without the end-users in
sight. Of course, money can be earned by selling things nobody wants, but it takes more to
be really successful.
Applying simulation or modeling results calls for a sturdy approach in which the end-users
play an important role. In case the end-users are the operators of a paper machine, too often
developers think that they will not like the change or are unable to understand the physics
involved, or are not willing to co-operate. This is not true in most cases! Operators like new
things, they like changes in the process because it makes or keeps things exciting. What they
dont like are sudden changes or things changing without being informed or without them
being involved in the process of change.
In the past decades too much new technology has entered the field of operations in a way
that doesnt fit the above description. This has in a way damaged the belief operators have
in management. Implementing simulation results or modeling efforts therefore has to be
done very carefully with a lot of operator involvement; otherwise what really is a carrot will
be seen as a stick.
26
If the models built should be physical, grey or black box model has to be discussed. If
the process is well known, a physical or grey box model may be the best. If the
process is less well understood but good measurement data from the process are
available, a black box model may be a good choice.
Pure white (physical) models may have to strong limitations on computer power
today and thus grey or black models should be used today. Still, you should aim at
using physical models as far as possible, but the limitations have to be considered
and will lead to simplifications, and lead to grey models. The best case is if we can
fit a physical model with real measurements of physical properties. We can also use
experience like knowledge about the heat transfer as a function of different
situations. This can lead to a white box model. Grey box models is when we tune
the physical or white model with process data.
Different quality of data input is needed for different models, and different types of
models ( linear, dynamics, steady state etc). The data can be discrete or continuous,
and it can be on-line measurements as well as lab measurements taken
intermittently.
27
When you start considering using simulation you first have to consider what the purpose of
the simulation is. Is it for optimizing a design, study performance of a process during
different operational situations or to use on-line for diagnostics or control purpose? The
purpose of the simulator will determine how detailed it should be modeled.
If the purpose is to simulate a complete process from a design point of view it may be
enough to have a steady state model that can be use as a support for optimization
calculations. If you also want to test operational aspects of the process a dynamic model
may be needed, where you also include controls like PID controls, inter-locking etc. Perhaps
a complete DCS code is to be tested towards the model? Then we have to include all aspects
of response to the DCS system from the simulator, which may include also a number of
digital signals.
If the purpose is to optimize the design of specific equipment like a gas turbine, a screen, a
flash tank, a reactor, a digester or a boiler, then we may need a very detailed model
describing all physical means of the process including all dynamic aspects. On the other
hand we can then skip the control aspects and perhaps also the interaction with other
equipment. Instead it may be necessary to model the equipment with a complex geometry
in a reasonably accurate way.
If the purpose is to use the model for on-line purpose the major issue may be the calculation
speed. You have to perform the full calculation perhaps every second, every minute or every
hour, depending what the model shall be used for.
This shows that it is very important to discuss the usage of the model before starting the
development of the model. Otherwise we may end up with a model that is too expensive or
not good enough to solve the problem you have.
Example: Retention, formation and wet-end parameters as a function of primary inputs
Design of a new process (need many components and use of a commercial library
may be useful at least as a starter. May include some dynamics and simplified
controls with controls of specified flows but without PID controls)
28
information is needed as input and output in detail). This may include both physical
models and statistical, data-driven models.
-
Use of the model for diagnostic purposes (the model can be detailed or overview,
physical or data-driven. Need very thorough consideration how to model different
tasks.).
Simulator for operator training (need a simulator interacting with the DCS system
including all process displays. Need full dynamics in the simulator but not
necessarily very detailed physical modeling of the different equipment).
The next step is to identify how we shall model specific equipment and how this shall
interact with the outside world that is what input and output are needed to the model.
This may include vectors with flows, concentrations, temperatures and other physical data,
but also other type of information like dimensions, parameter values like valve openings,
speed of a motor etc. When we include dynamics we also need to handle buffer volumes and
similar from one time step to the other. All these values may be included as a data set or just
interact on-line with a data base when needed, which may be only as initialization when the
simulator starts for parameters as dimensions and starting values.
For more detailed physical models we may also need to include boundary conditions.
If we have a running process it may be possible to build a model from process data. We then
get a statistical model with just a general structure like a polynomial or a physical model
with tuned parameters.
For a new process or if principal aspects need to be modeled we can not use process data.
Instead we try to include all important physical aspects including chemical reactions like
combustion processes, separation mechanisms and similar. We may also need to include
complex geometries in the equipment and complex fluid dynamics.
In this case it is important to start from the complex equations and then try to simplify in a
logical way. If we know what is happening from earlier experience we can simplify special
function. We can also make assumptions on what are realistic values for the operation, like
29
reasonable flow ranges, temperatures of operation, gradients in the process like how fast a
change can be allowed to be etc.
If we want to model a combustion process in a boiler for instance, you can chose between
following a specific particle and what happen to this or look at flows and temperatures in
specific volume elements in the combustion chamber, if you use one of the commercial CFD
programs like Fluent or CFX. Then you have to consider what you really want to study. Is
full geometrics needed? In this case you need to have a support tool for achieving the
geometry. On the other hand, we then also may need a very detailed grid, to really get the
influence of the full geometry. Just as an example: a mixer between hot and cold water was
to be modeled. When the grid had 300 000 volume elements the temperature difference just
after the mixer was only half of the one when 600 000 volume elements were used. The
conclusions from the experiment in the first case would then give the result that the
temperature difference would not cause technical problems, while when the temperature
difference was twice as high the mixer had to be rebuilt. As this was in a nuclear power
plant and exhaust in the material could be dangerous, it shows that the choice of the grid
was very important. If we had been looking at an on-line application, we would probably
not have been able to use more than perhaps 100o volume element from a calculation point
of view, as the calculation will take to long otherwise. This is to show how important it is to
identify what the purpose with the simulation is.
If the simulation result is to be used for important decisions it is also important to have a
possibility to tests towards real experiments. This may be directly in a pilot plant or a full
scale plant, but may also include testing the model towards similar applications, or compare
to how other researchers have verified similar models towards experiments in literature.
If we go back to the problem with the combustor, and if we can follow both particles
(Lagrangian) and the volume elements (Eulerian) it may be possible to do some kind of
interaction between these two approaches. We can first make a model determining the
temperature and flow rates and directions in the boiler, and then follow a particle moving in
this environment. For each time step we move the particle in the direction and with a
reasonable distance from where it is in the combustor for each time step. Then this
environment in the volume element where it is with oxygen, temperature etc will be used as
boundary conditions when calculating what happens with the particle.
When we in this way model particles more in detail we may find that the Eulerian solution
is not giving correct values, as it is not handling the particle interactions in a correct way,
and we may have to modify the equations for the volume elements in some way and do a
new iteration. Etc.
If we on the other hand only want to study the control of the boiler, we can have a very
simple split of the boiler into perhaps five volume elements, where we take overall
reactions only. With a certain amount of fuel and air we will get a certain reaction rate and a
certain composition of the exhaust gases from each volume element. This can be used for
adjustment of the feed of fuel or air, to see the dynamics of the whole boiler without
knowing the details of what is happening. On the other hand we can get very fast
calculations like every second a new value.
30
For a series of discrete events like cutting paper and forming rolls, transport of the paper
rolls to local storage and finally reaching a converter or printing mill, it is better to use
discrete events to follow the fibers. The principles are mainly the same, but the models
much simpler.
When we have selected what type of model to use it is time to put it into an environment.
We then can select a commercial environment with existing solvers, libraries etc, or we
make our own model in e.g. FORTRAN, C++ or some high level language like Modelica or
gPROMS. At least the FORTRAN and C-codes can easily be accessed from the commercial
simulator environments if we just have done a good job identifying what signals to send in
and out of the model.
It is useful to have a good interface for input and output, and the graphical interface is
especially important when other people are to use the model. An intuitive design of this is
the best. We also have to consider if the model shall be used only on one single computer, in
a network or if it even should be accessed from the internet. The signal capacity between
the model and e.g. a DCS system may be a limiting factor, especially for a larger model like
when a complete process is to be modeled like in a training simulator system or for on-line
applications.
In another chapter different simulator environments are discussed.
As discussed above it is important to verify that the model is giving realistic results. This
can be achieved either directly in the process by checking both on-line measurements and
by examining samples from the process at lab. This is for the complete process or for
specific equipment in the process.
In some cases it may be possible to build an experimental set up as a pilot plant specially
made for testing purpose. This is common for suppliers of special equipment to use for
development purpose. The problem is then that only some specific experiments can be
performed, but at least we can verify some specific conditions. From this we can know that
the model is giving reasonable results and if the model is physical- first principles- we can
know that the results also for similar conditions will be reasonable.
If we have a statistical, data driven model we know it is giving reasonable values from an
historical perspective, but have to compare the predictions for new values and compare to
lab measurements also for the future. When we see the model is deviating too much from
measurements in lab afterwards we can update the model with new data. This type of data-
31
driven models can also be used together with physical models in a larger model for a
complete process.
When we verify a model it is favorable if we cam measure important variables and results
during tests using a factorial design of the experiments. Then we know that we cover the
whole operational range in a good way, and can find also interactions between the
different variables.
In many cases it is very difficult to measure all variable we want to follow. If we have a
boiler for instance, it is not enough to measure the temperature at a few spots. The
temperature may vary 3-400 oC just from the wall to a position 3 dm into the boiler, and the
flow through the exhaust gas channel may have a temperature profile, as well as the solids
may be unevenly distributed over the area. It is thus very difficult to verify the performance
without extensive and well designed sampling at many positions.
On the other hand we can make use of overall mass and energy balances and to verify the
complete balance also for the model. We can then make changes in the model that can give
directions for how sampling should be performed to study selected issues and problem.
Another aspect is to test the model code so that it is stable under all conditions. This could
be seen as part of the debugging, but in reality it is very important to find the limits for
where the model can be used. Some models may only be useful up to a specific limit with
respect to concentration, temperature of flow rate, and then this should be noted.
2.1.6 INITIALIZATION
When we start the simulation we need to have initial values. These should be values at
steady state conditions. It may sound simple to achieve this, but in reality tuning of a
complex dynamic model for a complete plant is non-trivial. Usually the tuning to find a
steady state condition is the most time consuming task in the whole simulator project.
If it is possible to have the same model in two versions, one steady state and one dynamic,
the steady state version can be used to produce values for the dynamic model.
If we have a simple model it may be easy to find all variable to initialize, while if we have a
DCS system as well, all values may not even be measurable, as all values are not stored in
the DCS system. Then we may need to introduce dummy values, just to have some
reasonable value as a starter.
When we have the system up and running we need to have some way of doing e.g. dynamic
upsets or if the simulator is used for optimization interaction with other software like
solvers. Upsets can be ramps, sudden increase or decrease in specific values etc. It is
important to consider what situations you really want to test with the simulator at an early
32
stage. Otherwise this may be very tricky as you have to rewrite the code to fulfill demands
you had not though of earlier.
If the simulator is used for on-line purpose it is important to have a good interaction with
the DCS system, where filtering of signals as well as data reconciliation is important for the
input data. It may also be interesting to update the model with new data on-line. Still, this
really needs a good control of the input data.
It is also important that the operators and engineers get some time to train
at the simulator. This has to be planned in a robust way in advance, and may
need extra staff for some time period.
be valid also outside the area where we have measured data earlier. Physical models are
easy to use if good physical principles are known. Still we need values of parameters that
have to be taken from general knowledge or from process data. An advantage with physical
models is that you can use them to evaluate new processes with similar components, which
is normally more difficult for black box models. Still, you may lose the advantage of a
physical model if the simplification is driven to far. On the other hand it is easier to include
dynamics in a physical model normally.
Physical models can be as Finite element models or as Difference models. The finite element
models are normally used for stress analysis when designing equipment that need certain
mechanical properties like bending strength and similar. This type of models is normally
not used for process simulation purposes, and will therefore not be discussed further here.
Difference models are the type of models normally used to for process simulation. They can
be steady state or dynamic, depending on the purpose of the simulation. If the major
purpose is process design, we can start with a steady state model. When control aspects are
included normally we need a dynamic process model that can show how different process
changes affect the process performance with respect to different variables. Here we
normally try to identify the basic physical principles behind a certain phenomena.
If we talk about fluid flows, the pressure-flow relation is modeled, with pressure drops etc.
If we want to model chemical reactions, kinetics and equilibriums of different reactions are
included. Etc. The physical principles normally are formulated as a set of DifferentialAlgebraic Equations, DAE. These are then solved by using different solvers. Examples of
simultaneous solvers are e.g. Modellica and gPROMS. Earlier simulators usually had
sequential solvers, where reactions in one block were first calculated, and the output sent to
the next block in a predefined order. A separate pressure-flow network calculator gave the
flows in the network as a function of pressure drops and sources. The two solvers were then
used iteratively. This technique still is used in many applications, as it has also advantages
when a DCS-system is included in the system. With this approach it is normally easy to keep
track of the time steps, while the simultaneous solvers normally have different time steps
depending on the degree of activity at a certain time. During a fast grade change for
instance, the time step becomes much shorter, to get converging solutions.
The physical models make use of Navier Stokes equations, Energy equations and continuity
equations. Depending on the application these can be simplified. Finally key-parameters in
the equations are tuned with process data, to get the real quantitative effects. We then form
so called grey models from the principally white models, or pure physical models. As
long as the computer power is not without limits, pure physical models will not be possible,
but more different levels of grayness.
2.2.1.1. PHYSICAL MODELS AND THEIR VALIDATION FOR PULP AND PAPER
APPLICATIONS.
By Erik Dahlquist, Mlardalen University, Vasteras, Sweden
34
2.2.1.1.1 ABSTRACT:
In pulp and paper industry there are a huge number of different types of process
equipments. There are digesters, screens, filters, hydro cyclones, presses, dryers, boilers etc.
In many cases the equipment suppliers want to consider single equipments as unique, and
thus a special model is needed. This can give hundreds or thousands of different models to
keep updated in a simulation package, and when it comes to model validation and testing, it
becomes almost impossible to handle in reality.
If we instead try to identify the basic physical principles of each unit, we can start from that
point and then just add on special extra features. E.g. screens, filters and presses all have
similar basic principles. In this way it is possible to reduce the number of modules for a
complete integrated mill to some 20 models. These are then tuned with existing data from
literature on real performance, and configured with rough geometric data. For a screen for
instance there is normally data available on separation efficiency when a certain mixture
has been operated under certain conditions, like flow rate, geometric dimensions of the
screen with the screen plate etc. Still, seldom fiber size distribution is included, as this has
not been measured. The same is normally the case for e.g. cleaners and other types of
centrifugal separation devices.
We start from first principle models and then tune these for different operational
conditions, where once the size distribution was measured and some variables varied. In
another case study other variables or conditions were investigated. A generalization of the
model can be done by combining all this information, covering to at least some extent all the
different operational conditions, and all fiber sizes and concentration ranges. By then only
fitting the model with the existing simple mass balance data for a specific equipment, you
can get a reasonably accurate model for all kind of operations for this and similar types of
equipments. In this paper a description is made of how a number of different equipments
have been modeled in this general way. Tuning and the result of model validation for
different operational modes are also shown.
2.2.1.1.2 INTRODUCTION:
The reasons for using a dynamic simulator system may be many, but mainly fall into seven
categories of use:
1) To train operators before start up of a new mill or to introduce new employees to the
process before starting operating the real plant. [Ryan et al 2002]
2) To use the dynamic simulator for optimization of the process, in the design phase of a
rebuild or expansion of an existing mill, or for a completely new green field mill.
3) To test the DCS functionality together with the process before start up of the real plant.
4) To optimize an existing process line, by testing different ways of operation for process
improvements.
35
5) On-line prediction and control of a process line or part of a process line [Persson et al
2003]
6) Use in combination with an optimization algorithm for production planning or on-line
optimization and control [Dhak et al 2004]
7) For diagnostics purposes [Karlsson et al 2003]
8) For decision support [Bell et al 2004]
It should also be noticed, that a simulator system can be anything from a small test model of
a specific equipment, where the engineering and programming effort can be a couple of
hours to get it into operation, to huge systems with thousands or even tens of thousands of
DCS signals connected to a model of a whole factory. Here perhaps 10.000 engineering
hours or more are needed for the project. Therefore you have to be sure to understand what
you are really out for, before starting to discuss costs and time schedules for a simulator
project!
To build the simulator, we need a model for every single equipment in the plant. Some of
these models can be very simple while others are very complex. In most cases we use a
physical model as the basis and tune this with process data. This gives us a reasonably good
model over a large operational area, and it does not collapse even if we go significantly out
of the normal operational area, which can be the case for a pure statistical model.
In reality we do not principally have that many different physical mechanisms in the major
equipment. Most common is filtration or screening, where the mechanism is mechanical
removal of particles on a mesh, wire or porous media.
2.2.1.1.3.1 SCREENS
The basic separation is done where fibers or particles are separated depending on mainly
the ratio between the particle size and the pores they have to pass through. In filtering
almost all fibers are separated, and then primarily the water is passing through the pores
with a flow rate depending on the driving pressure as well as the pore size.
Often there is a concentration gradient, and thus it would be good to have the filter as in a
thin vessel, and filtration/screening take place along the surface, with an ever increasing
feed concentration.
Conc in
Conc out
36
When we configure a screen we first select holes or slots. The dimensions of these, the total
hole-area/m2 and the total screen area are inputs, and gives the total nominal hole area.
The ratio between the active hole-area and the nominal is giving the average pore size and
the pressure drop over the screen.
Active hole area=total nominal hole area*rpm_par*clog_par*conc_par
where
rpm_par=rpm_COF(6)+ ((1-COF(6)) * rpm/rpm_max) ; COF(6) = 0.3 as default.
This gives a realistic impact of the rotor for an average screen, reducing the separation
efficiency from 90 to 71 % by increasing the rotor speed by 30 %.These are impact values
reported from experiments for a typical screen. With COF(6) = 0.2, the impact will be going
from 90 to 79 % separation efficiency, which is a bit more conservative.
Clog_par= short_clog_COF(3)/(short_clog_time+short-clog_COF(3)* (long_clog_COF(4)/
(long_clog_time+ long_clog_COF(4))
conc_par= 1- ( concentration in reject/COF(5))
where COF(3)=0.06 as default. The COF(3) is chosen so, that the maximum value of the
reject before clogging the screen is used as COF(3). This may be 0.06 for a typical screen,
giving the right effect on e.g. a screen going from 0.5 to 1.5 %,with an increase in separation
efficiency from 58 to 81 %. COF(3)= 0.15 is the very maximum value of any screen. COF(4)
and COF(5) are 1.0 as default, but can be calculated from experimental data.
Area_par= active_hole_area/ total_nominal_hole_area
Dh = hole diameter or slot width in mm.
Fiber/ particle: lengths ,diameters and heights are also given in mm; virtual radius is also in
mm.
The separation efficiency (SepEff) for each particle size is calculated principally as the ratio
between the weighted fiber size and the hole/slot diameter, compensated for the clogging of
the pores by multiplying the hole diameter with the Area_par. To avoid division with 0, the
hole diameter is added to 1.0, and the COF(7) is used for the tuning to different screen types
SepEff = (fiber_COF(20)*length + 10*fiber_COF(20)*(width + height))
SepEff = (virtual radius/accept flow rate)^COF(21)* SepEff* COF(7) /( (1 + Dh* Area_par )
For slots Dh is calculated as Dh= slot width * 7.0. This has been seen as realistic from
experimental data.
COF(7) has to be calculated (see below) while COF(20) =0. 0350 and COF(21) = 0.08
as default values. Default value for COF(7) = 3.44 for flow rates 10-150 l/m2.s.
The mass balance between feed, reject and accept now is calculated as:
38
Mass_flow_reject =
SepEff * kg/h_each_fiber_fraction_in_feed*(m3/h_reject_flow/m3/h_feed_flow )^COF( 22)
(kg/h).
Mass_flow_accept= Mass_flow_feed - Mass_flow_reject
The concentration of each fiber fraction in the accept is the
Concentration_accept= Mass_flow_accept/ (m3_accept_flow_per_h).
The second most common type of equipment is the Hydro cyclone, or cleaner. This also
includes a number of similar equipment like the deculator.
Cleaners are looked upon as a vessel that is either full (= separation working), or as not full
(= separation is not working). Calculation of the liquid level in the cyclone or the common
vessel for several cyclones is first done, and if positive the separation is calculated according
to the following procedure:
Principally the deviation of particles from the stream lines during the rotational flow will be
related to the volume of the particle divided by the friction of the particle surface relative
the water, the density difference between particles and water, the rotational velocity, the
cyclone diameter ( giving the rotational velocity-higher for small diameter cyclones) and the
viscosity of the water.
First we calculate a shape factor = (1+ fiber diameter)/ (1+ fiber length).Both lengths in
mm. This compensates for the fact that an elliptical particle moves in a different way than a
spherical particle.
We then calculate an adjusted particle radius : First we calculate the volume of the
particle. For rectangular pieces Volume V = H*W*L (Height*Width*Length). For fibers the
volume is calculated from V= *R^2 * L. Hereafter the radius R for the sphere with the same
volume is calculated as
R = ( V*3/ 4) ^1/3.Length in meter.
The cyclone volume is calculated from the geometric inputs, to give the residence time =
volume/feed_flow_rate.
An adjusted cyclone radius is calculated as the average from swirl zone and bottom of
cone.
The basic equation for gravitational and centrifugal separation matches lifting forces with
buoyancy forces + forces due to liquid motion: (4/3)*R^3 *s*g=(4/3)**R^3 *liq*g+
39
6***v(d)*R [Bird et al 2002] . Solving for deviation velocity due to gravitational forces
becomes
v(g)= (2/9)*R^2 *(s -liq)*g/
while the corresponding velocity for centrifugal forces becomes principally for a sphere:
v(c)= (2/9)*R^2 *(s -liq)*v(r)^2/(r*) .
The deviation v(c) is in m/s from the stream lines in radial direction due to the liquid
turning around in the cyclone ( radius r) with velocity v(r). It is calculated in our algorithm
using an adjusted radius to the sphere for other particle shapes like fibers, and with a
correction for the larger drag forces due to long, thin fibers compared to spheres, by
multiplying with the shape factor:
v(d)= COF(11)* (adjusted particle radius)^2 *Shapefactor* (density difference waterparticle)* [v(r)^2/r]*[1/ viscosity]
v(r) = Qin / Area of inlet pipe to cyclone (m/s)
= viscosity, 10^-3 Ns/m2 for water.
Principally we can also include the effect of higher consistency and temperature effects on
separation in the viscosity term. This means that corr=*0.02*(conc/2.0)^-2.2 for zero to
3 % consistency. The temperature effect on the viscosity is directly related by T1 =
1.002*10^-3 * (T1/20)^-0.737, where the temperature is given in oC, and compared to the
viscosity at 20 oC, where it is 1.002 * 10^-3 .
The total distance for the particles will be given by multiplying with the residence time in
the cyclone, and will depend on the liquid flow as well as the volume of the cyclone.
The shape factor takes into account freeness (surface roughness) as well as the shortest
particle diameter. High surface area and long fibers will go preferably in the top or centre,
compared to spheres and short fibers, assuming the same density. High density particles
will go towards the wall, and downwards.
The absolute separation will also depend on the split rate between flow upwards (
Qupaccept,center) resp downwards( Q down,wall,reject).If we assume 50 % volumetric
flow in both, the separation will be 50 % of the fibers in each stream, if the density is the
same of the fibers as for water, for average sized and shaped particles. To give the mass
separation we calculate the part of the incoming (inject) mass flow that goes to the
reject(wall) as Minj*Qrej/Qinj.
If the density is higher for particles, the relative distance compared to the distance of the
radius of the cyclone, will give the extra separation efficiency of particles of a specific size
towards the bottom compared to the top outlet. Where the flow rate is very low, also the
gravity is considered, but then in relation to the liquid level in the vessel. This is giving the
gravimetric separation efficiency as:
40
g= COF(11)* (adjusted particle radius)^2 *Shapefactor* (density difference waterparticle)*g*(1/ viscosity)*(residence time/liquid level)
The separation factor for centrifugal forces will be calculated as :
c = v(c)* residence time in cyclone/ cyclone radius ( upper part)
where v(c) = radial velocity and the particle residence time in the cyclone is Qin/volume of
cyclone.
The mass flow (M(I)) in the reject for each particle fraction (I) is calculated by:
Mrej(I) = Minj(I) * ( (Qrej/Qinj) +c + g) for the reject and
Macc(I) = Minj(I) - Mrej(I) for the accept.
The mass separation efficiency of the cyclone then is s = Mrej/Minj= (Qrej/Qinj)+ c + g
The concentration concerning particles of a certain size/shape going to the wall/bottom
will be calculated according to:
Conc rej,bottom(I)= Mrej(I)/ Q bottom
For the top or centre we will get correspondingly:
Conc acc,top(I) = Macc(I)/ Q top
To get the negative effect of fast increases or decrease of the incoming flow Qinj on the
separation, this is decreased according to adding a turbulence effect factor
(c+g) (t) = (c+g) (t) - v/v(t)
The mass balance is calculated, to give the concentration of fibers for each fraction (I) , in
the top respectively bottom.
The pressure flow network makes use of the Bernoulli equation, which is principally:
v12/ 2g + p1/g + h1 = v22 / 2g + p2 / g + h2 + friction losses
where v1 and v2 is the velocities , p1 and p2 the pressures and h1 and h2 the liquid heads
upstream resp downstream an entrainment like a valve, or in both ends of a pipe etc.
If we just look at a valve with the same liquid head on both sides, we can simplify this to
principally
v = Constant * SQRT(( p1-p2)/)
for a fully open valve, where the constant relates to the friction losses caused by different
geometries. The velocity v multiplied with the open area of the valve, will give the flow
through the valve, for a given pressure difference.
By using this technique and making a number of equations, one for each node, and then
solve this set of equation simultaneously, we will determine the pressure and flow in the
41
whole network of pipes and process equipments. For the different process equipments,
pressure losses are determined due to the operating conditions, and thus are included in
the calculation, as well.
Separation of fibers and other particles for the different process equipments is also
calculated in each equipment algorithm. This gives the material balance over the
equipments and for the whole network, for each time step, considering also dynamics. This
is useful, when you want to test new advanced control algorithms , where you dont have
the DCS code for them, but can write them in Fortran, C++ or make use of Matlab instead.
With a good process model, it is possible to test the control strategy before implementing it
on the real process.
Other examples
There are many different models developed for both paper and pulp mills. For pulp mills
the major focus has been on the digester, and several examples exist on good models for a
number of different applications here, like [Bhartiya et al 2001],[ Wisnewski et al 1997] and
[Jansson et al 2004].
Examples of process equipments mentioned earlier can be the screen (to the right) and the
hydro cyclone (= cleaner, to the left)
For the screen algorithm we assume a plug flow from the top and down wards, but with
total mixing in radial direction. The rotor rotate with a relatively high velocity normally,
giving shear forces at the screen surface. Fibers are mechanically separated at the screen if
they are larger than a certain size in relation to the hole area or slot size, but also depending
on the concentration, shape, flow rate through the holes or slot, the rotor speed,
temperature and volume reduction factor (reject rate).The model gives the pressure drop
over the screen, as well as the mass and energy balance. At low reject flow, the
concentration in the reject goes up, and if it becomes too high, the motor stops, the fibers
42
accumulate and eventually the whole screen plugs up. The pressures around the screens are
calculated, but also the fiber size distribution of the fibers in the different streams, as well as
the amounts and concentrations.
In the table below results for the model algorithm calculations are compared to data from
experiments done by technical institutes in Canada( Paprican) [Gooding et al 1992]and
Sweden [STFI 1999]. The model used is the one described earlier.:
Table 1 For a 1.4 mm hole screen, the following results can be seen:
Qrej/ accept Separation efficiency
Qfeed l/m2.s 0-0.5
0.5-1
2-5 mm fibers
Exp
Calc
Exp
Calc
Exp
Calc
0.4
99
74.8
77.1
80.5
80.6
92.0
91.8
0.7
16.5
89.3
89.0
92.3
93.0
97.6
100
Tbale 2: For a 0.4 mm slot screen, assuming 0.7 mm long , 0.025 mm wide fibers:
Experimental
Accept l/m2.s 50
100
50
100
0.63
0.52
200
Q_rej/Q_feed
0.4
0.61
0.52
0.29
0.33
0.09
0.27
0.18
0.12
0.35
0.25
0.18
0.15
Table 3. For a cleaner( =hydrocyclone), the corresponding figures are shown below:
FlowQrej/Qinj 0.25 mm 0.75 mm 1.75 mm
3.5 mm fibers
l/min
Calc
270
0.10
0.32
0.33
500
0.26
0.77
0.71
As can be seen, the prediction of the separation efficiency can be quite good, although not
absolute. The reason is both the difficulty to make the experiments totally controlled ( see
e.g. the first row of the cleaner experiment, where the separation efficiency is not
correlated to fiber size for high reject ,low flow) , and to catch all possible effects in one
single model, based on first order principles. The tuning / configuration of the models are
43
done to fit the actual equipment and the normal operating range, but will give reasonably
good result also outside this area.
In reality data from many more papers were used to build the models, where different
variable were varied, and different types of equipments tested. Unfortunately few of these
include both variation of flow rates, concentrations, reject/accept ratio, different fiber size
distributions etc, but mostly only one of these, and normally not including looking at the
effect on different fiber sizes. Still, together all these data bring new elements into the
puzzle to build a good physical model. An example of starting with a physical model and
tune it with statistical plant data was shown by [Pettersson 1998]
When we look at a new type of screen we use normally simple mass balance data to tune at
some few operational conditions. These data will be used to tune some of the coefficients
mentioned earlier, while the rest are used as default values until new data comes up that
may be used to proceed with more detailed tuning for the specific equipment. Some of this
has been collected from [Nilsson 1995] and [Jegeback 1990, 1993]
After the start up, or for an existing mill directly, the dynamic simulator can be used for
optimizing the process. This is done so, that different ways of running the process can be
tested by the process engineers on the simulator, to see how e.g. water flows or dry solids
contents etc are influenced, e.g. during a grade change.
By collecting data through the information management system, and sending them to the
simulator, it may be possible to use the data for more advanced diagnostics [Karlsson et al
2003]. This will be for both the process and the sensors, by making use of the expert system
functions residing in the simulator. With a model, that shows how different sensors and
process parts are correlated to each other, predictions of performance can be made as well.
By comparing to the real process signals, deviations and drifts can be diagnosed, to alarm
the operator before it is possible to see the faults manually.
For the process optimization, higher fidelity models (the fidelity of the models can be
selected for the most important equipments) may be needed[Bell et al 2004],[Dhak et al
2004][Hess 2000] and [Morari et al 1980]. The communication speed with the DCS systems
does normally not need to be considered, as the process engineer can work on the simulator
without the real time DCS system connected. With interaction between a simulator and an
optimization algorithm the communication still may be the limiting parameter.
2.2.1.1.6 CONCLUSIONS
44
A model has been developed for major pulp and paper equipments. It has been shown by
comparing experimental results to model predictions that a reasonably good prediction for
separation efficiency for many fractions can be made for separation equipments operating under a
wide operational range. The same model can be used for many different types of process
equipments and models, where only a number of parameters (constants) have to be configured
with data from relatively few experiments, or principally normal mass balance data that can be
achieved from vendors or mills.
2.2.1.1.7 REFERENCES:
NOPS Paper Machine Operator Training System, ABB Process Automation, 1990
Bird R.B. ,W.E.Stewart and E.N.Lightfoot: Transport Phenomena, by,John Wiley & sons, 2nd
edition 2002.
Gooding Robert W. and Richard J.Kerekes: Consistency changes caused by pulp screening, Tappi
Journal, Nov 1992, p 109-118.
Nilsson Anders: The simulation resource Extend , Licentiate Thesis Pulp&Paper Tech Dept
,Royal Inst of Technology in Stockholm, TRITA-PMT Report 1995:14.
Jegeback M. and B.Norman: Fextend- computer simulation of paper machine back water systems
for process engineers, STFI report A 987, 1990.
Jegeback M.: Dynamic simulation of FEX back water system by Fextend, ,STFI report A 996,
1993
Data from experimental reports from STFI on screening and cleaning 1999.
Bell J., Dahlquist E.,Holmstrom K.,Ihalainen H.,Ritala R.,Ruis J.,Sujrvi M.,Tienari M:
Operations decision support based on dynamic simulation and optimization. PulPaper2004
conference in Helsinki, 1-3 June 2004.Proceedings.
Dhak J.,Dahlquist E.,Holmstrom K.,Ruiz J.,Bell J.,Goedsch F: Developing a generic method for
paper mill optimization.Control Systems 2004 in Quebec City, 14-17 June 2004.
Proceedings.
Ulf Persson, Lars Ledung, Tomas Lindberg, Jens Pettersson, Per-Olof Sahlin and ke
Lindberg: On-line Optimization of Pulp & Paper Production, in proceedings from TAPPI
conference in Atlanta, 2003.
Wisnewski P.A, Doyle F.J and Kayihan F.: Fundamental continuous pulp digester model for
simulation and control. AIChE Journal Vol 43, no 12, dec 1997, pp 3175-3192.
Pettersson J. (1998): On Model Based Estimation of Quality Variables for Paper
Manufacturing. Tech LicThesis , KTH
Bhartiya , Dufour and Doyle ( 2001) : Thermal- Hydraulic modelling of a continuous pulp
digester, in proceedings from Conference on Digester modelling in Annapolis, June, 2001.
45
Hess T. (2000): Process optimization with dynamic modeling offers big benefits,
I&CS,August, p 43-48.
Karlsson C., Dahlquist E., Process and sensor diagnostics - Data reconciliation for a flue gas
channel, Vrmeforsk Service AB, 2003, (in Swedish).
Morari M, Stephanopolous G and Arkun Y: Studies in the synthesis of control structures for
chemical processes. Part 1: formulation of the problem.Process decomposition and the
classification of the control task. Analysis of the optimizing control structures. American
institute of Chemical Engineering Journal , 26(2), 220-232.1980.
Ryan K. and Dahlquist E.: MNI Experiences with process simulation. Proceedings Asia Paper,
Singapore, 2002.
Jansson and Erik Dahlquist : Model based control and optimization in pulp industry,
SIMS2004,Copenhagen Sept 23-24,2004.
If we dont have a good process understanding, but there are good process measurements
available, data driven models are a good choice. You collect data from the actual process and
correlate these to different quality variables that can be achieved from lab measurements,
normally. In this way we can get prediction models for many properties on-line that are not
possible to get otherwise, or where on-line instruments are far too expensive to be motivated.
Still we must be aware of the fact that shit in- shit out. We need to have process variations that
are significant and uncorrelated to build a good statistical model. This is sometimes not realised,
and the operators will be very disappointed with the quality of the predictions. Often also very
important variables were kept constant during the data collection, like the wire speed of the paper
machine, and thus the most important variable is not included in the model.
To avoid this type of poor models, you can preferably make use of a factorial design of
experiments Here different variables are varied in a structured way independent of each other.
Even if you can not perform all the experiments you would like to, you can at least get a number
of them, and then probably a significantly better model than if you just used the randomly
collected data from the process. The model most probably also will be more robust with respect
to different process variations not directly encountered in the model.
When a statistical model is built you often start with including any kind of measured data, but
later it is normally best to reduce the number of variables in the prediction model to the most
important ones. This will give robustness and will be easier to update continually. Statistical
models can handle dynamics, but normally physical models are better on that. Examples of
statistical models are ANN, MVDA, BN, PLS, PCA. These techniques are presented later on in
this book.
One important application of statistical models is as soft sensors, and several examples of such
are given later on.
46
Correlation between attributes can be measured in a number of different ways, the perhaps
most commonly used being Pearson's correlation coefficient and the covariance. These are
very useful, robust measures that often give a good indication about the dependency
structure of the sequences. If, for some reason, we can be sure that there are only linear
dependencies and independent samples in the data, they are also optimal measures. When
non-linear dependencies are present or samples are not independent, as in a time series, the
measure might be fooled into either giving a low value of the correlation for two highly
correlated variables, or to significantly overestimate the actual correlation. The use of
information theory [4, 5], or more specifically the concept of mutual information, can
provide a solution [3].
47
The entropy of a stochastic variable [1, 2], which can be thought of as a measure of the
amount of uncertainty or the mean information from the variable, is defined (in the discrete
case) as
(1)
If two variables X and Y are independent, H(X, Y) = H(X) + H(Y). If the variables are not
independent, H(X) + H(Y) will be larger than H(X, Y), that is some of the information in
H(X,Y) is included in both the marginal entropies. This common information is called the
mutual information, I(X; Y). It can also be thought of as the reduction in the uncertainty in
one variable due to the knowledge of the other, or more formally
(2)
The mutual information is symmetrical and is always larger than or equal to zero, with
equality only if X and Y are independent. It can therfore be viewed as a measure of the
dependence between two variables. If the variables are independent, the mutual
information between them will be zero. If they are strongly dependent, the mutual
information will be large. Mutual information is a general correlation measure and can be
generalised to all kinds of probability distributions. It is also, given an appropriate model of
the distributions, able to detect non-linear dependencies between variables.
To be able to calculate the mutual information, we have to know both the variables marginal
distributions and their joint distribution. If the parametric forms of the distributions are not
known, it is still possible to calculate the mutual information by quantising each variable
into discrete values, or bins. Each marginal is discretised, and the joint distribution is
modelled by the grid resulting from the two marginal distribution models. Then histograms
are constructed from the data using this discretisation and from these the probabilities are
estimated.
The number of discrete values in the grid is critical. Choosing a too fine grid with more bins
than data points results in an estimate of mutual information equal to the logarithm of the
number of data points. If too few intervals are chosen, again only part of the correlation will
be detected. When discretizising the marginal distributions, it is useful to set the limits
between bins so that all bins contain the same number of data points. This gives us a much
more stable measure. It is less sensitive to both skewness and outliers and the resulting grid
uses the highest resolution in the areas where there are most data points.
Another common assumption about the distributions is that they are Gaussian. This results
in a measure where a linear correlation is assumed, and it is only possible to detect the
linear part of the correlations in the data. To derive an expression for the mutual
information between Gaussian variables, we can start from an expression of the entropy for
a Gaussian distribution. The entropy of an n-dimensional Gaussian distribution can be
written as
48
1
h(X1, X 2 ,K, X n ) = log(2e) n C
2
(3)
where |C| denotes the determinant of the covariance matrix. Using the equations presented
earlier, it is easy to calculate the mutual information for Gaussian distributions. When
calculating the mutual information under these assumptions, the only parameters needed
are the means and variances of the two variables and the covariance between them, all
easily estimated from data.
The measures described above are general correlation measures. If we are working with
sequential data, the correlation is usually measured as a function of the time shift between
the series. This can then be plotted as a correllogram for visual inspection or used for
automatic generation of dependency graphs.
Let us now have a look at how we can extend the notion of mutual information to efficiently
find dependencies between time series [6]. We start from an expression for the uncertainty
of a sequence, corresponding to the entropy of a single variable. Loosely speaking, if we
have a sequence of n random variables, this uncertainty can be defined as how the entropy
of the sequence grows with n. This is called the entropy rate of the process and can be
defined as
1
H r (X) = lim H(X n | X n1,K, X1)
n n
(4)
Based on the entropy rate, we can then construct a measure of the mutual information rate,
(5)
49
(6)
We will now present a couple of examples, both taken from a chemical plant application.
Only the linear versions of the measures are shown since they are more stable than the
binned versions.
The top left diagram of figure 1 shows the linear correlation between the measured variable
X3 and the controlled variable C63. It is an example of a well behaved, linear correlation
with a short and reasonable time delay. The correllogram shows just one clear peak at delay
-5 minutes, indicating that this probably is a real correlation between the attributes and not
produced by artifacts in the data. The mutual information rate correllogram for the same
attributes in the top right diagram of figure 1 shows the same behaviour. It is a bit more
peaky and shows much lower correlation, but the peak is at almost the same place, delay -6
minutes, as in the mutual information correllogram.
In the lower left diagram of figure 1 the linear mutual information between attribute X48
and Y is shown. The correllogram is very smooth, although somewhat low, but the measure
is obviously fooled by some general trend in the data since it is constantly increasing with
decreasing values of the delay. The maximum is at -400 minutes, simply because that is the
chosen plot range, and is a too long delay to be considered reasonable in this case. The
mutual information rate on the other hand, shown in the lower right diagram of figure 1,
shows a clear peak at delay 2. That is a plausible value of the delay between the sequences,
although the value of the correlation is rather low. The information rate diagram is not at all
as smooth as the mutual information, showing several small spikes which are most likely
effects of noise in the data.
50
Using mutual information or the normal correlation coefficient between time series tends to
give a too high value of the correlation. This happens because that if the time series moves
slowly enough, the relation between the series at one point in time is likely to be maintained
for several time steps. This means that pure random coincidences between the series gets
multiplied with a factor depending on how slow the series are, making that correlation
seem more significant than it is. The mutual information rate on the other hand, which only
considers the new information in every step, correctly compensates for that effect but
instead requires a more complicated model to estimate, which makes it more sensitive to
noise.
All in all it seems that the linear information rate is the measure that gives the most reliable
indications of correlations between the time series. However, due to the different trade-offs,
a general rule is that a feature, a peak, should appear using at least two of the methods to be
considered significant.
2.2.2.7 REFERENCES
51
[1] Shannon C. E. (1948). The mathematical theory of communication. Bell Syst. Tech. J.
27:379-423.
[2] Shannon C. E. (1951). Prediction and entropy of printed English. Bell Syst. Tech. J.
1951:379-423.
[3] Li W. (1990). Mutual information functions versus correlation functions. Journal of
Statistical Physics. 60:823-837.
[4] Cover T. M., Thomas J. A. (1991). Elements of Information Theory. John Wiley and Sons,
New York.
[5] Ash R. (1967). Information Theory. Interscience Publishers, New York.
[6] Gillblad D. and Holst A. (2001). Dependency derivation in industrial process data. In
Proceedings of the 2001 International Conference on Data Mining, pp. 599602. IEEE
Computer Society Press, Los Alamitos, CA.
52
Different types of ANNs can be defined depending on the structure and way of working.
Neuronal nodes are usually distributed in layers, which are interconnected node by node
with specific connection weights. Connections can exist only from a specific layer to the
following, or between nodes in the same layer, even backwards. Those different connections
determine the different types and applications of ANNs.
Figure X shows one of the simplest cases, the multilayer perceptron or feed forward ANN,
which has not got both lateral connections between neurons in the same layer and back
connections to previous layers. Each node sums its inputs, applies a transfer function and
obtains an output. Transfer functions in each node can vary and be step functions, linear
ones, sigmoid and so on.
Output Layer
Hidden Layer(s)
Nodes or
neurons
Input Layer
Widrow-Hoff ANNs: They are networks with applications mainly for linear systems
like the ADALINE network. They use algorithms based on minimisation of mean
squared error and most famous applications have been the echo removal in acoustic
signals.
Multilayer Perceptrons: They are the most used for nonlinear pattern recognition.
They consist of several simple processing nodes placed in a multilayer distribution.
They commonly use different back propagation algorithms including heuristic and
numeric modifications.
53
Associative ANNs: They include logic expressions in their calculations, like the simple
associative network or the recognition network. They are based on different
training error reduction rules like Kohonen rule or Hebb rule. Their main
application is the classification of different groups or classes.
Grossberg network: this sort of ANN is based on the filling of incomplete data, just as
our brain does with, for instance, visual imperfections.
Other ANNs: There are lots of ANN types and modifications of each type.
Probabilistic ANNs, Radial Basis ANNs, General Regression ANNs, with dead times
(Elman ANNs, autoregressive non-linear moving average ANNs, real-time recurrent
ANNs, polynomic ANNs, etc.).
Prediction error reduction in an ANN takes place through a training-learning process, which
consists of modifying internal weights in order to reach an output that should be as closer
as possible to training data. There are different types of algorithms that perform such
54
weight tuning like back propagation ones and their heuristic modifications, those based on
genetic algorithms, fuzzy logics, bayesian fusion, and so on.
The most used training algorithms in pulp and paper industry according to existing
literature are those based on back propagation weight changing rules like the gradient
descent rule. They are usually applied with a minimum of three layers (1 hidden layer)
(Aguiar H.C. and Filho R.M., 2001; Broeren L.A. and Smith B.A., 1996; Campoy-Cervera P. et
al., 2001; Kumar A. and Hand V.C., 2000; Masmoudi R.A., 1999; Miyanishi T. and Shimada H.,
1998; Vaughan J.S. et al., 1999).
Training of an ANN is structured with the following steps:
1. Creating random initial weight matrixes with values between 0 and 1, either
positive or negative. Matrixes created with constant values usually do not fit the
patterns as well as random ones. In order to assure good results, several initial
weight matrixes must be trained in the development of a robust model.
2. Training pair selection. A training pair is the input vector together with the output
vector taken at the same time. This selection is carried out with all data selected for
training. When working with vectors (the most usual way) there will be one vector
with one element per input.
3. With obtained weights (initial in step 1 and calculated in step 5) ANN outputs are
calculated following the same procedure in each neuron (summing weighted inputs
and applying transfer functions), from the first layer and concluding with the output
layer, obtaining the ANN responses.
4. Obtained responses are compared with target values from experimental data. Errors
are calculated as the difference between the response and the target.
5. Back propagation algorithm, which will be explained in detail, is applied in order to
adjust all the weights in the ANN and reduce training errors.
6. Steps 2-5 are repeated for each pair of input-output data until obtained errors in
training or validation, depending on the case, are considered optimum and low
enough.
When training, there are two main time-consuming calculations. The first one is carried out
feed forward in order to calculate ANN outputs. The second is calculated backwards in
order to adjust weights depending on errors, trying to reach experimental targets. All
calculations are made layer by layer, and the output in one layer becomes the input in the
following one.
A description of back propagation algorithm is presented by descriptive equations that
show the procedure step by step.
First of all, the input to one neural node is calculated as the sum of all weighted outputs
from the previous layer, as can be seen in equation 1
55
I = x1 w1 + x 2 w2 + ... + x n wn = x i wi
i =1
(1)
Where wi represent the weights applied over xi, which are the outputs from neurons of
the previous layer. After summing, an activation function is applied. This function varies
depending on each specific case. Usual functions are the step function, logistic sigmoid
(quite usual in hidden layers, hyperbolic tangent, linear or any other type with similar
characteristics.
There are many different transfer functions to be selected in each neuron of the ANN. They
are usually selected by layers, and can be step functions, linear functions, hyperbolic
tangents, logistic sigmoid, etc. In equation (2) logistic sigmoid with flexible parameter is
shown as an example of usual transfer function.
( I ) =
1
1 + e I
(2)
With those transfer functions, which just convert a single input into a single output, the
output of a neuron is obtained, and this output will be properly weighted and inputted in
each neuron of the following layer, where the same process (summing, applying transfer
functions and going forward through weighting the outputs) will take place until the output
of the ANN is obtained.
The use of non-linear transfer functions in the layers is usually due to the complexity of
studied process, which needs the introduction of non-linearities in the model. The output
layer makes final summing and often applies linear transfer functions in order to speed up
iteration procedure.
After obtaining the output of the ANN, it is compared with experimental values, which are
the target during training. Errors are calculated as the difference between experimental
data and ANN outputs in each output variable. Errors are usually squared to remove the
effect of signs.
Once obtained the errors different custom rules could be applied. The most commonly
accepted one is the Delta rule, which considers that the change in a connection weight is
proportional to the change in the squared error and inversely proportional to the change in
corresponding weight in previous iterations, as shown in equation 3.
w pq , k = p , q
q2
w pq , k
(3)
where the proportionality constant p, q is called learning rate, whose subscripts refer to
two different neurons p and q, respectively located in layers j and k, being k the last
56
layer, placed just after layer j. Note that calculations are slightly different to change
weights in previous layers, but they have the same theoretical basis. Connection weight w
is thus named according to the two neurons that it links.
Through mathematical developments that are available at the literature (Tsoukalas L.H. and
Uhrig R.E., 1997), by evaluating equation 3, an indication of the new weights to be applied in
different weight matrixes can be calculated with equation 4, expressed in two different
forms.
wpq ,k
q2
= p ,q
= p ,q pq ,k p. j
wpq ,k
wpq ,k ( N + 1) = wpq ,k ( N ) p ,q pq ,k p. j
(4)
pq ,k 2 [Tq q ,k ] q ,k [1 q ,k ] = 2 q
q , k
I q , k
(5)
where is the same parameter than in equation 2, q, k the output from neuron q in layer
k, and I q , k the input to the same neuron, calculated as the summing from the outputs from
previous layer, once weighted. Tq is the target value, with which ANN response is compared.
All the steps that have been mentioned are carried out in order to calculate the new weights
between last layer and previous one. The other weights can be calculated using the same
basis. After some calculations equation 6 is achieved.
r
(6)
where xh are the inputs of layer i, placed just before layer j and two layers before layer
k, and r is the number of neurons in layer k. Neuron that is considered for calculation of
the new weights in layer i is called neuron h.
A representative scheme with ANN structure and nomenclature is shown in figure X1.
57
wpl ,k
xh
whp, j
I p , j p, j
wpq,k
wpr ,k
Il ,k l,l k
I q , k q,k
Ir , k r,k
Comp
Tl
Comp
Tq
Comp
Tr
l
q
r
Figure 2 Scheme of a back propagation ANN. Source: Tsoukalas L.H. and Uhrig
R.E., 1997.
The calculation sequence previously mentioned is repeated, and training errors become
smaller with each new iteration. A low training error does not guarantee better predictions,
because it can be due to over fitting in the ANN. A proper validation procedure in
supervised learning should test the ANN responses with independent data for different ANN
architectures and different number of training iterations. Thus, the training-learning
procedure must be optimised in order to reduce validation errors and achieve robust
results.
There are numerous modifications available in the literature (Hagan et al., 1995). Most
common ones are the heuristic modifications like the momentum factor and variable
learning rate, and numerical optimization techniques like the Levenberg-Marquardt
algorithm.
Example of a heuristic modification: Momentum
Momentum has the same meaning as inertia in physics. The value of this parameter will
determine the weight of previous changes in internal weights when calculating that change
in current iteration. Equation 7 shows the influence of momentum over the change in the
weights in the last iteration.
w pq , k ( N + 1) = p ,q pq ,k p . j + w pq , k ( N )
(7)
where is the momentum factor and its values can oscillate between 0 and 1. The main
utility of this factor is the capacity of avoiding local minima of training errors that it
58
provides to the ANN, allowing a more probable finding of global minima, which would
probably allow to achieve more robust models. Nature of each system will determine the
optimum value of this parameter, or if it is better not to use it.
Therefore, a first stage when developing ANNs is to make different trials with different
types of algorithms, or even different values for other parameters like the learning rate, the
transfer functions or the number of hidden layers and neurons. There are also algorithms in
which learning rate varies automatically as weight changes also does, with custom
increasing and decreasing factors.
There are quite comprehensive works describing all possible variations in this back
propagation algorithm and even describing in detail other types of ANNs. In this chapter a
description of the most usual algorithm according to literature has been carried out.
Applications that have been analysed are shown in the following chapter.
Feed forward ANNs with back propagation algorithms are the most used in the pulp and
paper industry as previously mentioned. There are general issues that are commented
about constructing ANNs. Screening outliers and pre-processing data are essential in order
to achieve good results. These models have been used for many different purposes. The
most common ones are:
Product quality properties prediction (Broeren L.A. and Smith B.A., 1996; Joeressen
A., 2001; Kumar A. and Hand V.C., 2000; Alonso A. et al, 2006)
Going more into the detail of some published works, it can be seen that they have achieved
in general good results. ANNs have been used to optimize and control brightness (Broeren
L.A. and Smith B.A., 1996; Vaughan J.S. et al., 1999), for paper pulp inspection (CampoyCervera P. et al., 2001), to predict DBO (Masmoudi R.A., 1999) or to optimise wet-end
(Alonso A. et al., 2006).
They have also been used combined with deterministic models/equations to predict
pulping degree (Aguiar H.C. and Filho R.M., 2001) and corrugated box compression values
(Joeressen A., 2001); combined with genetic algorithms (Kumar A. and Hand V.C., 2000),
and also with MVA tools like PCA for fault detection (Miyanishi T. and Shimada H., 1998).
Moreover, there are some explicit references to economical benefits when applying ANNs.
total chemical cost reduction of US$ 138000 annually (Smith B.A. and Broeren L.A.,
1996).
Lime kiln control resulted in 7% specific heat energy consumption savings, and a
subsequent reduction in the heat losses (Jrvensivu et al., 2001).
When predicting corrugated box compression values, 5% of total costs were saved
(Joeressen A., 2001).
DBO prediction allowed US$ 1 million savings during 1998 (Masmoudi R.A., 1999).
Web-breaks diagnosis with ANNs and PCA (Miyanishi T. and Shimada H., 1998)
resulted in significant improvements: first-pass retention and first-pass ash
retention improved by 3% and 4%; sludge generation from the effluent clarifier was
reduced by 1%; alum usage decreased and the total annual cost reduction was
estimated about US$ 1 million.
Application of fuzzy logic and ANNs in order to control the kappa index resulted in a
reduction of deviations in kappa index by 25% (Pulkkinen et al., 1997).
Stapley et al. in 1997 experienced several benefits when using ANNs in wet-end. In a
substantiation trial, chemical costs were reduced by 32% (which means annual US$
800000 savings) and production was increased by 6%. Brightness was increased by
10% and that fact allowed a reduction in the TMP bleached chemical consumption
required to achieve specifications (which supposed additional savings between US$
100000 and 200000 per year).
Finally, the development of an ANN soft sensor for chlorine dioxide stage brightness
control (Vaughan J.S. et al., 1999) reduced the frequency of off quality rounds by
13%, the brightness variability by 50%, and it also reduced the amount of dioxide
used at the stage by 33%.
These applications are just some examples in a great amount of works that cover almost all
the steps in papermaking, from raw materials until the final product. There is a continuous
flow of new papers dealing with the use of ANNs in the pulp and paper industry with many
different purposes.
An important challenge when using such models will be converting the concept of being
black-box models into grey or even white box models, through different analysis of the
internal structure.
1.
Aguiar H.C. and Filho R.M., Neural network and hybrid model: a discussion about
different modelling techniques to predict the pulping degree with industrial data,
Chem. Eng. Sci., Vol. 56, No. 2, pp 565-570. 2001.
60
2.
Alonso A., Blanco A., Negro C., Tijero J. and San Po I., Application of advanced data
treatment to predict paper properties, Proceeding of the 5th MathMod Conference,
Vienna, Austria, 8-10 Feb., 7pp. 2006.
3.
Broeren L.A. and Smith B.A., Process Optimization with Neural Network Software,
Prog. Pap. Recycling, Vol. 5, No. 2, pp 95-98. 1996.
4.
5.
Chitra S.P., Use Neural Networks for Problem Solving, Chem. Eng. Prog., Vol. 89, N 4,
pp 44-52. 1993.
6.
7.
8.
Kumar A. and Hand V.C., Feasibility of Using Genetic Algorithms and Neural Networks
to Predict and Optimize Coated Paper and Board Brightness, Ind. Eng. Chem. Res., Vol.
39, No. 12, pp 4956-4962. 2000.
9.
Masmoudi R.A., Rapid prediction of effluent biochemical oxygen demand for improved
environmental control, Tappi J., Vol. 82, No. 10, pp 111-119. 1999.
10.
Miyanishi T. and Shimada H., Using Neural Networks to Diagnose Web Breaks on a
Newsprint Paper Machine, Tappi Journal, Vol. 81, No. 9, pp 163-170. 1998.
11.
Pulkkinen M., Saastamoinen M. and Skytt M., Kotka fuzzifies its kappa control, Pulp
Pap. Eur., Vol. 2, No. 10, pp 30-32. 1997.
12.
Smith B.A. and Broeren L.A., A tool for process optimization: neural network software,
TAPPI Proceedings of the 1996 Recycling symposium, New Orleans, LA, USA, ISBN 089852-657-4. 3-6 Mar, pp 163-187. 1996.
13.
Smith G.C., Wrobel C.L. and Stengel D.L., Modelling TRS and Sulphur Dioxide Emissions
from a Kraft Recovery Boiler Using an Artificial Neural Network, Tappi J., Vol. 83, No.
11, p 69. 2000.
14.
Stapley C.E., Butner R.E., Kangas M.Y.O., Broeren L.A. and Smith B.A., Tools for strategic
analysis: neural networking, Pulp Pap., Vol. 71, No. 1, pp 89-90,93-96. 1997.
15.
Tchaban T., Taylor M.J. and Griffin J.P., Establishing impacts of the inputs in a
feedforward neural network, Neural Comput. & Applic. Vol. 7, pp 309-317. 1998.
16.
Tsoukalas L.H. and Uhrig R.E., Fuzzy and Neural Approaches in Engineering, Ed. Wiley
Interscience, . 1997.
61
17.
Vaughan J.S., Gottlieb P.M., Lee S.-C. and Beilstein J.R., The Development of a Neural
Network Soft Sensor for Chlorine Dioxide Stage Brightness Control, TAPPI 99
""Preparing for the next millennium"", Atlanta, GA, USA, Book 1, ISBN 0-89852-734-1.
1-4 Mar, pp 147-159. 1999.
In manufacturing industries and for applications in logistics event driven simulation models
are common. Very often you have a number of actions to take in a serial way, but slightly
different actions depending on what product that is coming on e.g. a conveyor band. If a
piece is broke, it should be put in one box, otherwise in another. If a car is blue it should be
directed to one line, but if it is red to another, as examples.
If we have a logistics problem the simulation may be used to plan e.g. distribution of a
number of products to a number of sites. The models then are used together with an
optimization algorithm to determine a best distribution where storages, delivery times,
capacity etc are constraints.
In pulp and paper industry this type of models may be used to handle overall planning like
orders coming in, storage in a warehouse, distribution by boat or lorry etc.
A suitable simulation tool for this is e.g. EXTEND, which has a number of predefined blocks
to use, and is easy to learn how to handle. It also contains easy to use animations.
With hybrid models we normally mean grey box models, or combinations of physical
models with statistical data. This type of models has the advantage of physical models
giving reasonably robust solutions in a broad area of operations together with adaptation
with real process data. The negative aspect is that the physical models have to be simplified,
and there is always a risk that you loose important features of the process behavior. It is
often also more complex to tune a physical model than to build a completely data driven
model like a neural net or a PLS model.
Until we have computer power enough to model physical processes in very high resolution
the grey box models will most probably be the most used models for on-line applications, as
they are easier to transfer from one machine to another, and normally needs less
maintenance work after installed.
62
From what has been said already we can conclude that principally white, physical models
are the best from a process modeling point of view, at least if we understand the processes
to be modeled. On the other hand, we seldom have the computer power needed to model a
process in enough detail to really operate it as a white model.
The black box models have the advantages of principally no need for a process
understanding. On the other hand the models need good feed data to be good and robust,
and the verification of the process results used for building the model need to be very good,
to give a good model. Shit in shit out is very relevant for this type of models. To gather a
good model we thus need e.g. a factorial design for experiments in the process equipment to
be modeled, to really get information needed to build a model that is valid in a reasonable
large operational area. For different product grades it may be needed to have several
models built. Still, black box models like ANN, PLS, PCA and similar can be very good for e.g.
building soft sensors for important quality variables to be used on-line. This has been
shown in many examples, and will be covered later on in this book.
The grey box models or hybrid models have the advantage of having a better robustness in
a larger operational area even if we dont have process data covering the whole area. From
physical relations reasonable assumptions can be made, at least if we know the process
well. By tuning the model with available process data, we still can get reasonable accuracy
of the predictions. This type of models are gaining more interest for many different
applications, and are used to build standard libraries for simulators, where different
machines can be tuned with suppliers data, but the general equations are the same. They
have the advantage to be possible to expand with more variables without having to redo all
the tuning work from scratch, as is the case for statistical models. Grey box models can also
be used for soft sensors if there is a process understanding of the factors affecting the
quality.
Uncertainties are something we have to live with. Still, often this is not considered in many
models, and thus the models may be very accurate sometimes, but totally unreliable at
other times. If we can determine when it is accurate and when it is not, we could take this
into account also in the control and optimization.
63
We then can identify different types of uncertainty. One type is measurement inaccuracy.
E.g. consistency sensors often are quite unreliable if we have air bubbles in the flow, but
may be much more reliable if the air is removed. A temperature meter may be very accurate
just after calibration, but is drifting away as it is fouled. And so on
Another type of uncertainty is if equipment will be in operation or not. If we make an
optimization calculation assuming that the plant will be performing well, and then a pump
is cavitating, and thus the pump flow significantly lower than the set point from the
production plan, the whole production plan may be upset.
There also may be planned upsets like service of equipments, but it still will be uncertainty
as if we run to long the performance of equipment may be lower sometime in the future if
you neglect the service right now to increase the production short term.
It is not evident exactly how to handle the uncertainty but one way is to put different
uncertainty levels on different equipments or process parts. The constraints then may have
different intervals depending on this. It may also be an alternative to make a number of
different optimizations with different constraints to see how much the different alternatives
will affect the final production result. We get a sensitivity analysis, which can give a decision
support for the one making the production plan.
By collecting data from previous operations and feeding these into a BN, Bayesian net, it
may also be possible to create a tool for how to make a good decision if we have a similar
situation to compare with. If we know how we acted earlier and the outcome then, we can
get a statistical best alternative from earlier experience. The back side is that we can not
be sure if the situation really is the same, but at least we have some reasonable statistics to
fall back on. We also should look at not only the most favorable alternative, but also those
close to this, and try to analyze the worst scenario if you take the decision recommended by
the decision support system. By a good structure of the decision support system you may
get a very valuable tool for the decision, and a possibility to really make use of not only your
own experience, but also the whole organizations.
If we have uncertainties in the actual models these can also be decreased by updating the
models on a frequent basis. To do this we need to have a good control over the quality of the
input data, and an understanding of what variables to include and possibly also what extra
measurements that are needed! Dynamic calibration with a moving window technique is
one way and dynamic data reconciliation another. These techniques are presented later on
in this book. First we will discuss the data preprocessing, which is a key for successful
updating of models in an automated way.
2.3.2.1 SUMMARY
64
2.3.2.2.1 INTRODUCTION
Measuring properties are essential for plant control and monitoring. Constant strife for
improved plant efficiency demand better control, which often means more instrumentation.
The sensors are used for on-line computation of e.g. plant operation cost, emissions, heat
power, etc. These applications need reliable on-line measurement data. A known problem is
65
that sensors in the same location show different values. Regarding one as faulty or not
reliable often solves this or an average computation is made. The use of simulators make it
possible to both measure and compute the same state variable, but it makes the situation
complex when it comes to selecting reliable sensors and what sensor readings to believe in.
These problems can be solved by a method able to handle redundant information and
estimate the state of the process. The three components; gross error detection, gross error
isolation and data reconciliation can be put together to form a system. The research area of
data reconciliation and gross error detection spans over many other research areas, such as
signal analysis, statistics, optimization, process control, information theory, etc. Complex
industries have much to gain from data treatment. The development in the field is fastest in
the nuclear power generation and the chemical process industry. An accident in these
industries may be disastrous with consequence in both causalities and company
bankruptcy, [Hoo04]. Other areas using this kind of data treatment are the mining industry
and the conventional power generation industry.
The chain from sensor signal to operator interaction must be considered to find optimal
solutions. The economic considerations on number of sensors and maintenance costs must
also be included when choosing if, and by what degree the technique is to be implemented.
The core of data treatment in this text is data reconciliation (DR). DR algorithms can
handle random noise in the measurements, but they cannot handle systematic and large
errors. A protective layer of gross error detection and gross error isolation are therefore
vital parts of a data treatment system. The goal of the data treatment is to deliver a data set,
which is closer to the true process state than raw data. This data set is ready for use in top
applications, such as those in the bottom of Figure 1.
66
Energy and environmental taxes and fees have pushed power plant owners to
experiment with new fuels to lower costs. Fuel flexibility is more important than ever
before, and it is possible to meet this requirement with improved boiler technologies such
as fluidized beds and circulating beds. Fuels earlier used only for heat water production is
now introduced in heat and power plants, and these plants operate at higher temperatures.
Also new chemical reactions are activated. The experience of combusting these new fuels
and mixes of fuels in heat and power plants are increasing. Normal degradation of sensors
and process in heat and power plants are mainly due to:
Corrosion
Fouling
Erosion
67
High temperature
Unbalance in plant heat balance or flue gas flow pattern cause problems. This can be due
to use of fuels not intended or flaws in the plant design. The degradation causes listed above
can force a shut down of the plant for maintenance. However, normally a certain degree of
degradation is accounted for. Faster degradation in sensors is possible if they are positioned
where they are exposed to erosion, corrosion or fouling. The same reasoning can be applied
on the process if fuel quality differs from the expected, regarding for example moisture,
alkali, sulphur, metals and properties such as; size distribution and burn-out rate.
Temperature sensors have a drift due to ageing and the high temperature they are
working in. Positioning is important for temperature sensors in sections where large
temperature gradients may occur. Pressure sensors are sensitive to clogging and also need
careful positioning. Most plants are shut down and revised annually. During periodic
maintenance and revision the process is controlled and the sensors are calibrated. Not all
sensors are calibrated annually.
Damages on process.
Initial degradation can cause delayed alarms due to change in dynamics or because of a
biased sensor. Nuisance of false alarms due to degrading sensors is important when it
comes to the plant operator situation. In power plants, clogging of pressure sensors and
mass flow sensors is an effect of fouling. Erosion is due to streaks of flue gas transporting
bed sand or fuel particles hitting boiler walls or other components and thus causing
damage. Heat exchangers are also exposed to erosion, often in combination with fouling.
Some parts of the boiler are exposed to corrosion and are protected, but corrosion can
appear on surfaces not designed for aggressive chemicals, due to e.g. fuel mix properties. All
the mentioned phenomena can cause plant shutdown, and need to be monitored to take
preventive actions and plan maintenance. It is difficult to measure when this degree of
degradation moves into the field of loss of control, especially when the degradation is slow
and controllability is lost within the alarm limits.
There are tools for detection of degradation using statistics, but they need to be
performed off-line by an engineer. To solve the problem by increased maintenance is costly.
68
Sensors showing the same value over time can be faulty even though quality measures like
standard deviation and drift trends show excellent values. This kind of degradation is
causing loss of control and prevent maintenance because no fault is indicated, the fault
remain hidden until extensive data mining have been performed to find such sensors.
Optimization, financial reports, environmental reports and other top applications using
measurements also suffer from degrading sensors. Examples of how degrading sensors and
process can affect a process can be found in reports on recovery boilers used in pulp and
paper industry [Doy72] and [Kle98]. The Black Liquor Recovery Boiler Advisory Committee
has documented 156 explosions and 450 near-miss incidents in the last 35 years. Following
quote from Lefebvre and Santyr [Lef92] comment the report. In addition to equipment
damage, some of the more severe explosions have resulted in injury or even death of
operating personnel. There have also been several hundred emergency shutdowns where
the fear of an explosion has led to a forced outage. The frequency of explosions has
remained relatively constant over the years, and the problem cannot be considered solved.
More examples are found in [Doy72] and [Kle98].
Data reconciliation (DR) is a special case of the general estimation problem. DR is the
problem of having measurements in an over-determined system satisfying all process
constraints. Unmeasured variables are either estimated simultaneously with DR in a general
estimation problem or as a separate step after the DR problem is solved. Solving DR and
estimation simultaneously is in some literature called data coaptation. Data reconciliation
has been reported in applications for mineral industry processes [Sim91], chemical
production plants [Abu03], and nuclear power plants, [Sun03]. The most active application
field for research in data reconciliation is in chemical engineering. The first method of data
reconciliation for a chemical process was proposed in 1961, [Kue61]. The methods involved
linear system equations under the assumption that all process variables were measured and
absence of gross errors. The described problem can be solved analytically by least squares
methods. Complementary methods as gross error detection and isolation where developed
parallel to DR in order to handle real DR application problems. Graph theory was connected
to the DR problem formulation by Mah et al., [Mah75]. The assumption of unmeasured
variables was not efficiently solved until Crowe [Cro83], presented the projection matrix
method. Swartz [Swa89], showed that the projection matrix operation could be solved by
the robust QR-factorization.
Nonlinear data reconciliation was first proposed to be solved by successive linearization
by Knepper and Gorman, [Kne80]. Liebman et al., [Lie91], showed that nonlinear
programming methods gave improvements in accuracy, but had long computation time.
Liebman et al., [Lie92], later proposed a time window approach for dynamic data
reconciliation. Russo and Young implemented a dynamic DR algorithm in 1999, [Rus99].
Recently Soderstrom et al., and Gatzke et al., reported industrial implementations in
chemical plants, [Sod00] and [Gat02], using nonlinear dynamic DR strategies.
69
The general data reconciliation problem formulation for linear equations solved for I
samples by a least squares method [Rom00] is here formulated as:
I
min ( y i xi )
x ,u
Eq. 0.1
i =0
s.t. A1 x + A2 u = 0
Where y is the measured state, x is the estimated state, u is the unmeasured state, A1 is
the columns of A with only measured variables and A2 is columns of A with unmeasured
variables. The simultaneous solution of data reconciliation and estimation of unmeasured
variables in Eq. 0.1 is in many cases a large task. To reduce the size of the optimization
problem the DR problem can be divided into two sub-problems, [Cro83]. After the division
of the total problem, the next task is to solve the over-determined equations. The following
sub-problem is to estimate the unmeasured variables with data from the solution of the
over-determined equation system. The simplest DR problem with only measured variables
presented below is with linear equations with no gross errors present and the process in
steady-state:
I
min ( y i xi )
x
Eq. 0.2
i =0
s.t. A1 x = 0
The linear process model A1, containing only measured variables can for example be the
conservation laws for a system with mass and energy streams. The following sub-problem
solves the linear equation system of unmeasured variables.
A2 u = 0
Eq. 0.3
This equation system is solved with any linear equation solution method, for example LUdecomposition or QR-decomposition. The sensors have different measurement error; this
can be implemented in the problem formulation as a weight, w, on each sensor. How this
weight is determined does not follow any standard procedure. In [Sod00] an estimated
accuracy of the sensor is used, which can be modified by the plant operation engineers on
the basis of past measurements and their experience and knowledge about the considered
instrument. Define a matrix W, where the weights w is put in the diagonal. Now we can
formulate the weighted least squares problem,
I
min Wii ( y i xi )
x ,u
i =0
Eq. 0.4
s.t. A1 x + A2 u = 0
Where x and u are computed to minimize the objective function value. In a one-sample
approach it is difficult to distinguish a large random noise from gross error. Data
reconciliation for the present vector y in Eq. 0.4 can be improved by computing a time series
70
of samples. A time window makes more data available for the problem solving and noise
can efficiently be smoothed. The cost for the decreased measurement error is increase in
computational load, which at least is proportional to the number of time steps in the time
window. The DR problem formulation in Eq. 0.4 can be extended with a time window, T:
T
min Wt ,ii ( y t ,i x t ,i )
x ,u
t = 0 i =0
Eq. 0.5
s.t.
A x+ A
t =0
u = 0 , t = 1, , T.
The solution x and u from Eq. 0.5 is ready to use for applications depending on measured
data as, plant optimization, control purposes, diagnostics, economical and environmental
reports.
We continue on the example used throughout this thesis beginning with classification
and ending with data reconciliation. Earlier a gross error was detected in the data set and
was isolated to sensor 5. The size of the gross error was estimated to -2.97 kg/s. Use the
data set given in
Table 1 and correct the measured value in sensor 5 with the estimated gross error. Solve
the data reconciliation problem Eq. 0.5 for the corrected data set in
Table 1. Observe that there are no unmeasured mass flows in this case.
S2
S3
S4
S5
S6
S7
1.9934
2.0062
4.0233
0.9807
5.0696
3.0382
8.1164
2.0315
2.0170
3.9402
0.9887
4.9836
3.0055
7.8238
1.9589
1.9771
3.9818
0.9987
5.0722
3.0144
7.9316
1.9978
2.0097
3.9323
1.0165
5.084
3.0143
7.8115
2.0312
2.0110
4.0403
0.9796
4.9512
2.9813
7.8665
2.0333
2.0339
3.9903
1.0228
4.8818
3.0555
8.1078
1.9944
2.0121
4.0903
1.0070
4.9068
2.9779
7.8751
71
1.9991
1.9909
3.9927
1.0055
4.8628
2.9721
8.0434
1.9675
2.0121
3.9492
1.0044
4.9715
2.9947
8.0166
2.0442
1.9842
3.9804
0.9904
4.9479
2.9679
7.9577
Table 1. Sensor readings with corrected gross error in sensor 5 equal to -2.97 kg/s.
In this case the least squares problem as in Eq. 0.2 is possible to solve analytically,
resulting in the vector x = [2.0119 2.0247 4.0367 1.0126 5.0493 3.0526 8.1020].
Define the residual r, as:
r = x y true
Eq. 0.6
where x is the vector of estimated values and ytrue is the flow vector without noise and
gross errors. What effect does the data treatment have on the errors and noise in the
measured data? Two-norm of the residual vector r is a measure of how far the true and
estimated values are from each other. The two-norm is denoted r . By computing the twonorm, it is possible to follow the contribution of each data treatment step in
Table 2.
Comments
Two-norm of r
True measurement
3.0722
3.0722
Isolation
0.1431
Data reconciliation
0.1336
Initialization
72
To simulate a measurement vector, noise and gross error is added to the true measurement.
The three first rows illustrate the two-norm for true vector, vector with only 5% noise and
vector with gross error of 3 kg/s. The gross error was detected by the global test. It was
isolated to signal 5 and estimated to 2.97 kg/s in the isolation algorithm. After signal 5 was
corrected by 2.97 kg/s (and thereby the gross error was deleted) the two-norm decreased
from 3.0722 to 0.1431, a significant improvement. However, the added noise still remains.
Using a least squares method for data reconciliation, the noise is reduced and the two-norm
is further lowered to 0.1336, which is the remaining error. This value is lower than any of
two-norms for the simulated signals. This result implies that the treated measurements,
output of the data treatment, are closer to the true measurement vector than input
measurement vectors. Thus, the data treatment has in this example shown to improve the
measurements on the whole.
References
Avelin A, Jansson J, Dahlquist E. Use of Modelica For Multi Phase Flow In Complex systems,
With application For Continous Pulp Digester. APMMCT, Ukraine, Khmelnitsky, 2005
Crowe C.M., Campos Y.A.G., Hrymak A., Reconciliation of process flow rates by matrix
projection. Part I: Linear case, American Institute of Chemical Engineers Journal, vol. 29. pp
881-883, 1983.
Crowe C.M., Reconciliation of process flow rates by matrix projection. Part II: The nonlinear
case, American Institute of Chemical Engineers Journal, vol. 32. pp 616-623, 1986.
Karlsson C., Widarsson B., Dotzauer E., Data reconciliation and gross error detection for the
flue gas channel in a heat and power plant. Conference on Probabilistic Methods Applied to
Power Systems, PMAPS2004, USA, Ames, 2004.
Karlsson C., Arriagada J., Genrup M., Detection and interactive isolation of faults in steam
turbines for maintenance decision support, submitted to Journal of Modelling and
Simulation, Practice and Theory in February 2007.
Karlsson C., Kvarnstrm A, Dotzauer E., Estimation of process model parameters and
process measurements a heat exchanger example, Conference on New Trends in
Automation, Sweden, Vsters, 2006.
Sanchez M. A., Bandoni A., Romagnoli J., PLADAT A package for Process Variable
Classification and Plant Data Reconciliation. Journal of Computers and Chemical
Engineering (suppl. 1992), pages 499-506.
Sunde, S., Berg, ., Data reconciliation and fault detection by means of plant-wide mass and
energy balances, Progress in Nuclear Energy, Vol 43, no , 2003.
Wang Y., Rong G., Wang S., Linear dynamic data reconciliation: refinery application, 6th
IFAC Symposium on Dynamics and Control of Process Systems, Korea, pp 650-655, 2001.
73
2.3.2.3.1.1
INTRODUCTION
2.3.2.3.1.2
General case
We shall denote the input signals to the soft sensor by vector s, the output by a scalar x, and
model parameters by a vector . The stochastic dependence describing the soft sensor
model and the uncertainty of the soft sensor is the conditional probability density function
(pdf)
74
f X |s , ( x | s, )
(1)
where the capital letters refer to stochastic variables and small letters to their values. The
soft sensor output is the maximum likelihood value:
x = arg max f X |s , ( x | s, )
(2).
We shall consider the model parameters also as random variables. After model
identification the uncertainty in parameters is described by the joint probability
distribution
f B( 0, + ) ( )
(3).
Now let us assume that we know the pdf of parameters after (n-1)th laboratory reference
measurement xn-1 to be f B( n 1, + ) ( ) . Furthermore, let us assume that after this reference
measurement the (unknown) changes in model parameters are described with a stochastic
differential equation
d
= F ( , )
dt
(4),
where is some stochastic process. We can use (4) to solve for the pdf of at later time
instants, in particular at time when we get the nth laboratory reference measurement,
f B( n, ) ( ) (the minus sign in superscript denotes the distribution before reference
measurement is taken into account).
By using Bayes formula, by assuming that in case of several simultaneous reference
measurements the measurements are independent, and by using f B( n , ) ( ) as a priori
information, we get
N
f B( n, + ) ( ) = N 1 * f B( n , + ) ( ) * f X , S | ( xi , si | )
(5).
i =1
f B( n, + ) ( ) = N 1 * f B( n , + ) ( ) * f X |s , ( xi | si , )
(6).
i =1
( n, + )
B
actual
are
A = | f B( n, + ) ( ) < f 0 ( p 0 )
p0 =
( n, + )
B
(7)
( ) d
f B( n , + ) ( ) < f 0 ( p0 )
The new best estimates of parameters to be implemented at each step or when need for
recalibration is detected are given as the maximum likelihood values:
n = arg max f B( n , + ) ( )
(8).
cT ] .
T
f X |s , ( x | s, ) = (2 e2 ) 1 / 2 exp
( x 0 c s) 2
2
2 e
(9).
Hence
N
f
i =1
X |s ,
( xi | s i , ) =
N
N 2 * exp
0 0
2
2
] 1
cT cT
B.0
B0. 0 0
=
B.. c c
(10)
N 2 * exp ( ) T B( )
2
With
B.0 = B0T. =
1
N
1
si s i
N i
1
0 = xi
N i
1
c = xi si
N i
B.. =
(11).
Let us assume that after the nth update the distribution is normal and thus characterized by
the mean n(+) and covariance matrix n(+):
f ( n, + ) ( ) = N 3 exp ( n( + ) ) T ((n+ ) ) 1 ( n( + ) )
2
(12).
76
d
= (t )
dt
< (t ) >= 0; < (t )(t ' ) T >= D * (t t ' )
(13).
Then
(14).
Inserting (10) and (14) into (6), we see that f B( n 1, + ) ( ) is Gaussian:
(15)
with
(n++)1 = ( B + ((n+ ) + D * (t n +1 t n )) 1 ) 1
(16).
( )
(17).
with k as the dimensionality of .
The best estimate for the parameters during t [t n , t n +1 ) is
2.3.2.3.1.3
n(+).
EXAMPLE
We discuss models of plug pipe reactors in the continuous processes. In the plug reactor
the fluid is assumed to move as a plug and it is assumed that the fluid is not mixing in the
axial direction and the volumetric flow out of the reactor is the same as the flow into the
reactor.
Simple plug-flow reactor can be used e.g. as a simple model for pulp bleaching in the
PGW process. The target of the study was to estimate the amount of the total organic
carbon (TOC) after the bleaching. The TOC is needed to be estimated because high levels
cause disturbances to the process.
In the bleaching mechanical pulp, suspended in water and containing some TOC, and the
bleaching chemical are fed to the bleaching tower. In the tower the fiber and the chemical
react, increasing pulp brightness but at the same time producing TOC.
Let us note the position in the flow direction in the reactor by x. The chemical reactions
at each position are described with simple kinetic equations
dc fiber ( x )
dt
= k1 c fiber ( x ) cchem ( x )
dctoc ( x )
= k1 k 2 c fiber ( x ) cchem ( x )
dt
dcchem ( x )
= k 3 c fiber ( x ) cchem ( x )
dt
(Eq. 7)
where c fiber ( x ) , ctoc ( x ) and c chem ( x ) are the amount of fiber, TOC and chemical at position x,
respectively. The model parameters of interest are k1, k2 and k3. However, we shall assume
all material removed from fibers to turn into TOC, and thus k2 = 1.
The movement of material is described by writing the mass balance in the form:
t
tot
(Eq. 8)
t (t )
where ftot is the total volumetric flow and (t) is the residence time of material coming
out of the reactor, to be solved from (Eq. 8).
System Eq. (7-8) is solved in two steps within each time step: first the concentrations are
updated according to reaction dynamics at each discretized position and then the material
at each position is updated according to the mass flow condition. Figure 3 shows an
example of simulated total dissolved/colloidal organic carbon from a bleaching reactor.
78
In this simulation the parameters are kept constant and the input flow measurements are
assumed are exact. Before the simulation starts, the reactor inflow has been 0.97 volume
units/time unit (VU/TU) of water, 0.03 VU/TU fibers and 0.001 VU/TU Total Organic
carbon (TOC). No chemical has been fed to the reactor prior to the zero time of the
simulation. The volume of the reactor is 30 VU. Reaction kinetics is simple conversion of
fibers to TOC under the action of bleaching chemical, see below. The simulation runs
500 TU.
Three actions are made on the bleaching tower during the simulation period. At t=0
chemical flow of 1 CVU/TU is introduced into the inflow while all the other conditions
the same. At t=250 chemical flow is increased to 2 CVU/TU, and at t=450 inflow is
increased by 2/0.97.
Reference measurements and parameter updating
Figure 4 outlines the TOC and parameter estimation of the model (Eqs. 7-8) based on the
method given in Section 2.1.
79
TOCmeas
L()=
TOCpred
()
exp(-(TOCmeas
TOC ())2/(2*2))
Scan
PGWmodel
f()*L()
f()
Normalize
PGWmodel
f(TOCpred)
Whenever the TOC needs to be estimated the on-line flow condition history, is applied as
the model input, the model parameter uncertainty is increased according to the diffusion
model, (Eq. 2), and then the TOC estimate and its uncertainty taking into account
parameter uncertainty are calculated.
When reference data is obtained, the corresponding likelihood function is calculated, (Eq.
3). In calculation of the likelihood function the parameters need to be varied and model
predictions with all the parameter values calculated. Then the product of likelihood
function and the pre-existing information about parameters is calculated and normalized
as the updated parameter information established. We applied the GMM approach
outlined in section 2.2 for these computations.
Because of the characteristics of the bleaching process model parameters cannot be
estimated using only TOC reference measurement but also the reference measurement of
the outcoming bleaching chemical is required. If only the reference of the TOC is used,
the parameter distribution is spreading and the TOC estimates are useless. This is showed
in the figure 5.
80
When the reference measurements of TOC and chemical are available, the parameters
can be updated by combining the old distribution with the reference distributions. The out
coming TOC-concentration is updated combining the estimated distribution with the
reference distribution.
Figure 5. On the left the parameter distribution without the output chemical and on the
right the distribution with both the output toc and the output chemical. The real values for
the parameters are k1=0.0005 and k3=0.005.
Results from simulation
In this Section we show results from a simulation study on dynamic validation method.
We have one simulator representing the true behavior. This simulator is an
implementation of models, (Eq.7-8), but with drifting k1 and k3. In parallel with the true
behavior we are running another simulator model, predictor, that with earlier input
history estimates the current output TOC with constant values for k1 and k3. Occasionally
we get information about the true output as uncertain observations of TOC and
chemical concentration. On the basis of reference measurements, predictor k1 and k3
are updated with the methods of Section 3.2. We compare the true drifting parameters
81
k1 and k3 and their estimated values and uncertainty in parameter estimates, and the effect
of parameter uncertainty on TOC estimate uncertainty
Figures 6 and 7 represent the estimated values of the parameters k1 and k3. The upper
illustrates the true (black) and the expected value (grey) and the lower the standard
deviation of the value. Black circles represent the reference values. The expected value is
kept constant between the reference measurements. The variance increases between the
reference measurements, but decreases when a new reference measurement is available.
Figure 8 represents the estimated value of TOC. Again the upper illustrates the true
(black) and the expected value (grey) and the lower the standard deviation of the value.
Black circles represent the reference values.
Figure 6. The expected value and the uncertainty of the parameter k1. Circles show
points of reference measurements.
82
Figure 7. The expected value and the uncertainty of the parameter k3. Circles show
points of reference measurements.
Figure 8. The expected value and the uncertainty of the output. Circles show points of
reference measurements.
83
2.3.2.3.1.4
CONCLUSIONS
2.3.2.3.1.5
REFERENCES
[2] Jokinen, H,. Latva-Kyr, K., Pulkkinen P. and Ritala R. 2006. Modeling of
uncertainty: case studies on operation of papermaking process. Proceeding os 5th
MATHMOD Conference, Vienna, Austria, 8th -10th February, 2006, Argesim report 30,
Eds. I. Troch, F. Breitenecker, Session on modeling in pulp and paper industry pp. 4-1..410 (abstract p. 242)
[3] Latva-Kyr, K. and Ritala, R. 2005. Dynamic validation of multivariate linear soft
sensors with reference laboratory measurements. In: Kappen, J., Manninen, J. & Ritala,
R. (eds.). COST ACTION E36, Modeling and simulation in pulp and paper industry.
Proceedings of Model Validation Workshop. 6 October, 2005, Espoo, Finland. VTT
Symposium 238. p. 57-64.
[4] OHagan, A. and Forster J., 2004. Bayesian Inference, second ed., Kendalls
Advanced Theory of Satistics, Arnold.
[5] French, S and Rios Insua, D. 2000. Statistical Decision Theory. Kendalls Library of
Statistics. Arnold 2000. 301 s.
[6] Nabney, I.T, 2002. Netlab: Algoritms for Pattern Recognition. Springer.
84
When we want to use process data of different kind for building simulation models
or for verification of models, it is important to have input data that is as correct as
possible. If a signal is very noisy it may have to be filtered. If the values a spreading a
lot it may be that we have to make a decision on what signals are really values that
can be used, respectively what values can be considered as outliers.
When it comes to filtering this can be performed by just taking the average value of
a number of measured values, or we can use a more advanced moving window
type of filtering. Here we may take the average of e.g 10 values, but then next
timestep add the new value and remove the oldest one, before calculating a new
average. This type of method is very useful when we want to collect data on-line.
Still it is crucial not to make too strong filtering. It is common that noisy signals are
filtered so strong, that the true information is lost! If you have a very noisy signal it
may be relevant to first ask your self why the signal is so noisy. Is it e.g. a
consistency or flow meter it may be that it is placed to close to a pump outlet, and
thus a lot of bubbles are disturbing the measurement. To move the sensor some five
meters down stream may then solve the problem. Otherwise it may be of interest to
compare the result if we use different filtering, to find some condition that is not too
noisy, but still keeping the important information.
When it comes to outlier removal, this is even trickier than to do the filtering. It is
not that easy to judge if a value that is far away is it because something actually
happened in the process causing this, or if it was just a disturbance in the actual
measurement.
One way of automatic removal of outliers is to just make a decision on what
deviation from an average we shoul accept, and then remove all values outside this
window. Still, we then also should remove all other measurements at the same time
step, and not use any of these data for the model building or model verification. If
we only have a limited amount of good measurements, this may cause you a
significant uncertainty about the reliability and usability of the model, but by adding
more reliable measurements later on the model quality will be improved.
Another aspect is to have values that are not only noise, but also gives true
information. This means that you need to look at the correlation between different
variables. Ideally the different variables should be varied totally independent of
each other, but as this is normally not possible, we at least should try to get as close
as possible. We also have to vary all variables of importance, and not just a few. For
e.g. a paper machine you have to vary also the wire speed, if you want to get the
influence of this on the quality variable you want to model, e.g. some strength or
printability property. This of course complicates the task, as you normally dont
85
want to do this, but keep most important variables constant. Still, if you dont do it,
you will never ever find the optimum conditions as well! To avoid problems it can
actually be wise to make use of a simulation model to evaluate the probable impact
of the variations, if reasonable information and knowledge is available from the
actual machine, or some similar.
It has been a dream of control engineers to update models on-line in an automatic way.
Still, this is easier said than done. Mostly the problems relates to poor control of the
status of the process and the sensors. Still, if we can keep control of these in an efficient
way by the methods presented above, it is possible to update parameters in different
models and control algorithms on-line. In self-tuning adaptive controls this has been
performed for primarily linear systems [strm K-J]. Here predictions are made for a
number of timer steps ahead, and the difference between the predicted value and the
measured when we come to the actual time is used to update the parameters of the
controller using a forgetting factor lambda. In this way a slower or faster adaptation
can be achieved. This was implemented in from 1982 by ABB in the NovaTune
controller. For slow moving processes like the active sludge process in waste water
treatment this was working very well [Rudberg and Olsson, 1984] and the controller was
in operation for 13 years without any significant problems. For paper winders there was a
significant problem in the beginning, as the controller was reset to 0 for all parameter at
start, but when the parameter values from the last run were stored, the application was
working nicely also for this very on-linear application. ABB later applied this technique
also for refiner controls as part of the Auto-TMP package together with a moisture sensor
using NIR-measurements (now within Metsos scope of supply). Adaptive PIDcontrollers is a commodity today, and is possible for at least simple control loops.
When it comes to on-line adaptation of process models it has still been most common to
do the updates as a batch action, as the risk of including irrelevant data often is too high
compared to the advantage of getting the update automatically. If a process engineer
make quick look at the data very often irrelevant time periods can be removed to avoid
rebuilding the model in the wrong way.
Still, as the advancement of different diagnostic methods is proceeding, the automation of
the data handling becomes more robust, and in the future we can be prepared to see many
more automatic updates.
86
Figure 1 Illustration of how different sampling intervals can give very different
information
87
Here we can see that if we only consider the squares the trend is that the process is
operating at steady state conditions. No variations are seen. If we instead look at the
circles we see a quite strong variation. This shows that we have to define what process
variations we need to consider, and make sampling in accordance to this. It is normally
good to start having a high frequency of the sampling to learn about the process, and
from this make a judgement of what frequency would fulfil our real needs. It is not of
interest to get too much information either. Signal filtering is all about making this in a
balanced way. Not too much, not too little!
Especially when we build statistical models it is very important to keep track on the time lag
between different measurements. If we only measure all variables at the same time, but it is
e.g. 8 hours for the pulp to move through the digester, it is evident that this is needed.
One way of collecting all measurements for a certain volume of pulp is to use a plug flow
model, like the Quality Foot Print (PQF) made by ABB. Here we follow the pass of e.g. 5 ton
of pulp all the way through the fiber line to the final paper product. All new measurements
along the process line will be attached to the same volume of fibers, and thus prediction
models will be much easier to build with some robustness.
For time matching between different signals there are tools looking for patterns and
comparing curves to each other. In this way we can see if different signals are connected to
each other in some way. This is important to evaluate the performance of control systems.
Sometimes one control may give oscillations that are significantly enhanced in other parts
of the process, and not directly understood until a thorough analysis has been performed.
88
A soft sensor is a transformed signal from the process. The transformation can be as simple
as multiplying the original value with a constant value, like used in currency calculations,
but can also be as complicated as calculating the total actual running costs of the process
and comparing this to some key performance indicator. In both cases the result of the soft
sensor is recognizable to the end-user, instead of some less meaningful value like
temperature, flow or pressure.
The most viable places to use soft sensors are in the control-room of the papermaking
process and on the desk of the operations manager. All other places need detailed, inside
information of the actual process or sensor-readings and transforming the original signals
from the process will only lead to confusion or possible errors. Of course this is a shortcornered statement as, for instance, the technology department may need some sort of
performance indicator as well, but as a general rule-of-thumb soft sensors should be used as
close as possible to the decision makers, be it in the field (operators) or be it in the
operations (manager).
3.2.1 INTRODUCTION
In processes, there are variables that are difficult or even impossible to measure. If such a
variable is crucial for process control or holds somehow profitable information, there is a
definite need to measure it indirectly. The term soft(ware) sensor basically refers to a
measurement that is not taken directly from a process but that is produced in some
alternative way based on other information, typically other measurements. In process
industry, the alternative way refers almost exclusively to modelling. Soft sensors can be
utilized similarly as direct measurements in process monitoring and control. However, the
variable produced through modelling often holds information beyond the computed
numeric value. That information can be particularly useful in advanced process control
tasks, like controller adaptation or process monitoring. Several other terms, having close
89
resemblance in action, are in use: indirect measurements, smart sensors, sensor fusion.
These names, more or less, emphasize the technology applied.
Software sensor has been defined as the association of (a) sensor(s) (hardware), which
allows on-line measurements of some process variables, with an estimation algorithm
(software) in order to provide on-line estimates of the non-measurable variables, model
parameters or to overcome measurement delays (de Assis and Filho 2000, Chruy 1997).
There are several estimation techniques and four of them have been recognised to have
strong potential in the on-line estimation of bioprocesses, namely, (1) estimation through
elemental balances; (2) adaptive observer; (3) filtering techniques (Kalman filter, extended
Kalman filter); and (4) artificial neural networks (ANN). A review of soft sensor applications
is also available in McAvoy (2002).
A generalized model for smart sensors is given in IEEE 1451 standard. It complements the
transducer with a smart transducer interface module (STIM), which in turn communicates
with a network-capable application processor (NCAP) over a transducer-independent
interface (TII). It is the basis of NASAs intelligent rocket test facility (IETF, (Schmalzel et al.
2005)). Their approach is based on hierarchical systems approach, largely autonomous
sensors, Gensyms G2 software as an expert system development environment, and network
structures.
According to Luo et al. (2002), multi-sensor fusion and integration refers to the synergistic
combination of sensory data from multiple sensors to provide more reliable and accurate
information. The potential advantages of multi-sensor fusion and integration are
redundancy, complimentarily, timeliness, and the cost of the information. It can reduce
overall uncertainty and thus serve to increase the accuracy and reliability in the case of
sensor error or failure.
According to Juuso (2004), intelligent analysers are software sensors combining on-line
analysers and measurements to predict output, and detect input changes or trends. They
are used in quality prediction and control and in applications in the detection of operating
conditions.
In Control Engineering laboratory, indirect measurements have been developed and used
for a long time originating from 1980s. This paper reviews some of the applications
developed for the pulp and paper industry, together with a short general survey of methods.
3.2.2 METHODS
A good collection of modelling methods has found use in soft sensors. According to Luo et al.
(2002), four groups of methods exist: estimation methods, classification methods, inference
methods and artificial intelligence methods. A recent paper concerning the architecture and
standardisation aspects of smart sensors is available in (Schmalzel et al. 2005).
90
The simplest way to implement soft sensors is to built a steady-state or dynamic estimate of
the non-measurable process variable, y, using one or more measured variables, x
y(t)=f(x, p, t),
where p includes the estimated parameters that are usually defined in the least-squares
sense. The biggest problem is the variation of parameters in time, and several solutions
using adaptive observers exist. It requires the dynamic model of the process and in its
simplest form implements in the following equations (de Assis and Filho 2000, Chruy
1997)
dxe
= g (xe , u, p e ) + K1 ( y f (xe , pe , t ) )
dt
dp e
= K 2 ( y f ( xe , pe , t ) )
dt
Above e refers to the estimate and u are the control variables. K1 and K2 are the observer
parameters.
If the system can be described with a linear model and both the system and sensor error can
be modelled as white Gaussian noise, a Kalman filter provides unique, statistically optimal,
estimates. Extended Kalman filters (EKF) can be used where the model is nonlinear, but can
be suitably linearised around a stable operating point. Using conventional notation, the
state equations are
91
K (k ) = P (k k 1) H T (k ) H (k ) P(k ^ k 1) H T (k ) + R (k )
x (k k ) = x (k k 1) + K (k ) y (k ) H (k ) x (k k 1)
P(k k ) = K (k ) H (k ) P (k k 1)
J m = ( ik ) m Dik2
k =1 i =1
where Dik denotes the Euclidean distance between the data point z and the cluster centre, m
(>1) is the fuzziness parameter and c is the number of clusters. In the beginning of the
algorithm the memberships ik are initialized with random number [0 1], so that the
following condition holds
c
i =1
=1
ik
vi =
(
k =1
n
ik
(
k =1
) m zk
ik
, i = 1,..., c
ik =
Dik m 1
i =1 D jk )
c
92
P(b | a ) =
P(a and b)
P(a )
Above P(a)>0. For pattern recognition, machine learning and classification, we can use the
formulation given in Gama and Castillo (2002). Suppose that P(Cli | x) denotes the
probability that example x belongs to class i. Any function that computes the conditional
probabilities P(Cli | x) is referred to as discriminant function. In this case it is
P(Cli | x) =
r
P(Cli | x) log( P (Cli )) + log( P( x j | Cli ))
j
Gama and Castillo (2002) have also introduced adaptive nave Bayes, iterative Bayes, and
incremental adaptive Bayes.
The belief function theory of evidence aggregates degrees of belief with new pieces of
evidence (Hong et al. 2003). When a new piece of evidence T is observed, the belief function
Bel() is updated to the conditional belief function Bel(/T). The conditional belief function
can be calculated by using prior belief functions
Bel ( S | T ) =
Bel ( S ~ T ) Bel (~ T )
.
1 Bel (~ T )
93
3.2.2.4 AI METHODS
3.2.2.5 REFERENCES
Ainali, I., Piironen, M., Juuso, E.: Intelligent Water Quality Indicator for Chemical Water
Treatment Unit. Proceedings of SIMS 2002 - the 43rd Scandinavian Conference on
Simulation and Modelling, Helsinki, Finnish Society of Automation SIMS, 2002, pp. 247-252.
Alcaraz-Gonzlez, V., Harmand, A., Rapaport, A., Steyer, J.P., Gonzlez-Alvarez, V. and PelayoOrtiz, C.: Software sensors for highly uncertain WWTPs: a new approach based on interval
observers. Water Research 36(2002) 10, 2515-2524.
An, W.S. and Sun, Y.G.: An Information-Geometrical Approach to Kernel Construction in SVM
and its Application in Soft-Sensor Modeling. Proceedings of 2005 International Conference
on Machine Learning and Cybernetics, Volume 7, Guangzhou, China,18-21 Aug. 2005, 4356
4359
Chruy, A.: Software sensors in bioprocess applications. Journal of Biotechnology 52(1997),
193-199.
Choi, DJ. and Park, H.Y.: A hybrid artificial neural network as a software sensor for optimal
control of a wastewater treatment process. Water Research 35(2001)16, 39593967.
de Assis, A.J., Filho, R.M.: Soft sensors development for on-line bioreactor state estimation.
Computers and Chemical Engineering 24 (2000), 1099-1103.
Dufour, P., Bhartiya, S., Dhurjati, P.S. and Doyle III, F.J.: Neural network-based software
sensor: training set design and application to a continuous pulp digester. Control
Engineering Practice, 13(2005) 2, 135-143.
Gama, J. and Castillo, G.: Adaptive Bayes for user modeling. Eunite Annula Symposium 2003.
Albufeira, Portugal, September 19-21, 2002, 6 p
Haataja, K., Leivisk, K. and Sutinen, R.: Kappa-number estimation with neural networks. In:
Proceedings of IMEKO World Congress, Finnish Society of Automation, Tampere, Finland,
Volume XA, pp. 1 - 5.
Hadj-Sadok, M. Z. and Gouz, J. L.: Estimation of uncertain models of activated sludge
processes with interval observers. Journal of Process Control, 11(2001)3, 299-310
94
Hong, X., Liu, W. and Scanlon, W.: Integrating belief functions with model-based diagnosis
for fault management. Eunite Annual Symposium, Oulu, Finland, July 10-11, 2003, 6 p
Jrvensivu M., Juuso E. and Ahava O.: Intelligent control of a rotary kiln fired with producer
gas generated from biomass. Engineering Applications of Artificial Intelligence 14(2001),
629-653.
Juuso, E.K.: Integration of intelligent systems in development of smart adaptive systems.
International Journal of Approximate Reasoning 35 (2004), 307337.
Leivisk, K.: Kappa number prediction with neural networks. Control Systems 2006,
Measurement and control Applications for the operator. Tampere, Finland, June 6-8, 2006,
pp. 135-140. ISBN 952-5183-26-2.
Leivisk K. and Juuso E.: Modelling of Industrial Processes Using Linguistic Equations: Lime
Kiln as an Example. In Proceedings of the Fourth European Congress on Intelligent
Techniques and Soft Computing -EUFIT'96, Aachen, August 28 - 31, 1996 (H.-J.
Zimmermannn, ed.), volume 3, pp. 1919-1923, Aachen, 1996. Verlag und Druck Mainz.
Leivisk K., Juuso E. and Isokangas A.: Intelligent Modelling of Continuous Pulp Cooking. In
Leivisk K., (editor): Industrial Applications of Soft Computing. Paper, Mineral and Metal
Processing Industries. Physica-Verlag, Heidelberg, New York, 2001, 147-158.
Luo,R.C., Yih, C-C., Su, K.L.: Multisensor Fusion and Integration: Approaches, Applications,
and Future Research Directions. IEEE Sensors journal, 2(2002)2, 107-119.
Mc Avoy, T.: Intelligent "control" applications in the process industries. Annual Reviews in
Control, 26(2002)1, 75-86.
Murtovaara S., Juuso E. K., Sutinen R.: Fuzzy Logic Detection Algorithm. In Mertzios B.G. and
Liatsis P., (editors): Proceedings of IWISP96 - Third International Workshop on Image and
Signal Processing on the Theme of Advances in Computational Intelligence, 4-7 November
1996, Manchester, UK, 1996, 423-426.
Murtovaara, S., Juuso, E. K., Sutinen, R., Leivisk, and K.: Neural Networks Modelling of Pulp
Digester. In: Dourado, A. (Ed.): Proceedings of CONTROLO'98, 3rd Portuguese Conference
on Automatic Control, APCA, Portugal, pp. 627 - 630.
Rao, M., Corbin, J. and Wang, Q., Soft sensors for quality prediction of batch chemical pulping
process. Proceeding of the 1993 international Symposium on Intelligent Control, Chigago,
U.S.A., 150-155.
Sasiadek, J.Z.: Sensor Fusion. Annual Reviews in Control 26 (2002), 203-228.
Schmalzel, J., Figueroa, F., Morris, J., Mandayam, S., Polikar, R.: An Architecture for Intelligent
Systems based on Smart Sensors. IEEE Transactions on Instrumentation and Measurement,
54(2005)4, 1612-1616.
Yliniemi, L., Koskinen, J. and Leivisk K.: Data-driven fuzzy modeling of a rotary dryer.
International Journal of Systems Science, 34(2003)14-15, 819-836.
95
It is advisable to try to identify the basic functions in certain process equipment. If we look
at a screen in a digester, in the stock prep or deinking they all have basic function similar.
There is a relation between fibers or particle how they are separated and the size of the
wholes they have to pass. As a fiber web is built up, this will give the actual hole size. The
control of the size of this will be depending on shear forces at the surface, flow rate of the
particles also through the screen plate, which in turn depends on the pressure drop over the
screen and the web. Of course there are differences between a screen in a continuous
digester and a rotating screen, but the basic principles are the same, and thus the model can
be very similar.
If we look at an heat exchanger in a boiler or a digester house, the basic principles are also
the same, although one may be counter flow and the other co-current flow, and one may be
exposed to much more fouling than the other. Still, the basic functions are the same, and can
be modelled in a similar way.
When we have identified the basic principles of process equipment we need to understand
how it really behaves. What factors are impacting the operations, and roughly how much.
Here it is very valuable to discuss with suppliers, operators and process engineers. They
will probably have different views on it, as they see it from different angles.
From the discussion we can identify more in detail what parameters that need to be tuned
to fit different operational conditions.
96
Model based controls are used for multi-objective on-line optimization of processes. The
models can be statistical or physical. Normally optimization is performed and set points
sent to a number of parallel PID-controls. There also directly linked algorithms to the
control equipments like valves or pumps. The most common version of model based control
is the so called MPC, or Model Predictive Control.
We can also have a general type of model based control, which is more like a production
plan. If we have a process and make a dynamic optimization of this, we can get a number of
set points for a number of different process variables. These can be sent down to local
ramps or PID-controls, but will have only a feed forward action and not a feed back to the
control algorithm. This is normally implemented as a production plan with a time step of 15
minute up to hours, and is often more open loop control than closed loop control.
As an example an installation is implemented for a continuous digester at Korsns pulp and
paper mill in Sweden [Avelin et al 2006].
5.1.1 INTRODUCTION
5.5.1.1 BACKGROUND
Feedback control loops, in the form of individual PID loops, are ubiquitous in the pulp
and paper industry, as well as other industrial processes. The benefits of control in terms
of product quality as well as efficient running of the plant are well understood. The
advent of computer control has vastly increased the potential functionality of control, but
according to Maciejowski (2002) Predictive Control, or Model-Based Predictive Control
(`MPC' or `MBPC') as it is sometimes known, is the only advanced control technique
that is, more advanced than standard PID control to have had a significant and
widespread impact on industrial process control. According to Isaksson & Weidemann
(2006), in the context of the pulp and paper industry, it is only in the last decade that
model-based multivariable controllers have started to spread more widely. At the last
97
three Control Systems Conferences (Stockholm, Sweden, 2004; Quebec City, Canada,
2004; Tampere, Finland, 2006) successful applications of MPC were reported for
Dry weight, ash, machine speed and moisture control (Kosonen, Fu, Nuyan, Kuusisto
& Huhtelin, 2002), (Kuusisto, Nuyan & Kaunonen, 2006) as well as moisture control
with multiple actuators (Korpela & Mkinen, 2006).
Wet end consistency (Kokko, Lautala, Korpela & Huhtelin, 2002), (Hauge & Lie,
2002), (Austin, Mack, Lovett, Wright & Terry, 2002).
Bleaching (Shang, Forbes & Guay, 2004), (Dinkel, Villeforth, Mickal & Sieber,
2006) and brightening (Mongrain, Fralic, Gurney, Singh, Shand & Valle, 2004),
(Major, Bogomolova, Perrier, Gendron & Lupien, 2006).
Single array cross-directional control (Fan, Stewart & Dumont, 2002),( Shakespeare
& Kaunonen, 2002) and multiple array cross-directional control(Backstrm, He,
Baker & Fan, 2004), (Fu, Ollanketo & Makinen, 2006), (Fan, Stewart, Nazem &
Gheorghe, 2006).
Steam header pressure control (Gough, Kovac, Huzmezan & Dumont, 2002) this
paper also discusses consistency and bleaching applications.
Chip level control (Lindgren, Gustafsson, Forsgren, Johansson & stensson, 2004)
and holding tank level control (Sidhu, Allison & Dumont, 2004).
Our aim is to give a brief overview of what MPC is, why it may be beneficial to install
MPC, and what design choices are required for the successful implementation of MPC.
98
MPC was pioneered in the 70s. Two early developments were model predictive heuristic
control reported by Richalet, Rault, Testud & Papon (1978) and dynamic matrix control
reported by Cutler & Ramaker (1979). These included quadratic cost functions based on
linear models and constraint handling capability. Much of the early development of MPC
was associated with the oil industry.
Since then the use of MPC has become widespread in the chemical process industries, to
the extent that discussion is included in undergraduate textbooks such as Seborg, Edgar &
Mellichamp (2004). Many companies and consultants provide MPC technology; Qin &
Badgwell (2003) give a useful survey of current practice.
Meanwhile interest in academia has been steadily growing. This is surely inspired for the
main part by MPC's practical success. Important links have been established between
MPC and optimal control theory, and in particular LQG control. MPC can also be related
to traditional Dahlin controllers and Smith predictors. Many embellishments are proposed
to tackle nonlinear, constrained and robust control problems.
Widespread interest in the pulp and paper industry kicked off in the 90s. This was largely
a direct influence from other process control applications. However, there are two other
important inspirations for our particular industry:
The cross-directional control problem is specific to pulp and paper production (and
also a handful of other industrial processes such as plastic film extrusion and steel
rolling). The control loop requires the co-ordination of a large number of actuators
subject to constraints and it has been recognized since the 90's (Heath, 1996),
(Rawlings & Chien, 1996) that constrained MPC is an appropriate technology for this
application. See also Heath & Wills (2004) for a more recent discussion of the
suitability of MPC in this context.
MPC usually sits somewhere between low level (fast) PID controls and high level (slow)
scheduling in a control hierarchy (see Fig 1). Thus it performs a dual role:
99
MPC co-ordinates low level PID loops. These may be interacting, so MPC may be
viewed as a large scale multivariable controller.
MPC achieves steady state values that fulfil (or are designed to fulfil) some economic
goal. MPC does this via on-line optimization.
Traditionally MPC has been implemented with large sampling times. Nowadays it is
possible to solve convex optimization problems in micro-seconds; computation time is no
longer an impediment for reducing the sampling speed. Nevertheless the slow sampling
time means that the dynamics of low level loops are (for the most part) simple which
makes control design and tuning easier.
Optimization and scheduling are crucial to any large scale processing plant. Small
fractions of percentage differences in yield can translate into millions of dollars profit or
loss and it is no coincidence that MPC was pioneered in the oil industry. MPC algorithms
themselves often have an internal hierarchy of optimization, distinguishing steady state
values from transient behaviour.
The clear trend is for MPC to become the control technology of choice in all chemical
process industries, including pulp and paper. With increased availability of efficient
computation the technology will be applied to faster loops (with shorter sampling times);
it will be used to replace PID loops as well as to co-ordinate them.
Optimization
and
scheduling
Model
predictive
control
PID
loops
Figure 2: MPC is typically used to co-ordinate many lower level PID loops, with set
points determined by optimizers and schedulers.
MPC is gaining wider and wider acceptance in the pulp and paper industries. Although
fashion plays some part in this trend, there are tangible benefits that MPC brings. The
bottom line is that payback times for projects introducing MPC are often short.
100
As with any successful control scheme, MPC has the following features:
It is cost effective.
Its widespread use means it has industry support and is relatively easy to maintain.
It can be used to co-ordinate lower level PID loops and achieve optimal steady state
trade-off between loops.
Receding horizon. At each time step a sequence of future inputs (i.e. set values for the
manipulated variables) is computed, but only the first element is implemented.
This is analogous to human decision making at any particular time we may
formulate a plan that involves both immediate and future actions; but as the
scenario unfolds we update the plan and modify our actions accordingly.
Optimization. The sequence of future inputs is computed as the solution (or possibly a
heuristic approximation to the solution) of an optimization problem. The
optimization can usually be solved using standard software tools.
Prediction. The optimization problem is usually expressed in terms of predictions on
future outputs (i.e. measured variables). The relation between future inputs and
future outputs is determined via a plant model.
101
The combination of optimization and prediction implies that the cost function
corresponds to some control performance specification evaluated over some future
horizon.
The specification of the model becomes crucial: it should be sufficiently good that the
performance criterion corresponds to economic or quality requirements, but sufficiently
simple that the optimization can be solved in a straightforward manner.
In what follows we set out a simple MPC scheme based on a linear state space model of
the plant. We point out some of the alternatives that are available to enhance MPC
design.
PLANT MODEL
There are many forms of discrete plant model. We will consider a linear state space
model of the form
x k +1 = Ax k + Bu k
d k +1 = d k
y k = Cx k + d k
Here x k represents the state which may or may not correspond to some tangible
quantity, y k represents the output(s), uk represents the input(s) and d k represents output
disturbance(s).
A popular alternative to a state-space model is an input-output transfer function model of
the form
y k = G (z )uk + d k
where z is the forward shift operator. The two model types are equivalent if we put
G (z ) = C (zI A) B
1
PREDICTION MODELS
We assume the current input u k is known and the current output y k is measured.
However present states and future inputs are unknown.
One-step-ahead predictors of the states ( x k +1|k ), the disturbance(s) ( d k +1|k ) and the
output(s) ( y k +1|k ) are given by
x k +1|k = Ax k |k 1 + Bu k + J x y k Cx k |k 1 d k |k 1
d k +1|k = d k |k 1 + J d y k Cx k |k 1 d k |k 1
x k + 2|k
x k +3|k
=
=
=
=
Ax k +1|k + Bu k +1|k
Ax k + 2|k + Bu k + 2|k
M
x k + j |k
j 1
A j 1 x k +1|k + A j 1i Bu k +i |k
i =1
We have introduced predicted inputs u k +1|k ,K, u k + j 1|k which we are (for the moment) free
to choose arbitrarily.
Similarly further output predictions can be made as
103
y k + 2|k
y k + 3|k
= Cx k + 2|k + d k + 2|k
= CAx k +1|k + CBu k +1|k + d k +1|k
= Cx
+ d
k + 3|k
2
k +1|k
= CA x
M
y k + j |k
k + 3|k
= CA j 1 x k +1|k + CA j 1i Bu k +i |k + d k +1|k
i =1
With our simple disturbance model further disturbance predictions take the form
d k + j|k
= d k + j 1|k
= d
k +1|k
Notice that all the predicted values are expressed in terms of the current one-step ahead
predicted state x k +1|k , the current one-step ahead predicted disturbance d k +1|k and
predicted inputs.
COST FUNCTION
The control action is determined by minimizing a cost function. A typical cost function
takes the form
N
j =1
j =1
y
u
2
~
J k = y k + j |k rk + j|k + u k + j|k u k + j 1|k
with
u k |k = uk
The cost function penalizes both deviations of the predicted output from the set-point
and control moves. It has a number of elements:
Prediction horizon N y . Generally speaking a longer horizon ensures better performance.
Control horizon N u . Historically control horizons have been chosen to be short to
reduce computational load. If computation time is not an issue then choosing the
control horizon approximately equal to the prediction horizon is a sensible choice.
With our choice of indexing a suitable choice of N u is N u = N y 1 ; we will
104
Reference trajectory rk + j |k . This may be chosen equal to the desired set point, or may be
some filtered version of it. For some MPC schemes the design emphasis is on
choosing suitable trajectories for rk + j |k ; this is called reference governing.
With our choice of indexing, y k + j |k is independent of future inputs u k +1|k , u k + 2 k , K .
Hence minimizing J k with respect to future inputs is equivalent to minimizing
Ny
Nu
j =1
A key advantage of model predictive control is that constraints can be included in the
control objective. For example, actuator constraints of the form
u min u k + j |k u max
with j = 1,2,K, N u may be imposed. Similarly output constraints can be imposed, for
example of the form
y min y k + j|k y max
Constraints on the state can also be included.
It is important to note a distinction in category between input and output constraints.
Input constraints can be guaranteed. Output constraints may be violated if the prediction
model is insufficiently accurate. On a similar note, model predictive controllers often
distinguish hard and soft constraints, where hard constraints are guaranteed but soft
constraints may be violated where necessary.
There are two common techniques associated with output constraints:
1. Zone. The output is constrained (with either a hard or soft constraint) to lie within
some fixed bounds, but otherwise there is no set point.
2. Funnel. The output is constrained to lie between some bounds, which converge
over the horizon to the set point.
105
where
y k + 2|k
Yk = M
y k + N |k
y
u k + 2|k
and U k = M
u k + N 1|k
y
and where
CA
CA2
=
M
N y 1
CA
CB
CAB
CB
and k =
M
O
N y 2
N y 3
B CA
B L CB
CA
J k = U k HU k + 2U k f + {terms independent of U k }
T
where
rk + 2|k
dk +1|k
Rk = M and Dk = M
rk + N +1|k
dk +1|k
y
and also
1 L
1
1 1
0
and E =
M =
M
O O
1 1
and where
106
H = T + M T M
and
f = T x k +1|k + T (Dk Rk ) M T Eu k
u max
I
I U k u
min
where
u max
u max
u min
= M
and u min = M
u max
u min
y max
I
I Yk y
min
with y max and y min defined similarly. This becomes
y max I
I
I kU k y I x k
min
General linear constraints including all input, output and state constraints, can be
written in the form
AU k b
for some matrix A and some vector b .
U k*
arg min U kT HU k + 2U kT f
s.t.
AU k b
Uk
)
107
This is in the form of a convex quadratic program with H fixed, but f dependent on the
measured data. With only input constraints A and b fixed, but with output or state
constraints, they may vary with the data. Such a quadratic program has many nice
properties. In particular:
1. The map from f to U is continuous. If the cost is replaced by a linear cost the
resultant optimization is a linear program which often results in nasty
switching.
2. Since H is fixed it may be computed off-line. The matrices H and A are
usually highly structured.
3. Quadratic programs may be solved efficiently using either interior point
algorithms or active set algorithms.
The presence of state or output constraints means the control algorithm should
include a check for feasibility (and a strategy for dealing with infeasibility). Soft
constraints, nonlinear models or nonlinear constraints usually require more intensive
computation. Generally speaking it should be possible to solve an optimization if it is
convex. Many controllers compute only a suboptimal solution.
PRINCIPLE OF RECEDING HORIZON CONTROL
uk +1 = u k*+1|k
This value is delivered to the plant as the next input.
MPC has traditionally been driven by industrial requirements and developed in tandem
by industrialists and academics. Meanwhile there has been growing interest in the
development of the theory of MPC within academia. This is in part due to the success of
MPC and a recognition of its further potential, and in part a recognition that there are
many aspects that we simply do not understand. A good summary of our current
understanding can be found in the survey by Mayne, Rawlings, Rao & Scokaert (2000).
Nominal stability. First and foremost, there are no guarantees of nominal stability
(stability under conditions where the model exactly fits the true plant) for our simple
MPC construction example. This is despite the usual experience that such schemes can be
tuned to work well. It is possible to test for nominal stability when there are no
constraints, but even here the relation between the tuning parameters (such as horizon
108
length, choice of observer and value of the control weighting) and the performance is not
straightforward.
If the horizon is allowed to go to infinity and the full state is known (or there are no
constraints) then the controller is guaranteed to be stable. Furthermore, under certain
circumstances, it is possible to augment a finite horizon cost function J k with a weight
on the terminal state x k + N y |k such that the control action is equivalent to that with an
infinite horizon. This is known as quasi-infinite horizon MPC, and the augmentation to
the cost function is known as a terminal weight. Generally such a terminal weight makes
the control tuning more straightforward.
Unfortunately the result only holds when the steady state values are away from the
constraint boundaries this is often not the case in practice. Furthermore the required
horizon length varies with both initial conditions and the steady state values. Thus in
practice the terminal weight does not guarantee even nominal closed-loop stability.
Many researchers propose including a terminal constraint that ensures the final predicted
state x k + N y |k is sufficiently close to the steady state value. Theory shows this gives
nominal stability, but in practice it can de-tune the controller and be highly detrimental to
performance, especially for short horizons. The idea is analogous to applying deadbeat
control.
Robustness. There are almost no useful robustness results for MPC. In practice, of
course, most MPC schemes work very well despite the presence of significant model
errors. It is possible to guarantee robustness for certain MPC schemes where there are
only input constraints, including the cross-directional control problem (Morales & Heath,
2006).
There are many proposals in the academic literature for augmenting MPC in order to
guarantee robustness. A typical example is to replace the simple optimization with a minmax optimization (so that the worst-case scenario is taken into account). All such
schemes add considerably to the computation. At this time there is no consensus that such
schemes improve performance, or even improve robustness in practice.
Feasibility. If there are hard state constraints then it is possible for there to be no feasible
solution to the MPC optimization problem at a specific interval k . Any practical MPC
scheme must either have a hierarchy of constraints (where constraints may be dropped
when necessary) or a contingency strategy for when infeasibility occurs.
Nonlinearities. It is possible to incorporate a nonlinear model into the MPC prediction
and optimization. Similarly it is possible to include nonlinear constraints, and even make
the cost function nonlinear. All such modifications increase the necessary computation,
often considerably.
109
Models and simulation. The most important requirement for an MPC system is an
appropriate model. Often this is acquired from plant data via system identification
experiments (termed an empirical model). Nevertheless it should be consistent with both
physical principles and any simulations available from existing plant models. For some
applications it may be advantageous to use a model based on physical principles (a
mechanistic model) directly. Hauge, Slora & Lie (2005) discuss the advantages of
mechanistic models in MPC roll-out. In Table tab:mechVSemp we compare the relative
advantages of empirical and mechanistic models for MPC.
The model should also be transparent enough to be updated over the years as plant
conditions change. There is a useful discussion of control loop performance monitoring
by Mitchell, Shook & Shah (2004).
Tuning. The lack of theoretical results for MPC means controllers should be verified by
extensive simulation. It is good practice to ensure unconstrained performance is
(robustly) stable with good performance.
In general the horizon should be chosen to be long, and the control weighting should
be chosen large if necessary.
Computation time. Historically the application of MPC has been limited to loops with
slow sample times due to the heavy computational requirement. Nowadays this should no
longer be an impediment successful application of MPC to fast vibrating systems with
5kHz sampling rates have been reported (Wills, Bates, Fleming, Ninness & Moheimani ,
2005).
Nevertheless it may be necessary to perform the computations on a dedicated processor,
or still limit the application to slow sampling times if the computation is to be performed
on existing shared hardware.
Good design. Although MPC is a useful and versatile tool, it should not be seen as a
substitute for good control design. Success remains dependent on good control and plant
insight for example observing which loops may be effectively be decoupled, or which
variables are best controlled by feedforward or feedback loops.
MPC algorithms are usually implemented by vendors or consultants. Even if the
controller has been developed internally within a company, it is likely to be implemented
by engineers from a separate division from the process operations. Vendors usually spend
a period of time (typically between two days and a week) making tests on the plant. They
will be seeking the following information:
1. Which manipulated variables should be associated with which measured
variables? Tools such as the Bristol gain array method are used.
110
2. Simple empirical dynamic models. Early MPC algorithms were based directly on
step or impulse response models. Nowadays transfer function models are often
(but not always) used. Step input tests remain typical, but other excitation signals
may be used.
3. What are the constraints on the manipulated variables? What are the state and
output constraints necessary or desirable for the operation?
4. Over what range are linear models with fixed parameters appropriate?
There is always a temptation to use MPC's constraint handling techniques to satisfy
criteria that would otherwise be achieved by careful closed-loop design. Generally better
performance is achieved by not being over-reliant on such functionality of MPC.
Operator and engineer re-education. It goes without saying the success of any new
technology is dependent on its acceptance by both plant operators and plant engineers.
Table 1.1: Mechanistic versus empiric models. Reproduced from Hauge (2003).
Properties
Mechanistic
Empiric
yes
no
yes
no
low
high
high
low
no
yes
low
high
yes
yes
Applicability to design
yes
no
Extrapolation properties
good*
bad
yes
no
Complex
yes (non-linear)
no (often linear)
Simulation
long/difficult
quick/easy
yes
no
111
5.1.7 CONCLUSIONS
MPC is already in widespread use throughout the process industries, and is gaining more
and more acceptance in the pulp and paper industries. We have given a brief overview of
the state-of-the-art, both in terms of basic MPC algorithms and where it is being applied
in pulp and paper industries.
We envisage that MPC's use will grow, and it will soon be seen as a standard technology.
Furthermore we expect that its role will evolve and expand in future years. With
reference to Figure Fig:hierarchy:
It is likely that MPC will begin to replace low level PID loops. This will be driven by:
Computation time ceasing to be a limiting factor for application of MPC.
More widespread understanding of MPC tuning rules (just as PID tuning rules are
now understood by operators).
The recognition that MPC can bring improved performance and reliability even at
the level of low-level controls.
We expect to see MPC being used at the level of optimization and scheduling. MPC
should become an enabling tool for more understanding and improved dialogue
between production planners, control engineers and plant operators.
There is already a differentiation between linear MPC for standard operations and
nonlinear MPC for batch processes. There will be a drive to improve and tailor plant
models, as these are key to the successful implementation of MPC. In this context we
also envisage increasing integration of MPC and simulation software to provide decision
support for operators. Similarly the monitoring of control loops will become high
priority.
112
We recommend the survey paper of Qin & Badgwell (2003) as a useful introduction to
MPC. There is also a useful chapter in the undergraduate textbook by Seborg, Edgar &
Mellichamp (2004).
The book by Prett & Garca (1988) is by now quite old and some of what it says about
MPC is out of date. Nevertheless it is probably the best book on MPC from the
perspective of industrial requirements and motivation. More recently there have been a
number of books on MPC. The best of these is probably by Maciejowski (2002).
5.1.8 BIBLIOGRAPHY
Alexandridisa, A., Sarimveisa, H., Angeloub, A., Retsinab, T. & Bafasa, G. (2002). A model
predictive control scheme for continuous pulp digesters based on the partial least
square (PLS) modeling algorithm. Control Systems 2002. Stockholm, Sweden, June
3-5.
Austin, P., Mack, J., Lovett, D., Wright, M. & Terry, M. (2002). Improved wet end stability of a
paper machine using model predictive control. Control Systems 2002. Stockholm,
Sweden, June 3-5.
Backstrm, J., He, P., Baker, P. & Fan, J. (2004). Advanced multivariable CD control delivers
improved quality, runnability and probability for different CD processes. Control
Systems 2004. Quebec City, Canada, June 14-18.
Bhartiya, S. & Doyle, F. J. (2002). Modeling and control of grade transition in a continuous
pulp digester. Control Systems 2002. Stockholm, Sweden, June 3-5.
Bling, J., Forsman, T. M. I. & Lnnberg, B. (2002). Modeling, identification and
multivariable control of wood grinding. Control Systems 2002. Stockholm, Sweden,
June 3-5.
Bozin, A. S. & Austin, P. C. (1995). Dynamic matrix control of a paper machine benchmark
problem. Control Engineering Practice 3(10), 1479-1482.
Chmelyk, T., Ip, T., Sheehan, C. & Korolek, J. (2004). White liquor pressure filter effciency
improvements through the application of model predictive control. Control
Systems 2004. Quebec City, Canada, June 14-18.
Chow, C. M., Kuznetsov, A. G. & Clarke, D. W. (1995). Application of generalised predictive
control to the paper machine benchmark. Control Engineering Practice 3(10), 14831486.
Cutler, C. R. & Ramaker, B. L. (1979). Dynamic matrix control a computer control
algorithm. AIChE national meeting, Houston, Texas.
113
Dinkel, M., Villeforth, K., Mickal, V. & Sieber, A. (2006). Measurement and control strategies
for bleaching. Measurement and control applications for the operator. Tampere,
Finland, June 6-8.
Fan, J., Stewart, G. E. & Dumont, G. A. (2002). Model predictive cross-directional control
using a reduced model. Control Systems 2002. Stockholm, Sweden, June 3-5.
Fan, J., Stewart, G. E., Nazem, B. & Gheorghe, C. (2006). Automated tuning of multiple array
cross-directional model predictive controllers. Control Systems 2006.
Measurement and control applications for the operator. Tampere, Finland, June 68.
Fu, C., Ollanketo, J. & Makinen, J. (2006). Multivariable CD control and tools for control
performance improvement. Control Systems 2006. Measurement and control
applications for the operator. Tampere, Finland, June 6-8.
Fu, C., Ye & Dumont, G. A. (1995). A generalized predictive control design for the paper
machine benchmark. Control Engineering Practice 3(10), 1487-1490.
Gough, B., Kovac, S., Huzmezan, M. & Dumont, G. A. (2002). Advanced predictive adaptive
control of steam header pressure, saveall consistency, and reel rightness in a TMP
newsprint mill. Control Systems 2002. Stockholm, Sweden, June 3-5.
Hagberg, M. & Isaksson, A. J. (1995). Preface to the special section on benchmarking for
paper machine MD-control. Control Engineering Practice 3(10), 1459-1462.
Hauge, T. A. (2003). Roll-out of Model Based Control with Application to Paper Machines.
PhD thesis, Norwegian University of Science and Technology, and Telemark
University College, Norway.
Hauge, T. A. & Lie, B. (2002). Model predictive control of a Norske Skog Saugbrugs paper
machine: Preliminary study. Control Systems 2002. Stockholm, Sweden, June 3-5.
Hauge, T. A., Slora, R. & Lie, B. (2005). Application and roll-out of infinite horizon MPC
employing a nonlinear mechanistic model to paper machines. Journal of Process
Control 15(2), 201-213.
Heath, W. P. (1996). Orthogonal functions for cross-directional control of web forming
processes. Automatica 32, 183-198.
Heath, W. P. & Wills, A. G. (2004). Design of cross-directional controllers with optimal
steady state performance. European Journal of Control 10, 15-27. With discussion
pp. 28-29.
Isaksson, A. J., Hagberg, M. & Jnsson, L. E. (1995). Benchmarking for paper machine MDcontrol: Simulation results. Control Engineering Practice 3(10), 1491-1498.
Isaksson, A. J. & Weidemann, H.-J. (2006). Control systems for pulp and paper production
challenges and next steps. Control Systems 2006. Measurement and control
applications for the operator. Tampere, Finland, June 6-8.
114
Kokko, T., Lautala, P., Korpela, M. & Huhtelin, T. (2002). Comparative study of consistency
control strategies and algorithms. Control Systems 2002. Stockholm, Sweden, June
3-5.
Korpela, M. & Mkinen, J. (2006). Moisture control with multiple manipulated variables.
Control Systems 2006. Measurement and control applications for the operator.
Tampere, Finland, June 6-8.
Kosonen, M., Fu, C., Nuyan, S., Kuusisto, R. & Huhtelin, T. (2002). Narrowing the gap
between theory and practice: Mill experiences with multivariable predictive
control. In Control Systems 2002. STFi and SPCI, pp. 54-59. June 3-5, Stockholm,
Sweden.
Kuusisto, R., Nuyan, S. & Kaunonen, A. (2006). Multivariable predictive grade change
controller. Control Systems 2006. Measurement and control applications for the
operator. Tampere, Finland, June 6-8.
Lindgren, T., Gustafsson, T., Forsgren, H., Johansson, D. & stensson, J. (2004). Model
predictive control of the chip level in a continuous pulp digester, a case study.
Control Systems 2004. Quebec City, Canada, June 14-18.
Maciejowski, J. (2002). Predictive Control with Constraints. Prentice Hall, Harlow, England.
Major, D., Bogomolova, O., Perrier, M., Gendron, S. & Lupien, B. (2006). Control and
optimization of hydrosulphite brightening for a three-adition point process.
Control Systems 2006. Measurement and control applications for the operator.
Tampere, Finland, June 6-8.
Makkonen, A., Rantanen, R., Kaukovirta, A., Koivisto, H., Lieslehto, J., Jussila, T., Koivo, H. N. &
Huhtelin, T. (1995). Three control schemes for paper machine MD-control. Control
Engineering Practice 3(10), 1471-1474.
Mayne, D. Q., Rawlings, J. B., Rao, C. V. & Scokaert, P. O. M. (2000). Constrained model
predictive control: Stability and optimality. Automatica 36, 789-814.
Mercangz, M. & Doyle, F. J. (2006). Mathematical modeling and model based control of a
pulp mill powerhouse. Control Systems 2006. Measurement and control
applications for the operator. Tampere, Finland, June 6-8.
Mitchell, W., Shook, D. & Shah, S. (2004). A picture worth a thousand control loops: An
innovative way of visualizing controller performance data. Control Systems 2004.
Quebec City, Canada, June 14-18.
Mongrain, A., Fralic, G., Gurney, C., Singh, S., Shand, D. & Valle, M. (2004). Model predictive
control of newsprint brightening. Control Systems 2004. Quebec City, Canada, June
14-18.
Morales, R. M. & Heath, W. P. (2006). Numerical design of robust cross-directional control
with saturating actuators. Control Systems 2006. Measurement and control
applications for the operator. Tampere, Finland, June 6-8.
115
116
Production planning can be with different time horizons. It can be for the next few hours,
for the next few weeks but also for the next years. The purpose of course is quite different.
In the first case it may be just to get the plant running producing the right amount and
quality according to the order list and priority list. In the second case we would include
need for service and maintenance as well. For the long time perspective we also have to add
need for new investments, rebuild, long term contract with suppliers and customers.
A typical example of a production plan for a paper machine may be how to run a sequence
of products with different grammage, widths and lengths of the roll. For this purpose
different optimization models have been implemented, and the most economic alternative
is used for the production plan.
A problem occurs if e.g. an important and demanding customer demands a faster deliver
than you have planned for, or something happens in the production. Normally you optimize
for a sequence, and then try to deliver as fast as possible. The reprioritization aside of the
plan is normally not done automatically, but by a decision from the production manager and
the planner together.
In the future it may be possible to re-plan on a more frequent base and also to include
optimal time for service and maintenance, as well as risks for shut-downs due to too tough
production during a limited time period, to fulfill some demanding customers needs.
We have already been mentioning how dynamic optimization can be used for both
production planning, scheduling and on-line control using model based control. In the
example below we will show how it was implemented in the EU DOTS project for Dynamic
Optimization of Paper mills.
117
6.2.1.1 ABSTRACT
The production processes are becoming increasingly complex and the responsibilities of the
operators and the engineers are widening. Tools to manage this complexity under dynamic
conditions are needed. The dynamic optimization integrated to operator and engineer
decision support has a high potential in everyday use. As a result of an EU Commission
funded research project a toolset for supporting continuous decision making has been
developed. In this study the toolset is used to reduce grade change time on a paper machine.
6.2.1.2 INTRODUCTION
The main objective of the paper machine staff is to maintain acceptable machine run ability.
In practice this means high utilization degree and machine speed as well as meeting high
quality requirements. The paper is usually produced on order, and thus frequent grade
changes are inevitable. The time spent in grade changes must be minimized in order to
maximize the production time and minimize the costs. An important constraint for the
grade change operation is to avoid web breaks due to fluctuations in key variables during
the grade change.
The minimization of the grade change time while avoiding web breaks forms a challenging
optimization problem. In this paper the objective function for optimization is formulated,
and evaluated through running dynamic simulation. As the optimizer may need hundreds or
thousands of iterations to reach the optimal solution, the simulation time in practical
applications has to be as short as possible. The simulation model was implemented in
Matlab/Simulink. To reduce the simulation time the model is as simple as possible while
being realistic.
The optimization is done with the toolset developed in an EU Commission funded research
project DOTS (G1RD CT 2002 00755). The present implementation of toolset is
within Matlab which is easy to bring to mill environments either as a full system, or as
embedded in other systems, e.g. the process analysis system KCL-WEDGE (KCL, 2004). The
DOTS toolset offers inbuilt stochastic optimization methods and utilizes Tomlab
environment (Holmstrm, 2004) as an external optimizer. In this paper we present the test
the Sequential Quadratic Programming (SQP) method provided by the Matlab Optimization
Toolbox in comparison with the stochastic methods offered by the DOTS Toolset. The
objective functions are kept unaltered so that the results shown are comparable.
118
When the model is driven by the optimizer, the response time of the simulation is essential.
All extra features like displays and irrelevant outputs are removed. The model is
parsimonious in that it captures only the features relevant for a realistic grade change, and
nothing else. The parameterization of the model must be done so that the optimizer can run
the model. The problem formulation sets requirements for the simulation model
(Pulkkinen, et al., 2004). The inputs and outputs must be chosen and communicated
correctly to get the information the optimizer requires.
The paper machine simulator developed is a combination of blocks based on the basic
elements of a paper machine: tanks, valves and controllers (Pulkkinen, et al., 2003). The
basic idea of the simulation has been to formulate a simplified model of what happens in a
paper machine when a grade change is made. The main attention has been focused on the
examination of the quality variables of the end product.
Figure 1 presents the simulation model. The simulator can be divided in three main
elements: short circulation including the wire section, the press section and the drier
section. In real life the optimal operation of the short circulation including wet end
chemistry has been ranked to be one of the hardest challenges of the papermaking process.
In the simulator the simplified short circulation model consists of the machine tank, wire
section and the white water tank. The accuracy of parameters required is not high because
of the nature of simulator use. The flow of material has been divided into four categories in
the simulator: the flow components are water, filler and both short and long fibers. The
retention at wire section is taken into account by applying component-specific, yet constant
in time, retention coefficients. The main emphasis is put on the fiber, filler and water flow
after each block of the simulator.
119
When changing to a grade with higher basis weight the moisture of the paper produced
increases if no action is made on the dryer section. Even with moisture control on at
constant set point, there is a momentary increase in moisture, as the heating of the dryers is
120
slower than the dynamics of fiber retention. An opposite variation happens during a change
to a grade with smaller basis weight. The former case is used as an example in this paper.
The idea of the optimization of the grade change is to minimize the fluctuation of the
moisture as well as the fluctuations of basis weight and filler content.
The combination of control actions that results in smallest cost value is the answer to our
optimization problem. The total cost in this case is a sum of the costs calculated for each key
variable: moisture content, basis weight and filler content. Here the cost is calculated by
squaring the variation from the grade-specific set point and integrating over the duration of
the grade change. In practice the variance is minimized while the expected values are
changed. This approach has been used in the following examples due to the better success
in optimization.
The overall cost can be expressed as
(1)
where M, B and C are weighing factors with a relation 1:1.7:0.08. The cost caused by the
variations in each quality variable is quadratic
n2
(2)
where x is trajectory of the variable and xtarget is the grade-dependent target for the variable.
n1 and n2 represent the selected start and end times of the cost calculation.
An alternative approach of cost calculation is as follows. The cost is zero when the quality is
within grade-specific quality requirements and a positive constant when the quality is not
within the specifications (quality pipe). Hence the minimum cost is reached when the time
that quality variables are outside the quality pipe is minimized.
In order to find a minimum cost we are seeking optimal set point trajectories. This leaves us
with infinite choices: the trajectories as functions of time. It is however justified to simplify
the optimization task by parameterizing the trajectories appropriately.
In practice, a step down and up again is parameterized into the moisture set point time
series. The optimization algorithm is used for finding the optimum for the three
parameters: the two timing parameters and the step size.
The timing of the grade change is fixed, so the trajectory can be formed knowing one
constant parameter. The last parameter needed is the one for filler content. The change in
filler content is stepwise and its size fixed, again only the timing parameter is needed. The
parameterization reduces the search space down to four dimensions.
121
The parameters that need to be evaluated are shown in Table 1. The initial values are given
for each optimized variable and the timing of the grade change is a constant. The
optimization algorithm then finds the combination of values of the variables that result in
the least cost.
Table 3 The evaluated parameters of the simulator
Parameter
Type
Range
Fixed
Gc
Optimized
Gc +/100
Optimized
0.01-0.05
Optimized
Gc +/100
Optimized
Gc +/100
6.2.2 OPTIMIZATION
DOTS toolset offers an easy-to-use configuration for the dynamic optimization problems. In
this paper stochastic methods have been used from the portfolio of optimization methods in
the toolset. The parameters of the problem and the algorithm are specified through a
graphical user interface in Matlab environment.
It is known that the algorithm performance is case dependent. We will show that choosing
an algorithm plays an important role in starting with a new optimization problem (Dhak, et
al., 2004). The minimum of the smooth quadratic cost function can be found by almost any
optimization algorithm, which is also shown in (Ihalainen and Ritala, 1996). In this case the
cost function is more complicated and differences in performance can be shown.
The optimization process was repeated using the stochastic methods in the DOTS Toolset
and the SQP in Matlab Optimization Toolbox while the objective was kept the same.
Perhaps the simplest way of optimizing is a blind random search of right parameters. The
parameter combination with best found cost is then selected.
Genetic algorithms (Goldberg, 1989) consist of following steps:
122
1. Start
2. Fitness
3. New population
a. Selection
b. Crossover
c. Mutation
d. Accepting
4. Replace
5. Test
The idea of the simulated annealing algorithm (Otten and van Ginneken, 1989) is described
with following steps:
1. a feasible set of movements is generated randomly
2. the standard deviation of the cost function values in the feasible set is estimated and
the estimate as a starting temperature T is chosen
3. steps 3-6 are iterated n times
4. a feasible new point according to the previous procedure is chosen
5. the new point is accepted as a current point certainly if C<0 and with probability
of:
exp(-C / T) if C>0, where C is the difference of cost function values
between the new point and previous point
6. If the new point is the overall best found so far, the best point is updated
7. the temperature is reduced by 0.01*T2 / , where is the estimate of the standard
deviation over the previous iteration path
8. If the best score has not improved during the last 50 iteration paths, the
optimization is stopped; otherwise continued from the step 3
123
Tabu Search is also an iterative stochastic procedure designed for solving optimization
problems. Tabu Search keeps a list of previously found solutions so that re-finding solutions
in subsequent iterations is prevented. SQP, the sequential quadratic programming method,
is a smooth nonlinear optimization method. It is a generalization of Newtons method for
unconstrained optimization in that it finds a step away from the current point by
minimizing a quadratic model of the problem. In its purest form, the SQP algorithm replaces
the objective function with the quadratic approximation and replaces the constraint
functions by linear approximations.
The simulation results with the initial values for parameters to be optimized are shown in
the upper right corner of the Figure 2. It can be seen that the grade change actually results
in a large peak in the moisture. When the set point of the moisture is not manipulated and
the filler content is changed at the same time as the basis weight set point is changed, the
cost caused by the grade change is 373.2 units (costs after optimization range from 31 to
35). It is obvious that with a step in the moisture set point we can stabilize part of the effect
of grade change on the moisture of the final product. The three moisture set point
parameters determine how the stabilization is done. The filler content change time is
chosen so that the all the quality variables behave optimally. The lower right corner of the
Figure 2 presents the fluctuation of the moisture after the optimization
An example of the behavior of the key variables and the set point trajectories in the
optimized grade change situation are shown in the left side of the Figure 2. It must be
remembered that the results depend on the parameters given for the algorithm, and the
comparison between the algorithms was carried out with default values.
Figure 2 Example of an optimization result in the left side of the figure. The behavior
of the key variables and their set points are shown in the following order: moisture,
124
basis weight and filler content. The grade change starts at 75500. The right side of the
figure presents the fluctuation of the moisture before (upper corner) and after the
optimization.
Different methods of the toolset have been used in this optimization task. The simulator and
functions used with DOTS Toolset have also been used with an SQP function in the
Optimization Toolbox. It should be noted that the SQP algorithm gives local minima and the
result depends on the initial values given for the function/algorithm. The results of each
method used are shown in Table 2. Table 2 shows that the best result is achieved with
genetic algorithm.
Table 4 The optimization results
Method /
Tabu Search
Simulated
annealing
Blind Random
Genetic
algorithm
SQP
Grade change
75000
75000
75000
75000
75000
75015
74980
74985
74990
75005
74965
74965
74975
74980
74975
75070
75055
75055
75050
75050
Variable
33.90
31.57
31.32
34.34
Noticeable in the table is also the good result of the blind random search. This indicates,
that the search space is rather flat, which gives no grip for the more advanced methods.
However, as the Figure 3 shows, the optimization reduces the fluctuation effectively. Similar
effect can also be seen in the behavior of basis weight and filler content. Reduced variance
improves the stability of the process and speeds up the grade change.
6.2.3 CONCLUSIONS
By observing the cost values achieved for solutions with each method we conclude that the
best result in this grade change optimization is achieved by using genetic algorithm. The
genetic algorithm provides best results also for the quality pipe cost function. Despite the
fact that the methods give results close to each other, the genetic algorithm with proper
parameters should be preferred because of better optimized cost and lower number of
iterations. However, the performance of the methods depends on the cost function used and
125
the parameters of the algorithms. The algorithms have been applied with default
parameters, if the parameters were tuned, the results could be somewhat different.
Compared with the functions in Optimization Toolbox, DOTS Toolset offers an easy way of
solving an optimization problem. The graphical user interface makes the toolset somewhat
more desirable tool to use.
6.2.4 REFERENCES
Dhak, J., Dahlquist E., Holmstrm K., Ruiz J., Belle J., Goedsch F. (2004), Developing a Generic
Method for Paper Mill Optimization, Control Systems Conference, Quebec City.
Dhak, J., Dahlquist E., Holmstrm K., Ruiz J. (2004),Generic Methods for Paper Mill
Optimization, Simulation and Process Control for the Paper Industry, Munich.
Goldberg, D. E (1989). Genetic Algorithms, Addison-Wesley.
Holmstrm K., Edvall M. (2004), The Tomlab Optimization Environment, in Kallrath J. (ed)
Modelling Languages in Mathematical Optimization, Chapter 19, Kluwer Academic
publishers.
Ihalainen, H., Ritala R. (1996), Optimal Grade Changes, Control Systems Conference.
KCL, http://www.kcl.fi/wedge
Otten, R. H. J. M, van Ginneken, L. P. P. P. (1989), The Annealing Algorithm, Kluwer
Academic Publishers.
Pulkkinen P., Ritala R., Tienari M., Mosher A. (2004), Methodology for dynamic optimisation
based on simulation, Simulation and Process Control for the Paper Industry, Munich.
Pulkkinen P., Ihalainen, H., Ritala R. (2003), Developing Simulation Models for Dynamic
Optimization, SIMS Conference, Mlardalen.
126
We already have been discussing diagnostics in some forms in previous chapters. Data
reconciliation is a method that may be used for diagnosing sensors in a network. Variance
analysis may be a method for evaluation of noise level of a sensor or a rotating machine.
Development of heat transfer coefficient by a material balance over a heat exchanger is
another method. Loop tuning diagnostics is a fourth type of diagnostics.
After these types of diagnostics we want to give the process engineers, maintenance people
and operators an evaluation of what is coming out from the different diagnostic systems.
This can be performed e.g. through a Bayesian net, giving relative probability for different
types of faults, or e.g. Case Based reasoning (CBR) giving a distinct value fault or no fault.
Other methods also can be used, like look up tables etc.
Related to this we also have maintenance on demand. Also what-if scenarios may be used
for decision support, to determine how to operate the process in best possible way, e.g. to
avoid bottle necks.
Commercial diagnostics systems of different kinds are available for pulp and paper
industry, like from the major vendors like ABB, Honywell, Metso and Siemens, as well as
specific tools like WEDGE from Savcor Oy.
Early warning systems are something not used so much yet, but will be used much more in
the future, probably. One example of this is given under applications for utilities later on
in the book. Here an early warning system was implemented to determine probable faults in
a recovery boiler, like leakage in tubes in the combustor part respectively in the ECO-part,
as well as different sensor faults. The implementation was performed at Rottneros Vallvik in
Sweden, where a gas explosion followed by a steam explosion occurred in 1998.
Other types of early warning/diagnostics/decision support systems can be exemplified with
the system implemented by e.g. ABB for analysis of vibrations in rotating machines. By
having wire stretch meters mounted on motors or drives the vibrations can be determined
and analyzed. By variance analysis a specific value can be determined and warnings sent as
a certain level is passed. If nothing is done, the level will increase and a new, stronger
warning is sent to the operator displays. At a third level an alarm is sent to the alarm
system.
Depending on where the sensor is located different most probable cause of the vibrations
are given to the operator/maintenance people, so that they can take the most adequate
action.
127
128
129
9.1.1 INTRODUCTION
Examples where tables are used is when AutoCook is used to change production or
change wood species without producing more off-spec pulp than necessary. Each wood
specie/pulp grade has its own set of pulp tracking parameters.
The trend today is to optimize the whole mill not only towards production and quality,
but also towards for example minimization of energy and chemical consumption, and
effluents. Most of the models are first principles process model, which is physical
models. These are used to simulate the process as part of the optimization. With a good
process model it is possible to implement a more advanced control system like model
predictive control (MPC) and 2D or 3D dynamic online models for diagnostic purpose.
E.g. hang ups and channeling are of interest to identify in a continuous digester.
For the on-line simulation of the whole mill ABB has developed a Pulp and Paper Online Production Optimizer and this optimizer works like an overall mill management
system giving set points to each process section. The models will be on the same
software platform as the more advanced models, for example the digester model
presented in this paper.
130
On the overall mill level we want to optimize and control the chemical balance in a complex
pulp mill. The first application is the control of Sulfur in an integrated Kraft and NSSC
(Neutral Sulfite) pulp mill. This type of integrated recovery systems is called cross
recovery, as sulfur is passed over between the two lines. It is important to adjust the
dosage of sulfur to adjust the sulfidity, so that there is enough but not too much, and to keep
the balance between different parts of the plant. This has been done by physical models that
have been combined to form a plant model. This model is then used to make a production
plan for keeping the sulfur balance at its optimum. Implementation has been done recently
at a Billeruds Gruvn Mill in Sweden, which typically produces 630 000 metric tonnes of
paper a year.
Modeling
The modeling is based on a description of all components in the system as objects. The
object type groups are streams, production units, measurements and calculated properties
in streams.
Some examples of production units are digester, pulp tank, black liquor tank, lime kiln
and paper machine. Stream types may for example be pulp, white liquor, black liquor or
wash liquor. Examples of measurements are pulp flow, tank level, sulfidity, effective
alkali, reduction efficiency, black liquor density or alkali charge.
The streams are described by the total flow and by the concentrations of the different
components that are relevant for each stream. Examples of stream components used are,
fiber, dissolved organic solids, hydroxide, sulfide, carbonate, sulfate, sulfur. These
relations may be considered as a dynamic mass balance and are a system of differential
and algebraic equations of the form
131
The levels in buffer volumes are changing as a function of changes in the flow in and out.
Tanks are rigorously modelled assuming ideal mixing as
dM
= Fin Fout
dt
dcout Fin
=
(cin cout )
dt
M
where Fin and Fout denotes incoming and outgoing total flows. M denotes total mass in
tank, cin and cout denotes incoming and outgoing concentrations. One equation like the
second is used for each component. More about this is described in Persson et al, 2003
[1]
The optimization of the plant is aiming at minimizing:
Variation in sulfidity
Cost of sodium hydroxide and sulfur usage
Production losses due to up and down stream disturbances
Variation in sodium stock
Variation of the distribution of sodium between white and black liquor
The optimization is done by estimating what value each variable should have during each
time step during the time horizon we want to control. We are then not allowed to pass
limits like high or low level in tanks, feasible flows in pumps and valves etc. This is
where we really need a computerized system to keep control of all these constraints. This
optimization schedule is then updated on a frequent basis, to handle new situations
coming up like need for changing priority in the production at the paper mill; some
equipment has to be replaced etc. Each time we make a new optimization calculation we
need to have a starting point for where we are right now. This is not as clear as you might
believe, as the process seldom is as stable as you would like it to be. To handle this a
technique called Moving-horizon estimation (MHE)[2] can be used.
This means that we include a number of time steps and make an average for these. For
the next time step we then take out the oldest, and add the newest, and make a new
average of these. This is repeated in the same way every new time step. In this way we
can filter out noise from both sensors and actual upsets in the process. This is normally
positive as we do not use the optimization for doing changes on the very short time basis.
We then instead get a better view of stable values, so that the optimization is done from
more correct values than if only the momentarily measured values were used as starting
points. This calculation also can be used to determine process upsets and faulty sensors, if
the data are processed further.
132
Optimization
The production-planning problem is formulated similar to a model predictive control
problem [3]. The optimization criteria consist of different terms, quadratic and linear, and
include terms to enable the following functions included in the objective function:
1. Minimization of the deviation from set-point trajectories of any of the manipulated
variables. One example is the paper machine production plan that is incorporated in the
optimization with production rates and recipes. These set-points are automatically
imported from the paper machine scheduling system. The pilot mill very often re-plan the
paper machine every second hour.
2. Minimization of the deviation from set-points trajectories of any of the measurement or
soft-point variables. One example is the white liquor sulfidity.
3. Minimization of the deviation from set-point trajectories of any of the state variables.
One example is the total sodium stock in the recovery cycle and the part of the sodium
stock distributed on the white side. Another common example is a tank level set-point.
4. Minimization of the rate of change of the manipulated variables.
5. Minimization of linear costs. One example is make-up chemical cost
6. Maximization of linear revenues associated with a state variable. One example is
revenue by producing pulp during periods of overcapacity in pulp production compared
to paper production.
This optimization criterion is minimized subject to the following constraints:
1. Process model
2. Sensor model
3. Upper and lower bounds on state variables (may be a time dependent trajectory)
4. Upper and lower bounds on manipulated variables. A maintenance stop is introduced
in the optimization by adjusting the upper bound of the production rate during the stop
time for a production object e.g. digester.
5. Maximum and minimum sensor values.
6. Upper and lower bounds on change of manipulated variables. Changes in the
manipulated variables are sometimes bounded by hard constraints to avoid that the
resulting trajectories will be too aggressive. One example is that a digester should not
change the production rate faster than a certain rate per hour.
7. The production planning should start from the current estimated value from the state
estimation.
The results will be the set points for all important controlled variables for the coming
hours.
During steady-state periods, the buffer tank volumes are most likely not needed to even
out disturbances. It is then preferable to adjust tank levels in a way that statistically
133
minimize the production loss around each tank by considering the intensities of stops that
propagate through buffer tanks to adjacent production units. Long term set points for all
levels in buffer tanks are computed that minimize the production loss considering the
time between failure and time to repair for each process section.
Simulation
To check the feasibility of the way the process is operated for the moment, as simulation is
made to see how long it can be operated until some limits are violated. This is used as a
support for the operators.
Results
An example of a process display from the mill is seen in figure 1. Here we can see the history
of the levels in the tanks in the fiber line to the left; the optimized values recommended for
the future production strategy at the right of this graph and far two the right the actual level
right now. This figure gives the operator a very good quick overview over the process status
and recommended operations.
Figure 1. Historic trend of levels, optimized values for the future as well as the actual level
right now, seen from left to right. This for the fiber line.
In figure 2 we can see a bottle neck analysis. This gives the operator a quick overview of
were the bottle necks will be if we precede operations without any changes of the set points.
Finally we look at a single variable, in this case the sulfidity of the white liquor, in figure 3.
To the left of the dotted, vertical line we have the estimated value of the sulfidity.
Horizontally at 35 % we see the actual last measurement of the sulfidity. The difference is
that the concentration
134
varies in different parts of the system. To the right we have first a strait, dark line, which
predicts what the value would be without any actions. The brighter curve shows the
optimized value, if actions are taken for control.
Figure 3. Actually measured value of sulfidity is 35 %, in one spot. To the left we have the
estimated real sulfidity in the system, and to the right predicted value if no action is taken (
dark) respectively the optimized value( the bright).
A number of different approaches for the simulation of the digesters can be applied. A
number of these are discussed in this section.
136
In the third step we use the actual temperatures and concentrations in each volume
element to calculate chemical reactions, diffusion, transfer between phases etc. These are
then used to calculate the new concentrations of all the flows in and out, as well as the
new concentrations in the different phases in each volume element. In the figure to the
right we have real or faked valve between all the arrows, giving the pressure at all
points separating the arrows. Real valve data are used for the dimensions, but tuning has
to be done to give correlations between the actual flows as a function of chip size
distribution, viscosity of the mixtures, temperature etc.
A variant of this approach was used by Bhartia and Doyle ( 2004) to model plugging in
the digester, and also other types of engineering use is common, to study specific design
or operational problems.
This first approach gives dynamics for the complete mill, and interactions between all
equipments. It is very useful as a training simulator and teaches you how the DCS
functions like interlocking, PID controls etc respond. An example of a response to a change
in a flow rate is shown in the figure below. Here we get the process values directly at a
process display, just like in the real plant. Trends can be shown of variables, PID control
response etc.
Figure 2. A process display from a training simulator, where a DCS system is interacting
with a dynamic process simulator on-line.
137
138
139
Optim ized
Description
Decision Variables:
SI103A
FI107A
FI104B
FI217B
FI216B
TI216A
TI217A
TC102A
FI216C
FI217C
FI212H
FI104C
FI102A
22,421
2,499
4,463
2,701
6,578
147,387
153,017
94,000
6,467
6,363
42,237
0,108
1,220
22,421
2,370
4,240
2,836
6,907
136,874
160,274
99,000
6,790
6,048
42,238
0,102
1,354
6,56
26,067
5,680
0,713
0,406
(0,069)
1,653
(0,130)
16,241
55,175
51,521
89,56
57,45
0,047839
6,62
25,976
5,495
0,698
0,405
0,001
1,499
0,003
16,353
55,178
51,545
89,56
58,22
0,074172
Dependent Variables:
Pulp Produced
Chip Utilization
Dissolved Lignin
NaOH Consumption
Na2S Consumption
LP Steam Utilization
MP Steam Utilization
Steam Condensate
Total White Liquor Utilization
Total Wash Liquor Utilization
Black Liquor
Kappa Number
Pulp Yield
Objective Function
kg/s
kg/s
kg/s
kg/s
kg/s
kg/s
kg/s
kg/s (negative value indicates net production)
lit/s
lit/s
lit/s
%
US$/s
In figure 4 below we can see what this means for the temperature profile from the top to the
bottom of the reactor. The dark (blue) line is how the set-points would aim at if we were
operating according to the normal strategy for this type of wood. The brighter (red) line
shows how the MPC proposes to operate, to get lower energy - and chemical consumption,
but with the same kappa number (remaining surface lignin) out of the reactor.
Temperature Profile
180.0
170.0
160.0
150.0
140.0
Base Case
130.0
Optimized
120.0
110.0
100.0
90.0
140
Figure 4. Temperature profile in the digester from top to bottom according to normal
operation (dark blue) respectively as proposed by the MPC ( brighter red)
In this approach with 2-D we have a fake valve between each volume element, giving a
complex net inside the digester, but also including real valves in and out of the digester.
With this approach we calculate the flow between the volume elements, and can simulate
hang ups, channeling, different packing etc. The model is used together with
measurements of real flows, chip size distributions, chemical additions, concentrations in
black liquor in extractions, temperatures, final kappa number of the chips etc.
141
This gives the opportunity to use the model for different type of diagnostics as well as for
testing of different design and operational options.
The digester model will also be used for sensor validation for the measuring apparatus
around the digester. If the calculated value will start to differ to much from the measured
value, that will indicate in an early stage that there are some problems with that measure
point. An early indication that there are some problems will save production, quality and
money. This type of sensor validation is important as a pre-step to the actual optimization.
The 2-D physical model of the digester has been made taking into consideration the
pressure drop inside the digester due to the channels between the wood chips. When there
are mostly large chips, the channels between the chips are large, with low pressure drop for
fluid flowing between. When there are a major amount of fine pins etc, the channels
between the chips will decrease, and the pressure drop increase. This is what happens in
reality if we have different chip size distribution or different packing in the digester. One
reason for this may be that the chip size distribution is inhomogeneous or the chip screw in
the stack may not feed in a constant way. When we have a lot of flakes these may adhere to
the screens and cause hang ups. Aside of causing an increased pressure drop in the screens,
also the chips will get different residence times and contact with liquors of different
concentration of both chemicals and dissolved organics. This may cause a significant
variation in the kappa number of the final fibers. By identifying pressure drops, residual
concentration of chemicals in the liquors, temperatures and flows and compare actual
results to those predicted by the model, we can tune the model to match reality. This is
under the assumption that we first have achieved a good process performance. The model
can then be used both to optimize the performance by adjusting e.g. temperature and
chemicals dosage, as well as back flushing screens to avoid hang ups before the problems
become severe. There is also a potential for finding channeling to have a chance to go in and
adjust, although this demands regularly measurements as well.
This approach with a 2-D model is best to use for diagnostics purpose as to determine
channeling in the digester, or hang ups. Channeling will mean that liquid is not getting into
the chips, but passing in channels between the chips. This means that we will get less
reaction between chemicals and lignin, resulting in less dissolved lignin (DL) in the
extraction liquor, as well as more residual alkali (NaOH), as not all got the chance to react.
An example of how this can look is seen in the simulation in figure 6.
CHANNELING
9
% DL,NaOH*0.1
8
7
6
NaOHpred
NaOHmeas
DLpred
DLmeas
2
1
0
1
TIME
10
11
12
142
So far we have tested this in a simulation environment with good results. Next step is to
make real application tests in a mill.
When it comes to production planning and scheduling, including tank-farming, we have to
model not only the digester, but also all other equipment surrounding this. There are
several approaches for the modeling here. We can use the approach with models for tanks
and some other features directly in the formulation of the Objective function and the
constraints, and solve the dynamic optimization problem directly. Another alternative is to
use a dynamic simulation model, and use this to calculate the result for the whole time
horizon using a branch and bound type of solver. Then communication is intense between
the optimizer and the simulator, but on the other hand we know we get a feasible solution.
A third approach is to use the optimization using the first alternative, and then test this
solution towards the more detailed simulator, to avoid running into possible problems
implementing the optimal schedule.
All these approaches has been implemented, but for different applications. In the EU DOTS
project ( Bell et al ,2004 and Dhak et al 2004), we tested all these approaches, but for paper
mill applications. At Gruvon pulp mill, the approach was the first alternative. Here the
problem was very large and complex, where a large number of tanks and processes had to
be included. To avoid problems with unreliable sensor measurements the signals were
filtered through a moving window approach, which turned out to be quite successful. This
was done by ABB ( Persson et al 2003).
143
To do this we first have to consider what input data to use to update the model. Which
values are important for the model? Which data are reliable, so that we do not feed in scrap
information?
When it comes to the digester application, we need information on kappa of the fibers as a
function of operational data, as well as data on what fibers are introduced (chip size, type of
wood etc). The operational data include the capacity, concentrations of chemicals and
temperatures primarily. From these data we can calculate among other residence time for
the chips.
We can get the kappa number for what is coming out of the digester, but normally not from
inside the reactor (although there are some digesters having sample points inside the actual
digester as well).
Liquors from the extraction lines still are available, and the composition with respect to
residual NaOH, NaHS and dissolved lignin may be measured on-line or in lab.
Average values will give a filtered value, but also the standard deviation, which is of great
value to determine the process performance.
When we want to update the reaction rates with respect to the dissolution of lignin from the
fibers we have to simplify some factors, as it will be too complex otherwise. First we can
group the collected data into classes. The first class level will be the quality of the pulp. If we
operate short fiber and long fiber, we keep these two classes as the primary classes. If we
also have different qualities of short and long fibers, we make sub groups also for these. In
the long run it may hopefully be possible to determine on-line what type of fibers we feed
by measuring e.g. the NIR spectrum of the chips at the conveyor belt feeding the digester. A
lot of research has been performed to find good correlations between NIR spectra and chip
quality ( e.g. Axrup [4]) a well as NIR pulp quality (e.g. Liljenberg [5]). Principally we
calculate the H-factor:
H-factor = e (43.2 16113 / T) dt
Still, this is not including the effects of varying concentrations of chemicals, and thus we use
the following expression instead:
dLignin/dt = a * [OH] b * [HS] c *e ( d f/ T)
Here we note that there is normally approximately 21-27 % lignin (see table 1 below) in the
wood chips. The constant a is representing the wood properties including density and
size. A large chip will have a lower a than a small one.
By integrating over the whole digester we will principally get the total lignin dissolution,
assuming the washing is perfect. If the washing is less good we have to adjust for that as
well. We can also just make a digital notation that the washing is working as it shall or not.
Thus we can neglect it in the model as an analogue feature that needs to be updated.
Table 1. Wood composition
Cellulose Hemi cellulose
Lignin
144
Spruce 42
28
27
Pine
41
28
27
Birch 41
34
21
From this we can estimate that Kappa number to lignin relation is like the equation below:
% Lignin in dry pulp = kappa number * 0.18
Softwood Pulp that should be further bleached normally is cooked to approximately kappa
30, with a yield of approximately 40, while unbleached Kraft for liner is cooked to kappa 90110, with a yield around 58-60 %. For sack paper and some other applications, kappa 40-50
is the target, and the yield becomes around 50 %. These are the rough values.
What we do now is to calculate the lignin dissolution rate over the whole digester, and then
divide by the residence time in the cook part of the reactor. From this we can then calculate
the constants a,b,c,d and f. This in practice can be done in an iterative way, to get updated
constants.
For pines specie we got a reduction of lignin + hemicelluloses by 40 % from the original
55% with a residence time of 3.3 hours and 145 oC. We then have a reduction rate by 12.0
%/hour. This we use as dL/dt = 0.120. In this case we started with 0.9 moles/l of NaOH and
0.45 moles/l NaHS.
The constants b =0.5 and c = 0.4 are taken from literature as suitable, d= 18.33 and f= 5000
we calculate from process data when operation was done at different temperatures. The
constants here are for fast reacting wood specie. From this we can calculate a =
2.986*10^-4, and we have the formula:
dL/dt= 2.986^10-4*[0.95]^0.5*[0.45]^0.4* exp(18.33-(5000/T)),
where T is in Kelvin (K)
This we feed into each volume element in the model and calculate for the whole digester.
Here the actual temperatures are used, as well as the actual calculated concentrations. If the
overall rate simulated is 0.125 instead of 0.12, we reduce the constant a, for the
temperature we have decided should be the normal set point. When we have got process
data for a number of different temperatures, we recalculate the constants d and f. The
constants b and c we keep as constants.
What we have to be careful about is the specie going into the digester, while collecting
tuning data. Preferably we would like to have an on-line measurement of the species by e.g.
NIR spectrum, and correlate spectrum characteristics to final result of kappa under certain
conditions. Then we just keep the set of constants to be used and switch to the correct ones
as a consequence to what NIR spectra is measured at the feed point. A time delay is used to
compensate for the time from measurement until the fibers have passed the different
sections in the digester. When we do not have on-line NIR measurement, we need at least a
good guesstimate of the mix as manual input by the operators. The difference between the
guestimate and the resulting kappa will be used by the operator to train on how to make
better guesstimates from the information available.
145
In reality we tune to kappa number and not lignin, as the kappa number is the measured
variable, and not the lignin. In reality this kappa number is a mix of lignin and hemicellulose, as it just measures material that is oxidized by an oxidant in the standard method,
which is in reality affected by many different factors not always that well controlled.
By storing operational data from different runs we can both calculate averages for each run,
as well as an average for many runs during a longer time period.
So from these calculations we will both get a tool to predict what value the lignin should
have at the fiber surface, as well as in the liquor. We then can compare to measured values.
The difference between predicted values and measured can then be used for both control
and diagnostics.
If we combine the physical model for the digester and an optimization algorithm we can
optimize the operation of the digester in such a way that the yield can be optimized,
chemical consumption and energy minimized and quality variations minimized. The MPC
then gives set points for the different circulation loops, which may be controlled by PI or
PID controls with respect to flow rates. Under the assumption that also the other parts of
the plant are coordinated with the production in the digester we can get a significant
improvement of the economy.
If we just replace PID loops with MPCs without coordinating the operations of the different
equipments Castro and Doyle [2004] got the result that the difference was small, and thus
recommended that coordination of the production in the whole plant should be done.
22 43 64
130.00
125.00
120.00
115.00
110.00
105.00
100.00
95.00
90.00
85.00
80.00
85 106 127 148 169 190 211 232 253 274 295 316 337 358 379 400 421 442 463 484 505
Production
Kappa test
Figure 7. Typical variations of kappa number and actual production rates for a digester
during a week.
It is not enough to only control flows, temperatures and pressures. Also measurements on
quality variables are needed. Still, it is custom that all these variables are moving around all
the time, and steady state conditions are seldom reached. E.g. screens are clogged and have
to be back flushed. Because of this flow rates are changing, and with them also chemical
concentrations and temperatures.
Typical variations in Kappa number and production rates are as shown in the figure 4.
Kappa trends are shown as solids lines. Temperature and flow are changing in a similar
way. To use these values for tuning the model we test different averages. A problem is that
we often do not get lab measurements more than every two hours. On-line measurements
may be every 5-10 minutes up to every 30 minutes. These measurements may be usable.
146
Now we have to try to fit the kappa number measured to the average digestion
temperature, time and average concentrations of hydroxide and sulfide.
As can be seen the variations can be significant, although the wood is principally of the same
quality all the time!
The flow rates often drop as the screens are clogged. When they are back flushed the flow
goes up again. The problem is that the pumps cannot overcome the pressure drop over the
screens fully. The temperatures are collected in each zone and the average multiplied by the
residence time, which is calculated as the volume of the segment divided by the flow rate of
chips including liquor per second.
So far we have tested this in a simulation environment with good results. Implementation is
being done summer 2007 at Korsnas mill in Sweden.
For a simulation of the variations in the process we can see that variations occur in the feed
chips with respect to reactivity (lignin dissolution rate), moisture content and chips size
distribution. The variation will depend on where in the stack material is taken. A heavy
rainfall may decrease the DS content a lot at the surface and a bit into the stack. Chip density
may vary depending on how the screw moves in the stack. At the end positions, the bulk
density may go up. The wood quality depends on not only species, but also how long it was
since it was cut, if the trees were growing in a valley or at a hill side etc.
When it then comes to the digester performance, we have to consider these variations. We
also will often have a clogging of the screens by time, and thus need back flushing now and
then. The frequency needed will depend on the chip size and type. When the screen starts to
clog, the recirculation flow goes down, as the pressure drop goes up. This may affect both
temperatures and concentrations in the liquor. We also may have an in-homogenous
reaction due to unequal distribution of chips, with different size distribution both vertically
and horizontally in the digester.
9.1.3.4 CONCLUSIONS
What we have given here is a discussion on how physical models tuned with process data
can be used for several different purposes. First we discuss overall mill optimization.
Thereafter the optimization respectively diagnostics in the digester operations is discussed.
It is important to fit process data into the models in a good way and how this can be done in
an automatic way is discussed. An on-line measurement of chip quality fed to the digester
with e.g. NIR would principally have a major impact on the overall process performance
including the yield. A 1-D model was used for MPC and a 2-D model for diagnostics of hang
ups and channeling. There are advantages to have different degrees of complexity in the
models depending on what it should be used for. Long term when the computer power is
significantly better than today, it may still be possible to have the same model for many
purposes.
In the presentation we have discussed different means for modeling digesters to use for
MPC and optimized scheduling. The first approach with a dynamic simulation model with
iteration between pressure-flow net calculations and reactions in each volume element is
suitable for dynamic simulation, were also control actions through the DCS system has to be
147
encountered. This method is suitable for detection of different faults, testing of what if
scenarios and can be used in optimization interactively with an optimization algorithm.
The second approach without a pressure flow net work solver, but with a simultaneous
solver for calculation of all reactions taking place inside the digester is well suited to use for
MPC applications, where set points are to be given to the control loops. The third approach
with more detailed models in 2-D ( or even 3-D) is best suitable for detection of hang ups,
channeling and other type of faults in the process. In the future, when computer capacity is
significantly higher, it should also be possible to use this instead of the two other types of
models.
9.1.3.5 REFERENCES
Axrup L.,Markides K. and Nilsson T.: Using miniature diode array NIR spectrometers for
analyzing wood chips and bark samples in motion, Journal of Chemometrics
2000;14:561-572
Bell J.,Dahlquist E.,Holmstrom K.,Ihalainen H.,Ritala R.,Ruis J.,Sujrvi M.,Tienari M:
Operations decision support based on dynamic simulation and optimization. PulPaper2004
conference in Helsinki, 1-3 June 2004.Proceedings.
Bhartiya S and Doyle F J (III): Mathematical model predictions of plugging phenomena in an
industrial single vessel pulp digester, will be published in Ind&Eng Chem Res during 2004.
Castro JJ and Doyle F J III: A pulp mill benchmark problem for control : problem description,
p 17-29, Journal of process control 14, 2004
Castro JJ and Doyle F J III: A pulp mill benchmark problem for control : application of plant
wide control design , p 329-347, Journal of process control 14, 2004
Dhak J.,Dahlquist E.,Holmstrom K.,Ruiz J.,Bell J.,Goedsch F: Developing a generic method for
paper mill optimization.Control Systems 2004 in Quebec City, 14-17 June 2004.
Proceedings.
Liljenberg T, Backa S.,Lindberg J.,Dahlquist E.: On-line NIR charactization of pulp.Paper
presented at ISWPC99, Japan, 1999
Persson U, Ledung L, Lindberg T, Pettersson J, Sahlin P-O and Lindberg .: On-line
Optimization of Pulp & Paper Production, in proceedings from TAPPI conference in Atlanta,
2003.
Rau, Cristopher V, Moving horizon strategies for the constrained monitoring and control of
nonlinear discrete-time systems, University of Wisconsin-Madison, 2000
Maciejowski, J.M., Predictive control with constraints, Prentice Hall, 2001
Wisnewski P.A, Doyle F.J and Kayihan F.: Fundamental continuous pulp digester model for
simulation and control. AIChE Journal Vol 43, no 12, dec 1997, pp 3175-3192.
148
1)
An example of measurements from a batch digester with respect to this is shown in the
figure 1 below. The interesting thing is to see, that dissolved solids increases even after
the bulk delignification has stopped (90-110 minutes). This gives yield losses without
much improvements of kappa.
In figure2 kappa measured in lab is plotted (upper curve) together with the production rate,
as an example. It gives a typical figure of what variation in kappa there can be, and also
typical frequencies for production disturbances in many digesters ( lower curve).
Another example is from a Swedish mill with a batch digester (Obbola) , where the kappa
standard variation with advanced control together with conventional free alkali
measurement was 4.0 % before introducing CLA2000. After the installation and using only
the EA (free alkali) sensor, the kappa standard deviation was decreased to 2.6 %, but when
using all three sensors as shown in figure 3, the standard deviation dropped to 2.06.
149
100 %
170 (338)
Temperature
CLA 2000
150 (302)
130 (266)
50
C
Dissolved
Dissolved
Residual
20
40
60
80 100
Cooking Time In
12
140
160
Figure 1. Dissolved lignin, dissolved total solids and residual alkali during a batch cook as a
function of time.
As a complement also new software was introduced, to make use of the new function,
where a kappa prediction model was introduced. Similar results have been achieved in a
continuous digester in Finland, as well.
30 day Production Rate and Kappa laboratory tests
Data sample rate 30 min.
2000.00
1900.00
1800.00
1700.00
1600.00
130.00
125.00
120.00
115.00
110.00
1500.00
1400.00
1300.00
1200.00
1100.00
1000.00
105.00
100.00
95.00
90.00
85.00
80.00
145 217 289 361 433 505 577 649 721 793 865 937 1009 1081 1153 1225 1297 1369 1441
73
Production
Kappa test
Figure 2. Typical variation in kappa and production . Kappa normalized to 100 as the
average (right)
150
If we go into some details in the figure 2, we can see significant kappa shifts. This is very
common, and can be due to wood variation as well as problem in operation of the digester like
channeling. The shifts are seen more clearly in figure 3.
22 43 64
130.00
125.00
120.00
115.00
110.00
105.00
100.00
95.00
90.00
85.00
80.00
85 106 127 148 169 190 211 232 253 274 295 316 337 358 379 400 421 442 463 484 505
Production
Kappa test
Figure 3. Kappa shifts indicated with solid lines during a 10 day period.
BDW%. % alkali savings based on 15 % A/BDW% means 15*0.2 = 3 % alkali savings. With
an alkali cost of 95 USD/t alkali, we get a reduction ( assuming 57 % yield) : 3/100 *
15/100 * 400 000/0.57 * 95 USD/t alkali = 300 000 USD/y.
Washing soda loss reduction:
With better washing control, the washing result can typically be improved without
increasing the water usage. This can be translated into reduced soda loss. Typically the soda
loss in washing is 2.5 10 kg/ton pulp when running a high capacity. The wash water can
be approximately 4 -6 twater/tpulp. For a reduction of wash water by 2%, the dilution
factor should go down by 0.8-0.12 .We say 0.1. If we assume 15 kg alkali/tp and the same
price 95 USD/t alkali, we get the annual savings of:
0.1 * 15 kg/t * 95/1000 USD/kg* 400 000/0.57 tpy= 100 000 USD/y
Reduction in defoamer usage:
Often defoamer is added as a constant flow. By measuring and modeling the process more
accurately, also control of the defoamer addition is meaningful. This can mean a reduction
of defoamer cost by 10-20 %, or approximately 10 000 to 20 000 USD/y.
Reduced load on the evaporators and recovery boiler:
In very many mills today, the recovery line is the bottleneck for the production.
By increasing the yield by e.g. 1 %, the load on the recovery boiler will also go down,
depending on what yield the plant is running at. This means that the production could be
increased by approximately another 1 % in the whole plant, which means a gross
additional income of:
400 000/0.57 tpy*0.01* 600 USD/t = 4 210 000 USD/y
if the recovery line is the bottleneck.
For kappa 25 at 45 % yield the corresponding would be
400 000 / 0.45 tpy * 0.01 * 600 USD/t= 5 333 000 USD/y
With a direct cost of chemicals, wood and energy of 180 USD/t, the corresponding net
income increase would be (600- 180)/600 = 0.7 times the above figures. This means 2 950
000 respectively 3 730 000 USD/y.
Bleached pulp:
For bleached pulp the effect can be different. If we can drive the delignification towards only
bulk delignification, and avoid the negative effects of the residual delignification phase, the
bleaching should also be better. By reducing the kappa variation, we can also save chemicals
in the bleach plant.
Normally the consumption is approximately 2* kappa number kg active Chlorine per ton
pulp. The price for the Chlorine as ClO2 would be around 0.35 USD/kg active Chlorine. For a
kappa 30 pulp this means 60 kg/active Chlorine per ton pulp.
152
If the kappa variation into the bleach plant is reduced by 1 kappa unit due to better control
of the digester, it means a kappa shift of also 1 kappa unit. The chemical savings then will
be:
400 000 * (2* 1 kg active chlorine/ton) * 0.35 USD/kg = 280 000 USD/y
Energy savings:
If we assume the following heat balance:
1.42 GJ steam/ tp as flash steam and 0.61 GJ/tp as fresh steam for impregnation.
1.63 GJ /tp for heating to boiling temp
0.61 GJ/tp for hi-heat wash
-
______________________________________________-
10 000 USD/y
153
____________________________________________________
Total
For a kappa 30 mill with 400 000 tpy, the net earning will be :
Increased yield by 1 %
10 000 USD/y
____________________________________________________
Total
Conclusions:
From this we can see that there is a significant potential for improvements in digester
control, in relation to what we have in most mills today. Still, it must be recognized that it
is the combination of the right measurements together with the model based control,
which will give the benefits. If we only install the sensors or only introduce new models
to be used with the control, it will only give parts of the benefits (less than 50%), while
the combination is the driver.
The use of model based control will be a good method to achieve optimal income at
every moment, especially if we can measure the incoming wood properties on-line for
feed forward control.
References:
Axrup L. : Determination of wood properties using NIR. Journal of Chemometrics (2001)
Dahlquist
E.,H.
Ekwall,J.
Lindberg,S.Sundstrom,T.Liljenberg,S.Backa:
characterization of pulp stock preparation department, SPCI ,Stockholm (1999)
On-line
Lindstrom M. : Some factors affecting the amount of residual phase lignin during kraft
pulping. PhD thesis Royal Inst of Tech, Stockholm, 1997.
MacLeod M. and T. Johnson: Kraft Pulping Variables,TAPPI Short Course inj Savannah 1995.
154
Svedman M. and P. Tikka : Effect of softwood morphology and chip thickness on pulping
with displacement kraft batch cooking, 1996 Pulping Conference p 767-777.
Vroom K. E. : The H factor: A means of expressing cooking times and temperature as a
single variable, Pulp and Paper Magazine Canada (,1957) 58,228
Kappa number cannot be measured on-line in batch digesters, but there are many
approaches to model, or predict, it from the existing measurements, and use the predicted
Kappa to complement the H-factor control, as mentioned before. Both mechanistic and databased models are in use, but due to the complexity of analytical models, current industrial
practices rely almost exclusively on simple empirical or semi-empirical models. Several
intelligent methods (neural networks and fuzzy logic) together with advanced controls have
found applications also on this arena. There are also applications of intelligent soft sensors,
extended Kalman filters and fuzzy neural networks. Conventionally, these models utilize online alkali measurements done at the particular time instant of the cook, but lately also
other composition measurements (solids, lignin) have gained ground. It is essential to use
the information from the beginning of the cook, because it leaves more time to make
corrections as the cooking proceeds.
Leivisk (2006) reports on the use of Elman network as the soft sensor for the final Kappa
number after batch digesting using short sequences of the measured alkali content and
calculated H-factor from the very beginning of each cook. The sequences include 70-85
measurement values (corresponding 30-40 minutes in time), where H-factor is between
200 and 300, and the alkali content decreases fast. The selection of training data is crucial
there are several factors later on during the cook that can effect on Kappa number. The
training strategy was to train the network as good as possible to correspond to the training
set. The final judgement needed visual comparison, because e.g. correlation coefficients do
not tell the whole truth. After testing several training algorithms and network
configurations, Levenberg-Marquardt training showed the best results. Training continued
until the correlation was at least over 0.8. The tested configurations differed in the number
of neurons at the recurrent level. It seemed that the training performance improves when
the number of neurons at the recurrent level increases. In practice reasonable small
networks worked well.
The final test of the network consisted of introducing 50 unseen cooks to the network and
recording the network outputs in these cases. Even though the training performance
differed, there was not too much difference in actual testing. About 70-80 percent of points
155
were inside 1 Kappa number that is a good result as such. Calculating the mean of all
network outputs seemed to work better than calculating the mean only from the 15 last
outputs. Also, the case where only the declining part of alkali trend was used in training,
gave clearly worse performance than the original way of using the alkali trend starting from
the very beginning of the cook.
The network worked well with low Kappa numbers and there were so few observations at
the high Kappa levels that it is dangerous to draw any conclusions in this region. Also, the
whole number of cooks tested was so low that it is too early to decide on the applicability of
the method in on-line use. On-line use, of course, will require more careful determination of
the starting point of the trends together with the optimal sequence length, and also defining
the way, how to eliminate occasional outliers.
Soft sensors have also been used in the quality prediction of the batch digester (Rao et al.
1993).
Haataja et al. (1997) reported on the software sensor using cooking liquor analyzer for
Kappa number estimation in the continuous digester. By measuring cooking liquor contents
on-line all variations inside the digester can be monitored. The most interesting and useful
variables that can be analyzed from cooking liquor are alkali, solids content and dissolved
lignin. In this case these variables were measured on-line from the different liquor
circulations of the continuous Kraft pulp digester by the cooking liquor analyzer (CLA
2000).
Artificial neural networks (ANN) were used in the soft sensor to map input-output
relationships of measurements. In this study, the feed forward networks with back
propagation training were built and tested. Kappa number measurements that were used in
training as outputs came from an on-line device located in the blow line. Tested networks
were quite small in the sense of number of parameters. The number of hidden layers was
one or two and the number of neurons in a hidden layer two to five.
Estimators performed very well. It was encouraging that a good estimate of the final Kappa
number was given already after the impregnation of pulp.
Using wavelets for the data pre-processing gave more benefits, because more stable data
resulted in (Murtovaara et al. 1998). This way the data corresponded better the real
measurements than by using the median filter. Several techniques in building the software
sensor for Kappa number were also tested (Isokangas et al. 2001): partial least squares,
neural networks, linguistic equations, fuzzy logic and ANFIS (adaptive fuzzy-neural
inference system). The best results were obtained with linguistic equations.
Another approach is in (Dufour et al. 2005).
156
9.4.1 ABSTRACT
The dynamic optimization integrated to operator and engineer decision support has a high
potential in everyday use. As the production process and the structure of its material flows
is becoming increasingly complex, and as the scope of responsibilities of an operator or an
engineer are widening, tools to manage this complexity under dynamic conditions are
needed.
9.4.2 INTRODUCTION
This article presents designing decision support system for thermo-mechanical pulp
production for papermaking. In general the case is about continuous decision support for
running a plant of identical on/off processes under time-variant production costs, timevariant needs of the product and limited intermediate storage capacity.
The decision support system consists of dynamic simulation for predicting future
evolvement of the system, dynamic optimization based on simulations and a proper
software toolset to gain the user acceptance.
The task is challenging in several ways. The optimization methods require a huge number of
iterations. The objective function is evaluated at every iteration step which requires
running a dynamic simulation. The challenge is to have fast simulations or, equivalently,
simple enough models. The decision support environment must also support the user in
defining which operational scenarios are acceptable, and allow the end user to improve
upon the optimal scenario found by revising the scenario.
In the production of TMP, one of the primary costs is the electricity consumption of the
refiners. The market price of the electricity fluctuates. As the energy usage of the refiner
plant is huge, an optimal production schedule generates inevitably savings.
In order to minimize the operating costs, the TMP production schedule needs to be
optimized around the market energy costs and constraints of the paper mill. This is not as
simple as shutting down refiners while the electricity costs are high and running all of them
while energy prices are low.
157
The support tool performs dynamic optimization over a rolling horizon and determines a
reliable refiner schedule e.g. for the next 50 hours of production at e.g. 60 minutes decision
intervals. The decision is based on information obtained through measurements and
predictions.
The goal of the optimization is - with the dynamic simulation model of the process - to
determine which TMP production scenario and way of dividing TMP production between
paper machines will have the lowest electricity costs for a given TMP demand and
development of the free market electricity price. The schedule must satisfy the TMP
demand at all times and may not exceed the intermediate storage volumes.
The area of simulation consists of two refiner plants, TMP1 and TMP 2, and three paper
machines, PM 4, PM 5 and PM 6, Figure 1. Each refiner plant has five refiner lines of two
refiners each, in a two stage refining process called Tandem system. TMP 1 also has four
reject refiners, extensive screening and bleaching with hydrosulfite or hydrogen peroxide.
This plant feeds two paper machines, PM 4 and PM 5. TMP plant 2 also has five refiner lines,
three reject refiners and screening and feeds one paper machine, PM 6.
Mixing
Tank
Paper Machine 4
TMP 1
Wet
Broke
Paper Machine 5
Mixing
Tank
Wet
Broke
TMP 2
Dry
Broke
Dry
Broke
Paper Machine 6
Mixing
Tank
Wet
Broke
Dry
Broke
Figure 1. The area of simulation: two refiner plants and three paper machines
9.4.4 PHASES OF THE OPTIMIZATION
In our approach the optimization begins with the management of operational scenarios.
Scenario management means different things to different people. Economists use scenarios
for long range planning, management scientists use them for strategic decision making and
policy makers use them to weigh the consequences of their actions. Here an operations
scenario is a set of actions on the manipulatable variables of the process (process set
points) [1]. The task is to search the optimal scenario amongst all possible scenarios. The
optimization problem to be solved is initially very high dimensional, but can be reduced to a
158
more manageable one by limiting the scenarios to consist of small set of waveforms, such as
steps, ramps, and first order exponential dynamics.
Simulation
Scenario
Management
Objective &
Data
User
Structure
Interface
Optimizer
Constraints
Figure 2. The optimization loop (grey arrows) and the interaction with the data
structure (white arrows)
Figure 2 shows a flowchart of the optimization environment. The procedure is controlled by
the user via user interface. The set points of the initial scenario are given to the simulation
model. Objective function and constraints are derived from the simulated data, as well as
other parameters needed by the optimizer. The optimizer makes a decision towards a
better scenario and transmits a new set of parameters further on. The loop continues until
the solution that full fills the criteria is found. On top of everything is the data structure. The
data structure collects, stores and delivers all the information needed.
A software environment that realizes the functionalities presented in Figure 1 has been
developed. The present implementation of toolset is within Matlab which is easy to bring
to mill environments either as a full system, or as embedded in other systems, e.g. the
process analysis system KCL-WEDGE [2]. For more information about generic
methodologies for dynamic optimization and toolset, see [3, 4].
The simulation model will be driven by the optimizer. It is controlled by the inputs, decision
variables and some specific parameters. Inputs can be controllable or uncontrollable. The
decision variables usually measure the amounts of resources, such as the number of
products to be manufactured, the amount of a chemical to be blended and so on. Some of the
elements in a system are dynamic in nature, in the sense that they move through the system
causing other entities to react to them in response to some signal. Other elements are static
in nature in a sense that they wait for the dynamic entities or resources to act upon them
and pull or push them through the system. The complexity, size and detail level of the model
define which elements should drive the system.
159
In this case, multiple levels of simulations are used, to have fast simulation within
optimization and to verify the optimal outcome with a more detailed simulation model. The
typical optimization methods will require thousands, if not tens or possibly hundred of
thousands of simulation runs to find an optimal solution. Therefore, it is imperative that the
simulations take very little computing time, in the order of tenths to thousandths of a
second.
To speed up the simulation time, first run simulations are developed on the basis of
simplified process. They only consider mass balances of one component, TMP, and follow it
mostly through its storage locations. Chemical pulp and its mass are accounted for in the
wet broke tanks to include for their impact on total volume and wet broke availability. The
different areas in the refining plants are considered storage areas that are only based on the
total volume of the tanks and set point consistencies.
Simplifying the simulations leaves out important interactions, such as intermediate process
waters, intermediate tank levels, etc. Because of the absence of these interactions, a second
level simulation will be used after the optimization. This detailed process simulator assists
the end user in assessing the detailed feasibility of the optimal operational scenario
determined through a simplified process model.
The level 1 simulation is based on mass flows and storage only. The tanks after the refiners
are not defined by their volume, but by the mass of pulp that they can hold at their set point
consistencies. For example, a 2000m3 tank at 4.5% consistency has a storage capacity of 90
dry tons of pulp. Figure 3 presents a section of the schematic of level 1 simulation with
defined variables.
TankPM6(t)
TMP 2
5 lines
RefinersTMP2(t
)
FlowToPM6(t)
DemandPM6(t)
PM 6
GradePM6(t)
BrokefeedPM6(t)
BrokeTankPM6(t)
flows could be included in the constraints of the first level simulation. If it is know, for
example, that turning on a certain number of refiners in a short time period causes a
shortage of dilution water, which can become a constraint of the system. For more complex
interactions, a level 2 simulation is needed that takes into account water flows and
intermediate tanks and their levels.
The level 2 model incorporates both pulp and liquid flows. It also contains all the
intermediate storage tanks. The primary purpose of this model is to verify that the solution
determined in the level 1 model and optimization is reasonable. The level 2 model can be
used to ensure that there is not a need for large volumes of extra fresh water or that there is
not excessive overflowing of filtrate tanks.
TMP2 to TMP1, yet cannot be pumped from TMP1 to TMP2. Therefore, the optimization
must avoid scenarios that underutilize production in TMP2 for favor of production in
TMP1.
There are costs and constraints associated with startup and shutdown of refiner lines.
However, these costs are rather subtle as they involve e.g. the cost of wear and tear and
utilizing maintenance personnel. These actions require operator involvement and thus
their number is constrained. Instead of including these effects explicitly in the
optimization, we limit the number of refiner startups and shutdowns over the
optimization horizon. Furthermore, this aids the optimization. We allow only for twelve
state changes in each refiner schedule over 50 hours. A process state change, for
example, would be changing from three refiners operating to only one refiner running in
TMP1.
The goal of the optimization is - with the dynamic simulation model of the process - to
determine which TMP production scenario and splitting of TMP production between
paper machines will have the lowest electricity costs for a given TMP demand and
development of the free market electricity price. The schedule must satisfy the TMP
demand at all times and may not exceed the intermediate storage volumes.
In principle, this optimization problem can be solved with deterministic methods, such as
linear and nonlinear programming methods, or stochastic methods. Based on the size of
the search space and the type of the problem, we have chosen stochastic optimization
methods, genetic algorithm [6] and simulated annealing [7] to be implemented within in
the DOTS toolset. For the deterministic methods we mostly use the Tomlab environment
[8] for advanced optimization.
The results of the optimization are displayed to the end user either as direct suggestions
on the action to be performed next, or through more detailed graphical representations,
such as Fig. 4.
162
9.4.7CONCLUSIONS
We have introduced our vision of the decision support system. The case example was the
production of TMP but the methodology and the applications are valid in a broad range of
processes.
With presented methodology and tools we can manage complex dynamics of wide process
areas through optimizing the future actions with respect to an objective derived from the
economy of the production line, and to repeat this optimization with a rolling optimization
horizon.
The tools presented can be applied in on-line mode to improve production efficiency and
supply chain efficiency. They can also be used in off-line mode to provide a tool for setting
up and developing operational policies. This is very useful in investment projects. Finally,
tools can be used to analyze the effects of production equipment and their dimensioning on
operational efficiency and thus to make more educated investment decisions.
163
We are convinced of the potentiality of the toolset and are looking forward to utilizing it in
further projects.
9.4.8 REFERENCES
1. PULKKINEN P., RITALA R., TIENARI M., MOSHER A., Methodology for
dynamic optimisation based on simulation, Simulation and Process Control for the
Paper Industry, Munich 2004, PTS Manuscript PTS-MS 441.
2. http://www.kcl.fi/wedge
3. DHAK, J., DAHLQUIST E., HOLMSTRM K., RUIZ J., BELLE J., GOEDSCH F.,
Developing a Generic Method for Paper Mill Optimization, Control Systems
Conference 2004, Quebec City.
4. DHAK, J., DAHLQUIST, E., HOLMSTRM, K, RUIZ, J.,Generic Methods for
Paper Mill Optimization, Simulation and Process Control for the Paper Industry,
Munich 2004, PTS Manuscript PTS-MS 441.
5. RITALA R., BELLE J., HOLMSTRM K., IHALAINEN H., RUIZ J., SUOJRVI
M., TIENARI M., Operations Decision Support based on Dynamic Simulation and
Optimization, Pulp and Paper Conference, Helsinki 2004.
6. GOLDBERG, D. E. Genetic Algorithms; Addison-Wesley, 1989.
7. OTTEN, R.H.J.M; VAN GINNEKEN, L. P. P. P. The Annealing Algorithm;
Kluwer Academic Publishers, 1989.
8. HOLMSTRM K., EDVALL M., The Tomlab Optimization Environment, in
Kallrath J. (ed) Modelling Languages in Mathematical Optimization, Chapter 19,
Kluwer Academic publishers, 2004.
9.5.1 ABSTRACT
In 2005, a three year optimisation research project was finished. The project studied how to
connect dynamic optimisation tools with existing simulator software. During the project
several cases were implemented for testing the usability of the tools developed during the
164
project. One of the cases was a TMP (Thermo-Mechanical Pulp) production scheduling. In
that case TMP demand of mills paper machines was estimated based on paper machine
production schedules and grade dependent TMP consumption for all paper grades. Based
on the estimated TMP demands the optimal production schedules for TMP plants were
calculated. In the calculation several constraints needed to be taken into account e.g. limits
for level variation of the TMP intermediate towers. By using well-defined optimisation
problem it was easier to include all relevant information from several data sources and run
the mill wide process more efficiently.
The results from the TMP case were so promising that after the project the mill decided to
extend the case to cover another mill. This extension made the application even more
useful. It seems that there exists great saving potential in TMP production scheduling and it
can be achieved by decision support tool combined with optimisation. The implementation
procedure of optimisation tools on top of Savcor-WEDGE process analysis system is
described in this presentation.
9.5.2 INTRODUCTION
At the beginning of year 2005 ended a three year EU-funded project called DOTS. The
abbreviation DOTS refers to Dynamic Optimisation of Operational Tasks and Scenarios.
The objectives of the project were to develop operations scenario management and
dynamic optimisation tools to be used with process simulators, and to combine scenario
management, dynamic optimisation and dynamic simulation into a process operator
decision support system for greater efficiency.
During the project several cases were implemented to test the usability of the tools
developed during the project. One of the cases was a TMP production scheduling case. The
case consisted of two refiner plants, TMP1 and TMP2, and three paper machines, PM1, PM2
and PM3. The TMP1 plant feeds pulp to paper machines PM1 and PM2 while the TMP2 plant
feeds pulp mainly to PM3 but also some pulp to TMP1 screening section and from there to
PM1 and PM2.
165
The extended optimisation application consists of two paper mills. The mills share a joint
electricity power with maximum usage limit and therefore the mills need to be handled as a
single unit. The application includes
166
one paper machine not using TMP at mill A, but using electricity and therefore it was
included in the optimisation
At mill A both refiner plants have five refiner lines. At mill B the first refiner plant has also
five refiner lines and the second plant four refiner lines. So totally there are 19 refiner lines
at the mills.
The refiner lines need to be scheduled to produce a right amount of TMP to fulfil TMP
demands all the time but not let the intermediate TMP towers to overflow. Yet this is not as
simple as shutting down refiners while the intermediate towers are high and running more
of them while intermediate towers are low. The mill must predict electricity usage for next
couple of days and there are penalties for under- and overconsumption compared to the
predicted usage. Because the mill wants to avoid penalty costs, the number of refiner lines
in use is more or less fixed for the near future.
The refiner lines optimisation is done only for the time period following a certain fixed time
horizon. After the fixed time period starting up and shutting down the refiner lines are
mainly used to control the levels of the TMP towers. Because of the TMP quality issues
starting up or shutting down refiner lines should be avoided. The paper machines TMP
demands vary a lot based on grade they are producing. Another problem is that the input
information can change frequently. For example paper machines production schedule can
change every day based on new orders from the mills customers or based on some
production problems at the paper machine. Overall electricity consumption must be taken
into account as well because there also is a maximum usage limit that cannot be exceeded.
In the electricity consumption calculation also the seventh paper machine not using TMP
need to be included because electricity demand of paper machine changes whether its
running or not. As a consequence of all these factors, it is a complicated task to take into
account all relevant information coming from several sources when doing decision of TMP
production schedules.
After several discussions with mills personnel the problem statement was defined as:
Given initial TMP tower volumes and TMP-flows to the six paper machines, find the
optimal refiner schedules for the four TMP plants to minimize start-up/shutdowns of
refiner lines and to keep intermediate tower levels within certain limits and predict
electricity consumption for next two days and keep it below its limit.
From this decision support tool the process operators should get suggestions when to start
up and shut down refiners. Also a reasonable prediction of electricity usage is given, which
they can follow to avoid penalties of electricity consumption forecasting.
9.5.4 APPLICATION
167
The application was implemented into the mills readily available KCL-Wedge (referred
from now as Wedge) process analysis system. The application was configured in Matlab
environment and after the configuration the optimisation was used as a background tool in
the Wedge.
168
Paper machines TMP demand is not constant. Actually it can vary a lot based on paper
grade type and of course the biggest change in TMP demand is when paper machine starts
up or shuts down. The optimisation calculation needs to know the TMP demands over the
optimisation horizon. The first step of this is to calculate how much each paper grade has
consumed TMP per production hour on history. In the optimisation it is assumed that the
consumption will be the same for the same grade in the future also.
The Wedge system has a connection to production schedule system so the current
production schedules for each paper machines are always available. From production
schedules and assumed TMP demands, paper machines TMP demands are easily calculated.
By establishing these flows, it is possible to then optimise the number of refiners running
and the other flows. The optimisation occurs within the constraints around the maximum
and minimum tower levels, the maximum electricity consumption, the maximum number of
refiners running and maximum flows between towers.
9.5.4.3 IMPLEMENTATION
To make the optimisation system as user-friendly as possible, the case was implemented
into mills readily available Wedge system. Wedge is on-line connected to the mills process
databases and also to the production plan system and it can call Matlab-functions, like
optimisation algorithms in this case. The application was configured in Matlab and after the
configuration the optimisation was used as a background tool in Wedge.
Based on grade specific average TMP consumption values and the current paper machines
production plans, the total TMP demand for near future is calculated. Because Wedge is
connected to mills databases it can automatically get the current status of the process like
current TMP tower levels and the number of TMP refiners running at each TMP plant. These
numbers are set as initial values to the simulation and optimisation. The user can run the
calculation at any time just by one mouse click on the process diagram. To make the use
even easier this application was configured to run automatically once per hour. With this
configuration the operator can easily and quickly check the latest suggested control actions
to TMP plants without waiting the optimisation calculation to proceed
9.5.4.4 RESULTS
The results are presented with tables and trends. The main results are the optimised refiner
schedule tables telling when to start or stop refiner lines. To increase the user confidence
towards the results several trends are also presented, e.g. forecasted TMP demands in the
paper machines and forecasted levels of the TMP towers. This decision support system
169
helps the operators to take into account all relevant information when deciding when to
start-up or shutdown refiner lines.
In the Figure 3 an example of optimisation results are shown. In the first and the second
graph at figure 3 the TMP demands of PM1 and PM2 are shown. In the third graph the
suggested production schedule to TMP1 plant is presented. To increase user confidence
towards the results, the predicted TMP tower levels are also presented, graphs 4 6. At the
end of time period there is a shutdown period at PM1, therefore its TMP tower is run to low
level (graph 5). During the PM1 shutdown the suggested schedule has swapped between 2
ad 3 refiner lines. That is because the PM2s TMP tower level is kept near the target value,
around 50 %, while the intermediate tower dont have target value and hence can vary
freely between its minimum and maximum values.
Figure 3. Example of future time series for one TMP plant and corresponding paper
machines.
The application is in use in the mill and it works well. With the tool, the mill wide TMP
balances can be handled more efficiently and electricity consumption penalties can be
avoided.
170
9.5.4.5 CONCLUSIONS
With the decision support tool it is possible to run the mill-wide system from TMP
production to paper machines more efficiently and keep the intermediate tower within
certain limits. It is possible to modify the application so it would minimise electricity cost of
the TMP production. This kind of application requires that the electricity price has some
fluctuation and that the near future fluctuations are known or can be adequately predicted.
These requirements are met at least in Finland. Then the optimiser could propose a higher
TMP production when the electricity price is at the lower level and less when the price is
high, but still keep the tower levels within their limits and also avoid start-ups and shutdowns. The question is how to weight the different factors to be optimized.
In the production of TMP, one of the primary costs is the electricity consumption of the
refiners. Therefore there exists a great electricity cost saving potential in TMP production
by combining electricity cost to operator decision support tools.
171
Janice Dhak, Erik Dahlquist, Kenneth Holmstrm, Mlardalen University, Jean Ruiz, Centre
Technique du PapierDomaine Universitaire, Juergen Belle,Frank Goedsche,Papiertechnische
Stiftung
10.1.1 ABSTRACT
A generic method for formulating pulp and paper optimization problems is presented.
Two ongoing projects in the framework of the DOTS project illustrate the method:
optimization of sizing quality at a specialty paper mill, and optimization of the water and
broke systems at a coated paper mill. Explicit and implicit formulations are compared, and
different usages of external simulators in conjunction with optimization are discussed. The
problems are solved using MATLAB/TOMLAB. Some results from different optimization
algorithms are also presented.
10.1.2. INTRODUCTION
All pulp and paper mills strive to consistently produce high quality products and
schedule production runs and maintenance shutdowns to maximize mill output. Also all
mills must be ecologically and economically responsible when using resources and utilities
such as fiber, water, chemicals and power, and when handling waste products such as
effluents and off grade production. Clearly, a single mill can have a diverse range of
optimization problems. Also, the detail level of the optimization problems can vary
significantly, ranging from a complete mill to a single piece of equipment. There are a
number of references regarding optimization of production schedules, and managing mill
inventories [1], [2]. There are also a number of references regarding optimization of energy
and water systems [3], [4].
Optimization tools for production planning, and utilities consumption are most often
used by mill management to get an overview of the whole plant. However, there is also a
need for optimization tools for operators and process engineers to formulate and solve
area specific problems, in addition to the more general mill overview problems.
172
Two very different optimization projects are described in this paper. The first deals with
optimization of a mills white water and broke handling systems. The second project
involves optimization of paper sizing quality. Both projects are part of an EU funded
project, the DOTS project. The goal of this project is to develop a toolset for operator
decision support based on dynamic optimization. The toolset developed in the project will
be tested and implemented at four partner mills; however it will be generic in nature so it
can easily be applied to other types of problems, and other mills. This paper deals with the
part of the toolset regarding a generic method for formulating and solving optimization
problems. Other papers have been written describing other components of the toolset [5],
[6], [7].
The main tasks in formulating the optimization problem are defining the objective
function and constraints and formulating them according to the chosen optimization tool.
TOMLAB [8] has been selected as one of the tools for solving the optimization problems
in the DOTS project. TOMLAB is a commercially available optimization environment
developed primarily for use in MATLAB. The environment consists of over 70 different
algorithms for various types of optimization problems, as well as a large number of C+ and
FORTRAN solvers implemented as MEX interfaces. The guidelines for formulating the
problems are independent of the optimization package or algorithm used to solve the
problem. However, one of the advantages of using TOMLAB is that it provides a
standardized platform for formulating problems, the Prob structure (refer to TOMLAB Users
Guide [9]). Also, once a problem is formulated using the Prob structure, it is possible to test
and compare different types of optimization algorithms, without reformulating the problem,
in order to determine the most appropriate algorithm for a specific case. There is also a
standardized Result structure. The Prob and Result structures provide convenient interfaces
for incorporating the optimization platform into the complete toolset, which includes user
interfaces, scenario management tools, simulators, and a data structure that collects, stores
and delivers all pertinent information.
Following are some guidelines for formulating optimization problems:
System boundaries
Review the process and define the physical boundary limits of the system to be
optimized. If relevant, prepare a process flow diagram showing all parameters within the
boundary limits. Also, define the time frame for the optimization, bearing in mind that a
long time frame may take an unrealistic time to solve and produce results which have little
relevance to the actual process because there are many unpredictable factors, such as paper
breaks, which can occur during an extended time period. Also, identify which parameters
can be modified by the optimization (optimization variables), and which parameters cannot
be modified by optimization (inputs to the optimization problem).
Objective function
173
Based on mill operating experiences, describe the objective of the optimization. For
example:
-
maximize output from a certain piece of equipment, or a particular area of the mill;
etc
Each term in the objective function should also be assigned a priority or weight to indicate
its relative importance. For example, weight factors can be the unit costs of chemicals, or
energy prices during peak and off peak hours. However, in some cases, the choice of weight
factor is not so obvious. For example, what is the cost associated with minimizing the
variation in a tank level? Optimizing the problem with different weight factors is one way
the user can test different what if scenarios.
Constraints
Determine all of the constraints on the problem. Some examples of constraints are:
-
When defining constraints it is important to review the process and understand why
certain constraints exist. For example, if a mill is accustomed to operating a tank at a
certain level, based on trial and error, this does not necessarily mean that this is the most
appropriate level. If the level is allowed to be modified by the optimization, perhaps a
better overall operating strategy will be found.
Constraints can be formulated implicitly or explicitly. With implicit formulation a
simulator generates the parameters required to formulate the constraints. For example,
174
when dealing with quality parameters which cannot be measured directly, a simulation
model is used to predict the parameter. With explicit formulation, simulation is
accomplished by converting recursive equations into equality constraints. Recursive
equations are equations that relate the present to the past. An example of an explicit
constraint is the mass balance around a tank:
Define the upper and lower bounds on the optimization variables; n x 1 vectors x_L
and x_U
For MIP problems, define a 0-1 vector, IntVars, where nonzero elements indicate the
integer variables
Define the upper and lower bounds on the constraints; m x 1 vectors b_L and b_U
Define bounds on the nonlinear constraints; mN x 1 vectors c_L and c_U, where mN=
number of nonlinear constraints
The objective function may be expressed in different ways, depending on the type of
problem. For example, for linear programming problems, define an n x 1 vector of objective
function coefficients. For nonlinear problems, write functions to compute the objective
function and the gradient of the objective function (Jacobian of the objective function).
Also, define a matrix, JacPattern (similar to ConsPattern). Some solvers may also require
the Hessian of the objective function. For the least squares solvers in TOMLAB, the user
defines the residual vector and the Jacobian matrix of derivatives, and the objective
function, gradient and Hessian are computed from these, by the optimization algorithms.
Depending on the optimization tool used and the type of problem, other parameters,
such as initial values for the optimization variables, maximum number of iterations,
feasibility tolerances for the constraints and / or the objective function, weighing factors,
etc can also be defined in the formulation.
An important aspect in this project is the use of external simulators in conjunction with
optimization. Simulators can be used in various ways:
-
to validate optimization results. The boundaries of the optimization problem and the
simulation may be the same or, a mill may have a detailed simulation program for the
entire process or a large part of the process, while the optimization problem focuses
on a specific area. In this case the simulator can be used to verify how the optimized
solution will affect other parts of the process that are outside the scope of the
optimization, before implementing changes in the real process.
Drewsen. Initially a partial least squares (PLS) model based on 20 different process
parameters (total input matrix) was formulated. After examining the PLS model it was
determined that the quality parameters could also be predicted using simpler linear
regression models. A linear regression model based on the total input matrix was
formulated, and then another linear regression model with only 7 input parameters
(reduced input matrix). Table 1 summarizes the parameters in the total input matrix and
the reduced input matrix. The parameters with the suffix _x are the optimization
variables for the two different variants of the problem. The quality parameters W(t), M(t),
and A60(t) are the other optimization variables in both variants.
Reduced Input
Matrix
Resin size
Int1_x(t)
Int1_x(t)
PAC
Int2_x(t)
Int2_x(t)
Int3_x(t)
Int3(t)
Retention Aid
Int4_x(t)
Int4(t)
Microparticles
Int5_x(t)
Int5(t)
Starch solution
SP
Surf1_x(t)
Surf1_x(t)
Surface Sizing
Polymer SP
Surf2_x(t)
Surf2_x(t)
Specific refining
power
u1(t)
Chalk
u2(t)
Broke Input
u3(t)
Finished Stock
u4(t)
Consistency
Finished Stock
u5(t)
Dilution water
size press
u6(t)
Moisture before
u7(t)
177
size press
Grammage
u8(t)
Ash
u9(t)
Consistency
broke
u10(t)
Steam
consumption
u11(t)
Dilution water
headbox
u12(t)
Web speed
u13(t)
Objective Function
The objective of the optimization is:
-
minimize variation of the sizing parameters W(t), M(t), and A60(t) from reference
values REF_W, REF_M, and REF_A60
Fobj = w1 [W (t ) REF _ W ] +
2
t =1
hp
w [M (t ) REF _ M ]
t =1
hp
[Int
hp
j =1 t =1
where,
2
hp
(t ) + 2 [Surf i (t )] +
l =1 t =1
2
(
)
Surf
(
t
)
/ (Int j (t ) ) +
3
1
t =1
j =1
l =1
hp
[Int
J
hp
j =1 t =1
L hp 1
(t + 1) + 5 [Surf i (t + 1)]
l =1 k =1
hp
, w
= weight factors
178
=2
Constraints
1. The ratio of surface sizing to internal sizing must be within certain limits defined by Rmin
and Rmax:
(b * Int
i
i =1:5
j =1:5
(t )) +
(b * Surf
i
i =6:7
j =1:2
(t )) +
(b * u
i
(t ))
i =8:20
j =1:13
Results
Appendix 1 shows representative optimization results using the 3 different models to
predict the product quality parameters. The solver used in each case was NLSSOL. The
figures also show how the simulated quality parameters from each model compare to actual
process data.
Table 2 also compares the 3 different problem formulations, and shows that the
optimization using the nonlinear PLS model takes longer to solve, compared to the linear
models, and that the reduced input linear model with fewer input parameters and fewer
optimization variables takes less time to solve than the total input linear model. Another
advantage of using the simpler linear model with fewer inputs is that the model may prove
to be more robust over time compared to the other more complicated models.
TABLE 2: ALTERNATE FORMULATIONS OF
DREWSEN PROBLEM
179
PLS Model
Linear
Regression
model
total input
Linear
regression
model
reduced
input
Number of
variables
300
300
210
Number of
residual
elements
533
533
356
Number of
linear
constraints
263
353
266
Number of
nonlinear
constraints
90
CPU time,
seconds
124.14
3.52
0.8590
f_k
5840000
5831415
127810
x_k
Fig. 3, 4, 5
Fig. 6, 7, 8
Fig. 9, 10
Table 3 compares the performance of some different optimization solvers in TOMLAB for
the linear regression model with the total input matrix. All 3 solvers find optimum
solutions in a reasonable amount of time, with NLSSOL taking the least amount of time.
Optimum
Objective
function
Solution time,
seconds
NLSSOL
5831415
3.52
SNOPT
5831415
21.28
MINOS
5831415
855.59
180
Figures 1 and 2 show the water and broke handling systems at Papeteries de Lancey, a
coated paper mill near Grenoble, France. The variables beginning with Q or U represent
flows; variables beginning with C represent consistencies; SW, HW, GW, and EUC
represent softwood, hardwood, ground wood and eucalypt compositions respectively. Tank
levels in the water system are designated as T1, T2,T4, and tank levels in the broke system
are designated as X1, X2X7. Variables that are optimized have the suffix _x; other
variables are inputs to the optimization. Optimization variables include those that are part
of the constraints and the objective function, and those that are only part of the constraints,
and are manipulated as a result of the optimization.
fresh water
Qe92_x(t)
Qe91(t)
Qe41(t)
Qe42(t)
Qs9o_x(t)
Ue43_x(t)=Us91_x(t)
T4_x(t)
fresh water
Qe102_x(t) Ue101_x(t)=Us92_x(t)
Ue111_x(t)=Us93_x(t)
Qs4
1(t)
T9_x(t)
disc filter
Qs101(t)
chemical pulp prep
Qs111(t)
T10_x(t)
fresh water
Qe112_x(t)
effluent
Ws97_x(t)
to broke system
(repulping)
Qs10o_x(t)
effluent
T11_x(t)
Qs96_x(t)
Qs11o_x(t)
effluent
Qs95(t)
to broke dilution
to fosse
181
Ue11_x(t)
Ce11_x(t)
Coated broke
Uncoated broke
Qe13(t)
Ce13(t)
Ue12_x(t), Ce12_x(t)
X1_x(t)
Cs1_x(t)
Ue21_x(t)=Us12_x(t)
Ce21_x(t)=Cs12_x(t)
Ue51_x(t)=Us21_x(t)
Ce51_x(t)=Cs21_x(t)
Us52_x(t), Cs52_x(t)
X5_x(t)
Cs5_x(t)
Qs51(t)
Cs51_x(t)
Mechanical pulp
Chemical pulp
Ue41_x(t)=Us31_x(t)
Ce41_x(t)=Cs31_x(t)
Qe32_x(t)
Ce32(t)
X3_x(t)
Cs3_x(t)
Us4o_x(t), Cs4o_x(t)
X2_x(t)
Cs2_x(t)
Ue55_x(t), Ce55_x(t)
Us22_x(t), Cs22_x(t)
effluent
Qs11_x(t)
Cs11_x(t)
Qe31(t)
Ce31(t)
Ue52_x(t)=Us41_x(t)
Ce52_x(t)=Cs41_x(t)
Ue54_x(t)
Ce54_x(t)
Us62_x(t)
Cs62_x(t)
X6_x(t)
Cs6_x(t)
Ue73_x(t)=Us61_x(t)
Ce73_x(t)=Cs61_x(t)
Sw_e73(t)
Euc_e73(t)
From Fibermaster Hw_e 3(t)
7
Gw_e73(t)
Qe72_x(t)
Ce72(t)
Gw_e72(t)
Qe71_x(t)
Ce71(t)
Sw_e71(t)
Euc_e71(t)
Hw_e71(t)
X4_x(t)
Cs4_x(t)
From Fibermaster
Qe74(t)
Ce74(t)
Sw_e74(t)=Sw_e71(t)
Euc_e74(t)=Euc_e71(t)
Hw_e74(t)=Hw_e71(t)
X7_x(t)
Cs7_x(t)
Sw_s7_x(t)
Euc_s7_x(t)
Hw_s7_x(t)
Gw_s7_x(t)
Qe75(t)
Ce75(t)
Sw_e75_x(t)=Sw_s71(t)
Euc_e75_x(t)=Euc_s71(t)
Hw_e75_x(t)=Hw_s71(t)
Gw_e75_x(t) =Gw_s71(t)
papermachine
Qs71(t)
Cs71_x(t)
Sw_s71_x(t)
Euc_s71_x(t)
Hw_s71_x(t)
Gw_s71_x(t)
182
The repulping of off spec paper is actually a batch process. The optimized solution
should provide a schedule for repulping. Formulation and solution of this problem,
which is a nonlinear least squares problem with integer variables is ongoing. The
objective function and constraints described below are relevant to the case where Ws97
and Qe32 are treated as continuous variables, and to the case where repulping is treated
as a batch operation. However, there will be additional constraints involving integer
variables to describe the batch operation of the repulper.
Objective Function
stabilize the process in terms of variations in transferring flow rates and tank levels.
minimize variations in the composition of stock in the tank feeding the paper
machine (Tank X7)
minimize the difference between the actual quantity of off spec paper repulped,
compared to a required quantity, Qe32_des
minimize variations in the ratio of fresh pulp to broke pulp, RFP/BP, feeding the paper
machine
) (
) (
2
2
2
hp
2
2
2
Fobj = 1 Qe9 (k ) + Qe10 ( k ) + Qe11 (k )
k =1
2
2
2
2
hp
o
o
o
6
+ 2 Qs9 ( k ) + Qs10 (k ) + Qs11 ( k ) + Qs9 (k )
k =1
((
hp
) (
) (
))
) (
)]
2
(
(Hw _ s
(Qs (k ))
)
(k 1) )
+ 1 11 Sw _ s 7 (k ) Sw _ s 7 (k 1) + 12 Euc _ s 7 (k ) Euc _ s 7 (k 1) +
1
k =1
13
+ 2
hp
1
1
k =1
hp
1
7
(k ) Hw _ s 7 (k 1) + 12 Gw _ s 7 (k ) Gw _ s 7
1
+ 3 Qe3 ( k ) Qe3
k =1
2 _ des 2
hp
+ 4 (R Fp / Bp ( k ) R Fp / Bp ( k 1) )
k =1
7
hp
2
+ 5 j Diff (X j (k ) )
j =1
k =1
Constraints
All of the constraints for this problem are expressed explicitly. Following is a description
of the constraints for the problem:
183
1. Constraints which recursively define the mass balance around each tank. For example for
Tank T4:
1
2. Constraints which only allow manipulated controls in the water system to be changed
during a control horizon of duration hc, rather than over the whole prediction horizon, hp.
3. Constraints on flows and tank levels. These constraints are actually upper and lower
bounds on the optimization variables.
4. Constraints that define the consistency of stock in each tank in the broke system, and the
composition (%SW, %HW, %GW, and %EUC) of stock that is fed to the paper machine.
These are nonlinear constraints. For example, for Tank X4
1
5. Constraints regarding the ratios of chemical pulp to mechanical pulp, and broke pulp to
fresh pulp.
10.1.7 CONCLUSIONS
For the Drewsen project it has been determined that explicit formulation results in a
faster optimal solution than implicit formulation. Initial results from both projects have
indicated that the optimizations can be performed within a reasonable time, and
TOMLAB has proven to be a useful tool in comparing different types of optimization
algorithms. Also, concurrent development of the simulator and optimizer can be useful
in verifying each of the components, as in the case of Lancey.
10.1.8 REFERENCES
1. PERSSON, U., LEDUNG, L., LINDBERG, T., PETTERSSON, J., SAHLINE, P., LINDBERG, A.,
On-line Optimization of Pulp & Paper Production, TAPPI 2003 Fall Technical Conference,
October 2003, Chicago, Illinois, USA.
2. SARIMVEIS, H., ANGELOU, A., RESTSINA, T., BAFAS, G., Maximization of profit through
optimal selection of inventory levels and production rates in pulp and paper mills, TAPPI
Journal, Vol.2: No 7., pp 13-18.
184
185
186
187
10.2.1 INTRODUCTION
With the introduction of advanced computer based control systems and reliable on-line
sensors much of the work of running a modern paper machines have been automated. For
188
example, in the wet end of the paper machine the amount of pulp used is measured by
consistency and flow meters and kept on its set point by local regulatory controllers
controlling the pulp pump speed. On a modern paper machine, several hundreds of these
local control systems control properties like flows, consistencies, tank levels, pH, pressures
etc.
However, even though much of the control of a modern paper machine is automated, paper
quality is still controlled manually by the machine operators. The main reason for this
situation is that paper quality is often characterized by specific testing methods, most of
them being destructive in their nature, usually performed in a laboratory with specific
testing equipment. Attempts have been made to transfer these measurement methods to
on-line sensors, and although some methods are possible to transfer, most laboratory
testing methods still have no on-line equivalent. This results in that the operator(s) controls
the paper quality based on laboratory measurements, which are usually done on samples
taken from the end of each jumbo roll. Besides being rather infrequent (approximately 1
sample per hour), the results from the laboratory are usually delayed 20-30 minutes due to
climate conditioning. This means that if the quality suddenly goes out of the specifications,
the operators might not know until a lot of paper has been produced. This is of course very
unsatisfactory, and a lot of research and development have been done trying to solve this
problem. Most solutions are based on the use of a mathematical model that describes the
relation between signals measured in the paper manufacturing process, i.e. on the paper
machine, in the stock preparation and in the pulp mill, and the corresponding laboratory
measurement. Because most of these signals are measured continuously, such a model
would make it possible to on-line calculate the current quality, so that the operator can
continuously monitor the quality. It also opens up for the possibility to automate the control
and optimization of paper quality and production.
This paper is devoted to the development of such a model for the on-line prediction and
optimization of the quality for two-ply liner. The quality variables that are of interest are
primarily Ring Crush (RCT) and Burst Strength, but we have also studied Tensile Stifness
(TSI) and Tensile Strength (TStr). The last three variables are measured by an Auto- Line
300 Profiler [8], while RCT is measured by a stand alone Crush Tester. All variables are
measured as profiles in the cross-machine direction and then averaged. Also, TSI and TStr
are measured with a machine direction (MD) and cross-machine direction (CD) component.
We thus want to predict a total of 6 quality properties.
The studied paper machine is a two-ply fourdrinier kraft liner machine, using unbleached
kraft pulp and recycled pulp. The recycled pulp enters the mixing tanks almost directly from
its storage tower, while the unbleached pulp is dived on two stock preparation lines, one for
each ply respectively. The bottom wire line includes two refiners in series, while the top
wire line has a single refiner. The intention is to use only bottom wire pulp in the bottom
ply, while the top ply contains a mixture of recycled pulp, broke pulp and unbleached pulp
from both refiner lines. It can be mentioned that the recycled pulp mainly consists of
189
container board recycle. Besides ordinary sensors like flow and consistency meters, the
stock preparation area is equipped with some more advanced sensors: the ABB Smart Pulp
Platform (SPP), Near Infra-Red sensor and a Pulp Expert lab robot. It has been shown in [5]
that the combination of measurements from the SPP (Kappa number and fiber size
distribution) together with NIR and other process variables can predict paper strength. On
the paper machine, there is a press section consisting of two shoe presses, which are
followed by the drying section and a soft nip hot-roll calander. Before the paper is rolled up
on the jumbo reel, a Quality Control System (QCS) measures and controls properties like
basis weight, moisture and caliper on the paper. It also measures fiber orientation and angel
on both sides of the paper web [7]. When a jumbo-reel has been produced and removed
from the paper machine, a strip of the outmost layer of the reel is cut out and taken to the
laboratory for quality analysis. The paper machine produces a number of different grades,
most of them with different basis weight but some grades differs in the pulp blending and
machine settings as well.
To be able to predict the quality of the paper, a suitable model relating measurable signals
from the process to the quality variables, must be found. This can be done by assuming that
the process can be described by the following equation
y(k) = f(,u(k))
(1)
190
Signal
unit
Pulp flows
ton/h
Refiner loads
kWh/ton
Sizing
ton/h
Starch flow
ton/h
ton/h
kPa
kPa
QCS Caliper
um
g/m2
QCS moisture
m/min
data was sampled with a 1 minute sampling period, while the data measurements were
stored at the time instant when they were collected. Samples with laboratory measurement
were removed together with samples where one or more of the measured signals where
outside pre-specified bounds. This resulted in a data set with laboratory measurements and
corresponding process data from 1780 jumbo rolls. For the identification purpose, the
laboratory measurements were normalized with the basis weight of paper. The reason for
this is to attenuate the dominating influence from basis weight on the quality variables,
which has a tendency of reducing the identifiability of other parameters.
Ident. Data
Valid. Data
RCT
4.1
4.7
Burst
5.5
7.6
TStr CD
4.7
7.1
TStr MD
6.8
10
TSI CD
4.6
6.8
TSI MD
4.1
6.8
Table 2: Prediction error [%] for the identification and validation sets, respectively.
To verify the accuracy of the identified parameters, a validation data set was created from
the time period of December 2001-February 2002. Cleaning the data in the same way as for
the identification data set, the validation data set includes 707 laboratory measurements.
The result of the prediction for both the identification and validation data set can be found
in Table 21. As can be seen, the prediction error is slightly larger for the validation data set,
mainly because of systematic biases in the prediction errors.
With the models structure and parameters identified from process data it would be straight
forward to implement the model as it is for on-line prediction. However, just running the
model on-line would typically result in a prediction that captures the large variations in
quality but with a low frequency bias in the prediction during in-grade production (cf. the
validation data set). The reason for this is of course that paper making is a complex process
and even though we have a large and reasonable model with a lot of inputs, there are always
things that are immeasurable or even unknown. To deal with this, some kind of adaptation
of the model to the changing process behavior is necessary.
The method chosen for this case is based on recursive batch identification [9]:
N
192
i=1
(2)
where N can be seen as the memory in the identification and Q is a matrix reflecting the
scale of the
1It should be mentioned that the sensor for measuring Burst Strength was broken during a part of both the
identification and validation data sets, which means that the statistics for the Burst prediction is based on a
smaller number of samples.
Figure 1: Prediction (-) and measurement (+) of Burst (Top) and Burst Index (Bottom).
signals. This gives the following on-line predictor:
y(k + 1) = f((k), u(k + 1))
(3)
The choice of N in Equation 2 is crucial for the performance of the on-line predictor. If N is
chosen too large, the adaptation of the predictor will become too slow to changes in the
193
process, if it is chosen too small the least-square minimization of Equation 2 becomes illconditioned with the subsequent loss of identifiability of .
To overcome this, the following was done:
Select N such that it uses only the last 3-5 laboratory measurements.
In Table 3 the result from a simulation with N = 3 and the recursive identification of two
parameters reflecting the gain of the process in the MD and CD direction, respectively, is
shown. As can bee seen, the bias in the prediction is removed and the total absolute error is
reduced (cf. Table 2). In Figure 2
Mean
Mean Abs.
RCT
0.0
3.1
Burst
0.1
4.0
TStr CD
0.4
4.0
TStr MD
0.4
4.0
TSI CD
1.3
4.7
TSI MD
0.2
4.7
Table 3: Mean of prediction error and absolute prediction error from the simulation of the
recursive identification.
The prediction is shown as time-series for 350 consecutive lab measurements. Notice how
well the model captures the grade changes. It is also interesting to notice that although both
Ring Crush and Burst correlates, there is almost no correlation between their indices. This
means that, besides basis weight, Ring Crush and Burst reacts very differently the machine
settings. In fact, Ring Crush is related to the CD component of the tensile strength, while
Burst relates mainly to the MD component of tensile strength.
On-line optimization
Besides the use of the model as an on-line predictor paper quality, the intentions are to use
it as an aid the operators to actively improve the quality. As first approach, methods for online optimization of quality were developed in [2], in which the problem finding good
machine settings was formulated a constrained optimization problem:
min c(u(k))
(4)
u(k)
s.t.
194
f((k),u(k)) = yopt(k)
(5)
g(u(k)) = 0
(6)
(7)
where c() is a cost function (typically c(u(k)) = c(k)u(k)), yopt(k) is the quality target at time
k, g()
process constraints (mass- and component balances) and ulo(k); uhi(k) are upper and lower
bounds for the inputs, respectively.
In [2] the optimization problem is solved using Matlab and the Optimization Toolbox [10].
One of the situations when this functionality is very useful is when the operator is to
perform a grade change. The operator can then select suitable initial values and bounds for
u(k), typically from some grade library, select the target quality and then launch an
optimization, which will find the cheapest
195
Figure 2: Prediction(-) and measurement(+) of Ring Crush (Top) and Ring Crush Index
(Bottom). Notice the obvious outlier at tambour nr 45.
In the figure 3 below we have another example of RCT for predicted versus measured
values, from another set of tambours..
Figure 3 Ring Crush Test (RCT) with measured values(stars) vs predicted values (solid line).
Arjo Sinon, Sappi Fine Paper Europe, Netherlands, D.M.R. Lo Cascio, TNO,
Netherlands
10.3.1 BACKGROUND
Poor performance of the paper machine always results in extra operating costs or lost
profits, which makes the outcome of the PDM a valuable tool for the operator of the
paper machine or for the technological staff.
196
Poor performance of the press section means that final press dryness is lower than
optimal meaning that more water has to be evaporated by the dryer section. This results
in extra (avoidable) steam usage.
Whenever the dryer section is performing inadequately the actual drying speed is lower
than optimal, which means that the machine speed could be increased, resulting in higher
output of the machine. In case no need exists for increased output of the machine,
improving the dryer performance allows lower steam pressures while maintaining the
current speed. Again this results in steam savings, mainly because of reduced blowthrough steam usage.
10.3.2 INTRODUCTION
The TNO Press and Dryer Performance Monitor (PDM) is an on-line software
application for paper and board mills, integrated into the existing mill information system
(e.g. PI, PIMS). The PDM is based upon a mechanistic model of the dewatering process
from the press- and dryer section of a paper machine. This model calculates the dryness
of the paper web as it travels all the way through the paper machine, based on actual
process data like speed, grammage, steam pressures, and so on. The calculated dryness of
the paper web is continuously compared to the best-practice value for the current grade
and running conditions. This way the performance of the press- and dryer section is
determined and presented.
Poor performance of the paper machine always results in extra operating costs or lost
profits. Poor press performance leads to an increased dryer load resulting in extra
(avoidable) steam usage and poor dryer performance leads to reduced machine speed.
Because all problems connected to press- and dryer performance will be made visible by
the PDM, it is a valuable tool for the operator or for the technological staff.
197
During the trials with the PDM we identified three sources of potential erroneous behaviour
of applications using on-line data:
These points, which seem obvious on first sight, can make life very difficult, especially
because the quality of data available in peppermills nowadays leaves plenty of room for
improvement. However, if good results can already be obtained based on the current state
of the on-line data, imagine what could be achieved if we could fully rely on the data being
correct all the time.
The TNO Press and Dryer Performance Monitor (PDM) is an on-line software
application for paper and board mills, integrated into the existing mill information system
(e.g. PI, PIMS). The PDM is based upon mechanistic models of the dewatering process
from the press- and dryer section of a paper machine [1, 2, 3, and 4]. These models have
been validated extensively [5] in close cooperation with the Dutch paper and board
industry.
The current version of the PDM [6, 7, and 8] determines the actual performance of the
press- and dryer section of the paper machine, based on real-time data from the running
process.
198
Top
1.5
Bottom
Top
Bottom
1.4
Performance ENP
1.3
1.2
1.1
1.0
0.9
0.8
0.7
0.6
0.5
0
200
400
600
Production time (h)
800
1000
Figure 2: Detail of one possible way to visualise the dryer performance. This one is like a
snapshot of the current production situation. Visualised is the dryer section (top) with the
199
thickness profile of temperature and moisture (middle), followed by the average dryness
and temperature of the sheet (bottom) from press to reel (left to right).
Example 1:
All dryer capacity problems are connected to poor performance of either press section or
dryer section or from the stock preparation or wet-end of the machine. The PDM actually
presents the performance of the press- and dryer section. This either tells you that the
origin of the dryer capacity problems you are facing is in the press- or dryer section: for
instance a bad pick-up felt, or flooded cylinders in the second pre-dryer; or it tells you
that the origin is not in the press- or dryer section, strongly reducing the amount of
searching necessary to pinpoint the origin.
Example 2:
Sheet breaks mostly occur at positions without web support, especially at low dryness.
Such a position often is an open draw between two presses. Low dryness of the web in
the open draw seriously increases the risk on sheet breaks. The PDM continuously
presents the operator with the dryness of the sheet all the way through the press section.
This information can be an early warning for upcoming sheet breaks, even when final
press dryness is normal. A dryness sensor after the press section would not have given the
operator this early warning.
Malfunctioning sensors:
Detecting malfunctioning sensors sounds rather easy but once we start defining
malfunctioning it might become clear that a whole world of interesting challenges
lies beneath this. It goes without further explanation that using drifted sensor
readings for model calculations will result in false results.
It is not very uncommon in the paper and board industry that parts of the process
are operated based on this kind of sensor readings. This is possible because the
operators translate such readings and compensate for it. Of course this is
intolerable when using the same sensor readings in automated model calculations.
201
The effect of noise on the information in a signal most likely is known. In most
cases any form of filtering will reproduce the original signal but filtering also can
introduce many unwanted effects. It gets even worse when this is combined with
measurement errors. It is not always possible to exactly measure what is needed,
so something else is measured instead that represents the quantity we need. This
we can identify as measurement errors. This may results in (transient or dynamic)
behaviour of the measured signal that does not belong to the original quantity. Of
course the application can compensate for this, but it must be known first.
It is very difficult to get to know all these measurement details, especially because
most of these details are so self-evident to mill staff.
These points show that the quality of on-line data available in peppermills nowadays
leaves plenty of room for improvement. It would be interesting to see how this is in other
industries. On the other hand we would like to stress once more that good results already
can be obtained, despite the current state of the on-line data. Imagine what could be
achieved if we could fully rely on the data being correct all the time.
10.3.4 CONCLUSIONS
The TNO Press and Dryer Performance Monitor (PDM) is an on-line software
application for paper and board mills, integrated into the existing mill information
system. The PDM determines the actual performance of the press- and dryer section
using mechanistic models of the dewatering process, based on real-time process data,
even if the quality of this data leaves room for improvements.
The outcome of the PDM consists of the actual final press dryness compared to the bestpractice value and the actual drying speed compared to the best-practice value. These
quantities give valuable information to the operator and technological staff, enabling
them to realise cost-savings in practice.
Three main sources of possible erroneous behaviour of any application using on-line data
have been identified. From these sources it becomes clear that the quality of on-line data
in the paper and board industry leaves room for improvements, presenting many technical
challenges to solve.
Even with the current state of the on-line data it is possible to achieve good results with
applications using mechanistic models. We can only dream about the possibilities with
this kind of applications when we can fully rely on process data being correct all the time.
202
10.3.5 REFERENCES
[1]
[2]
TAPPI JOURNAL May 2000; Vol. 83; No. 5 - Single Sided Steam Impingement
Drying Of Paper, A Modelling Study; M. Riepen , H. Kiiskinen , R. Talja and O.
Timofeev.
[3]
TAPPI JOURNAL October 2000; Vol. 83; No. 10 - An Inside View on Impulse
Drying Phenomena by Modelling; M. Riepen
[4]
Conference paper Paper Machine Technology, Lanaken, 7&8 februari 2001 Modelling the influence of the press felt on the moisture distribution in the paper
web; D. Lo Cascio
[5]
Desing and optimisation of the press and dryer section of Kappa Graphic Board
with the TNO Dewatering model as useful tool, Haanappel; Uil; Lo Cascio;
Riepen
[6]
[7]
TNO TPD Report 2003 Press- and Dryer Monitor: Specifications and
Implementation working plan; A.M.J. Sinon
[8]
[9]
[10]
TNO TPD Report 2004 TNO Press and Dryer Performance Monitor:
Applications & Revenues; A.M.J. Sinon
[11]
203
Where to put the effort to increase the performance of your press section or dryer
section
What set points will be required to run a new product on your machine
How effective a rebuild really will be above the machine builders guarantee
204
Break
Risk Indicator
+ : indication correct
and on time
! : indication correct
but late
: indication incorrect
0
200
400
600
800
1000
1200
1400
1600
Time [min]
Result:
Result:
205
Advice:
Result:
16 % production increase
Result:
Result:
Result:
Laboratory analyses have shown that the deposits found in various parts of the process
consist mostly of substances such as wood pitch, slime or bacteria, fillers, inorganic salts
and paper chemicals. Inherently, these substances are not sticky, but they may be turned
into a sticky form.
For example, wood pitch appears in process water as small droplets that are quite stable
due to steric stabilisation by glucomannan. Steric stabilisation is known to be affected by
pH, multivalent ions and temperature. Similarly, additives and binders used in paper
products may be turned into stickies in fibre recovery processes. Today chemical
additives such as fixatives are used to bind these substances on paper. Internal process
water purification methods are not widely used in paper industry due to their relatively
high investment cost.
206
10.5.2 TMP
A research group of bo Akademi lead by Prof. Bjarne Holmbom has studied water
chemistry related to mechanical pulping since the early 90s. Several publications and
doctoral thesis have been published. One of the post graduates, Anna Sundberg, has made
in her doctoral thesis a very thorough review on research concerning the chemical
composition of TMP pulps, effects of DisCo substances in papermaking and control of
detrimental substances.1 In the following, main items of the review are presented.
10.5.2.1 FIBRES
The TMP process results in considerable fibre damage. The middle lamella and the
primary wall are partly removed from most of fibres, revealing smaller segments of the
outer secondary wall. The secondary wall can also be loosened or ruptured into ribbon
like pieces, which remain attached to the fibre. The fibres therefore are not uniform but
will have areas with a high concentration of lignin (middle lamella) and other areas with
high concentration of cellulose (outer secondary wall). The TMP suspension will also
contain fines of different origin, i.e. ray cells, fines from the middle lamella and fines
from the primary wall.
The most abundant hemicelluloses in softwood are galactoglucomannans. In addition,
softwoods contain also arabinoglucoronoxylans, arabinogalactans and pectins. Wood
resin can be found inside unbroken parenchyma cells or is smeared onto tracheids and
parenchyma cell surfaces in patches.
The content of anionic groups in papermaking is important for paper strength. The
amount of anionic groups also affects fines retention, sizing and adsorption of wet end
additives. The amount of anionic groups in, or on, the surface of fibres can be considered
beneficial for the papermaking process. These groups can improve swelling and may
improve the retention of fillers by interaction with retention aids. Most anionic groups in
wood or mechanical pulp fibres are carboxyl groups of uronic acids that are units in
mainly xylans and pectins. Furthermore, wood resin components, i.e. fatty and resin
acids, also contain carboxyl groups. Carboxyl groups in uronic acids and in fatty and
resin acids are mainly dissociated in pH 5. The amount of anionic groups in unbleached
spruce wood is about 70 -100 meq/kg. The amount of carboxyl groups in uronic acids
corresponds to about 80 -90 meq/kg and in fatty and resin acids to about 9 -15 meq/kg.
Anna Sundberg: Wood resin and polysaccharides in mechanical pulps, Chemical analysis,
interactions and effects in papermaking, Thesis 1999, bo, Finland
207
Up to 5 % of wood substances are released from the fibres to the aqueous phase in
ground wood production. The main constituents of the organic compounds are
hemicelluloses, pectins, dispersed wood resin and lignin material. Smaller amount of
lignan, acetic acid, formic acid and inorganic constituents can also be found. The release
is affected by pH and temperature at different points of the process.
In softwood pulps, galactoglucomannans are the most abundant hemicelluloses followed
by arabinogalactans and pectins. The wood resin released from TMP is dispersed as
colloidal droplets with an anionic charge. The colloidal resin droplets released from
unbleached TMP are sterically stabilised and can only partly be destabilised by addition
of electrolytes. Microbiological activity during wood storage can change the composition
of wood resin, which, in turn, can affect the stability and the deposition tendency. The
lignin material and lignans do not interact very strongly with paper chemicals, nor do
they participate in deposit formation. However, they can accumulate in the process water
and may contribute to the decrease of the brightness of paper. DisCo substances with
anionic charge are sometimes referred to as anionic trash or detrimental substances.
These substances may interact with, and consume, added process chemicals without
beneficial effects to the papermaking process such as retention.
with conjugated double bonds, which are oxidised and degraded. Wood resin is no longer
sterically stabilised and coagulates upon the addition of electrolytes. Most of the lignans
are degraded during peroxide bleaching. More lignin material is released after peroxide
bleaching, and this lignin contains carboxyl groups.
10.5.3 DIP
De-inking process is used to recover fibres from waste paper. Different process steps
such as slushing, screening, cleaning, fractionation, flotation, dispersing, and beating are
used to separate fibres from the constituents of waste paper. In the course of processing,
deposits referred as stickies are formed that may reduce production efficiency and paper
quality. The stickies originate from wood pitch, coating binders, printing inks, adhesives
and papermaking additives.
The most common classification of stickies is based on size. Macro stickies remain as
screening residue after laboratory screening with a slotted plate equipped with a slot
width of 0.10 or 0.15 mm. Micro stickies can pass the screening slots. Substances in
recovered paper processing classified according to their size are shown in Figure 2.2
Fapet: Book: Recycled fibre and de-inking, Chapter 5, Unit operations and equipment in recycled
fibre processing, in Papermaking science and technology series edited by Johann Gullichsen and
Hannu Paulapuro
209
Stickies are hydrophobic, anionic and tacky in their nature and they have various shapes
and surface areas. Known factors for stickies to form deposits are among other things
their chemical composition, viscosity and surface area, and process temperature and pH.
High shear forces used in disintegration of RCF and coated broke may increase the
tackiness of detrimental substances.
The reason for this type of distinction is that macro stickies from recovered paper in
industrial processing systems can be removed from the pulp slurry by screens and
cleaners. Micro stickies are usually so small that they cannot be eliminated even by the
most effective screening units. If the macro stickies content in the head box of a paper
machine or in the produced paper is excessive, the screening of the pulp processing
system, the approach flow system, or both require improvement. Micro stickies that pass
the screens possibly agglomerate and lead to deposits on the paper machine or its clothing
or pass into the product as newly formed secondary macro stickies. The efficiency of
contaminant removal of different unit operations are shown in Figure 3.5
Figure 3, Efficiency ranges of unit processes for contaminant removal in recovered paper
processing.
A distinction is made between primary and secondary stickies. Primary stickies are
identical with the aforementioned macro and micro stickies and characterized as intact
tacky particles of adhesives such as hot melts or pressure sensitive adhesives, inks,
binders, waxes, plastics, or wet strength resins.
Shock-type, chemical-physical changes (temperature, pH value, charge, shearing forces,
or concentration) in pulp suspensions are potential causes for formation of secondary
stickies.
210
There can also be interactions that increase the tackiness of stickies and the problems
experienced with stickies. Such interactions can involve primary or secondary stickies.
Fig. 4 shows in a schematic form some possible interactions between virgin fibres,
chemical additives, and contents of the recycled fibres that can cause deposits. For the
formation of secondary stickies of Fig. 5, the cause is often a sudden change (shock) in
the physico-chemical state.3
Fapet Book: Recycled fibre and de-inking, Chapter 11, Stickies in recycled fibre pulp, in
Papermaking science and technology series edited by Johann Gullichsen and Hannu Paulapuro
211
Process measurements are chiefly needed for process control. Other application areas
include
212
The objective of process control is to lower production costs and to maximise the run
ability and to minimise the quality variations of the produced paper. Control methods are
used to minimise the fluctuation of the process. Consistency, ash content, pH,
temperature, conductivity, drainage and wet strength properties and charge are the main
items to be managed. Direct control methods exist only for flow, conc., ash, temp. and
pH. The significance of these control parameters is described in Table 6.4
Significance
Consistency
Ash content
Variations in the wet end affect paper quality such as strength and
porosity. Accordingly, uneven distribution of ash (both MD and CD)
generates problems on the paper machine, and in coating and
printing
pH
Temperature
Conductivity
Drainage
strength
Charge
and
wet Pulp freeness is the most watched quality variable for furnish
management. Drainage and wet end strength properties vary with
grade, fiber furnish and running conditions thus creating process
and quality control challenges to be tackled at the source.
The ability to control the interactions of charged particles, such as
fibers, fines and DisCo material is the cornerstone of the wet end
stability
Wet end chemistry plays a crucial role in water removal, retention and formation thus
affecting, more or less directly, numerous paper quality properties and PM run ability. A
4
TM
213
central factor in the interaction between fibres, fines, fillers and trash (DisCo) is the
charge of each component. The control and stabilisation of charge is one cornerstone of
modern, efficient process management.5
The DisCo substances can be removed from the white water system with the effluent,
with the final paper or by using an internal cleaning stage. Washing of TMP pulp after
primary-stage refining could provide an effective means for removal of DisCo substances
while taking advantage of the fast drainage characteristic of high freeness primary-stage
pulp.
The DisCo substances can also be aggregated and, subsequently, retained into the sheet
during papermaking. Fixing agents, i.e. cationic polymers with high charge density and
low molar mass, are often used to aggregate substances detrimental to the papermaking
process. These can be synthetic or starch-based cationic polymers. It has been suggested
that the mode of action of fixing agents is aggregation according to charge neutralisation
mechanism, although some evidence has also been found for fixation of wood
components to the fibres. The dosage and contact time are very important, and so are the
charge density, type and structure of the polymer. A retention aid compatible with the
fixing agents is a prerequisite for good control of DisCo substances. Other chemical used
to aggregate DisCo substances are alum and PAC. Examples on polymers used as fixing
agents are shown in Table 7.
(diallyldimethyl
Alum
Polyamindiamines
214
The organic polymers may be either linear or branched. The molar mass and charge
density of organic polymers may range from 20,000 Daltons to millions and from 2 to 8
meq/kg.
Enzymes can also be used to decrease the amount of certain DisCo substances or to
diminish the problems caused by them. The highly anionic pectic acids, released in
alkaline peroxide bleaching, can be degraded by enzymes so that they no longer interact
with cationic process chemicals. Lipases are used to degrade triglycerides to free fatty
acids in wood resin even in industrial scale. The paper quality is improved with fewer
holes and spots in the paper, and the deposits on paper machine equipment are reduced.
Fillers and special pigments, such as talc and bentonite, can also be used to adsorb DisCo
substances or to change the tackiness of wood resin and deposits. The formation of
deposits can also be controlled by the addition of dispersants. These are anionic polymers
with low molar mass that can change the viscosity of pitch droplets or may reduce the
machine surfaces affinity for pitch. Dissolved air flotation, evaporation or membrane
filtration techniques can be used as internal cleaning stages.
Although a lot of data are available for the highly instrumented process industry, finding
the essential information is difficult. Savcor -WEDGE is a process analysis system
designed to analyse fluctuations in continuous processes. The goal of using the system is
typically either to reduce quality variations or improve process efficiency, or both. Savcor
-WEDGE includes a wide range of tools for searching and locating origins of variations
and for determining the cause-effect relations present in a process. Continuous process
monitoring can be automated making it easier to identify problematic areas and making
the analysis results accessible for a wider range of users.6
KCL-WEDGE integrates with the existing systems that provide the information needed
for analyses. All measurement data can be combined into graphs for easy viewing and
editing. In order to generate reliable results of the process, the data must be pre-processed
before analysis. The application includes handy tools for pre-processing carried out by
user, which do not call for any knowledge of mathematical signal processing. The preprocessing procedures can also be automated in KCL-WEDGE and performed hidden
from the end-user.
The application enables users to quickly form a picture of the whole process and its
status. The following data analysis methods are implemented in KCL-WEDGE as easy6
215
to-use applications: statistics of the data, xy-plots, histograms, and correlations, delays
between measurements, principal component analysis, periodic fluctuation analysis,
spectra, MAR analysis and waveform matching. A user can also define with html- and
Matlab-language own additional analyses tailored for his/her needs.
216
The water properties that are considered to be important in papermaking are summarised
in Table 8.
Effects or interactions
Temperature
PH
Conductivity
Hardness
Alkalinity
Turbidity
Dissolved gases
Surface tension
Viscosity
217
knowledge. In order to improve understanding on the factors affecting the chemical state,
chemical measurements are needed.
The on-line measurements relating to the chemical state of the PM wet end can be
divided into two groups: Direct measurements and indirect measurements needing preprocessing.7
Term on-line measurement is often used to refer widely different measurements needing
different amounts of maintenance of device, pre-treatment of sample and interpretation of
results. On-line measurement refers here to a measurement that is automatic, measures
continuously and needs relatively little maintenance. Also the results of measurement
should be reliable. In the report on-line measurements are divided into direct
measurement and indirect measurements that need pre-treatment e.g. filtration. Direct
measurements are e.g. different on-line measurements. The chapter includes also a
category for possible future on-line measurements that are under development and a
category for other chemistry monitoring systems. A summary of information obtained
from Ecotarget partners (STFI, KCL, PTS and PMV) is shown in Appendix 1.
10.5.7.1 DIRECT ON-LINE MEASUREMENTS
pH
pH is a logarithmic value of the hydrogen ion activity and only slightly influenced by
external factors. Fluctuations in pH are undesired phenomena and lead to changes in
solubility of different substances, making the system much more sensitive to build-up of
different precipitates and causing web measurements. On-line measurement of pH is a
well-established technique. However, the measurements need to be cleaned and
calibrated regularly.
Conductivity
218
Temperature
Temperature affects many reactions occurring in the aqueous phase. Thus, solubility,
dewatering, reaction rates, microbial activity and effects of most chemicals are dependent
on temperature. Fluctuations in temperature are undesired and may cause precipitations
and run ability problems.
Redox potential
The redox potential is a measure of the affinity of a substance for electrons (its electro
negativity) compared with hydrogen. By definition, the redox potential of hydrogen is
zero. Substances more strongly electronegative (i.e., capable of oxidizing) than hydrogen
have positive redox potential and those capable of reducing have negative redox
potential. Oxidations and reductions always go together. Changes in redox potential are a
sign of the presence of oxidants or reductants in the process. Chlorine/bleaching
compounds, dithionite, biocides, etc and their stability can be followed. Redox potentials
are measured with electrode measurements.
Gas content
Air content in short circulation affects paper machine runnability and dewatering in the
wire section, as well as consistency measurements and control. Low air content also
improves formation and decreases pinholes and dirt deposits in the paper web. Air can
exist as bubbles or as dissolved air. Both can be measured on-line with several devices
based on ultrasonic waves or by measuring the volume decrease when the water sample is
compressed.
10.5.7.2 INDIRECT ON-LINE MEASUREMENTS
Most chemical analyses of the water phase of pulp suspension require filtration of the
sample as solid matter disturbs the measurements. Continuous separation of solid matter
from the aqueous phase is often difficult as the fine fraction in the pulp suspension clogs
the filter, sooner or later. This will naturally affect the functionality of the filter which has
to be cleaned or replaced. Today new techniques exist for continuous thick stock
filtration.8
Fapet: M. Holmberg, Paper machine water chemistry p.213-219, in Papermaking science and
technology series, book 4 Papermaking chemistry edited by Leo Neimo, pp. 219, 1999
219
A continuous filter with counter-current cleaning has been developed at KCL. A ceramic
tube filter with 4 or 6 mm ID channels is used. The pore size of filter is between 0.2 - 1
m. The filter is equipped with an automatic washing cycle.
BTG has a filtration sampler (Mtek TSS-70) specially developed for paper industry
application. The sampler automatically draws pulp samples and produces samples with
constant low fibre content. Sampler is cleaned automatically with rinsing water and highpressure air. The filter is developed for BTG charge and turbidity analysers.9
Metso Automation has also an application for thick stock filtration (kajaaniFSD100).10
BTG has a commercial charge analyser (Mtek PCT-20). The analyser continuously
monitors the total surface charges of all dissolved and colloidal substances in an aqueous
sample. The method is based on the streaming current principle, where a moving piston
in a measuring cell causes a liquid stream between the cell wall and the piston. This
distorts the surface charges of the colloidal substances so that a current is induced.
Polyelectrolyte titration is used to determine the charge level. 11
The Kajaani (CATi) cationic demand measurement is also based on titration technology.
The measurement principle is an industry standard that complements laboratory methods
for detecting anionic trash. Proactive sample handling is used so that always a fresh
sample is measured. Large sample volume enables representative samples and accurate
measurements. The measuring chamber is cleaned automatically by pressurized air,
water, cleaning chemical and ultrasound.12
Turbidity
BTG has a turbidity measurement device which can be included into the TSS-70 filtration
sampler. The filtrate turbidity measurement is based on white light that is sent through
the sample. Opposite the light source there is a sensor, which measures the transmitted
light.13
10
11
12
13
220
On-line titration
Applikon is an on-line process titrator for aqueous samples. Different methods can be
programmed for the device. In KCL methods have been developed for determination of
the following process components: dissolved calcium, silicate, aluminium, starch, COD
(Chemical Oxygen Demand), dithionite and dissolved sodium.
Multivalent cations with X-ray fluorescence
The Courier instrument is based on X-ray fluorescence. The sample is irradiated by an Xray source (55Fe) and the emitted secondary radiation, which is specific for each element,
is measured. The device needs calibration for each element. In papermaking the device is
mostly used for detection of dissolved calcium, but other multivalent cations can be
measured as well. Filtration is needed for reliable results.
TOC
There are commercial devices for continuous detection of TOC (Total Organic Carbon).
The sample is treated with phosphoric acid to remove inorganic carbon dioxide and then
burned in 900 C. The carbon dioxide formed in combustion is fed into an IR
spectrometer and the absorbance is detected continuously. Integrating over the sample
time gives the amount of carbon.
Flow cytometry
For decades, Flow Cytometry (FCM) has been used for characterization of blood cells
and bacteria and has become indispensable for medical and biological use. FCM is able
to count thousands of particles per second and simultaneously determine their type and
size, ending up in a statistically significant report within less than a minute. The principle
of FCM is based on a light excitation of a "lined up" particle stream and a multi-channel
determination of scatter and fluorescence stained particles. This rapid technology has so
far not been used to any greater extent within process industry, except for counting
bacteria in milk and beer.
The methodology and suitability of this method for paper mill applications was
developed and tested in a project in the national CACTUS research programme aiming at
221
A commercial capillary electrophoresis device has been tested for on-line monitoring of
individual ions in the wet end part of papermaking process.15 The sample has to be
filtered carefully before measurement. Two-stage filtration has turned out to be necessary
to prevent any precipitation of the sample. 4 ml of filtrate is pumped to the measurement
cell that is pressurized with the pumping system of the apparatus. Approximately 10 nl is
pumped to the capillary. The measurement cycle is about 15 minutes. The measurement
is based on so called indirect UV-detection principle. A buffer mixture of dicarboxylic
acid gave good separation and detection of anions. Respectively, an imidatsol buffer was
good for cations. The measurement cycle for cations is about 10 minutes. It is possible to
measure concentrations of selected anions (such as chloride, sulphate, oxalate and
formiate ions) and cations (potassium, calcium, sodium, magnesium and aluminum).
Interestingly, anions and cations can be measured from the same sample. For the moment
no commercial on-line CE measurement system is available because the system would
require too much maintenance.
NIR measurement of paper sample
PTS is developing within Ecotarget measurement methods for stickies that is based on
the spectrum analysis of radiation that is reflected from paper surface. The paper sample
14
Ralf Degerth, Final report of the CACTUS programme, bo Akademi, Laboratory of Forest
Products Chemistry
15
222
is illuminated with near infrared radiation between 1400 and 1900 nm.16 Characterisation
of detrimental substances is done by comparing the measured spectrum with reference
spectres (Indege).
A small company from Jyvskyl Finland has developed an on-line measurement system
for monitoring the chemical state of process water. The system is based on
electrochemical measurements.17 It has been used to improve the control of papermaking
by optimizing the dosing of chemicals and by observing the process load caused by
different ions. Measurement technology is based on measurements of reduction and
oxidation reactions. This is done by using a combination of ion-selective electrodes (or
receptors). The electrodes may be coated or pure metals/metal mixtures. Also
temperature and pH are measured as they will have effects on the activity and solubility
of most ions. Following receptors, depending on the application, may be used
Dithionite
Thiosulphate
Biocide
Peroxide
Fillers; kaolin and Talcum
16
Patrick Plew, PTS, Work report: Introduction to new at-line measurement system, Ecotarget SP5
meeting in Munchen 21-22 April 2005
17
Sakari Laitinen, Liqum Paper Oy, Ohjelmakaari 1 FIN-40500 Jyvskyl, Finland, Principle of
Liqum Chena measurements
223
Bentonite
Defoamers
Hydrex
Fixatives
Dispersing agents
Optical brighteners
Alum.
Sulphuric acid
EDTA
The results of measurements are formulated into an electrochemical taste index that is
followed by the operators. The index is followed in selected parts of the papermaking
line. Un-disturbing process conditions have a taste index of 100. Deviation from optimal
processing conditions is used as the guide for improvements in processing conditions.
The successful use of the methodology requires continuous learning on how to affect the
indexes in a desired way. Several mills have adopted the methodology.
A small company from Finland, WIC Systems, have developed a measurement system
for on-line chemistry monitoring of a paper machine. The system takes samples from
various parts of the process and pre-treats the samples for water quality analyses. The
sampling system is patented. The following analyses are available: charge density/
turbidity, starch, alkalinity, manganese, aluminium and COD. The analysis time is
usually 10 minutes, but two hours for COD. Altogether 4 sample lines are connected to
the measurement system. Over the past 10 years, more than hundred on-line analysers
have been sold across paper, tissue and board and pulp industry. The measurement
system can be provided with remote monitoring capabilities through a modem or
internet.18
18
224
Paper machines are normally equipped with on-line scanning devices for the control
grammage, ash and moisture content, and thickness of the paper web. Holes and specks
are also quite often monitored continuously by measuring the transmittance of light
through the web.
Pulp freeness is the most watched quality variable for furnish management. Drainage and
wet end strength properties vary with grade, fibre furnish and running conditions thus
creating process and quality control challenges to be tackled at the source. Both
laboratory and on-line methods exist for the measurement of pulp freeness. More detailed
analysis of fibre properties can be done with equipment such as FiberMaster19 and
FibreLab20.
Numerous paper quality parameters are measured from machine reel samples either in lab
or with automatic paper testing equipment. Automatic paper lab21 is capable to analyze
every machine reel. Paper quality parameters, especially when out-of-quality
specifications, can also be used for indication of process disturbances.
10.5.11 CONCLUSIONS
There are a number of measurements that would be useful to improve the control of the
papermaking process. Some of them (a few) are already in use and at-line and on-line
methods are available on the market. There is, however, a large number of parameters
that can be measured only by complicated laboratory methods. These parameters are
sometimes measured and it is possible to get some useful information from the values. It
is, however, difficult to prove that such parameters are useful and that they should be
measured frequently in the process.
Scanning devices are used to control the retention, grammage, ash and moisture content,
and thickness of the paper web. Holes and specks of the whole paper web can also be
monitored continuously.
Numerous paper quality parameters are measured from machine reel samples either in lab
or with automatic paper testing equipment. Automatic paper lab is capable to analyse
every machine reel.
Modern mills are equipped with automation systems that gather huge amount of process
data and highly sophisticated process analysis tools such as KCL Wedge are widely used
19
20
21
225
in paper mills. Extraction of data relating to chemical state of the process, and
verification and processing of the data into useful form for process control has, however,
turned out to be problematic.
A lot of laboratory research results and mill experience exist on factors that affect the
release of detrimental substances and their impact on runnability and paper quality
defects, but clear cause and effect knowledge on detrimental phenomena is not available.
According to literature, a large change in pH, temperature, conductivity and anionic trash
increases the probability of deposition incidences. From practical point of view, the
carry-over of detrimental substances from pulping plant should be minimised or
stabilised and on the other hand processing conditions leading to agglomeration and
adhesions on process surfaces should be avoided. The operator on the other hand needs to
have information about the changes of chemical state of the process and on the other
hand he needs clear instructions; should he react to this information and in what actions
he should take.
The control of the chemical balance of the wet end has turned out to be very difficult.
Occasionally problems occur that are known to be connected to fluctuation of the
process, but the mechanisms leading to process disturbances are poorly known. Higher
concentrations of detrimental substances may cause unstable physico-chemical
conditions, where even small variations, for example in pH or temperature, can disturb
chemical balance in the wet. On the other hand, a paper mill may very well operate the
process at low water consumption level successfully longer times and then suddenly enter
into troubles.
Even though the pulping plant is often integrated with the paper machine, they are not
operated as a whole, but have their own operators. Steady state operation of the mill is
very seldom possible due to grade changes, process shut-downs for maintenance and
reduced performance of equipment. The interactions of plant operations on web
measurements and paper quality defects are not known.
The overall objective of WP5.2 is to develop methodology that can be used to minimise
web measurements and paper quality defects caused by deposit formation originating
from changes in the chemical state of process water networks. The methodology will be
based on the utilisation of the present data system. The objective is to study what kind of
additional data would be necessary in connecting the deposit formation with the changes
in process operation. The methodology is described in Figure 7.
226
Complementary
measurements
Extraction and
analysis of
process data
Measurements
and
mill operations
Dynamic model
of the mill
Factors
leading to
disturbances
Case mills
Monitoring
system
Lab-simulation
of
mill operation
Samples
New
measurement
methods
Figure 7.
Methodology of WP 5.2 to control and monitor detrimental
phenomena in papermaking.
10.6.1 INTRODUCTION
The stock preparation plant and the paper machine are very well controlled and many data
are nowadays accessible at paper mills by means of data acquisition systems e.g. WinMOPS.
These data are used to find correlations between the different process and product
parameters and to build models that allow predictive and adaptive control of the processes.
Furthermore, external data, coming from laboratory or from other sensors installed at the
mill (e.g. on-line Focused Beam Reflectance Measurement probe) are also considered in
this modelling process to build better simulations.
Process modelling involves lots of different techniques that can be used to build soft
sensors, simulations and diagnostic tools to monitor the system, predict its behaviour under
different scenarios, and optimise the product quality or the machine runnability.
Results of process modelling depend on different variables. Data analysis, removal of non
significant or wrong data, the choice of an appropriate modelling technique and how to use
each technique are some of those variables. The quality of monitoring and product quality
227
Once the simplest methods are discarded (for instance, multiple regressions or simple
statistical analysis), another crucial step in a successful implementation is to demonstrate
that those black box, complex, advanced models are not a mystic tool, and to prove that
a soft sensor, built on the base of an advanced model, gives the results as easily as the most
simple one. The only difference reside on the modelling stage, which need more expertise in
the case of advanced models, but stages before and after modelling are not necessarily
different.
10.6.2 OBJECTIVES
A general aim of papermakers is to predict the quality of the product and/or runnability
problems based on available data in order to define potential improvements for the
optimisation of the process based on the increase of process stability. The final aim is to
increase the productivity of the mill and/ or to increase the stability of the product quality.
In this example, the optimisation of the wet end in order to increase the quality of the
product has been carried out. A model to predict paper properties based on wet end
parameters has been developed.
UCM approach implies building advanced models that are able to generalise the behaviour
of the wet-end, as well as give results with enough accuracy and robustness, in order to
optimise the wet-end of a paper machine. This optimisation will suppose an increase of
process stability as well as an increase in average quality values, through direct
recommendations that operators will be able to easily transfer into the paper mill. This
approach could be later on transferred to other paper machines.
At the moment main efforts have been forwarded to optimise off-line models, by modifying
the values of internal parameters and finding the optimum combination that may allow
creating robust on-line models in future.
Furthermore, an additional objective is to develop an automatic-updating system that may
allow facing process changes automatically, as well as selecting the Current Best Model.
228
That would allow operators a rapid monitoring of a great variety of changes in operational
states, and a periodic renewal of the models. Maintenance and calibration would also be in
that way automatic, representing important time savings when facing possible troubles.
10.6.3 METHODOLOGY
When a model is built to use it, for instance, as a soft sensor, objectives have to be clearly
defined. A model built to optimise the process or a model built to just predict or monitor
specific variables with highest accuracy have quite different needs. When an optimisation
has to be carried out, a model with reduced number of inputs is more useful, because during
optimisation, best values for each input variable are provided for each output. If those
recommendations are obtained over an excessive number of inputs, operating the system to
find the optimum range becomes almost impossible, because process variables are not
totally independent. With a reduced number of inputs, this problem is significantly reduced.
Theoretically, a model with only one input would make this problem disappear, but
accuracy would be significantly affected. Thus, a compromise solution must be adopted.
On the other hand, when a model is developed to have the best possible prediction,
selection of only a few inputs is not usually the best option. Anyway a definition of
interdependencies and avoiding the use of too correlated variables has to be taken into
account.
Therefore, the number of variables to include in the study and the compromise between
accuracy and robustness has to be defined as objectives. A model can be developed to have
a very accurate response over a certain time period despite robustness is lower, or can be
developed following the highest possible robustness in long term predictions. There are
infinite intermediate situations.
Modelling stage will slightly change depending on selected robustness and modelling
techniques. For instance, when an artificial neural network is trained and validated,
stopping criteria vary depending on robustness, that can be quantified through testing the
model with data from different periods of time than those used for training or validation.
Results are shown in quite different graphic forms, but they have to be compiled in order to
be easily understood, following the same structure than objectives. Final idea must be that
mathematics going beyond the models can be complex, but applying them at the mill can be
as easy as a straight line model.
The construction of the models has been structured defining a methodology that implies
different steps. There are parallel stages that are present during the entire research. One
example is the automation of each stage by developing user-friendly interfaces that allow a
rapid analysis or modelling. This task should be linked in further steps to operators
training and model implementation, once the utility of a model, to solve specific problems at
mill scale, has been proven and accepted. The main steps are:
229
Modelling scheme definition: For each modelling method, a scheme has been
developed to assure desired robustness.
Full study: After that, a full study is carried out to fulfil the objectives and to
evaluate the influence of each input over each output.
Second variable selection: With the results from the full study, further selection
of inputs and outputs can be carried out in order to build reduced models with
few selected inputs. These models will be called optimisation models, and they
are only used if process or quality optimisation are predefined objectives.
After the preliminary approach used to justify the selection of ANNs, another preliminary
study was carried out to select the optimum validation procedure, as well as to select a
range of ANN architectures and training algorithms to use in the full study. Since, there are
numerous factors that may influence over the accuracy and robustness of the results. ANNs
architecture, training/validation procedures and the range of time of data used for training
are some of these factors.
The full study has optimised the validation procedure through the analysis of different
groups of ANNs. The same procedure has been applied to optimise the interval of time
considered for training data.
230
Both preliminary and full studies follow specific methodologies based on the creation of
numerous models grouped by different parameters.
The selection of the optimum number of hidden layers has been skipped due to previous
experience that shows that only one hidden layer is enough in this specific case.
The selection of transfer functions in each layer has also been skipped. From previous
experience we know that logistic functions in both input and hidden layers and linear
functions in the output layer give best performance as well as an important time saving
when reducing training errors.
Thus, these steps involve the development of several models varying the number of hidden
neurons and the training algorithm. Levenberg-Marquardt algorithm (LMA) and the
gradient descent algorithm (GDA) with momentum and variable learning rate have been the
two proposed training algorithms. In the case of GDA algorithm, momentum values are
another parameter to be optimised.
5 ANNs have been developed for each combination of parameters. Developed programming
scripts allow an automatic selection of the models with less validation error in order to save
their parameters and analyse their results. The way of working in architecture optimisation
step is shown in figure 1.
231
Training algorithm
Alternatives
LMA
Model parameters
to be optimised
Number of
hidden neurons
Number of
hidden neurons
Momentum
Developing 5 ANNs
for each combination
of parameters
Optimum parameters
Once the ANN architecture and training algorithm have been optimised, in full studies, there
is another step that varies depending on the case. In a full study, several ANNs are created
to compare, for instance, different distributions of validation data over time respect to
training data. Each studied distribution implies developing ANNs with previously selected
optimum architecture that is usually a tight range of values. From all created models, the
best ones are selected in order to analyse validation and simulation errors. A scheme of this
selection task is shown in figure 2.
232
Optimum architecture 1
Developing 5 ANNs
Optimum architecture 2
Developing 5 ANNs
Optimum architecture 3
Developing 5 ANNs
Selection of best ANNs
Optimum architecture N
Developing 5 ANNs
Average
validation error
Average
simulation error
Analysis of results
Figure 2 Best ANN selection. Best models are selected through finding the minimum
validation error.
Training, validation and simulation
Validation error is used while training the models, to select the optimum number of
iterations. Validation data are not used for training the ANNs, but to check the results.
Simulation errors are calculated in the same way, but the purpose is to check model
robustness when modelling data from different periods of time.
In each step of the modelling scheme definition, training, validation and simulation
concepts have been constantly present. Validation and simulation involve only the
calculation of outputs with an ANN, by putting as inputs validation and simulation data
respectively.
On the other hand, training process vary depending on the selected stopping criteria. In this
case, stopping criteria at training is finding the minimum validation error. These validations
are carried out for different number of training iterations. The optimum number of training
iterations is the one where the validation error (difference between ANN responses and
validation data) reaches its minimum. A schematic view of this procedure can be seen in
figure 3.
233
Training data
Validation data
Training
X iterations
Validation 1
Validation
error 1
Training
2X iterations
Validation 2
Validation
error 2
Training
3X iterations
Validation 3
Validation
error 3
.
.
.
.
.
.
.
.
.
Training
NX iterations
Validation N
Validation
error N
Find minimum
validation error
Optimum number
of training
iterations
When training each model, a plot of both training and validation errors is carried out, as
well as an automatic record of them, in order to automate the selection of best ANNs and
the analysis of results.
UCM has also been working on this stage, planning possible alternatives in order to pass
from off-line modelling to on-line automatic decision support systems based on different
models that could be used as soft sensors or simulators.
Building a switching tool that may allow either the selection of best models
according to operating conditions or the update of a generic model that may be
used to generalise strange/new data.
234
In principle, the second one seems to be more accurate and useful, minimising the risk of
having a misleading model. However, this decision will be carried out when these studies
are in a more advanced stage.
The switching tool in the second option can be selected from many different techniques.
Dynamic validation has been combined with ANNs in a preliminary stage. Figure 4 shows
the concept of dynamic validation of an ANN output.
Outputs of ANN and corresponding reference measurements are taken as inputs of dynamic
validation block. Then, a simple linear model is formulated between outputs of ANN and
reference measurements (one model per variable (SISO) or combined model with all
variables (MIMO)).
Model (SISO) is the form of y = a*x + b, where x is the output of ANN and y is corresponding
laboratory reference measurement. In optimal situation, where output of ANN equates to
reference value, the parameter a is 1 and b is 0.
Then, there are two options:
1) To observe dynamically this linear model and when it differs statistically enough
from original model we produce a warning that ANN should be trained or the new
ANN should be taken on-line or,
2) to update dynamically the parameters of linear model. This is equivalent that ANN
acts as static part of the whole model and linear model is the dynamic part of the
whole model, as in dynamic linear model is in serial to ANN. Uncertainties of ANN
parameters are aggregated into the linear model parameters. This option using
dynamic validation has been discarded due to the high number of parameters that
an ANN has.
Figure 5 shows an application of practical use of dynamic validated ANN output. On-line
data is soft-sensed with a previously trained ANN. The output is dynamically compared with
reference measurements and if the difference is statistically enough the warning is
produced and soft sensor is updated. Here dynamic validation block act as a dynamic state
estimator. If none of the previously trained ANN will produce good enough results then
general ANN is used or a new one is trained.
235
Reference
On-line
ANN
Dynamic
2) Produce dynamically
validated ANN output.
Update ANN.
236
Reference data
Dynamic
Validation
Soft Sensor
ANN1
Online data
ANN2
Output
ANN3
ANN - G
The best ANN in terms of validation error has been analysed as an example of an ANN
showing good robustness in both simulation steps.
Figure 6 shows an example of simulation 1 for formation index. It must be noticed that
variability in that period is quite low and simulation errors are significantly low (See Table
1).
When the same ANN was simulated with data obtained three months after the training
interval, errors increased significantly, but still continued being acceptable (See Table 1).
When all simulation 2 data were plotted, it was encouraging to notice that the model, 3
months after developing, still performed following the majority of fluctuations in the
predictions, as shown in Figure 7.
237
69
Formation Index
67
65
63
61
59
Lab data
Predicted data
57
0
100
200
300
400
500
600
Table 1 Simulation 1 & 2 errors for each output variable by the best ANN in ANN-3
set.
Longitudinal
breaking
load
Longitudinal
breaking
length
Formation
index
Simulation
1
Averag
e error
() (%)
7.6
4.4
Simulation
2
Averag
e error
() (%)
13
17
238
2.9
2.8
2.7
2.6
2.5
2.4
2.3
2.2
Lab data
2.1
Predicted data
2
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
6500
7000
7500
8000
8500
Figure 7 Simulation 2 of longitudinal breaking load by the best ANN in ANN-3 set.
10.6.7 CONCLUSIONS
Paper quality optimisation may be carried out through building soft sensors.
ANNs have proven to be adequate tools in order to predict paper quality from wetend measurements.
Robust models to predict paper properties based on wet-end analysis, with errors
around 10-15%, have been developed. Robustness have been analysed with two
simulations for different time periods.
Wolfram Dietz, Jrgen Belle, Johannes Kappen, Desiree Somnitz ,Christian Mannert, Frank
Goedsche and Frank Brning, PTS, Munich
239
In the face of high machine velocities, increased use of secondary fibre sources, rising
levels of detrimental substances in the water circuits and high demands on paper
qualities, modern papermaking needs a comprehensive understanding and an efficient
control of the core wet end processes. The interactions between different factors
influencing process stability and paper quality are complex and could up to now not be
described on the basis of first principles. Reliable predictions of improvements or
pending problems are difficult. As a consequence suboptimal efficiencies of the additives
applied, reduced productivity and decreased paper quality may occur.
Data based modelling serves the aim of describing wet end interrelations without the
necessity to appraise all chemical and physical interactions. The procedure of setting up a
model is generic, but an individual model has to be identified for every application case.
This chapter will outline how data based modelling can be used to design a soft sensor for
sizing.
Many quality parameters can be determined only after a reel has been finished, by taking
samples and by conducting subsequent laboratory analysis. If quality problems occur
while production is ongoing, corrective action involves a delay. This inevitably produces
broke. This also applies to setting the required specifications after a grade change. It is
therefore desirable to be able to predict quality parameters that cannot be measured
online.
The objective of the project presented here was to develop a soft sensor estimating sizing.
The benefit aimed at was to adapt the sizing agent dosage to the desired paper quality,
based on the internal sizing / surface sizing ratio, resulting in better sizing stability while
reducing sizing agent consumption. The work was originally part of a research project
funded by the European Commission. The paper mill involved produced a wide variety of
different specialty paper. The project steps will be described as follows. All mathematical
analyses were conducted using MATLAB software
240
To start off, a brainstorming session involving the papermakers was held to compile all
conceivable factors that could correlate to sizing. In addition to data relating to sizing
agent dosage and sizing agent dilution for internal and surface sizing, all other additive
dosages were also regarded. These parameters were supplemented by settings from the
preparation process, such as refining settings or stock consistencies. Last but not least, the
calculations also had to include the quality data from online and laboratory quality
assurance. All these factors added up to more than 150 individual parameters that had to
be registered online or through lab work for the acquisition of calibration data.
In consent with the mill operator, machine trials with strong parameter variations were
conducted, even hazarding the consequences of out-off-specification production. In order
to limit the trials, it was decided that the system should be stimulated abruptly but briefly
by changing individual settings. The following parameters were altered:
the internal resin size was doubled,
the resin size dosage was dropped to zero with and without surface sizing, with and
without surface starch,
the retention additives dosage was dropped to zero
the broke pulp dosage was varied.
The trials were conducted during four days of production without causing a single broke.
Neither the grade nor the grammage were varied during the trial period.
To distinguish between internal and surface sizing, sizing measurements were performed
with an ultrasonic penetration measurement. The parameters W and MAX indicate the
degree of surface sizing, the parameter A60, which is comparable to the Cobb value,
represents the internal sizing degree.
The data from these trials laid the basis for extensive correlation analysis and initial
modelling.
To evaluate all the data quickly various mathematical tools and methods were applied.
One such method is known as correlation mapping. The strength of the correlations
between pairs of data time series is calculated and visualized. A positive correlation, i.e.
two variables reacting to a system stimulus by both rising or both falling, is marked red.
Blue is used to indicate a negative correlation, i.e. two variables diverging in their
241
reaction. The deeper the colouring, the greater is the correlation. The correlation map in
Figure 1 contains and illustrates more than 30 variables with respect to their paired
correlations. Variable p9 stands for the dosage of the sizing agent, p10 for the amount of
PAC. The map points out that there is a correlation between these two variables and p92
(steam consumption), p34 (moisture in front of the size press), and p179, p182 (A60, size
result).
This evaluation is important for subsequent mathematical modelling, since it makes it
easier to select the important correlating factors to sizing quality. It has to be kept in mind
that parameters that are kept constant through a closed loop control will not show any
correlations. They would fall out of a cause-effect interpretation of the correlation map,
even if in theory having an impact. For building a soft sensor it is not necessary to refer
to cause-effect-relationships. The goal of the correlation analysis is to assess correlating
factors, not necessarily influencing factors. But to be able to use the models for
optimization calculations, the inputs should be influencing and be manipulable.
Resin size
PAC
Chalk
Kaolin
Wet end starch
Retention agent
Microparticles
Broke treatment
Starch solution
Moisture before size press
Grammage
Moisture pope reel
Ash
Paper output
Stean consumption
Surface starch 1
Surface starch 2
Surface sizing
Sizing
Sizing factor
Cobb water 60 top side
Cobb water 60 wire side
W top side
Max top side
A60 top side
W web side
Max web side
A60 web side
Poly Dadmac 0.001n white water
Retention
Resin size of paper
Resin size of head box
Resin size of white water
Figure 1:
242
p9
p10
p11
p12
p13
p14
p15
p19
p25
p34
p35
p36
p37
p41
p92
p117
p119
p127
p146
p147
p148
p149
p177
p178
p179
p180
p181
p182
p183
p184
p185
p186
p187
0 .4
Loadings on PC 1 (69.76%)
0 .3
p009
p185
p 178
p037
p010
p 186
p181
0 .2
0 .1
-0 .1
-0 .2
p036
p 019
-0 .3
-0 .4
p092
p 034
p179
6
p 182
10
12
V a r i a b le
Figure 2 :
Loadings for modeling acc. to the PCA method (Principle
Component Analysis)
Like PCA, PLS (partial least squares) modelling operates with principal axis
transformation of the covariance matrix. Unlike PCA, however, the PLS method
allows models to be developed for selected output variables. Figure #3 shows the
loadings for a PLS model for the sizing parameter A60 (wire side). The loadings
for variables p009 (resin size), p010 (PAC), p019 (broke input), p034 (moisture in
front of size press), p037 (ash) and p092 (steam consumption) correspond to the
strong correlations of these variables with the output variable A60.
243
p 03 4
0.4
p 0 92
p 03 6
p 0 19
0.3
Loadings on LV 1 (73.24%)
0.2
0.1
0
-0 .1
-0 .2
-0 .3
-0 .4
-0 .5
p010
p 00 9
1
p037
3
V a ria b le
Figure 3 :
Loadings for modeling of variable p182 (A60 value wire side) acc.
to the partial least squares (PLS) method
The correlation studies were initially conducted individually for all paper machine trials
to obtain good individual models in this way. Once these calculations had been
completed, the models were merged to form an overall model that combined all system
reactions stimulated by the individual trials.
One simulation using this model is shown in Figure 4. The calculated value deviates only
slightly from the measured value. Two important points must be taken into consideration
in this context: The sizing rate was measured every 5 minutes while the trials were
running, the values in between these intervals are not known. The individual values were
analysed by linear interpolation. The calculated values result from the available online
data that was collected from the system at a sampling rate of one minute. This means that
one value can be calculated every minute. The fluctuations result from the system
fluctuations as a whole and might possibly have to be smoothed when the calculations are
refined and improved. In conclusion, it is definitely possible to calculate the sizing rate
on the basis of online data.
244
variable
p 182
wire
W e rtefor
n fr
V ariab le
p -1 8(A60
2 (A 6 0
-wireside)
-s id e )
1.5
Y Measured 1, Y Predicted 1
m eas ured
0.5
predic ted
-0.5
-1
-1.5
10
20
30
40
50
60
70
S a m p le S c a le
Figure 4 :
Comparison of measured and PLS predicted values for p182
(A60 value wire side). Bold line: predicted; thin line: measured value.
Other models were calibrated and validated using the trials' data. Figure 5 shows the A60
simulation results for all paper machine trials, i.e. data from various trials was used to
calibrate the PLS model. Validation (also shown) was based on steady state data from a
normal production series in calendar week 9 (KW9). The sizing rate A60 is plotted along
the y axis. It is possible to satisfactorily calculate the sizing rate with deviations. The
results were checked against data taken from different production runs. Other grammages
and one grade with lower additive requirements were used for this purpose. Calculation
of the sizing quality on the basis of online data is also possible using these data. The
fluctuations result mainly from the noise in the input channels. Similar simulation
curves could be calculated for all parameters of sizing quality.
245
Figure 5 :
Comparison of measured and PLS predicted values for p182 (A60
value wire side) for all PM trials with a steady state validation range.
Orange line: predicted; bold black line: measured value.
The soft sensor that was developed for evaluating sizing quality can also be used to
implement optimisation procedures. The optimisation algorithm is intended to provide
new targets for selected input variables (correcting variables) that are to be manipulated.
In the first optimisation trials, the correcting variables that were chosen were the input
dosage variables p09 (resin size) for internal sizing, p25 (starch solution size press) and
p127 (surface sizing polymer size press) for surface sizing. Input variable p10 (PAC) is
included in a fixed dosage ratio to p09. The aim of optimisation was to stabilise these
correcting variables with an optimum ratio of surface to internal size so as to minimise
the cost function K as defined by
2
K =
p 25 + p127
2 p09 ,
(the sum extends over all time periods of the observed time window).
246
This is intended to result in a reduction of the dosage of the surface sizing agent at the
expense of the lower-cost internal size without suffering any resultant loss in sizing
quality.
To provide a constraint with respect to optimisation, the optimisation algorithm requires
that the sizing quality characterised by the penetration values W, MAX and A60 must
remain within predefined acceptance limits. To simplify the optimisation process, a
multivariable regression model was used instead of the PLS model variant as the soft
sensor model. The cost function was minimised under constraints with the help of the
Monte Carlo simulated annealing search scheme.
Total amount of K
Initial results for sizing optimisation that were found in this way are shown in the
diagrams in figures. 6 to 9 for a time frame of 100 minutes.
Simulation
Optimisation
Time in minutes
Figure 6 :
Cumulated course of the cost function for simulated sizing (with
and without optimization)
The cost function that was reduced during the optimisation procedure corresponds to
intensified internal sizing and weakened surface sizing. In spite of this, the resulting
sizing quality (shown in Figure 9 based on the calculated A60 value) might even be
improved.
247
optimised
measured
Time in minutes
Figure 7:
Measured p09 (resin size) curve compared to optimized p09
dosage: optimization resulted in intensified internal sizing.
Manipulated input variable p 127 (surface sizing polymer)
optimised
measured
Time in minutes
Figure 8 :
Measured p127 (surface sizing polymer) curve compared with the
optimized p127 dosage: optimization causes reduced surface sizing.
248
A 60 values
Upper limit
optimised
Setpoint
Lower limit
Time in minutes
Figure 9 :
Sizing quality (A60 value) simulated with the soft sensor compared
to the optimized A60 curve:
Both curves are within the predefined acceptance limits. The surface to
internal size ratio, however, was reduced during optimization.
10.7.3 CONCLUSIONS AND OUTLOOK
The application example proved that parameters for sizing quality can be calculated with
the help of online data recorded by the process control system using a suitably calibrated
soft sensor. Actually, it does not make sense to include all available data in the
calculation. The most important correlating factors should be determined by correlation
analysis prior to modelling and then incorporated as input variables in the soft sensors
models.
Still many paper quality parameters can be measured only by laboratory analysis of reel
samples. Thus raw material and process fluctuations cannot immediately be balanced.
The consequence is off-specification production. Soft sensors for these quality
parameters are well suited for ensuring constant paper quality. In a next stage of
completion, the soft sensor can be further enhanced to calculate not only the quality
parameters but also the optimum process parameters required for compliance with set
values.
249
10.7.4 REFERENCES
Cho, B.-U., G. Garnier and M. Perrier
Dynamic modelling of retention and formation processes;
Control Systems 2002, Swedish Pulp and Paper Research Institute, Stockholm, 2002,
200-204
Austin, P., J. Mack, D. Lovett, M. Wright and M. Terry
Aylesford PM 14 Improved wet end stability using model predictive control,
Paper Technology 43 (5), 41-44 (2002)
Brosilow, C. and B. Joseph
Techniques of Model-based Control,
Prentice Hall PTR, 2002
Pickhardt, R. and K. Schulze
Verbesserte Stabilitt der Nasspartie einer Papiermaschine,
Allgemeine Papier Rundschau 126 (46), 25-30 (2002)
Hauge, T.A., R. Slora and B. Lie
Model predictive control of a Norske Skog Saugbrugs paper machine: Preliminary
study,
Control Systems 2002, Swedish Pulp and Paper Research Institute, Stockholm, 2002,
75-79
Yue, H., E. Brown, J. Jiao and H. Wang
On-line web formation control using stochastic distribution methods,
Control Systems 2002, Swedish Pulp and Paper Research Institute, Stockholm, 2002,
318-322
Dhak, J., E. Dahlquist, K. Holmstrm, J. Ruiz, J. Belle and F. Goedsche
Developing a generic method for paper mill optimization,
Control Systems 2004 Conference, 207-214
Fridn, H. and H. Nslund
Online decision support and optimization for kraft paper production
Control Systems 2006 Conference
Belle, J., J. Kappen, F. Goedsche and H.-A. Khrt
Simulation-based control of the sizing rate of resin sized paper,
professional papermaking, No. 1/May 2005, 23-29
250
10.8.1 INTRODUCTION
Modern papermaking industry is intrinsically complex in terms of both material and
mechanical processes. Vast amounts of paper are produced at high speed with specialized
machinery and different forming methods. However, no matter what types of paper the
paper manufacturers are producing, it is always important that the end product paper is as
uniform as possible. This requires the intensive monitoring and rigid control of the paper
machine forming section. In general, the various available types of forming section include
twin-wire, roll, blade, hybrid and fourdrinier.
Physically, paper is formed when water drains from a suspension of fibres on or between
specialised mesh fabrics. Formation is thus defined as the small-scale variation of fibre
distribution within a sheet. The term small-scale in this case implies that the size of
variation is less than 100mm. The ideal distribution of the fibres in the formed sheet should
be perfectly uniform [1]. However, this is commonly unattainable, because of the combined
additive effects of the stock approach system, the random nature of component fibres,
chemical additions, and the particular operating principles of the head-box and forming
section, etc. The evolving importance of formation has spawned a great deal of studies
concerning measurement and possible real-time control, so as to quantify and measure the
mass density of paper and make it as close as possible to uniform distribution. However,
due to the complicated nature of the model and the multiple effects of all the wet end
variables on the web forming process, the effective closed loop control of the formation is
not achievable at present.
Of course, one possible method that could be applied to achieve optimum formation
would be to use the well-developed principle of basis weight control to attempt to realise
a possible formation improvement. Basis weight is defined as the total oven-dry mass
per unit area and can thus be used to reflect the distribution of solid material, to some
extent. In recent years, this method has been utilised to gain some success in the
improvement of formation. However, since only a small portion of the input variables
have been considered in the process basis weight control, the actual improvements in
formation achieved using these methods, have been minimal to-date. For example, if the
retention polymers in the wet end are not precisely controlled, the subsequent retention of
the solids will have high variability. Such variability will lead to concurrent variations in
paper formation, which cannot realistically be corrected automatically by the existing
basis weight control systems on present paper machines. A further difficulty is that
current basis weight control relies on a scaling sensor that is installed after the drying
section of paper machines. This will cause a very long time-delay between the wet end
operation and the dry end monitoring. As such, it can be concluded that precise control
of paper formation is still an unresolved important problem, and hence, an effective
solution would be to
1) include all the possible variables in the wet end of paper machines (not only those
used in present basis weight control systems), and to
251
2) utilise the recently available, rapidly developing wet end sensor technology (such
as Webform) as an effective measurement device in the wet end, so that a local fast
closed loop system can be constituted.
Figure 1.1 shows a general diagrammatic structure of the wet end of a paper machine that
has a fourdrinier single wire forming process.
A = Head box
J = Vacuum box
C = Breast roll
D = Forming board
E = Hydrofoil assemblies
N = Stretch roll
F = Table rolls
P = Guide roll
boxes located underneath the wire. In most cases, a visible 'dry line' will be apparent on
the wire table, indicating the end of the primary filtration and drainage process. In this
context, the process taking place between the slice lip of the head-box and the dry line
should be regarded as the main paper formation process in papermaking. From this brief
analysis, it can be concluded that paper formation is affected by a large number of
variables in the wet end of paper machines. The major input variables that should be
considered should include:
1) the quality of fibres and fillers;
2) chemical inputs;
3) machine operation variables;
4) variables which characterise fluid dynamics in filtration and drainage processes;
5) slice lip geometry and thick stock input.
In terms of fibre quality, all the variables in both the refining stages and stock preparation
systems will have some effects (such as CSF, stock pH and conductivity, etc).
Considering chemical inputs, variables such as retention aids should at least be
considered, since too high a retention will adversely affect the formation process.
Machine operation variables to be taken into account, should consist of all the foil angles,
drainage parameters, vacuum power and machine speed. As for the slice lip geometry,
all the cross-directional actuators should be included. In addition, due to the high
operating speeds of commercial paper machines and the involvement of large scale
movement of huge volumes of water, severe turbulent flow has an important impact on
the formation process as well. The force generated by such turbulent flow is referred to
as shear force in papermaking. In general, the greater the degree of turbulent flow, the
worse will be the retention. This variation in retention will also have an effect on the
formation process. To this end, it can be concluded that the formation system is
subjected to a large number of variables during the wet end operation of paper machines.
As discussed, paper formation can be measured by the calculation of the solid distribution
of finished paper. Since the final association of fibres and fillers in the finished sheet
have a random nature as a consequence of their individual sizes and properties, this solid
distribution is a 2D random variable controlled by the above mentioned inputs.
The aim of this work is to implement a new on-line entropy control strategy for the wet end
section of the UMIST pilot paper machine, so as to demonstrate the achievable
improvement of the web formation through the control of an input retention polymer.
All experiments have been carried out on the paper machine at the Department of Paper
Science in UMIST. This is a typical Fourdrinier forming section with adjustable machine
speeds of 3-50 m/minute. There are two presses, seven pre-dryers, and three after-dryers
prior to reel up. For this paper machine, there is a beta gauge for basis weight measurement
and a BTG backwater consistency meter for on-line retention measurement. The schematic
of the real-time closed loop system for the web formation control is presented in Figure 2,
where a digital camera is used to grab images of the sheet as it travels through the machine.
Machine trials have been designed to study the effects of chemical addition (polymer) on
the formation process with all other inputs remaining constant. Experimental conditions
have been kept constant for all the machine trials to obtain consistent results. The pulp
blend used 70:30 blends of bleached sulfate chemical hardwood and softwood fiber
respectively, which was known to give a fairly well formed sheet. The hardwood was Eno
254
birch and the softwood was Lapponia pine and this blend is refined at 120 kWh/tonne. The
grammage was set to 60 g/m2 and the machine speed at 8m/minute. The polymer used is a
cationic polymer with concentration of 0.05%.
The best measure of the formation process is distribution of mass density, which is not easy
to obtain during on-line trials. Instead, the grayscale distribution of the paper image was
adopted as a means of assessing the web formation quality. It was therefore important to
find a way to obtain the information of the grayscale density on-line. The concept here was
to take images of the paper at each sampling time and then obtain the distribution of the
grayscale density by means of image processing. Using the obtained image, the entropy
calculation can be carried out within each sample period. Local variations in basis weights
can be determined by measuring the attenuation of visible light through the sheet,
providing the measurement occurs before the calendar and taking into consideration light
scattering effects of both fiber and filler [3]. A Kodak Professional DCS 520 Series Digital
Camera (based on the Canon EOS 1) was used for this purpose. It can provide a fast frame
rate (a continuous frame rate of 0.5 frames/second), a good resolution and an image size of
2.1 million pixels (1728x1152).
The camera was equipped with an IEEE 1394 (Firewire) interface, which can be plugged
directly into an industry standard interface available on most PCs, and facilitates high speed
transfer of images. This fast image transferring capability will satisfy the on-line control
requirements of the web formation process. The software development kit (PDC-SDK)
includes source code libraries supporting the direct use of a DCS camera without Twain and
Photoshop packages. The code libraries are in a variety of programming languages and can
access commands for image compression and storage. With certain image processing
programs and calculation techniques, the grayscale intensity distribution of the paper
image can be obtained within seconds and thus provide the on-line output measurement of
the control system. Positioning the camera near the end reel reduces the light scattering
effect due to moisture in the sheet, although large time delays would thereby occur between
control input and measurement output, due to the time taken for the sheet to travel down
the machine.
Some problems with the illumination system employed were initially encountered, when a
basic 500-Watt industrial halogen floodlight was tried. The floodlight provided a powerful
and robust light source and was also easy to install on the paper machine. However, it
provided an uneven illumination over the sample area being imaged, and allowances had to
be made, during subsequent image correction. The improved method then developed was to
add a piece of etched safety glass in front of the camera so as to produce a diffuse light
source. This effectively eliminated the uneven illumination. Figure 3 shows the schematic of
the optical sampling system ultimately devised
255
Modern paper machines have varied chemical additions for many different requirements,
such as wet strength, retention, and de-foamers, etc. Since it would be impractical to
include all chemicals in the initial trials, a cationic retention aid polymer has been used for
this work, so as to provide a simple single chemical input.
This polymer was added into the pulp by a pump whose input flow is determined by the
pump speed. As such, by setting the voltage of the DC-motor in the pump, the flow rate can
be adjusted. In this work, the voltage signal was generated from the PC platform and was
sent to the pump by a D/A converter card. A programme developed using Borland C code
provides data acquisition.
256
f (n) = 1
(1)
i =1
(2)
(3)
For the greyscale distribution of a paper image, its entropy is calculated from
255
H = f (n ) ln f ( n )
(4)
n =0
257
Once the entropy is calculated at each sampling instant, a pair of input and output
sequences can be obtained with the polymer addition as the input and the entropy as the
output. This is a single input and single output system, where the purpose of controller
design is to use the calculated entropy as a feedback signal so that the entropy of the closed
loop system can follow a given entropy value. For this purpose, the model describing the
mathematical relationship between the input polymer and the output entropy needs to be
established. In this work, the well known step response test is used, which is a convenient
way to characterise process dynamics because of its simple physical interpretation [5]. This
test can be easily implemented in industrial applications. Typical ways to obtain the
continuous-time transfer function models from a step response test include graphically
based methods, area-based methods and some least squares identification methods [6].
Different from the existing methods, the direct step-response identification method used
here has the following advantages:
1) Model parameters of the continuous transfer function are estimated directly with the
development of least squares regression form. No extra efforts are needed for model
transformation. The precision of the result is thus increased.
2) The time delay item is considered in the construction of a parameter vector and can
be estimated directly without prior knowledge.
3) The method is insensitive to data length. It does not need to wait for the process to
enter into the new steady state completely. This will reduce the time of step
response tests.
4) It is robust to measurement noise, because multi point integrations are included. The
identification algorithm for First-Order Plus Dead-Time (FOPDT) model is proposed
in [7]. Extensions to second-order system and other common systems are provided
in [6].
In terms of the current entropy control system, a FOPDT model
G p ( s) =
k
e Ls
Ts + 1
(5)
is sufficient to approximate the process dynamics within the operating range. Suppose the
process is under zero initial conditions before a step change with amplitude a is applied at
t = 0 , then the regression form of the FOPDT model can be formulated to give
k
[at a y (t )] Lk = A(t ) (t )
T
(6)
258
where A(t ) =
t
0
= +
(7)
for t = ti L , where
= [k Lk T ]T
atm
at
= m +1
M
at N
= [A(tm )
a y (tm )
a y (tm +1 )
M
M
a y (t N )
A(tm+1 ) L A(t N )]
and is the noise vector. Using this equation, the estimate for can be obtained using the
following least squares estimation
= ( T ) 1 T
(8)
Although empirically or optimally tuned PID controllers used without considering robust
performance may work well under certain nominal conditions, they cannot provide
consistent performance in the case of large model errors. This is true for the formation
control process where the model uncertainties can be caused by many variables, including
the variation of the stock chest level, the polymer concentration and the machine speed, etc.
As such, the control method used for the web formation entropy control is an IMC-base
robust PID controller. The inherently robust PID controller design combines robust control
technique with conventional PID tuning so that its ability to accommodate process
uncertainties is increased effectively.
The objective of the robust PID controller is to achieve a minimum integral squared error
(ISE) considering the entire closed loop response under the worst-case process dynamics,
and to satisfy the performance ratio requirement. The performance ratio r is employed to
adjust the speed of the closed loop system and its robustness.
r=
The optimization is based on the min-max rule, that is, to achieve the best performance for
the worst process situation. From the robustness point of view, the worst case means the
259
least stable case within an uncertain range. The solution of min-max design is less sensitive
to model errors and improves the overall control performance over the entire range of
model dynamics [8].
To solve this constrained optimization problem, a new method combining both numerical
search and explicit optimization is proposed in [9], where the design procedures include the
following steps:
1) Perform a general IMC-PID tuning for the nominal system without incorporating
specified uncertainty measurements, as occurs in the FOPDT model (5) with
standard PID controller,
Gc ( s ) = K p (1 +
1
+ Td s )
Ti s
(9)
Kp =
2T + L
K ( 2 + L)
L
2
(11)
TL
2T + L
(12)
Ti = T +
Td =
(10)
The key point is to design the tuning parameter according to the performance
ratio r . Since the IMC control can provide certain robustness [10], the pre-design
results are applicable in some situations when the requirement of robustness is not
very strict or the model uncertainty is not large.
2) Determine the worst-case situation within the specified model uncertain ranges. For
the FOPDT model, it is easy to find the worst case in terms of robust stability.
Suppose
the
parameters
are
within
the
range
of K [ K min , K max ] , T [Tmin , Tmax ] , L [ Lmin , Lmax ] , then the worst case should be
260
The dynamic of this process is determined from open loop step response tests within the
operation ranges. The system input is the polymer flow added into the pulp and the output
to be measured is the entropy of greyscale distribution within the paper image (see Figure
6).
Moter voltage
of Polymer
u
The open loop step response tests have been repeated many times for different amplitudes
of the input and at different operating conditions. The purpose of this was to observe the
uncertainty of the model and, at the same time, decide the input range of a linear model.
The data of the calculated entropy has been filtered with a time lag unit to reduce the
effects of noise pollution. The auto-correlation result presented in Figure 7 shows the
effectiveness of the filter.
261
The fitting curves from a step response test are shown in Figure 8.
G p ( s) =
0.03 150 s
e
45s + 1
with 25% of parameters uncertainty. This implies that the ( k , T , L) might change 25%
from the nominal model parameters under different operating conditions. The large time
delay is introduced by the position of the camera, as discussed previously.
262
At each sampling instant, an image of the selected paper area is produced by the camera,
transferred to a PC, and saved as a *.tif file. The entropy of grayscale distribution is then
calculated within the sampling time and applied as the on-line measured output. The
sampling time was 8 seconds for the pilot machine trial.
The closed loop result is presented in figure 5.4 with the tuning PID controller K p =17,
263
10.8.8 CONCLUSIONS
In this paper, a new entropy control strategy has been implemented for the web formation
at the wet end of a pilot paper machine. The innovative idea of entropy control for
stochastic system has been proposed and proved to be applicable for this system. The
effectiveness of the new modelling method and the robust controller design method has
been demonstrated through real-time tests. The current research concerns only the
influence of the polymer on formation quality, which may be insufficient for the
requirements of the real paper industry. Further research should consider the retention
system with fibre as well as filler. Therefore, other control inputs must be considered and
the control strategy will be considerably more complex.
10.8.9 REFERENCES
1. Trepanier, R.J., Jordan, B.D. & Nguyen, N.G. (1998). Specific Perimeter: a statistic for
assessing formation and print quality by image analysis , Tappi Journal, 81(10), pp.
191-196.
264
2. Smook, G.A., (1994). Handbook for Pulp and Paper technology, Ed.2, Angus Wilde
Publications.
3. Landmark, P. & Joensberg, C., (1984). Development of an On-line Formation Tester To
Determine Optimal Use of Retention Aids, Paper Trade Journal, No.9, pp.84-86.
4. Papoulis, A. (1991). Probability, Random Variables, and Stochastic Processes, Third
Edition, McGraw-Hill, Inc., Singapore.
5. Astrom, K. J. & Hagglund, T. (1995). PID Controllers: Theory, Design and Tuning, 2nd
edition, Instrument Society of America, USA.
6. Wang, X.Z., Yue, H., Gao, D. J. (2000). Direct Identification of Continuous Models with
Dead-Time/Zeros, Proceedings of the 3rd Asian Control Conference, Shanghai, China,
July 4-7, 2000, pp. 1709-1714.
7. Bi, Q., Cai, W.-J., Lee, E.-L., Wang, Q.-G., Hang, C. -C., Zhang, Y. (1999). Robust
Identification of First-Order Plus Dead-Time Model from Step response, Control
Engineering Practice, 7(1), pp. 71-77.
8. Honeywell Inc. (1995), Robust-PID User-Manual.
9. Yue, H., Gao, D. J. & Liu, S. (1999). A Method for Improving Robustness of Industrial PID
Controllers, Chinese Journal of Automation, 11(3), pp. 263-269.
10. Morari, M. (1989). Robust Process Control, Prentice-Hall Inc., New Jersey.
Additional references:
Austin, P., J. Mack, D. Lovett, M. Wright and M. Terry: Aylesford PM 14
Improved wet end stability using model predictive control, Paper Technology 43 (5),
41-44 (2002)
Austin, P., J. Mack, D. Lovett, M. Wright and M. Terry: Improved wet end stability
of a paper machine using model predictive control, Control Systems 2002, Swedish
Pulp and Paper Research Institute, Stockholm, 2002, 80-84
Belle, J., J. Kappen, F. Goedsche and H.-A. Khrt: Simulation-based control of the
sizing rate of resin sized paper, professional papermaking, No. 1/May 2005, 23-29
Brosilow, C. and B. Joseph: Techniques of Model-based Control, Prentice Hall PTR,
2002
Cho, B.-U., G. Garnier and M. Perrier: Dynamic modelling of retention and
formation processes; Control Systems 2002, Swedish Pulp and Paper Research
Institute, Stockholm, 2002, 200-204
265
One reason for using an operator training simulator can be to reduce the amount of e.g.
paper breaks or down time of the process during the start up phase of a new paper machine
or pulp mill. As often people with very little experience of paper machine or pulp mill
operations are hired for new green field mills, it will be very risky to start the new process,
if the operators do not get good training in advance. Here it can be interesting to refer to a
study done in the US on how much we remember of information we are fed with:
10 % of what we see
30% of what we see and hear simultaneously
70 % of what we also train at simultaneously
and close to 100 % of what we repeatedly train at.
This is the reason for training at a dynamic process simulator, as the operators can acquire
a very good knowledge even before the actual start up of the mill.
266
If we just make some rough estimates on what benefits a training simulator can mean, we
may assume 10 % higher production the first month after starting up a new process line.
For a paper mill this would mean approximately 400 USD/ton* 1000 tpd * 30 days* 10% =
1.2 MUSD in earnings. One paper break can be worth some 100,000 USD in lost production
etc.
This is no guarantee on earnings, but a qualified estimate, showing that this is not just a
game for fun.
At MNI, Malaysian Newsprint Industries, the start up to full production was achieved in a
20% shorter time period than normal although operators with no previous experience
from pulp and paper industries were recruited. They had been trained during a eight week
time period, half day in the simulator, while the other half being out in the mill looking at
the real hardware.[Ryan et al 2002], [NOPS 1990].
Both ABB with Simconx and Honeywell SACDA have simulator systems primarily for
operator training. First principles algorithms made in FORTRAN and C++ are covering
process equipments are linked in a sequence. Flows in and out of the equipments are
determined from a separate pressure flow calculation, which iterates every time step with
the material balance calculations inside the algorithms. In these chemical reactions,
separation etc is determined. Changes in inventory in each equipment is calculated, and
thus giving the dynamics. For valves, pumps etc the process model is interacting with DCS
controls. If the flow is too low, the control system sends a signal to open the valve. Next time
step the valve position then becomes a little larger, and the pressure drop becomes less, and
the flow increases. In this way the dynamics also includes the DCS interactions.
The DCS controls are run as emulated code in the simulator in these cases. This means that
FORTRAN code is produced for the functions, and a translator transfers the DCS
configuration into the simulator automatically. This means that in reality everything is run
in the simulator, and the actual Process Computers are not needed. In the modern versions
everything is on PC with DOTNET, while earlier everything was run on an Alpha computer,
for both simulators.
These simulators are the most advanced dynamic simulators on the market, where a
complete plant with full controls can be operated in real time!
Simconx uses G2 as the graphical interface. In this way full object orientation can be
achieved with inheritance of functions, although the code in each module is still old
fashion FORTRAN. Modules for different equipments from the 70th are operated together
with those produced today, which is a great advantage. Otherwise it is too expensive to
transfer code from one system to the other, having to rewrite major parts of the code.
When the system is used for operator training the teacher is sitting on the other side of a
glass window, so that the teacher can see the trainee, but not the opposite. The teacher then
267
can introduce actions without the notice by the operator outside what is seen on the
process displays.
Simulated Process
Operator Station
Operator Station
Instructor
Station
AS 515
AS 515
PC
The same
iGES
Graphical User
TCP/IP
MB 300
The same
Emulated
Controls
I/O
ABB Simconx
Process
Model
Figure 1 The configuration of ABBs training simulator including DCS system, process
simulator , iGES (G2) graphical configuration and engineering interface and the relation to
the process.
Typical process displays used in operator training looks like below for a digester plant in
pulp and paper industry (figure 2).
268
Figure 2 . Typical operator display from a process industry. In this case from a digester in
pulp and paper industry.
- Sequences
- Group starts
- VALVECON, the valve controller
- The same
- PIDCON, the PID controller
- The same
- Also trend log for e.g. Flow rate
- Let operator build a trend log, following a detailed instruction??
- Recipe
- Show how you bring up on screen
- Show how you enter new values
2. Select a sequence and run through
- Pulper 1
- Explain what happens, along the sequence
3. Trip a motor, e.g. the pulper rotor (stirrer)
- Show the alarms, and the alarm list
- Group start and sequence display
- Object display, for the tripped motor
- Show that motor stops because to heavy load, as too high concentration
- The fault can be level sensor gave wrong level. Assume fixing this.
- Add some water, and then restart pulper
4. Now go through the whole process from pulper to storage tower, for the pulp mill.
- Look at values and what the displays look like, when everything is OK.
- Explain what is happening on each display
- Explain, that on the displays, a lot can be seen, but not everything
270
- Exemplify the latter by shutting off the reject totally for a series of screens, and
show that actually not too much is seen. You have to go out in the mill to see that reject is
actually coming out. This should be done now and then.
- Explain the impact if no reject- a risk that shit is passing through, causing more
pin- holes in the final paper product, although far down the line.
5. Set a screen rotor stopping
- Look at the pressures and flows, what happens
- Tell that now you have to go out and clean the screen.
- Restart, by going back to an original initial condition.
6. Change the flow rate in the whole line slowly, by 20%.
- Look at some flow meter trend of some main stream, e.g. to flotation or in the
cleaner line.
7. Go back again
- Follow trend of some main stream, especially in to flotation and in the cleaner line.
8. Start drifting a flow meter. Dont tell what you are doing right now. After a certain time
period, we should have a look at what is going in to flotation, and what is
into the
cleaner line.
- Ask why one is flat, and the other goes down. Explain due to drift in sensor.
9. Shut down the process
- Start with the pulper. Dont feed anything more.
- Then stop the process feeding, filling in with dilution water for an easy start up
again, in pumps, cleaners and screens.
- Then stop according to written instruction
10. Start up a process from initial conditions, where everything is filled with water.
11. Do some more faults:
- Seal water failure - motor stops in e.g. screen
- Too high conc. in screen- clogs. Have to clean.
- One pump in flotation stops- all pumps stops.
- Clog reject of 15 % of cleaners in a battery. What happens?
- Too high level in a tower. What to do?( may need replay, to run in a different way?)
271
PAPER MILL
1. Start bringing up one display, e.g. the mixing chest. Explain the functionality of the DCS
display first:
- Alarmlists and acknowledgement
- Event list
- MOTCON, the motor controller.
- What tools?
- Object display
- Sequences
- Group starts
- VALVECON, the valve controller
- The same
- PIDCON, the PID controller
- The same
- Also trend log for e.g. Flow rate
- Let operator build a trend log, following a detailed instruction??
- Recipe
- Show how you bring up on screen
- Show how you enter new values
2. Select a sequence and run through
- Decide suitable.
- Explain what happens, along the sequence
3. Trip a motor, e.g. the agitator rotor (stirrer)
- Show the alarms, and the alarm list
- Group start and sequence display
- Object display, for the tripped motor
272
4. Now go through the whole process from pulper to storage tower, for the pulp mill.
- Look at values and what the displays look like, when everything is OK.
- Explain what is happening on each display
- Explain, that on the displays, a lot can be seen, but not everything
- Exemplify the latter by shutting off the reject totally for a series of screens, and
show that actually not too much is seen. You have to go out in the mill to see that reject is
actually coming out. This should be done now and then.
- Explain the impact if no reject- a risk that shit is passing through, causing more
pin- holes in the final paper product, although far down the line.
5. Set a screen rotor stopping
- Look at the pressures and flows, what happens
- Tell that now you have to go out and clean the screen.
- Restart, by going back to an original initial condition.
6. Change the flow rate in the whole line slowly, by 20%.
- Look at some flow meter trend of some main stream, e.g. from white water tank 1.
7. Go back again
- Follow trend of some main stream, as above.
8. Start drifting a flow meter. Dont tell what you are doing right now. After a certain time
period, we should have a look at what is going in to flotation, and what is
into the
cleaner line.
- Ask why one is flat, and the other goes down. Explain due to drift in sensor.
9. Shut down the process
- Start with the feed from the pulp mill.
- Then stop according to written instruction (feed to wire, steam etc)
10. Start up a process from initial conditions, where everything is filled with water or pulp.
11. Do some other failures, like
273
- Shower water failure. Explain that this will cause wire to clog faster if not
adjusted. Then the production will be influenced, and even the wire has to be
changed.
- No vacuum in deculator, indicated from instructor as too many pin-holesWhy?What caused the effect? Look at the vacuum etc.
- To high level in broke tower. How react. Balance to avoid (do replay).
This meant that MNI acquiring a completed design and all the major equipment
from Voith Sulzer and DCS from ABB. Consequently, this meant that the projected
time line till start-up was relatively fast. In a normal project the critical path is the
delivery time for the normal green field paper-machine is the paper machinery, but
in MNI case, this was the main process building)
In addition the senior Management of MNI required a world class start-up of the
mill. This target was to meet start up was based on the previous start-up of paper
machines
It was also soon clear that MNI could not hire enough experienced pulp and paper
operators from within Malaysia, so the only viable option is to train up a team to
meet the challenge of meeting the start-up curve.
The task of the Operation Mill Team (OMT) was to deliver an operating team to
meet the startup curve.
Two process engineering left for Sweden to assist in the development At the same
time they would gain the experience to operate the simulator during the later
operator training
274
As some of the trainee did not have exposure to a DCS system, the simulator was
used to train the operator in the basic of the ABB Advant system. This included
Navigation, Trending, menus etc, Meaning of Manual, automatic and E1 (Cascade)
Familiarise the trainees in how the process is represented on the DCS system.
E.g. what and where the controls for each item are located are located, control loops
are presented
For these reason a conscious decision was made to give priority to those element in
the DCS that were visible to the DCS operator.
20 P&IDs
500 valves
276
Figure 2 View of the engineering computer and Alpha Computer with the simulator models
One of the most important aspects being the customer was to define what the limits to the
simulation are.
This usually takes the form of what processes should we simulate and what information
about the streams should we include in the simulator model.
During the development of the simulator it quickly became apparent that it would be
impractical to simulate all the physical process especially for the paper machine compared
than the RCF plant.
The reasons behind being unable to simulate are varied:
Too complicated
Commercially sensitive
Information doesnt exist
Not cost effective
An example: The Simulation of the paper machine was complicated by the PLC control of the
certain items of the paper machine such as on line calendars or the PLC controlling the
277
amount of crown on the press rolls. In the end a decision was made to eliminate these from
the model.
The most important rule that we used to make these decisions was to determine if the
operator could observe the effect of the simulation. Effort was put into things that are
material to the operator e.g. alarms,
The other major area to consider for the simulation is the number and type of the stream
parameters that should be modelled. In general these are properties that are conserved
such mass, energy, etc are very easy. However simulation of properties such as pH, pulp
free, ink specks number are not conserved and are depending on other physical properties
that are not well defined.
Thus the choice of the stream parameters to be models needs to be balanced against the
advantages vs. the difficulty involved.
What is required for modelling both physical process and stream properties is that there
must be a reason for why you want to model this parameter. In other words - be realistic,
not everything is worth simulating.
In the end analysis, for the simulator to achieve the objective of training the
operators, details knowledge is not generally required.
We found a large number of errors in the configuration of the DCS system. These we
communicated back to the ABB/Voith DCS people for rectification. The importance
of correcting these errors is that the majority would have slowed down the
commissioning of the plant during the water run and or worse still the initial fibre
runs. Typically these error would be such things as Reversed controllers, reversed
logic connections
We also found that complete section of DCS configuration were also missing or did not
correspond with other sections. E.g. Bleaching tower controls etc.
Another aspect is that developing the DCS simulator forced the people involved such as the
pulp mill manager, myself as the technical manager and the process engineers working of
the development of the simulator developed a much deeper understanding of the way that
the control system was designed to operate. This information is simply not available on the
PI&D. It also forces you to determine exactly how a piece of equipment has been configured
to operate (almost a rigorous process to teach both the operators and the management
about the process)
Everyone involved in the development process gained a detailed knowledge of how the
plant would run before the plant start-up.
278
Flotation Cells
L.T
To Fine Slot
Screens
To Secondary Flotation
From White Water
FiltrateTank
Rerverse
Flow Alarm
Figure 3. Process layout for the flotation unit with the too large valve
When commissioning the simulator, the simulation would keep shutting down on a
reverse flow alarm.
The reverse flow alarm is to prevent stock from flowing back into the filtrate tank if
there should the amount of stoke flowing forward should exceed the pumping
capacity of the pump.
To control the level in the flotation cell, the valve to the suction of the screen feed
pump would open up, increasing the supply of stock. Since there is a large amount of
water in the floatation cell, the flow through the valve must increase too greatly to
bring down the level within a reasonable time. This would result in more stock to
the pump could handle and the excess would then flow back to the filtrate tank and
generate a reverse flow alarm.
The solution was to clamp the opening of the valve to such an opening. This
prevents a flow that is too high for the pump and hence prevents a reverse flow
alarm.
The training program start with a basic induction course to ensure that every one
have basic level of English and other education skills
This was followed by a introduction to pulp and paper .This was designed to give the
students a basic understanding of the unit process such as pulping screening and
cleaning etc
279
The people then spent approximately 6 weeks at one of the companies sister mills to
gain basic experience in their respective areas.
This program was then competed with a extensive program on the training
simulator
Simulator training was organised into small teams of three to four people to maximise the
ability to of each of the plant operators to use the simulator hands on experience.
To get through all the scheduled training, teams scheduled 24 hours per day
Things that were practices on the simulator:
A) Start-up and Shut down both lines or shut down line 1 only
B) Change from operating 2 lines down to 1 line only and vice versa.
C) Change the production rate by ramping up or down the production rate
The actual training for the operators was very well received. I (Kevin) believe that this was
because the simulator was absorbing the operators interest and it was interactive. When
using the simulator in fact they often requested additional simulator time
Simulated faults introduced into the system required a more thought than we originally has
estimated since
If the fault was too bad, the process interlocks would shut the process down immediately
And if the system fault was too minor, the DCS operator usually has very little or nothing to
indicate that there was a fault in the entire system
It was very clear that the combination of the training program and the simulator resulted in
a team of operator that was very well trained and prepared for the start-up
It was also clear that the simulator increased the level of the confidence of the operators. If
they could run the simulator then they could run the real thing
One of the definite benefits was the increased confidence of the operators when faced with
commissioning. This allowed the operators to keep up with the commissioning personnel
rather than play catch up. Usually the commissioning people are concerned with the process
of getting the plant up and running rather than training and coaching the operators.
Comments from the equipment suppliers indicate that they were impressed with the
understanding that the operators had of the process.
Conclusions:
Any design error that the simulator finds is likely to more expensive to fix as the
project progress
Decide what you want the simulator is needed for. This will impact decision on the
type an extent of the process units and stream parameters that will require
simulation.
To get the best use of the simulator required it to be incorporated into an extensive
program. The simulator is a very good way to bring all the training together before
starting to practise with the very expensive real and very easily damaged
equipment.
For MNI the process of development was important in fully understanding the RCF and PM
plants. It also meant that the production manager, the technical manager and two process
engineers learnt both the process and the interaction with the DCS system as they first
made the training program, and later performed the training with the operators.
MNI achieved the target Production Start-Up Curve. This meant to reach the full production
20% faster than is normal at an Australian mill with experienced staff!
Introduction
Paper web breaks account for 2-7 percent of the total production loss, depending on the
paper machine type and its operation. For a modern paper machine with production of 330
000 tons a year, this means 1.5-6 million Euros lost in a year (Ahola 2005).
Computational break prediction systems have been studied. Various signal processing
methods have been used in data classification: high and low signal levels, abrupt changes
and rapid trends and bumps (Suojrvi et al., 1996, Ihalainen et al., 1997). The main problem
with signal processing methods is that correlations change in time and from one machine to
another so some kind of adaptation is required. There are some reports (Hokkanen 1996;
Li and Kwok 2000; Ihalainen 2000) where PCA has been used successfully for the
classification of break situations. Independent component analysis (ICA) was tested in (van
Ast and Ruusunen 2004). Reported neural network applications utilise back-propagation
network (Miyanishi & Shimada 1998) and Kohonen Self-Organizing Map (Sorsa et al. 1992).
Sensitivity indicator
281
The web break sensitivity indicator was developed as a Case-Based Reasoning type
application with Linguistic Equations approach and Fuzzy Logic (Ahola 2005). The Case
Base contains case models with different number of breaks. (Ahola 2005, Ahola and
Leivisk 2005). The system requires a lot of data pre-processing and modelling calculations,
but these are introduced elsewhere. The hierarchical model structure is shown in Figure 1.
Figure 1. The model structure. Categories here refer to the number of breaks during the 24
hours modelling period. Category 1 means no breaks and Category 7 a lot of breaks.
The calculations start when a new set of measurements is coming from the paper machine.
The system defines the best fitting web break category and the break sensitivity. The best
fitting category is selected using a simple fuzzy strategy. If an equation is true, the degree of
membership for the equation is one. All deviations reduce this degree according to a
triangular membership function. The degree of membership for each case is evaluated, in
the original system, by taking the weighted average from the degrees of membership of the
individual equations.
nei
ci =
w
j =1
ej
nei
ej
0 wej 1
where ci refers to the ith case, ej to the jth equation and nei is the number of equations in the
ith case.
The degree of membership for each web break category is calculated from the degrees of
membership for all the cases included in the category. Once again, the weighted average is
used.
282
nci
( cat ) k =
w
ci
i =1
nci
ci
0 wci 1
Figure 2. Example of results using the weighted average method (Ahola 2005). White
vertical lines show the actual breaks, black fractional lines the break sensitivity calculated
from the actual breaks and black line the predicted breaks.
Three other methods were tested by simulation (Ahola 2005). In the first method, the
biggest membership was given the weight 1, the second a weight 1/3, third the weight 1/9
and so on. In the second method, only the biggest membership was used. Finally, the third
method based on the biggest adjoining memberships. This method starts from the biggest
membership and adds to it the effect of two closest to it. The biggest membership gets the
weight 2/3 and the average of the two adjoining memberships the weight 1/3.
283
The selection of the correct calculation method improved the results markedly, as Figure 3
shows. It should also be noted that in this case the output was the average of calculated
values for 120 minutes.
Figure 3. Results when using the maximum membership as the basis of calculations (Ahola
2005). Here vertical lines show the actual breaks, white fractional line the break sensitivity
calculated from the actual breaks and black thicker line the predicted number of breaks in
24 hours.
References
Ahola T. (2005) Intelligent estimation of web break sensitivity in paper machines. Acta
Universitatis Ouluensis C232. University of Oulu. http://herkules.oulu.fi/isbn9514279573
Ahola T., Juuso E. and Leivisk K. (2006) Variable selection and grouping in paper machine
application. Proceedings of ALSIS06 - The 1st IFAC Workshop on Applications of Large
Scale Industrial Systems. August 30-31, Finland. Also in: XXX
Ahola T., Juuso E.K. and Oinonen K. (1997) Data analysis for web break sensitivity indicator.
Proceedings of TOOLMET '97 - Tool Environments and Development Methods for Intelligent
Systems, Oulu, 1997, pp. 150-159.
Ahola T., Kumpula H. and Juuso E. (2002) Simulation of Web Break Sensitivity. In
Proceedings of SIMS 2002- The 43rd Conference on Simulation and Modelling. 26-27
September 2002, Oulu, Finland, pp. 178-184.
Ahola T. and Leivisk K. (2005) Case-based reasoning in web break sensitivity evaluation in
a paper machine. Journal of Advanced Computational Intelligence and Intelligent
Informatics. 9(2005)5.
Ihalainen H., Paulapuro H. Rislakki M., Ritala R. and Suojrvi M. (1997) Novel method to
analyse relationship between web breaks and process variability. 47th Canadian Chemical
Engineering Conference, Edmonton, Alberta, Canada, 5.-8.10.1997. 1997, University of
Alberta, Abstract.
284
Li I.S., Kwok K.E. (2000) Prevention of sheetbreaks using multi-variate statistical analysis in
an expert system environment. Pulp & Paper Canada 101(11):T336-340 (November 2000).
Miyanishi T., Shimada H. (1998) Using neural networks to diagnose web breaks on a
newsprint paper machine. Tappi Journal, VOL. 81:NO.9, September 1998.
Sorsa T., Koivo H. N. and Korhonen R. (1992) Application of neural networks in the
detection of breaks in paper machine. Preprints of the IFAC Symposium On-line Fault
Detection and Supervision in the Chemical Process Industries, Newark, Delaware, USA, April
22-24, 1992, pp. 162-167.
Suojrvi M., Ritala R., Ihalainen H. and Jokinen H. (1996) Finding significant changes from
process data. The yearbook of Finnish Statistical Society 1996, part I, pp 75-82. Helsinki,
Finland: Finnish Statistical Society, 1996, 255 pp.
van Ast J. and Ruusunen M. (2004) A guide to independent component analysis Theory
and practice. Report A No 23. University of Oulu, Control Engineering Laboratory. Oulu
2004. 53 pp.
285
Flotation is a wastewater purification process, which uses air bubbles to remove impurities
from water. Usually, pressurized air is supplied to reactor forming air bubbles due to
significant pressure drop. Impurities attach to bubbles and rise at the water surface. Some
chemicals are used to make process more efficient. Basically, those chemicals bind
impurities together, thus increasing the amount of impurities attached to an air bubble. In
such a water purification process, quality variations in inlet water have great influence on
process efficiency. It should also be noticed that used chemicals are often quite expensive,
thus overdose is not desired. (Ainali et al. 2002)
An indirect measurement was developed for a flotation unit that gave a quality index for
inlet water. The calculation of quality index was based on the deviation in process
behaviour compared to typical process model. The typical process model was developed
with linguistic equations. The indirect measurement not only gave valuable information for
process modelling but also was used in process control. A feed forward controller was
developed for the flotation unit. The controller strategy was solely based on the water
quality index and expert knowledge. (Ainali et al. 2002)
In the literature, the uses of interval observers and neural networks have been reported
(Choi and Park 2001, Hajd-Sadok and Gouz 2001, Alcaraz-Gonzles et al. 2002).
A software sensor was developed for the recovery boiler used in the pulp and paper
industry in (Murtovaara et al. 1996). The aim was to further improve the edge detection
capability, the image processing algorithms, and the tuning of the threshold parameters in
an existing commercial system for the image analysis of the recovery boiler. These directly
affect the searching for the contour of the bed. The resulting improvements were to be
embedded in the existing system.
The efficiency of the edge detection was improved by developing a fuzzy algorithm. It
presented two improvements compared to the existing system. The first was to utilise the
286
previous contours in the analysis. The algorithm generated and updated membership
functions for each contour point on the basis of history data. Then it defuzzified the
resulting fuzzy numbers into a new contour. Defuzzification was based on the centre of
gravity method. According to the tests, this algorithm filtered out fast changes of the
contour. The second improvement was to include the adjacent (neighbouring) points in the
analysis. This corresponds to the filtering in the space dimension. If the distance between
the predicted point and the point given from the image processing exceeded the limit value,
algorithm eliminated the point from the edge.
The horizon of the history data effects also on the performance of the algorithm. A suitable
number of contours in history data was 10 since the effective changes in the state of
burning processes are slow, a typical time constant being in the order of minutes. The fuzzy
algorithm was developed and tested in Matlab and later coded in C-language in the actual
commercial system as an individual module. It was tested and found to filter out fast
changes and to improve the edge detection capabilities.
Another indirect measurement application for pulp and paper industry is developed for a
rotary lime kiln. Lime is used in recausticising process for recovery of white liquor. A lime
kiln is a huge heat exchanger, where lime mud is burned into lime. The operation of lime
kiln has three stages: drying, heating, calcining and agglomeration of the formed CaO
powder. The main problem in lime kiln control is the lack of on-line measurements for the
product quality. Efficiency of the process depends on operating temperature and
temperature profile over the kiln. Too low temperature leads to product quality
degradation while too high temperature leads to a decrease in causticising efficiency.
Additional control problem is due to the use of sawdust as a fuel in burning. Sawdust quality
changes randomly leading to fluctuations in heat production and temperatures. (Jrvensivu
et al. 2001)
As the process is rather sensitive to occurring temperature changes, adaptive control is
needed. Adaptation has been implemented through scaling of control variable. Scaling
coefficient was determined based on an indirect measurement defining operating condition
changes. Indirect measurement was implemented as a LE model inputs being certain
important process variables.
287
Rottneros, Vallvik, is a kraft pulp mill in Mid-Sweden. In 1998 a gas explosion followed by a
steam explosion took place in the recovery boiler. The interest for an early warning system
that is removing the risk for this happening again thus is very strong!
In a project performed in co-operation with the branch organisation for heat production in
Sweden, Varmeforsk, and also co-funded by SEA, Swedish Energy Agency, a project was
performed where an early warning system was developed and implemented by Malardalen
University.
The focus of the system was to identify the sensors available for keeping control of the
energy and material balance for the recovery boiler. These were then logged and used to
calculate the energy and material balance on-line, where both the immediate value and the
trend for different time perspectives were analysed, and used as input to a Bayesian Net.
As a complement the blow down water is analysed with respect to conductivity. If there is a
break in a tube in the actual boiler, the conductivity will go down, as water is pouring out
instead of being evaporated, increasing the conductivity in the steam drum. If on the other
hand the break is in the ECO-part, the conductivity will not be affected. In this way an
indication of the position, and risk for explosions, can be achieved.
This information together with bed camera information on bed and combustion are all
given as input to the BN. This then gives a feed back to the operators about possible risks
for tube leakage, sensor faults, and poor performance of the boiler as such.
The system has been implemented taking data from the DCS system, but with the actual
calculation on a separate computer. The process flow sheet is seen in figure x.
288
Figure 1. Recovery boiler at Rottneros, Vallvik, where an early warning system has been
implemented to detect risk of tube leakage, sensor faults and other problems.
289
Working Group C (WG-C) of the COST Action E36 has been entitled Use of Simulation
Software in the Pulp & Paper Industry. This WG intends to bring software developers and
(possible) users together in order to come to agreements on contents, features, relevance and
performance of software products.
WG-C is specially intended to exchange knowledge between the WGs and to foster the
development of better simulation tools with a high compatibility across platforms used.
This WG has performed two different working paths in order to achieve the results:
The first one has been the elaboration of three simulation samples with
different complexity grade, starting with a steady-state one and continuing
with dynamics and process control. These sample models have been
constructed under different simulation software packages by WG-C
members. In addition, there is an open space for presentations of each ones
works before the end of the action.
This chapter will not focus on the first option, which is being reported in COST E36 Action
as additional information. Description of each software package is out of the scope of this
book, and last survey will be finished by the end of 2007. Thus, this chapter will cover
current use of software within COST E36 members, and the most important criteria to take
into account when referring to simulation software, according to people from the pulp and
paper sector.
First of all it must be strongly noticed that the conclusions have been made from 17
replies only from members of this COST E36 action. Thus, they can not be extrapolated
to the overall industry or any other group of companies, research centres, etc. Still it
gives at least some indication of what softwares are used by a relatively representative
group of users in European pulp and paper industry.
12.2.1 PREFACE
The purpose of this survey was to give an overview of the current use of software by COST
Action E36 members. All the questionnaires received from the different members have been
analysed, and the conclusions were considered as a starting point to clearly distinguish
every members needs and goals and translate them into the whole WG-C goals.
First of all, we would like to thank the members that have contributed with their replies:
CTP, France
ICP, Slovenia
KCL, Finland
STFI-Packforsk, Sweden
PTS, Germany
UCM, Spain
UdG, Spain
VTT, Finland
VPC, Slovakia
Some general questions were asked, to get an overview of current situation concerning the
use of simulation software.
By hand
22%
Other
2%
With specific
software
20%
With Excel
34%
292
No
41%
I don't know
41%
A high percentage (41%) of the replies doubted whether they should use other software
tools or not. A conclusion was that the WG-C would try to help all the members to solve
these issues in order to get maximum benefit from any software and for everyones specific
purpose.
One point to focus on was the strengths of each software package. These strengths,
according to the results, make the COST Action E36 members not to consider changing
software.
Changing software is quite difficult when users have already been learning one, since all of
them have tricks and details. Assessment can be mainly useful when there is a decision step.
In that way, it was decided to classify the most useful criteria to be evaluated for each
software package.
293
Yes
24%
No
47%
The importance of defining objectives is crucial because WG-C will make the assessment of
different software packages and this assessment must consider the specific environment of
this WG.
Members that work with mills or pilot plants can add more experience in complex data
acquisition systems, addressing the points to be boosted and those to be avoided.
294
Static
13%
Dynamic
0%
Both
87%
Dynamic is a simulation that takes time variable into account. In that case,
user can introduce changes and see both the effects and the time that these
effects need to take place.
Dynamic is a simulation in which the user can change parameters, but time
is not needed. This point of view can be seen as changing between different
steady states.
295
Other
50%
There is a great variety of used software packages. Table 2.1 shows the software used by the
participants, classified by application. It also shows the number of users within the COST
E36 participants.
Excel
Number of replies
using it
4
Access
Application
Name
Data acquisition
296
Application
Name
ADV V3.2
Data analysis
Number of replies
using it
1
KCL-Wedge
LabView
MS-SQL
NivuLOG
Optimas
PI (OSI-Soft)
WinMOPS
Excel
Matlab
Design Expert
KCL-Wedge
Simca
SPSS
Statgraphics
Voiths tools
Darec
297
Application
Name
Extract
Number of replies
using it
1
GPC
Maple
298
Data analysis
Simulation
Modde
Neuronline
NNModel
SAS
Unscrambler
WinGEMS
APROS Paper
Excel
Simulink
ABAQUS
Balas
Cadsim Plus
CASSANDRA
Chem Up
Dymola
FlowMac
gPROMS
IDEAS
KCL-ECO
KCL-PAKKA
299
Matlab
NNModel
ReKo
TNO Models
ABAQUS
AnovaTM(Taguchi method)
FEMLAB
OPTO
Solgas Water
Other
300
All software packages were analysed by their users. An average of the evaluations was made
for the most used ones. Results are available at the report in COST E36 page.
As a summary, a table can be made with these software packages and their main benefits
and drawbacks.
Benefits
Drawbacks
Data acquisition
Excel
Code accessibility
Easiness of use
I/O tools
Data analysis
Excel
Easiness of use
Code accessibility
Matlab
Code accessibility
Handbooks
Improvement possibilities
Easiness of use
Simulation
WinGEMS
Easiness of use
Code accessibility
Other
Aspen Plus - HYSIS
301
Benefits
Drawbacks
Complete
Code accessibility
It can be seen that easiness of use is a common strength of most used software, but their
code accessibility is pointed as a weakness. Thus ideal software should be easy to use, cheap
and with high code accessibility to introduce new features that may fit everyones needs.
The following questions were carried out as an open space to include different opinions:
Topics and themes in which you have more interest for Working Group
discussions.
Do you know other potential partners that could have interest in this
Working Group?
What do you think that this Working Group should achieve at the end?
12.2.6 CONCLUSIONS
Utility of this COST Action has been reflected in the responses. All possible doubts show that
a good assessment of all software should be of great value not only for members, but also
for pulp and paper mills and research centres.
302
According to the replies, there are lots of questions, topics, algorithms and software
packages. Thus, there is a hard work to be done in this WG-C, if a helpful assessment must
be carried out.
12.3.1 PREFACE
The purpose of this survey was to give an overview of the requirements that different users
groups have for simulation software. Different criteria have been analysed in order to
classify them by their importance to the users. The questionnaires have been received from
different people at DOTS-ATIP Conference held in Annecy (France) on April 2005; COST E36
Workshop on Validation held at Espoo (Finland) on October 2005, and COST E36 members.
We thank all off them their contribution.
A study over different criteria about simulation software, in order to classify the importance
that users give them in their work, has been carried out with 56 replies from people from
10 different countries:
Denmark
England
Finland
France
Germany
Norway
Romania
Spain
Sweden
303
The Netherlands
Replies have been classified according to different affiliations. Results are shown in figure
3.1.
Number of replies: 56
Suppliers; 5; 9%
Paper
Manufacturing;
16; 29%
Research
Institute /
University; 33;
58%
Engineering /
Consultancy; 2;
4%
Number of
replies
Matlab (Simulink)
14
304
Number of
replies
WinGEMS
Balas
FlowMac
Aspen-Hysis
Biowin
CadSim Plus
IDEAS
ACSL
APROS
C++
FcoSim Pro
Femlab
Fluent
Fortran
Modelica
PI
Siemens tools
Witness
305
The importance of each criterion has been analysed and the general results have been
summarised in table 3.2.
Results show that criteria concerning the supplier are the most important for people
working in the pulp and paper industry. Both reliability of the software and reliability of the
supplier are the most important criteria.
On the other hand, total cost of the system, licensing policy of the supplier, the possibilities
of handling both chemistry and chemical reactions and compatibilities with other software
packages are considered as the least important criteria. This points show that:
Total cost of the system usually does not matter too much.
306
no opinion
non important
(5)
Low (4)
Average (3)
High (2)
Criterias
Importance
NUMBER OF REPLIES
General
2.46
19
22
2.30
18
17
13
2.32
27
19
2.14
13
25
15
2.14
19
17
13
2.37
10
24
16
1.67
25
24
2.20
307
no opinion
non important
(5)
Low (4)
Average (3)
High (2)
Related to suppliers
Importance
NUMBER OF REPLIES
1.87
20
23
11
2.51
18
21
1.93
19
22
13
2.10
no opinion
non important
(5)
Low (4)
Average (3)
High (2)
Software possibilities
Importance
NUMBER OF REPLIES
2.00
19
22
10
2.00
18
23
10
2.04
12
28
14
2.44
10
18
17
308
2.05
14
26
13
2.04
13
26
13
2.28
11
23
14
2.48
10
21
12
10
2.17
The same calculation has been carried out according to affiliation. Results are shown in
table 3.3. People from research institutes or universities and people from paper
manufacturing think in general the same about the criteria list. There are some exceptions:
Payback time is logically more important for people from paper mills, while
the variety of applications is more important for research centres and
universities.
Criteria related to supplier are in general more important for people from
research centres and universities, especially in the case of the licensing
policy.
Criteria about software possibilities and some related general criteria (e.g.
the variety of applications and uses, compatibility with a standard PC, the
possibility of modifying existing blocks, handling chemistry and chemical
reactions, and the variety of importing/exporting tools) are important for
research centres and universities, while paper manufacturers give them less
importance, probably due to the fact that they are sceptic about complex
simulations and they do not expect too much when using them.
309
Suppliers
Paper
Manufacturin
g
Engineering /
Consultancy
Criterion
Research
Institute /
University
AFFILIATION
General
The total cost of the system, including training.
2.48
2.44
2.6
2.82
1.5
1.69
1.2
2.03
2.5
2.94
2.2
2.21
2.5
2.06
1.8
2.03
3.5
2.37
1.6
2.30
3.5
2.25
2.8
1.65
2.5
1.69
1.4
2.22
2.6
2.21
1.9
310
Suppliers
Paper
Manufacturi
ng
Engineering
/
Consultancy
Research
Institute /
University
Related to supplier
2.00
1.2
2.22
3.5
3.00
2.4
1.84
2.5
2.19
1.4
1.96
3.2
2.40
1.7
1.91
1.5
2.19
2.2
1.84
2.37
2.2
1.78
1.5
2.47
2.6
2.22
1.5
3.19
1.8
2.03
1.5
2.31
1.6
Software possibilities
Suppliers
Paper
Manufacturin
g
3.5
Engineering /
Consultancy
1.81
Research
Institute /
University
311
2.00
1.5
2.13
2.2
2.03
2.67
2.8
2.39
2.5
2.75
2.2
2.02
1.6
2.51
2.2
The other two affiliations (Engineering / Consultancy and Suppliers) are quite different in
their interpretation. Suppliers give importance to criteria concerning general aspects of the
software and the supplier, while the possibilities of the software are considered less
important. People from Engineering/Consultancy group have the opposite opinion. They
give far more importance to software possibilities than the other groups, and for them
general criteria and those related to supplier are far less important than for the other
groups.
A last question was made to consider suggestions to include new criteria to the list. The
replies have been the following:
12.3.4 REMARKS
312
In addition to these surveys, that also include the collection of specifications and
information from main software packages (provided by suppliers), WG-C is performing
several practical cases that have been exposed in different WG meetings, and are being
reported within the COST E36 Action.
Final aim is to create a final report that should help potential simulation software users to
decide within different available options, and even help current users, showing different
possibilities when a bottleneck appears.
313
13.1.1 OBJECTIVE
The objective of the exercise is to simulate the operation of brown stock washing process.
The washing sequence is formed by three units. The description of the washers is given
below. The stock is feed to a tank in which the level is controlled by PID controller to 50%.
The outlet of the tank is connected to the first washer mean while the washing water is feed
to the third washer (counter flow operation). The washing filtrate is collected in tanks,
maintained at constant level (80%), and used as washing water in the previous unit and
also to dilute the feed of the washers (consistency 1%).
To study the dynamic behavior of the system we can introduce a perturbation (linear or
conditional) for the treated brown stock flow rate.
13.1.2 DATA
Feed:
Fiber flow: 500 t/d
Consistency: 10 %
Organic Dissolved Solids concentration: 4 %
Inorganic Dissolved Solid concentration: 3 %
Temperature: 80 C
Wash water:
314
3 drum
12
0.002
0.08 %
concentration, S , over the ideal washing that could be achieved with infinite washing:
D
315
At very low dilution factors, before any breakthrough of the wash water, there will be
displacement washing. The displacement ratio for this segment of washing is equal to the
wash ratio. At some point the wash water will break through the sheet and the
displacement ratio curve falls away from perfect displacement.
The washers are modelled by entering two coordinate points to describe the curve:
- Measured washing efficiency point
- Breakthrough point
Note: The breakthrough point is in reference to the wash ratio axis as DR = WR on the
perfect displacement line.
The first point should come from measurements. The slope of the washing efficiency curve
can be modified to fit the curve to measurements at other conditions by sliding the
breakthrough point up and down the perfect displacement line.
Additional information
Dilution factor is the liquor in the shower minus that in the discharge:
DF = L L
S
where:
L = mass liquor / mass fiber in the discharge
S refers to shower, D refers to discharge
where:
L can be calculated from the discharge consistency, C, by:
D
L = (100 C)/C
D
Tests have been performed with the first stead state model with : Flowmac, APROS, BALAS,
IDEAS, WINGEMS,CADSIM
Later tests also have been perfoemd also with the dynamic model with : Flowmac,APROS,
BALAS, CADSIM,
Finally two of the tools have been tested with interacting PID controller : Flowmac, APROS.
The results of the tests will be published later, but were not ready for this book.
316
317
When we log sensor values it may be nessecary to filter the signals or check the signal
quality before you use the values in the simulation. There are different methods and tools
for this, but in reality this may be a quite tricky part in the chain from sensor to simulation
result. Especially out-liers may cuase problems. What values should be included or
excluded? This signal chack normally should be inbetween the DCS data base and the
simulation software.
Most major DCS vendors today also have solutions for this type of applications. As
mentioned earlier OPC-servers are the most common links between different softwares
today, but we also can communicate on a lower level like directly with TCP/IP, Ethernet ans
similar protocols.
If we want to consider how the user interface should look like and what properties there
should be, most users seem to want the simulator interaction interface in the same
environment as the DCS, or at least looking the same. If you have an information
management system, it may be sitauated there as well. To have a totally separated software
may be acceptable, but it is difficult to keep track on many different systems, why it is
considered as a less good solution if the system shall be used by operators in the daily work.
One interesting alternative is to just let the interaction show up if something happens on the
operator screens, but be hidden otherwise. If for instance an early warning is to be
submitted, it may pop-up at the screen at a relevant position. This is something vendors are
working on and will probably be standard in the future. It may also be of interest to send an
SMS or similar to process engineers, operators and/or service staff, preferably together
with some short message indicating what to do.
to provide parameter values for calculation of the objective function and constraints
to generate required parameters to calculate and formulate constraints, e.g. when
quality parameters which cannot be measured directly is handled a simulation
model can be used to predict the parameter.
to validate optimization results. If only a part of the process is optimized the
optimization results must be tested and validated in a simulation model for the
entire process.
There are certain requirements that must be met by the software used for optimization and
simulation if they should be able to interoperate and share data. They must provide
interfaces to set and receive data. For instance, the optimization software must be able to
set parameters in the simulation model according to certain values of the optimization
problem. Similarly, after a simulation is performed the optimization software must be able
to obtain output values from the simulation model.
The interoperation between the optimization and simulation software can be accomplished
using different techniques. In general, there are two different ways to achieve
interoperability between the optimization and the simulation software [4]:
One common technique is to use a file-based interface where the optimizer change
values of certain parameters in an input file to the simulation software, and where
the simulation software then prints the results of the simulation into an output file
readable for the optimization software.
Another technique highlighted in this paper is using component-based software
technology like Microsofts Component Object Model (COM). This component
technology allows a binary interoperation between the simulation and optimization
software which is a more direct method for sharing of data via a programming
interface.
To be able to send data from one application to another they have to interact using the same
interface type. The file-based interface is a more manual interface but is probably supported
by more optimization and simulation software than COM is. However, the file-based
approach is quite ineffective and impractical because of the risks for reading and writing
wrong data and it is also more time-consuming and inefficient than the faster and more
reliable interaction technique when using COM.
319
The Microsoft Component Object Model (COM) is the Microsoft standard for creating
software components. The Component Object Model was introduced by Microsoft in 1993, and
is today one of the most commonly used component technologies for developing componentbased applications. A component is defined as a specific part of software that does something
specific and predefined. Any program should be able to use the component. The component
will work by itself and should be easy to replace. COM is creating binary software
components that are available to interact with each other [6][7].
The Microsoft Component Object Model (COM) is a standard for specifying an object model,
and the programming requirements that are enabled for COM objects to interact with each
other. The programming can be done in many programming languages, and the objects can
be structurally different. This is why COM is referred to as a binary standard that applies
after a program has been translated to binary machine code.
The COM technology permit objects to interact with each other across process and machine
boundaries. A components object can never have direct access to another components
object. An object can only get access to another object through interface pointers. This is the
main feature of COM which allows it to completely defend encapsulation of data and
processing [8].
Figure 1: Two applications may connect to each other's objects, in which case they extend
their interfaces toward each other [6].
320
Figure 2: A COM Component that supporting three interfaces A,B and C.[6]
14.2.4 MARSHALING
The mechanism that let objects be used across threads, process and network boundaries, is
called marshaling. Marshaling also makes independency of the location possible. Marshaling
allows one objects interfaces that are exposed in one process to be used in another process
[8].
The client is any piece of code that uses another objects services by calling methods
through that objects interfaces. The only way that clients can interact with a COM object its
by using a pointer to one of the objects interfaces and calling methods through the interface
pointers. Encapsulation is achieved through this pointer.
There are two types of COM servers:
Dynamic Link Library (DLL) that is used for in-process servers that is loaded in the same
process as the client. The communication between the client and server handles through
function calls [8].
Executables Files (EXE) that are used for out-of-process servers. They are running in
another process on the same machine as a local server or in another process on a remote
machine as a DCOM component [8].
321
A standard COM component runs on the same machine and in the same process space as the
software that are using the COM component. The Distributed Component Object Model
(DCOM) usually runs on a different machine and it runs its own process space. The DCOM
component can also run on the same machine as the software that is using it. The
functionality of the component is available to programs or other components for all
computers on the network. This fact provides you with the ability to talk to services on a
machine other than the client computer. For example if the database is located on some
server some where, with DCOM you can place actual functionality on the server and access
it like it was local [10].
14.2.7 AN EXAMPLE
322
Buildings
Flue Gas to
the Stack
Flue Gas
~55 oC
Condenser
~48 oC
Turbine Generator
~170 oC
Boiler
170 MW
District Heating
System
Biomass
Coal
Condenser
70 to 100 oC
Figure 5: The combined heat and power production system to be optimized [5].
The interoperation between the optimization software and the simulation software was
accomplished using Microsofts COM technology.
323
MATLAB supports COM and can interact with contained controls or server processes, or to
act as a computational server controlled by a client application program. MATLAB can be
configured to either control or be controlled by other COM components. When MATLAB
controls another component, MATLAB is the client, and the other component is the server.
When MATLAB is controlled by another component, it is acting as the server. MATLAB
supports four different COM client-server configurations; MATLAB Client and In-Process
(Control) Server, MATLAB Client and Out-of-Process Server, Client Application and MATLAB
Automation Server, Client Application and MATLAB Engine Server [15].
The configuration used in this case is the second configuration, MATLAB Client and Out-ofProcess Server. This configuration runs the server in a separate process from that of the
client application, here MATLAB. Since the server runs in a separate process, it may be run
on either a local or remote system.
% Start IPSEpro
app = actxserver('PSE.Application');
Now IPSEpro has been started and it is time to open the simulation model that shall be used.
The MATLAB-function invoke invokes a method on an object or interface [15]. In the code
below the method openProject is invoked on IPSEpros COM object. The second input
parameter to the invoke function is the name of the simulation model that shall be opened.
Now IPSEpro and the correct simulation model are started. When loading a simulation
model into IPSEpro the simulation model already contains some initial simulation data. In
IPSEpro the simulation model consists of a collection of objects representing different
components in the real process; one object for the flue gas condenser, one for the steam
condenser, one for the generator, one for the district heating feed stream, objects
representing the biomass and coal combustors, and so on. The objects in IPSEpro are
represented by a name. Every object also has one or more parameters representing
different parameters to be adjusted on the specific object [11].
Figure 6 shows the graphical representation of the simulation model for the combined heat
and power plant process used in this example.
325
District heating
consumer
DH water pump
Flue gas
condenser
Flue gas to
the stack
Generator
Turbine
Condenser
Boiler
Biomass
Combustion air
Coal
Combustion air
Figure 6: The combined heat and power plant system in IPSEpro [5].
To be able to set the values of the parameters of the different objects in IPSEpro the method
findObject in the IPSEpro COM object is used. findObject gets the name of the object in
IPSEpro as input and returns a handle to that. With this handle it is then possible with the
method value to set the value of an objects parameter. In the following code the value of
the parameter mass of the object fuel_stream001b is given the value 7. The parameter
represents how much biomass is given to the biomass combustor every second.
326
% Do a simulation in IPSEpro
run = invoke(proj,'runSimulation',0);
The simulation is run for one time step and is then stopped. To get the value of an objects
parameter follows the same procedure as to give the parameter a new value, with the
difference of invoking the method value without an input parameter. In the code below the
MATLAB variable Q_heat is given the value of the parameter q_trans of the object
htex_counter001 in IPSEpro.
The optimization procedure starts with loading the initial values for the optimization
variables and parameters into MATLAB. MATLAB then starts TOMLAB and the optimization
procedure is performed so that TOMLAB calculates new values of the optimization
parameters, puts the new values into the simulation model in IPSEpro and then starts a
process simulation with IPSEpro. The results from the simulation are then transferred back
to MATLAB and used by TOMLAB to calculate the values of the objective function and the
constraints. The values of the objective function and the constraints are evaluated by the
optimizer. This sequence is then run again until the optimal value of the objective function
is found according to some criteria specified in the optimizer.
The results of the fuel mix optimization problem in [5] showed that it was preferable to use
coal rather than biomass if the electricity price was high, and vice versa. The results of the
optimization problem in [5] were that the method with an external simulation model clearly
works, but the optimization algorithm to be used depends on the type of optimization
problem and the structure of the objective function.
The differences between the algorithms tested were the speed of how fast they found an
acceptable minimum of the objective function. One optimization algorithm produced a
slightly more accurate result than the other, but took on the other hand considerably longer
327
time to find the optimum and required many more simulation runs than the other algorithm
[5].
The simulation model of the combined heat and power plant is flexible and easy to maintain
and change if the system is to be retrofitted in the future. As long the component names of
the objects in the simulation model, the flue gas condenser, the steam condenser, the
generator, the district heating feed stream and the biomass and coal combustors are not
changed, the user is free to change the simulation model without changing the optimizer
setup in TOMLAB [5].
The results and discussion in [5] showed that simulator-based optimization of industrial
processes is promising but needs to be further
investigated on a large-scale problem [5].
14.2.7.7 DISCUSSION
The aim of the simulation-based optimization approach is to separate the process model
from the optimization environment. This to be able to use the best possible optimization
software for the optimization procedure and the best possible simulation software best
suited for the specific process.
When traditionally trying to optimize an industrial process using computers and
mathematical algorithms the optimization has to be done in one specific environment suited
for the optimization. The simulation model of the process has to be built and implemented
in the environment suited for the optimization software.
However, the environments for the optimization software are not always well suited for
implementing simulation models. Simulation models can be easier to implement and tuned
for specific processes when using simulation software intended for specific processes. The
benefit of simulation-based optimization using external simulation models is the possibility
to use simulation models built in another software environment than the optimization
software environment. The simulation software is then controlled and run by the
optimization software.
When using an external and a more process-specific simulation model there is also the
advantage of having more functionality and process dynamics incorporated into the
optimization, such as the mass and energy balances of the process and the quality
characteristics of the product produced in the process. Using a more accurate process
model also gives the possibility to check the feasibility of the optimization results where a
larger scope of the process is simulated than is directly included into the objective function
and constraints of the optimization.
In many cases companies have invested much money in well developed and tuned
simulation model for their specific process. The simulation software often though lacks the
possibility to automatic optimize the processes with the help of optimization algorithms.
This means that even though a powerful simulator is used, the optimization of the process
still has to be made manually by changing values of different parameters in the simulator.
328
For many companies there could be much money to save if the already tuned simulation
models can be used in the optimization procedure.
As mentioned in the results of the example, the method of using an external simulation
model in optimization of processes is very flexible. Depending on the simulation software it
can be very easy to change the simulation model without having to change the setup of the
optimization software. This leads to the advantage of having experts on the specific process
tuning and updating the simulation model and the experts of optimization algorithms
tuning, changing and updating the optimization software and the optimization algorithms.
Optimization using external simulation models can though be extremely time consuming,
e.g. when using the simulation output to calculate the objective function and constraints in
each iteration in the optimization procedure. The quest for the optimal solution requires a
vast number of iterations and since the simulation is performed at each iteration step, the
decrement of the simulation time becomes crucial. The speed on the optimization
procedure depends in a high amount on the speed of the simulation software [16].
The limitations of this method can also be the dependent on the interoperation techniques.
The fastest one is to have an interoperation between software using a programming
interface like when using component-based technologies such as COM. However, this
method it though a little bit slower than having everything in one software package.
There might also be some limitations with the optimization algorithms possible to use. As
mentioned the optimizations might also take a bit longer when having to run a whole
simulation when trying new parameters.
The data format is not a specific one, this application is totally specialized to this
application. To use another simulation software in the optimization procedure another data
format must be used. You also have to adjust the optimization parameters to suit the
specific simulator. A general simulation data format for optimization with external
simulation models should make the process of changing simulation software much more
easy.
14.2.8 REFERENCES
329
[5] Hggsthl D., Kvarnstrm A., Dotzauer E. and Holmstrm K. (2004), Fuel Mix
Optimization of Combined Heat and Power Production Utilizing a Simulation Model, The
9th International Symposium on District Heating and Cooling, August 2004, Espoo, Finland
[6] MSDN, http://msdn.microsoft.com, 2005.
[7] David S. Platt. (2000), The Essence of COM, Prentice Hall Inc, ISBN: 0-13-016581-6
[8] Williams S., Kindel C., The Component Object Model: A Technical Overview, Microsoft
Corporation, 1994.
[9] Component Object Model Specification, Part I, Microsoft Corporation, 1995.
[10] Rofail A, Shohoud Y (2000), COM and COM+, ISBN: 0-7821-2384-8.
[11] IPSEpro, Manual PSE, version 4.4.001, 2003
[12] IPSEpro, http://www.simtechnology.com, 2005
[13] TOMLAB, http://www.tomlab.biz, 2005
[14] MATLAB, http://www.mathworks.com/products/matlab/, 2005
[15] MATLAB Help, MATLAB 6.5, 2005
[16] Pulkkinen P, Ihalainen H, Ritala R, Developing Simulation Models for Dynamic
Optimization, Proceedings from 44th Scandinavian Conference on Simulation and Modelling,
September 2003.
330