Outline

Linear Programming (LP): Formulating Models for LP

1

Straightforward Models for LP

Beno Chachuat <benoit@mcmaster.ca> t

McMaster University Department of Chemical Engineering

Approximations and Reformulations as LP Models

ChE 4G03: Optimization in Chemical Engineering

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

1 / 25

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

2 / 25

Straightforward LP Models Formulation

Everything in life is not linear and continuous! But an enormous variety of applications can be modeled validly as LPs: Allocation Models Blending Models Operations Planning Operations Scheduling

Allocation Models
The main issue in allocation models is to divide or allocate a valuable resource among competing needs. The resource may be land, capital, time, fuel, or anything else of limited availability Principal decision variables in allocation models specify how much of the critical resource is allocated to each use Example: The Ontario Forest Service must trade-o timber, grazing, recreational, environmental, regional preservation and other demands on forestland. The optimization seeks the best possible allocation of land to particular prescriptions (e.g., in terms of the net present value), subject to forestwide restrictions on land use. xi ,j = number of acres in area i managed by prescription j

For additional examples, see Rardin (1998), Chapter 4

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

3 / 25

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

4 / 25

Blending Models
The main issue in blending models is to decide what mix of ingredients best fullls specied output requirements. The blend can be from chemicals, diets, metals, animal foods, etc. Principal decision variables in blending models specify how much of the available ingredients to include in the mix Composition constraints typically enforce lower and/or upper limits on the properties of the blend Example: Gasoline Blending Process
Maximize: Prot Subject to: Product Flow = Octane No. RVP Rel. Vol. Component Flows
Beno Chachuat (McMaster University) t LP: Model Formulation
Reformate FC LSR Naphta FC AT nButane FC FCC Gas FC Alkylate FC FT Final Blend

Operations Planning Models

The main issue in operations planning models is to help a decision maker decide what to do and where to do it. The decision making can be in manufacturing, distribution, government, volunteer, etc. Principal decision variables in operations planning always resolve around what operations to undertake recall that, to gain tractability, decision variables of relatively large magnitude are best modeled as continuous Balance constraints assure that in-ows equal or exceed out-ows for materials and products created by one stage of production and consumed by others Example: Which amounts of Feed 1 and Feed 2 should we use?
Product 1 Feed 1 Feed 2
4G03 5 / 25 Beno Chachuat (McMaster University) t LP: Model Formulation

Product 2 Product 3
4G03 6 / 25

Operations Planning Models

Class Exercise: What is the Optimization Model?
Feed 1 Feed 2 Prod. Min. Prod. Max. Prod. Value Prod. 1 0.7 0.2 0 100 10 Prod. 2 0.2 0.2 0 70 11 Prod. 3 0.1 0.6 0 90 12 Feed Min. 0 0 Feed Max. 1000 1000 Feed Cost 5 6

Operations Scheduling Models

In operations scheduling models, the work is already xed and the resources must be planned for meeting varying-time demands. Principal decision variables in operations scheduling are time-phased time is an index and decisions may be repeated in each time period

Number of ball-bearings produced during period t Inventory level at the beginning of period t Number of employees beginning a shift at time t

max s.t.

[10P1 + 11P2 + 12P3 ] [5F1 + 6F2 ] P1 = 0.7F1 + 0.2F2 P2 = 0.2F1 + 0.2F2 P2 = 0.1F1 + 0.6F2 0 P1 100 0 P2 70 0 P3 90 0 F1 , F2 1000

Are there important issues not included?

Scheduling models typically link decisions in successive time periods with balance constraints of the form: starting impacts of starting level in + period t = level in period t decisions period t + 1 Covering constraints assure that the requirements over each time periods are met

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

7 / 25

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

8 / 25

Operations Scheduling Models

Example: Raw material deliveries are now periodic: We must decide when and how much raw materials to purchase in order to maximize prot while satisfying the demand
Warehouse Factory Periodic Deliveries Feed 1 Feed 2 Intermediate Storage Demand Product 1 Product 2 Product 3 Product 1 Product 2 Product 3 Transportation

Straightforward Models for LP

Fundamental Balances: Material: ball bearings, uid, people, etc. Energy: vehicle travel, processing, etc. Space: volume, area Lumped quantity: pollution, economic activity, etc. Time: utilization of equipment, people work, etc. Only retain key decisions as variables: Production rates Flows Investment Inventories Combine other factors in parameters (constants)
4G03 9 / 25 Beno Chachuat (McMaster University) t LP: Model Formulation 4G03 10 / 25

The model must consider the inventories in tanks, factory, warehouses Delays in transportation can be important What type of balances are needed?
Beno Chachuat (McMaster University) t LP: Model Formulation

Formulating Straightforward Models for LP

Example: Flow Splitting
F2 F1 F3
Feed and product mole fractions 1 , . . . , n

Formulating Straightforward Models for LP

Example: Perfect Separator
F2 1 , . . . , m F1 1 , . . . , n F3 m+1 , . . . , n

Material Balance: F1 = F2 + F3 Component Balance:

Material Balance: F1 = F2 + F3 Feed and product composition (mole fractions) cannot change Why?

F1
k=1

m = F2

Can we make the product mole fractions 1 , . . . , n variables? Could we model it dierently?
LP: Model Formulation 4G03 12 / 25

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

11 / 25

Beno Chachuat (McMaster University) t

Formulating Straightforward Models for LP

Example: CSTR Reactor
reactants products

Approximate Models for LP

Besides straightforward LP models, certain classes of nonlinear or multiobjective optimization problems can be reformulated or approximated as LP models: Base-Delta Models Separable Programming Minimax and Maximin (Linear) Objectives Goal Programming These approaches are usually reasonable when the uncertainties in the problem do not justify further model accuracy Otherwise, solve the nonlinear model using NLP!

Reaction system: Feed ow rate: Ff

A+B C B+C D

Product ow rates: FA , FB , FC , FD
coolant

Material Balances: FA = A Ff FB = B Ff FC = C Ff FD = D Ff
Beno Chachuat (McMaster University) t

Remarks: The s are for specic reactions, reactor temperature, level, mixing pattern, etc. If A, B, C, D are the only components, A + B + C + D = 1 The F s are mass units!
LP: Model Formulation 4G03 13 / 25

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

14 / 25

Base-Delta LP Models
Goal: Extend straightforward LP models to include nonlinear, secondary decision variables: 0 = f(x, y) f(x, y ) + f y (y y )T
x ,y

Formulating Base-Delta LP Models

Class Exercise: Pyrolysis of n-heptane
nominal severity

The base model, f(x, y ), describes the (linear) eect of the primary decision variables, while the secondary variables are kept constant at their nominal value y = y The delta model provides small corrections for deviations (deltas) in the secondary variables y around (x , y )
1

1.75

The accuracy of the solution depends on how well the approximation applies at (x , y ) = (x , y ) To improve the accuracy, the primary and secondary decision variables should be limited by upper and lower bounds
Beno Chachuat (McMaster University) t LP: Model Formulation 4G03 15 / 25

Develop a straightforward model that predicts the ow rate of methane from the reactor Enhance this model by adding a delta due to changes in severity. Recommend the allowable range for the severity variable Is the material balance closed in the base-delta approach?
LP: Model Formulation 4G03 16 / 25

Beno Chachuat (McMaster University) t

Separable Programming
Consider the following mathematical program: min
x n

Approximating Separable Programs as LPs

Piecewise ane approximation: (Nj intervals)
fj (xj )

Dene: cj
(k)

z=
j=1 n

fj (xj ) = f1 (x1 ) + + fn (xn ) aij xj bi , i = 1, . . . , m

fj

fj

(k) (k)

fj

(k1) (k1)

s.t.
j=1

(0)

xj

xj

0 jk xj
c2 =
(1) fj (2) (1) xj xj

(k)

xj

(k1)

xj 0,

j = 1, . . . , n
fj
(3) (1)

fj

(2)

Substitute each variable xj and function fj by:

Nj

The objective consists of n nonlinear, separable terms fj (xj ), each a function of a single variable only Examples? The m constraints are linear

fj (2) fj
0 j2 xj
(2)

xj
xj
(1)

(0) xj

+
k=1 Nj

jk
(k)

When each fj is convex on the feasible region, the separable program can be approximated with an LP An analogous situation exists when the objective is to maximize a separable concave function Why?
Beno Chachuat (McMaster University) t LP: Model Formulation 4G03 17 / 25

xj

(0)

xj

(1)

xj

(2)

xj

(3)

xj

fj (xj ) fj

(0)

+
k=1

cj jk

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

18 / 25

Approximating Separable Programs as LPs (contd)

Approximate Linear Program (on Nj Intervals):
n n Nj (k) cj jk j=1 k=1 Nj n

Minimax and Maximin Problems

Consider the case of multiple, competing linear objectives: f1 (x) = cT x + d1 , 1

min
j=1 n

fj

(0)

...,

fN (x) = cT x + dN N

Minimax problem: minimize the worst (greatest) objective: aij xj ,

(0)

min max{cT x + di : i = 1, . . . , N} i i = 1, . . . , m
x

s.t.
j=1 k=1

aij jk bi
j=1 (k)

Maximin problem: maximize the worst (least) objective: max min{cT x + di : i = 1, . . . , N} i

x

0 jk xj Important Remarks:

xj

(k1)

j = 1, . . . , n, k = 1, . . . , Nj
f (x)

Minimax Problem

f (x)

Maximin Problem

Convexity guarantees that the pieces in the solutions will be included in the right order This approach does not work if not all functions are convex! The solution can be made as accurate as desired by using enough intervals One pays the price in terms of increased problem size!
Beno Chachuat (McMaster University) t LP: Model Formulation 4G03 19 / 25

cT x + d1 1 max{cT x + dj : i = 1, . . . , 3} j cT x + d3 3

cT x + d2 2

cT x + d3 3 min{cT x + dj : i = 1, . . . , 3} j

xmin

cT x + d2 2

xmax

cT x + d1 1 xmin xmax

x
4G03 20 / 25

Beno Chachuat (McMaster University) t

LP: Model Formulation

Formulating Minimax Problems

Class Exercise: Two groups of employees in a company are asked to work on Sundays, depending on the actual plant production x, Group 1: f1 (x) = 5x Group 2: f2 (x) = 3x + 2

Reformulating Minimax and Maximin Problems

Idea: Introduce a slack variable, along with N inequality constraints Minimax Problem: min max{cT x + di : i = 1, . . . , N} i
x

Maximin Problem: max min{cT x + di : i = 1, . . . , N} i

x

The CEO wants to minimize the maximum number of employees working on Sundays in all groups. Formulate a model for this optimization problem. LP Model: min
x,z

LP Model: z cT x 1 + d1 max z
x,z

s.t. z . . .

s.t. z cT x + d1 1 . . . z cT x + dN N

z cT x + dN N

Additional linear constraints can be included in the formulations

Beno Chachuat (McMaster University) t LP: Model Formulation 4G03 21 / 25 Beno Chachuat (McMaster University) t LP: Model Formulation 4G03 22 / 25

Goal Programming
Consider the case of multiple, competing linear objectives: f1 (x) =

Formulating Goal Programming as LP Models

Soft Constraints
Target levels specify requirements that are desirable to satisfy, but which may be violated in feasible solutions
Idea: Introduce deciency variables, dk , representing the amount by which the kth goal is over- or under-achieved

cT x, 1

...,

fN (x) =

cT x, N

and corresponding target levels 1 , . . . , N Find a compromise between the various goals in such a way that most are to some extent satised

Lower One-Sided Goal: Achieve a value of at least k for the kth goal, fk (x) = cT x k k

Lower One-Sided Goal: cT x + dk = k , k + Upper One-Sided Goal: cT x dk = k , k

dk 0 + dk 0 + dk , dk 0

Two-Sided Goal:

+ c T x + dk dk = k , k

Upper One-Sided Goal: Achieve a value of at most k for the kth goal, fk (x) = cT x k k

Fundamental balances (such as material and energy balances) should never be softened! These must always be strictly observed Softening constraints may help debugging models in case infeasibility is reported

Two-Sided Goal: Achieve a value of exactly k for the kth goal, fk (x) = cT x = k k

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

23 / 25

Beno Chachuat (McMaster University) t

LP: Model Formulation

4G03

24 / 25

Formulating Goal Programming as LP Models (contd)

Non-Preemptive LP Model Formulation: All the goals are considered simultaneously in the objective function, min
x,d

T d k = 1, . . . , N

s.t. cT x dk = k , k

Ax b x 0, d 0 Determining the weights is a subjective step... Dierent weights often yield very dierent solutions! Preemptive LP Model Formulation: The goals are subdivided into sets, and each set is given a priority

The solution proceeds by solving a sequence of subproblems, from highest to lowest priority goals Goals of lower priority are ignored in a given subproblem Goals of higher priority are enforced as hard equality constraints
LP: Model Formulation 4G03 25 / 25

Beno Chachuat (McMaster University) t