Sie sind auf Seite 1von 13

Automatica, Vol. 32. No. 3. pp. 305-317.

19%

Pergamon Copyright 0 1996 Elsevier Science Ltd


ooos1098(%)00141-7 Printed in Great Britain. All rights reserved
ooos-lo98/% $15.00 + 0.00

State Estimation for a Large-scale Wastewater Treatment


System*
R. TENNO? and P. URONEN?

A filter developed for a large-scale wastewater treatment system permits


estimation of the concentration profile of biomass and organic matter in
any area of the treatment basin more precisely than a filter developed for
a completely mixed basin.

Key Words-Biotechnology; wastewater treatment; modelling; distributed parameters; partially


observable stochastic system; state estimation.

Abstract-A new distributed parameter stochastic model is biomass growth rate;


introduced for modelling a large-scale wastewater treatment
system with distributed water feeding, sludge recycling and maximum growth rate;
removal. The state estimation problem is formulated as a Michaelis-Menten saturation
biomass and organic matter profile estimation problem for constant;
the aeration basin. It is solved by approximation of the
original model with a finite-dimensional bilinear model and biomass yield;
the filtration distribution with the normal distribution. It is oxygen uptake rate;
shown that the profile can be estimated online using oxygen yield;
dissolved oxygen and gas analysis results measured from
several points over the aeration basin. The estimation quality mass transfer constant;
is higher in the case of a large-scale aeration basin than in the saturation constant for oxygen;
case of a completely mixed basin. The estimation algorithm is carbon dioxide production rate;
obtained in a computationally effective form. The model and
the estimation algorithm are tested in a simulation carbon dioxide yield;
experiment. It is demonstrated that the sludge and water endogenous respiration rates.
distribution has a strong effect on treated water quality. The
distribution is considered as a new control parameter. It can
be used in practice for improvement of treatment process
quality. Measurements
to dissolved oxygen concentration;
oxygen uptake rate;
NOMENCLATURE carbon dioxide production rate;
ic
PO oxygen volumetric concentration;
Process PC carbon dioxide volumetric
Sf organic matter concentration in concentration;
wastewater: readily biodegradable Q&T outlet gas flow rate;
component; Qr wastewater feed rate;
S organic matter concentration in sludge recycling rate;
treated water; :: excess sludge removal rate.
X activated sludge concentration in
the aeration basin;
X, return sludge concentration; Stochastics
(dX), biological growth (growth W, W, U$:.,W, W,, WC, L, V, and K are
difference), Wiener processes;
L+,ui, uj, aft a, and a, are standard deviations
* Received 9 July 1994; revised 18 March 1995; received in of errors caused by model inadequacy;
final form 7 August 1995. This paper was not presented at rdO, r, and r, are measurement accuracies.
any IFAC meeting. This paper was recommended for
publication in revised form by Associate Editor K. P. White
under the direction of Editor A. P. Sage. Corresponding
author Professor R. Tenno. Tel. +35804513845; Fax System
+358 0 462373. state;
fLaboratory of Process Control and Management,
Helsinki University of Technology, Kemistintie 1, SF-02150 F measurements;
Espoo, Finland. W,, W,, W, Wiener processes.
305
R. Termo and P. Uronen

Estimation recycling system. The latter two are known


m, online estimated current state; (Gray. 1990) as the step and tapered feeding
Yl covariance of the estimation errors. systems respectively. In this paper the general
configuration is considered in order to solve the
Lengths state estimation problem for any sludge or water
2 length of the aeration basin: feeding system.
c moving coordinate: A state estimation algorithm can be developed
1 index: I = 1, , L. more effectively for a large-scale aeration basin
than for a completely mixed basin. This is
Subscripts because of the use of more measurements, the
I, i biomass: favourable situation for identification of low and
J, j growth substrate high concentrations at the same time at different
f wastewater: points of the aeration basin, and the strong
0, 0 oxygen: correlation between local processes in near
c, c carbon dioxide: neighbourhoods. The estimation algorithm can
DO, do dissolved oxygen. bc developed even more precisely if there is no
distributed feeding along the aeration basin, for
A lower-case subscript is used for scalar
example if a conventional sludge and water
designation, and an upper-case one for a vector
feeding system is used.
or matrix designation.
The main reason to analyse the processes in a
Ek,, is an elementary matrix: the kth element large-scale aeration basin is the particular
in the pth row is a 1, with all the rest zeros: drop opportunities they present for stabilisation.
the second index if the indices are equal. Thus, for example, a troublesome waves of
organics in wastewater can be treated more
1. IN’I‘KODU(‘TION effectively (Tenno and Uronen, 1995c, d).
The purpose of this paper is to propose a new
Biological wastewater treatment is a relatively
method for online state estimation for a
slow process. Therefore a large-scale treatment
large-scale aeration basin. This is a method that
basins must be used for massive treatment of
can be used for estimation of biomass and
water. These basins are completely rear-mixed.
organic matter distribution using dissolved
The organic matter concentration at a basin inlet
oxygen and gas analysis results measured at
is about 10 times that at the outlet, as are the
several points over the aeration basin. The
load and loading characteristics. Because of poor
development of the method is as follows.
mixing, biological growth is highly disturbed, and
The model of the process is introduced in
the treated water quality is frequently low.
Section 2 as a system of stochastic first-order
A special treatment system with distributed
partial differential equations. It is obtained in
water feeding, sludge recycling and removal can
this form by extension of the lumped parameter
be used for improvement of quality. The most
model (Tenno and Uronen, 1995a) for the
general configuration is shown in Fig. 1. So
distributed parameter case. State, observation
general a configuration is rare in practice.
and control are all considered as distributed
although distributed water feeding is widely
parameter processes. The state estimation
used, as is, to some extent a distributed sludge
problem is formulated as a distribution estima-
tion problem. It is solved by approximation of
Wastewater feeding the original processes by finite-dimensional
111111111111 bilinear processes (see Appendices B and C) and
I I the filtration distribution as the normal distribu-
Aeration basin tion (Appendix D). The estimation method is
represented in Section 3 and tested in the
I * I I I I I I I simulation experiment in Section 4. It is
1TlTlTlTlT
demonstrated that the sludge and water distribu-
Sludge feeding and removal
tion has a strong effect on treated water quality.
The treatment process depends bilinearly on the
controls. The exact dependence of the model on
the controls is given in Appendices B and C and
summarised in Section 5.
Stochastic distributed parameter processes are
Fig. 1. Large-scale wastewater treatment system with relatively new in control theory, especially in the
distributed water feeding. sludge recycling and removal. partially observable case. Most results have been
Estimation for large-scale system 307

obtained for completely observable processes Here the biomass growth and substrate oxidation
and for the general linear operator case processes are considered as the coordinate-
(Benssousan, 1982; Rozovskii, 1983). They have dependent processes, unlike the situation in a
been applied for analysis of the filtration completely mixed aeration basin.
equation (Rozovskii, 1983) and for modelling The following notation is used in the model:
some random field processes, such as population
X biomass (activated sludge
genetics (Fleming, 1975). The process is much
concentration);
simpler in the pollution control application. To
S growth substrate (organic
some extent it is close to the fixed and fluid bed
matter concentration in
reactor model (Dochain et al., 1992). Similarly to
the aeration basin):
that case, the process here is considered in its
simplest form using a system of first-order partial sr wastewater organic matter
concentration;
differential equations, but it is stochastic and not
completely observable in this case. Thus, it can x, return sludge concentration;
(dxh, biological growth (versus
be considered more like a random transport
the mechanical growth);
process (Funaki, 1979) or a wave formation
/A(S) = /dI(Kj + S) specific growth rate;
process in a random medium (Sobczyk, 1985).
maximum growth rate;
Unfortunately, analysis of these processes is not
applicable in the partially observable case. :I Michaelis-Menten
tion constant;
satura-
The state estimation problem has been solved
Y biomass yield;
by Glonti (1985) in the general linear operator
case. These results are not used in this paper, {X, = Xg(t, l)}, {X = X(t, f)} and {S = S(t, e)}
because the treatment process is simpler than the are stochastic processes with values in a
general case, and therefore the estimation and separable Hilbert space, {X, = X,(t)} and {S, =
control algorithm can be developed more easily. $(t)} are stochastic processes in a Euclidean
It is natural to use a system of ordinary space, and {W = W(t, f)}, {W = W(t, f?)} and
differential equations for approximation of the {Wj = Wj(t, e)} are Wiener processes with values
state and observation processes and for solution in a Hilbert space* and with smooth covariance
of the estimation and control problems. This operators CT,(T;, and oj respectively;
approach is used in this paper. u, u;, a, standard deviations of errors
caused by model inadequacy,
2. MODEL e moving coordinate;
2 length of the aeration basin;
2.1. Treatment process Qr = Qf(t, e) wastewater feeding rate,
Removal of organic substances in a full-scale (distribution);
aeration basin with distributed water feeding, Qr = Qrk f> sludge recycling rate:
sludge recycling and removal can be described Qe = Q& 4 excess sludge removal rate.
by the following system of stochastic functional
The latter three flow rates are local. Total
differential equations (for the basic ideas of
feeding Qf(t), recycling Qr(t) and removal QJt)
stochastic partial differential equations, see
rates for the whole system can be calculated as
Rozovskii, 1983).
sums of local flows by integration over the length
(dX), = /_@)X dt + dW, (1) of basin, e.g.

dX = (dX), + 2 Q&f, - X) - Qo,, $ Q,r(t) = kyQ,N 4 de.

Q,,Jt, f) is the sludge and water flow inside the


- (Qf + 8.)X] dt + dW> (2) aeration basin. It can be calculated as the
moving coordinate sum
dS = - YP’(dX), + 2 Qr(Sf - S) - Q,,, $
1 Q,& 0 = 1’ W, u> du (4)
0
+ QJS(t, L) - S1) dt -t dW. (3)
of the feeding, recycling and removal flows:

This is the simplest description of the large-scale *For any orthogonal basis e E H the scalar product
aeration basin. It can be obtained (Appendix B) {(W(t, j), e(C))} = {W(r)} is a standard Wiener process,
with covariance M{(W(C 4, e(e))(W(s. 4, e(R))} =
in this form by extension of the lumped min (t, s)(cr’(e, R)e(e), e(R)), where u*(e, R) is the positive
parameter model (Tenno and Uronen, 1995a). self-adjoint nuclear operator.
308 R. Tenno and P. Uronen

U = Qr + Qr - Q,. For interpretation of these saturation constant;


flows, see Fig. 2. carbon dioxide production rate;
Y, carbon dioxide yield;
2.2. Observation process m,,, m, endogeneous respiration rates for oxy-
The dissolved oxygen concentration &<,= gen and carbon dioxide respectively;
&,Jt, f), oxygen uptake rate .&,= &(t, e) and
carbon dioxide production rate &. = &.(t, f) are {V,,>, KJ, K>, W1 and {WJ are Wiener
the main observable components of the process. process in a Hilbert space; it is assumed that the
The latter two can be simply calculated from the covariance operator is the identity for measured
local mass balance of the aeration basin, processes {V.} and smooth (a,, a,) for other
processes {W.};
.&,= 14.3(21 -pC,)QR, s&= 19.6(p, - O.O3)Q,.
standard deviations of errors
i.e. from the outlet gas flow rate Q, = QR(t, P) caused by model inadequacy,
and the composition measurements p,, =p,,(t, 6’) measurement accuracies.
rdcjt r,, rc
(oxygen) and pr =p<.(t, f?) (carbon dioxide) in
terms of volumetric concentration. These con- The perturbations W, W,, W,, W,, WC of a real
centrations in air are p,> = 21% and pC = 0.03%. treatment process are correlated with nearest
A volume index of 1 mol for 22.4 dm’ at 0°C is neighbours. In the simplest case this correlation
used here for determination of the constant ratio can be accounted for in the model using the
between mass and volumetric concentrations. exponential function
The observation field inside and above the
v’([, R) = a*R”-‘I, R < 1.
aeration basin can be described as the following
system of stochastic functional differential The model (l)-(B) can be expressed in
equations, which generalise those of Tenno and discrete form as a system of ordinary stochastic
Uronen (1995a): differential equations (see (B.l)-(B.5) in App-
(do&, = Y, ‘(dx), + m,,X dt + dW,, endix B). Each equation of the system can be
(5)
considered as a local description of the treatment
processes in a small aeration tank connected to
d&,<>= RQ,E* - t,,o) - ~Qou, %
the others as a unit in a sequence (Fig. 3). The
treatment and observation processes in the
- (Q.r + QrkL ] dt - (dOA + r,,, dx,,,, large-scale aeration basin are very similar to the
process in a chain of small treatment units. The
chain can be considered as the simplest system
for description of the processes in the large-scale
aeration basin.
(7) Remark. The model depends on units of
(dC),, = Y;‘(dX), + m,.X dt + dw, measurement. It is assumed that the flows are
measured in dilution rate units and calculated
(8)
per aeration basin volume. The physical
with the following notation: interpretation of local flows is special. They
(WC oxygen uptake rate; cannot be measured in the same units as the
uptake yield; total flows. In this paper the following units are
K
R mass transfer constant; used: h-’ for total and h-’ m-’ for local flow
rates. The flow inside the aeration basin QO,, is
considered as total flow, also in units of h-‘.
Standard units (m” h-‘, m* h-‘) can also be used,
but in this case the model has to be modified
with the replacement Q + Q/V, 2?+ l/A, where
V is the volume and A is the cross sectional area
Q.,(St-a) Q.&O of the aeration basin.
it-
2.3. Problem
The problem is to estimate the current
distribution of the state {X(t, e), .S(t, e), S,(t)} by
discrete observations of the random field
Fig. 2. Flows for a moving coordinate section of the aeration {td&? l>t t& 81, k($ e), 0 5s s th optimally

hasin. in the mean-square sense.


Estimation for large-scale system 309

Fig. 3. Decomposition of a large-scale treatment system into a sequence of small treatment units.

2.4. Solution 3. STATE ESTIMATION


The solution of the problem is summarised in The problem is to estimate the current
Section 3. It is obtained in this form by large-scale state 0, using available observations
approximation of the infinite-dimensional proc- &, 0 5 s zz t, optimally in the mean-square sense.
esses as finite-dimensional (Appendices A-C) An exact solution of the problem can be
and of the filtration distribution as the normal obtained, but is hardly compatible with real-time
distribution (Appendix D). Approximation of application. A relatively simple and at the same
the nonlinear distributed parameter processes time effective calculation rule can be obtained by
(l)-(8) in a simpler form is an essential step in approximation of the infinite-dimensional filtra-
the solution of the problem. It is shown that the tion distribution by the normal distribution. The
model in final form can be expressed as the algorithm obtained by this method has a form
following system of bilinear stochastic quite similar to the bilinear filter (Tenno and
differential equations: Uronen, 1995a), but the dimension of the filter is
larger and it contains a sum over the coordinates
de = [a, + a, 0, + $, a:S(t, fS)X(t, Is)] dt of state and observation processes. The algo-
rithm is in the following form:

1
+ b, dW, + bz dW,, (9)
dm = U, + u,m + i u:(mimmf+ rfj) dt + u dw,

(11)
[ I=1
d,$ = [A” + A,X, + & A:S(t, lS)X(t, [s)] dt

+ B, dW, + B2 dW,, (10) dy = (DuT + uD= + b, b: + b2b:‘ - acr’) dt

with the following notation: + i @(% d‘Q (12)


I=1
state, 8 = [X’ ST S,]‘;
where m, the estimation of the current state, -y,is
measurements. &=
the covariance of is the estimation errors, rij is
an element and y.j a column of the matrix ‘yI,
vectors, e.g. X =
[X(t, 6) . . . X(t, L6)lT; (“w;) is an innovation process with new
measurements,
discrete space interval,
with L an integer;
d‘IV=dt- Ao+A,mi+5Ai(mfmj+$j) dt,
Wiener processes for modelling [ I=1 I
uncertainties;
vector and matrix functions, (+ is the gain for correction of the state,
depending on sludge and water u = (b, BT + DAT)(BBT)-‘,
flows and the dissolved oxygen BBT = B, By + B,B;,
concentration, as well as on the
biological constants and the 9’2t! is the gain for correction of the covariance
standard deviations of measu- of the estimation errors,
rement errors and model in- @ = r!,ry + y!,yT, %!?= (BB=)-‘A;,
adequacy errors.
I@, d,?V) ‘,” the scalar yduct,L a =
For details of the system (9)-(lo), see al u2 . . . u2) and A=[A, A2 . . . AZ) are
Appendix C. block matrices composed of elements of the
310 R. Tenno and P. Uronen

system (9), (lo), D = [do d, . . . dL] is a block sludge and water feeding was concentrated at
matrix composed of estimation-dependent the beginning of the aeration basin and the
elements sludge removal at the end.
2. In distributed feeding the sludge and water
d,, = ^J/, d, = mfy!, + mjy!,, I= 1, , L.
feeding was distributed uniformly along the
With the proposed algorithm, one can aeration basin as well as the sludge removal.
estimate the current distribution of the state in 3. In rotated feeding the location of the feeding
any area of the aeration basin using commer- point was changed every hour: rotated to the
cially available equipment: a dissolved oxygen end and started again from the beginning of
analyser (multielectrode modification), the outlet the aeration basin. Location of the recycling
gas flow rate, O2 and CO2 analysers, and valves flow was changed similarly, but with sift back
for gas probe collection over the aeration basin. from the location of feeding point in a
The derivation of the algorithm can be found one-hour period. The removal flow was
in Appendix D, and its testing in the simulation concentrated at the end of the aeration basin.
experiment in Section 4.
The following was observed in the simulation
Remark. The dimension of the filter can be experiment. Variation of the profile is rather
reduced if it is used on the conventional aeration drastic for organic matter (Figs 4-6) and nearly
basin, because the sludge concentration there is constant for biomass. Both these characteristics
nearly constant along the basin. In this case the are strongly affected if the feeding or recycling
estimation accuracy is higher, since the number distribution is rotated (Figs 6 and 7). The sludge
of observations per unknown states is nearly two and water distribution has a strong effect on
times higher. treated water quality (Figs 4-6). The distribution
is an important parameter for process control.
4. SIMULATION
4.2. Estimation accuracy
A special simulation experiment was used for
The simulated and estimated processes are
testing the model and the state estimation
relatively close to each other. To see a small
quality.
difference, compare these processes in Figs 8 and
9. The estimation error characteristics are
4.1. Model quality
summarised in Table 1. The estimation error
The profiles of biomass and organic matter are
depends on the feeding mode. It is smallest in
shown in Figs. 4-7. They are shown along the
the case of uniform feeding, because the
aeration basin for a three-day period for three
variation of sludge is minimal and the model
different feeding modes.
inadequacy errors are strongly correlated with
1. In the conventional point-feeding mode the nearest neighbours in this case. Variation of the

1350-400

I
moo-350
8250-300
mzoo-250
~150-200
E9100-150
mo-100
m-50

mgth, m

1 6 I’, l6 *’ 26 3, & ’ I q
_. 4’ 46
Ime,n . 51 56 6; 7 l6
71

Fig. 4. Organic matter profile along the aeration basin when the conventional point-feeding mode was used.
Estimation for Iarge-scale system 311

Fig. 5. Organic matter profilewhen the distributed feeding mode was used.

sludge is drastic in the case of rotated feeding. obtained here and by Tennu and Uronen,
The estimation errors are largest in this case. (199Sa). A similar simulation experiment was
They are moderate in the case of the tested there. Xnlocal area the aeration tank was
conventional point-feeding mode. loaded s~m~lar~iy in both cases, The geometry of
The estimated state is about 1% biased for the the large-scale treatment system was chosen
conventional and distributed feeding modes and eight times larger in volume and length than the
l-10% biased for the rotated feeding mode. The geometry of tank considered earlier. For more
bias is arises from errors due to approximation details of the simulation experiment, see Tenno
of the original model with a suite-dimensionai and Uronen, (1995a).
bilinear model and the exact ~~tratjo~ d~str~bu”
tion with the normal dist~but~on.
The estimation quality is much better for the
large-scale aeration basin than for the compl- Sludge and water distribution is an important
etely mixed aeration tank. This can be concluded parameter for process control. ~~furtunate~y, it
from comparison of the estimation accuracy is not clear froqn the model (l)-(8) or from the

Fig. 6. Organic matter profile when tie rotated feeding mode was used.
312 R. Tenno and P. Uronen

16
4

1 6 1776 2, & m
3’ 36 4, &--
5’ 56 ;,
Time, h

Fig. 7. Biomass profile when the rotated feeding mode was used.

system (9), (10) how the treatment process and b2 are state-dependent vectors, b3 is a
depends on the controls. Therefore it is shown in state-dependent matrix, uK is the recycling
Appendices B and C that the only controllable distribution, with normalised values uR = QR/Q,,
elements of the system (9) (10) are the vector a,, Qx is the (unnormalised) recycling distribution,
and the matrix a,. a,, depends on the distribution Q,: is the wastewater feeding distribution, QE is
in the following linear manner: the excess sludge removal distribution, and U,,
and QE are vectors, e.g. U=
i;:;. o)Q:. . u(t, L6 - S)]‘.
The system is linear in the control variables. It
al has the following quasi-linear dependence is bilinear in the state variables and also
(strictly, the matrix a, depends nonlinearly on mutually in the state and control variables.
the controls): The observation process depends on the water
distribution through the washout effect of
a, 0 = a, + b,QF- + &QK + &Q,...
oxygen. This dependence is rather weak. It can
Here 8 is the state, a,) is a state independent be ignored as an optimisable parameter in a
vector, b0 is a state-independent matrix, a,, b, control problem statement.
a0

60

501
0 1 2 3 4
'0
Time, days

Fig. 8. Estimation quality for organic matter when the distributed feeding mode was used.
Estimation for large-scale system 313

7.2 ,

1=16m

5.6

5.2 m
0 2 4 6 6 10 12 14 16 16 20 22 2
Time, h

Fig. 9. Estimation quality for biomass when the distributed feeding mode was used.

6. CONCLUSIONS problem has been solved. The estimation


algorithm is the main result of the paper. It is
A relatively simple model can be applied for obtained in a computationally effective form. It
description of the processes in a large-scale can be applied to real processes for biomass and
aeration basin. Using the model and a simulation organic matter concentration estimation in any
experiment, it can be verified that the sludge and area of the aeration basin using dissolved oxygen
water distribution has a strong effect on treated and gas analysis measurements. The estimation
water quality. The distribution can be used as a accuracy is higher for a large-scale aeration basin
new parameter for wastewater treatment process than for a completely mixed basin.
control. Using the model, a new state estimation

Table 1. Mean and standard deviation of estimation errors

Distributed feeding

I=1 2 3 4 5 6 7 8

Biomass Mean -0.062 -0.079 -0.087 -0.073 -0.072 -0.058 -0.057 -0.054
Stand. 0.312 0.314 0.296 0.286 0.266 0.262 0.251 0.249
Treated Mean 0.096 0.068 0.054 0.096 0.109 0.119 0.121 0.108
water Stand. 1.527 1.668 1.527 1.504 1.488 1.485 1.4% 1.515
Wastewater Mean 2.21 Stand. 14.8

Conventional feeding

I=1 2 3 4 5 6 7 8

Biomass Mean 0.010 0.000 -0.011 -0.011 -0.015 -0.001 0.009 -0.020
Stand. 0.149 0.169 0.191 0.206 0.212 0.221 0.231 0.232
Treated Means -1.142 -1.026 -0.889 -0.706 -0.526 -0.362 -0.247 -0.169
water Stand. 8.672 6.870 5.197 3.728 2.578 1.747 1.173 0.804
Wastewater Mean 6.51 Stand. 17.8

Rotated feeding

I=1 2 3 4 5 6 7 8

Biomass Mean -0.402 -0.020 -0.153 -0.055 0.002 0.078 0.011 0.006
Stand. 0.922 0.539 0.507 0.494 0.434 0.474 0.427 0.431
Treated Mean -2.305 -1.166 -0.093 0.469 0.827 0.516 -0.559 -0.730
water Stand. 15.67 14.21 14.06 14.42 16.06 17.38 18.00 16.58
Wastewater Means 9.41 Stand. 24.2
314 R. Tenno and P. LJronen

‘l‘he lmear approximation coetbcients also depend on


REFERENCES
concentration through the function (A.2). An average value
Benssousan. A. (I 982). Stochustic Control hj, bitr~rronul of the stock : = MZ(t) and concentration .Y = MX(t, 2) can
Analysis North-Holland.
Methods. Amsterdam. he used for simplification of the model, i.e. for
Dochain, D., _I. P. Babary and N. Tali-Maamar (1992). approximation of the function (A.2) and specification of the
Modelling and adaptive control of nonlinear distributed linear approximation coefficients a,, = u,,(x), p = 1, 2.
parameter bioreactors via orthogonal collocation.
Automatica, 28, 873-883.
.Appro_~m~utiorf ol the growth rate fimction
Fleming, W. H. (1975). Distributed parameter stochastic
The nonlinear model (l)-(8) can be simplifed using a
systems in population biology. In A. Benssousan and J. L. piece wise-linear approximation of the growth rate function
Lions (Eds). Control Theory, Numerical Methods and
Computer Systems Modelling, pp. 179-191. Lecture Notes
in Economic and Mathematical Systems. Vol. 107. (A.3)
Springer-Verlag. Berlin.
Funaki, T. (1979). Construction of a solution ol random where CL;)and CL: are linear approximation coehicients.
transport equation with boundary condition. ./. Moth. Sot. In this case the model can be approximated as a bilinear
Jpn, 31. 719-744. model hy state variables.
Glonti. 0. A. (198s). lnuestigations in the 17frorv of
Conditional Gaussian Processes. Metsniereba. Tbilisi (in
Organic- mutter fktuation vi wastewater
Russian).
Gray, N. F. (1990). Actfvated Sludge. lleory and f’racfrw.
The lluctuation can be accounted for in the model (3)
Oxford University Press. using the relationship
Pugachev, V. S. and I. N. Sinitsyn (1987). Stochtr.\frc,
s, -- /*, t ab&(f ~ 24) ~~p, ] + s,
Differential Systems. Springer-Verlag. New York.
Rozovskii, B. L. (1983). Stocha.stic Evolution .S~.sfcm.~. d.S,, .- Cp,S,, dr + <r, d W,.
Nauka, Moscow (in Russian).
Sobczyk. K. (1985). Stochastic Wave Propagatum. Else&r. where CL, is the average organic matter concentration, a, is
Amsterdam. the fluctuation rate for short-term (sampling period) drift and
Tenno. R. and P. Uronen (199Sa). State and paramctci (I):, that for long-term (dailv) drift, W, is a Wiener process,
estimation for wastewater treatment processes using a and CT,is the standard deviation of the prediction errors.
stochastic model. Conrroi Engineering Practice, 3, 7’9%+Ik!.
Tenno. R. and P. Uronen (1995b). Stock and concentration
dynamics of activated sludge processes. In Proc. 6th APPENDIX B-FINITE-ELEMENT APPROXIMATION
International Conjl on Computer Applicutiom in
Biotechnology, Garmisch-Partenkirchen, pp. 3 10-3 14. A suitable model for control application can be obtained
Tenno. R. and P. llronen (199%). Optimal fcedinp by proceeding as follows.
distribution for large scale aeration basin. In Proc,. 7th
IFAC/IFORS/IMACS Symp. on Large Scule Systems : t~~inirr4emenr uppro.~imafion
Theory and Applications. London. Vol. 2. pp. 603-608. L,cr L be an integer and S a constant: 6 = Y/L. With the
Tenno. R. and P. Uronen (1995d). Stochastic control for a replacement
large scale wastewater treatment system. Automafictr.
submitted.

APPENDIX A-SPECIFICATION OF THE MODEL


dW(t. /)+ ,$, tr(l. hS) dW(t, hS) = rr’ dW,
The following steps are necessary for transformation of the
model in explicit form and for its approximation in a form
suitable for control application. Here the same approach can the state (l)-(3) and observation (S)-(8) processes can be
be used as in Tenno and Uronen (199Sa). but with approximated as the following system of ordinary stochastic
modifications due to distributed parameter processes. differential equations:

trcafmcnf process
Return sludge flux
The return sludge concentration can be eliminated from
(dX); - p(S(f, l ))X(t. 0 dt + (7’ dW(r, 0.
the model (2) using the following relationship between the
sludge concentrations in the aeration and settling basins dX(f> f ) ~ (dX):. + [Q,(r. ( ~ 6)X,(t) ~ Q&r, ( ~ 6) AX(r, l)
(Tenno and Uronen, 199Sb):
(Q, + Q, + Q,.,(r. 4 ~ s)X(t. C)] dt
Q,(t, P) X,(t) = Q,(t, i) X(f. Y) + {pa,cu + larQ,(f) pa,<

~ p(k,, + qQ,(t))[l ~ a>(1 ~ Z)]}X(f. Y)}v(f)u,(f. (). (A. 11


i CT:dW,(r. I), (B.1)

Here (Y is the specific settling velocity and c the reduction 01 d.Y(i. 0 = _~Y ‘(dX);+ [Q,(r, ( 6)&(r)
the velocity due to thickening, p is the volumetric ratio:
settling per aeration basin, k,, is the decay rate and 174, the ~~Q,,,,,(r. / 6)AS(t. 0 t Q,(r, & S)S(f. L6)
washout rate, n is the mass ratio of nonseparable per
~~(Q, + Q,)(f. C fi)S(r, t’,] dr
separable part of the total biomass, 2 is the stock per settling
basin capacity, v(t) is the recycling efficiency. (,, and ~2~arc + v; dW,(t. 0: (B.2)
linear approximation coefficients.
The biomass concentration in the aeration basin is affected rhwrurit~orr pr0ces.r
by the total recycling rate through the recycling efficiency
(dam,):, = Y,, ‘(dX),; t rrr,,X(r. f) dt + tr:,dW,,(r, /).
Q,(r)
dt,,,,(l, 0 = {RQ,(f. ( ~ S)[P ~~f<,,,(t. l)]

The latter depends on the sludge concentration according to Q,>,,,(f, ( 6) &xf* (1


the relationship

X(1. W) (cr),+ Vr)(f. i ~~sK,,,(r. t’))dt


u(X) z a 2 0. (A.2)
X’(f, Y) + p U, (WX, + r,,,,dV,,,,(f.0. (8.3)
Estimation for large-scale system 315

ko, &, W, W, W,, WC, r’,,, ib, vcand Qoor, e.g.


.&At,
e) = (4X,’ + Qoutk 1- 8) &,o(t, e) x = [X(t, 6) . x(t, L6)lT, and for elements of the
matrices o,, u,, uo and tro, e.g.
+ (Q, + Qr)(t, f - %,A e) + roi’,(t, e), 03.4)

&.(t, e) = Y,‘(X);‘+ m,X(t, e) + a,PJt, 6) + r,V,(t, e),

(W

where AX(t, f) = X(t, e) - X(t, 8- 6) etc. and Q&t, e) is


the flow rate inside the aeration basin, Numbers from 0 to L6 - 6 are used for coordinates of the
vectors U, QF, QR and Q, and for elements of the matrices
OouT, C7, 0+ and b. Here F and OouT are band-type
Qo& e) = QourO,e - 6) + W, e - 61, QwtO,0) = 0. matrices,
In this form the system has a simple physical interpretation.
It can be considered as a system of small treatment tanks in 0 ... 0 0
sequence (Fig. 3). Here the local flows and flow inside the
aeration basin are defined as individual flows for each tank,
in units of hh’. They are defined using the small aeration
tank volume (compare with the units in the distributed
parameter model). The finite-element model (B.l)-(B.5) can
be obtained on the basis of the distributed parameter model
F= I 1 .

0! .. .. .. 0i. 0
0.
7

0
(l)-(S) using the following replacements and renominations o = Qoutk 6) -Qout(t, 6)
out
of the specific flows:
b .
(2,,Jm3 h-9, C?outb3h-‘l~ Q lh-,l
Qdh-‘I = ,, out 7
K 0
0
zQ,,h-, m-,l = @Wh-‘l+ @.b*h-‘I ~ b,Ih-,l
V K
Qou-dt. L6 - 6)
Here the natural (m” hh’, m2 h-‘) and specific (hh’) units are
used, depending on the model, as shown in square brackets. 6 and 6’+ are diagonal matrices with elements {Q(t, e)} and
The specific flows are calculated on the basis of whole {J$rs ~,‘~~(;~x4] respectively, Qk e) = Qdt, e) + Q,k 4,
aeration basin volume V (distributed parameter model) and
small aeration tank volume V’,(finite-element model).
OQ=
Remark. Since the total feeding rate Q,(t) is defined on the
basis of the whole aeration basin volume and the local 0
Q,(t, h) feeding rate on the basis of the small aeration tank 0
volume, the sum over the feeding distribution is equal to L: ..
6 Q(t, L6-6) -Q,(t, t8-8)
5 u,(t, h) = L.
h=l and rDo, ro and r, are diagonal matrices, e.g. r,o = r,,l,
with r,, a scalar and I the identity matrix.
Here u,(t, h)
is the feeding distribution: u,(t, h) =
Qf(t, h)/Qf(t). A similar sum (equal to L) is valid for the
recycling and removal flows. Control-dependent elements
The return sludge flux can be eliminated from the model
Vector presentation
(B.6) using the relationship (Al) between sludge concentra-
The state and observation processes can be expressed in a tion in the aeration and settling basins. It can be expressed as
vector form very similar to the scalar case: a linear function either in terms of sludge concentration,

treatment process QRX, = g + f-X 03.8)

dX = (dX), + [QRX, + (0our - 0+)X] dt + a, dW,, (B.6) or in terms of distributed controls uR = QR/Q,,

dS = - Y-‘(dX)G + [Q& + (0our - @)S] dt + u, dW,; QRX~ = go+ 8 UR. (B.9)


observation processes Here

(do,), = Y,~‘(dy)~ + m,X dt + uo dWo, 9 = Q&R, G= [0 0 QR + (pvuR],

d5,, = [RQ&* - &A + (GOUT


- 015001dt - (dOJu + a, = WI av, g1 = (Q, + w)X(t, La),
rD0dV,o,
cp= azQr - w,c - dl - dl- z)l& + vQ,h
0 is the zero vector, and QR is the local and Qr the total
recycling flow.
flow rate inside the aeration basin It simple to prove that both vectors (0&r- 0+*-)Xand
(Ob”T - @)S can be expressed in terms of distributed
Qow = FQcur + u,
controls as linear functions. They can be written in the
u = QF + QR _ QE. (B.7)
following forms:
Here lower-case subscripts are used for scalars and upper
case for vectors or matrices. Coordinates of vectors and (Oou, - c)X = MQF + QR) + %QE~ (B.lO)
matrices are enumerated in two ways. Numbers from S to LS
are used for coordinates of the vectors X, S, (dX),, (do,),, (Gxrr - @)S = &QF + (4 + %~,)QR, (B.ll)
316 R. Tenno and P. Uronen

and W, are Wiener processes, and a,,, al,. ,& are vector
and matrix functions!

d, = 93, + i X(f. B)(E,,, , - E,)(I ~ F) ‘.


I- 2

%, = ~ i S(f, /S)E 1.1..


I-I

.oi; and !“A, are defined similarly to &, and &, but for S,
X(t. t8) is replaced with s(t. 16).

Proof: Obviously,

1I
C* := C + i Q,(r. 16 - 6)E,,, ,

where
I- I
A, = IL Y,, ‘4 -
Y,;‘Mo+m,
Y, ‘M,+w
wl
, A:=

o’= 2 [Q,(f, 16 - 6) + Q,(t, 16 -


I I
6)]E,.

El,,, is an elementary matrix (drop the second subscript if the


two are equal), Q&r, IS - 6) is the coordinate of the vector w, = [w;;
QooI-. The latter can be expressed as a solution of (B.7) in
the form dm(r - 24)
d = @z4
dt
- @,b, + @dmf(t - 24) - I+11
Qou.r = (I - F) ‘U. where L’ = Qt + QH Q,.

Here (I - F)- ’ is a triangular matrix: all of its elements are I 4 =i d,G is a diagonal matrix,
on or under the diagonal, with zeros above the diagonal. IL,
For both Z = X or Z = S the following relationship is valid:
M’, = p(E, is a vector,

%orfTZ= f: Z(t, WE,,, I - E,)Qo,, I. E, is an elementary matrix whose Ith diagonal element is 1,
I-2
with all the rest zeros, and /.L{, and CL{ are constant
parameters (see the piecewise-linear approximation (A.3) of
Obviously. the growth rate function.

P”X = i X(r. IS)E,[Q, + QR + Q,-1. Dependence on control


I I
The following decomposition of the system in controllable
and uncontrollable components can be expressed from the
PFS= i [.S(t.M)E,(Q,. + QR) + S(r, 6L)E:,,, Q,yJ (B.9)-(B.11):
I= I

Therefore (B.lO) and (9.1 I) are also valid.

APPENDIX-SYSTEM For any 0 (or estimated state), the following decomposition


is valid:
The model in its final form can be expressed as the
following system of bilinear stochastic differential equations:

de= u,+a,e,+ia$e{ei dt+h,dW+hldWz.


i I~-I

dt = (A,, + A, X, + i A;O$;) dr + B, dW, + Bz dW,.


I- I

Here 0 is the state, t represents the measurements. W,, WI


Estimation for large-scale system 317
APPENDIX D-SOLUTION OF THE ESTIMATION
PROBLEM
The state estimation algorithm discussed in Section 3 can
he obtained by specification of the general filtration equation = yA: + 5 (m:y!, + rr~jy!;)(A:)~ = DAT,
(Pugachev and Sinitsyn, 1987) ,=I
dm = (? dt + (bET + f)=(BBT)-’ dW,
A,(e,-m;)
dy = [g +g=+ bbT - bBT + f(BBT)-‘f7 dt + 5 ,& d”W,,
n=,
dW=dt-adt, +i A:(&ej - Me!6$)](BBT)-‘}“)

with the statistics


(Y!~Y:T+ Y:Y!?)[A:W*)-‘I,,
li =M(ao+o,B+,~,a:~~)=n,+n,m +&m$n;+
,=I
y:,),
=2

bbT = M[(b, dW, + b, dW,)(b, dW, + bz dW,)TI


a =M(A,+A,X+,$A:@$)=A,+A,m, = b, b: + b,b:,

+ 2 A:(m'$n; +y:,),
I=,
bBT = M[(b, dW, + bZ dW,)(B, dW, + & dW,)T] = bBT.

BBT = M[(B, dW, + Br dW-J(B, dW, + B2 dW,)T]


g=M = yaT
= B,Lq + B*BT,
calculated by approximation of the filtration distribution by
+ i (mfy’, + m;y:)(a:)’ = DA’,
,=I the normal distribution.

AUTO
32:3-B

Das könnte Ihnen auch gefallen