Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Fuzzy Control of the Activated Sludge Wastewater Treatment Process
Tong, R.M., Beck, M.B. and Latten, A.
IIASA Professional PaperNovember 1979
Tong, R.M., Beck, M.B. and Latten, A. (1979) Fuzzy Control of the Activated Sludge Wastewater Treatment
Process. IIASA Professional Paper. Copyright © November 1979 by the author(s). http://pure.iiasa.ac.at/1202/ All rights reserved. Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed for profit or commercial
advantage. All copies must bear this notice and the full citation on the first page. For other purposes, to republish,
to post on servers or to redistribute to lists, permission must be sought by contacting [email protected]
FUZZY CONTROL OF TEE ACTIVATEDSLUDGE WASTEWATER TREATMENTPROCESS
R.M. Tong11. B. BeckA. Latten
November 1979PP-79-7
Professional Papers do not report on work of theInternational Institute for Applied SystemsAnalysis,but are produced and distributed by the Institute asan aid to staff members in furthering their profes-sional activities. Views or opinions expressedarethose of the author(s) and should not be interpretedas representingthe view of either the Institute orits National Member Organizations.
INTERNATIONAL INSTITUTE FOR APPLIED SYSTEMS ANALYSISA-2361 Laxenburg, Austria
M.B. BECK is a researchscientistat the International Institutefor Applied セ ケ ウ エ ・ ュ ウ Analysis, Schloss Laxenburg, 2361 Laxenburg,Austria.
A. LATTEN is the managerof セ ィ ゥ エ ャ ゥ ョ ァ ィ 。 イ ョ TreatnentWorks, AnqlianWater Authority, Trowse, Norwich, United Kingdom.
R.M. TONG is with the Electronics ResearchLaboratory, Universityof California, Berkeley, California 94720, USA.
-ii-
PREFACE
Two previous papers (IIASA ProfessionalPapersPP-78-10and PP-79-3) have reported some of the results ヲ イ ッ セ a smallcollaborativeproject investigating the modeling and controlof the activated sludge processof wastewatertreatment.This brief paper provides a more detailed evaluationof afuzzy controller for the activatedsludge process. Such anapproachto processcontrol utilizes the empirical operatingexperienceof the plant manager. セ セ ッ ウ エ conventional controlsystemdesign procedures,in contrast, are basedupon analysisof a model of processdynamic behavior. Given the currentlimitations in understandingand instrumentationof the acti-vated sludge process, fuzzy control appearsto be a particularlyappropriateapproachto adopt for processcontrol.
iii
ACKNOWLEDGEMENTS
The authors are grateful to the Anglican water Authorityfor their financial support in this project. One of theauthors (RMT) would also like to acknowledgethe financialsupport provided by a SRC/NATO PostdoctoralFellowship.
v
SUMMARY
The activatedsludge processis a commonly used methodfor treating sewage and waste waters. It is characterizedby a lack of relevant instrumentation,control goals thatare not always clearly stated, the use of qualitative infor-mation in decision making and poorly understoodbasic behaviormechanisms. In this brief paper we examine the behavior ofan experimentalfuzzy control algorithm constructedto reflectactual operationalpracticp.. We conclude that this algorithmdoes rather well and that a fuzzy controller would be auseful and practical way of regulating the activated sludgeprocess.
vii
1. !ntroduction.
Fuzzy controllers have been successfullyused in a
variety of applications (seeTong, 1977 for a review). In all
of these applications,however, the control goals were clearly
specified, accurateand reliable measurementsof the relevant
processvariableswere available and, perhapsmore importantly,
none of the controlled processeshad more than two inputs or
two outputs. It would be hard to argue, therefore, that these
were "difficult" control problems. Nonetheless,the success
has encouragedthe belief, at least amongst its advocates,
that the fuzzy approachcan be used on a wide variety of pro-
cesses.
In this セ 。 ー ・ イ we report on some results ヲ イ ッ セ a con-
tinuing study of the role of fuzzy set theory in the control
of the activatedsludge wastewatertreatmentprocess (ASP).
The ASP is characterizedhy a lack of relevant instrumentation,
control goals that are not always clearly stated, the use of
qualitative information in decision making and poorly under-
stood basic behavior mechanisms. As such, it appearsto be
an ideal candidatefor fuzzy control and is certainly a more
severe test of the methodology than previous applications.
Section two of the paper outlines the behavior of the セ s p
and highlights the principal control problems. Section three
discussesone particular controller with which we have experi-
mented and analyzes some of the resulting closed loop re-
sponses. We then make some general comments on the design of
fuzzy controllers.
-1-
-2-
2. The Activated Sludge wastewaterTreatmentProcess.
The ASP is one of a number of unit processesused for the
treatnentof sewageand waste waters. The basic feature of the
processis the decompositionof complex dissolvedand suspended
organic substratesinto simple end-productssuch as carbon
dioxide and water. Decompositionis achievedby a hetero-
geneousculture of micro-organisms(the activatedsludge), which
in part utilize the waste organic substratesin the synthesis
of their own biological cell material.
Our studieshave been concernedprimarily with a particu-
lar ASP plant at the セ ヲ オ ゥ エ ャ ゥ ョ ァ ィ 。 ュ TreatmentWorks, Norwich,
England. This installation is shown diagramatically in Figure ャ セ
If we consider only that part of the diagram within the dotted
lines, we see that there are two stagesin the overall process:
an aeration tank followed by a clarifier/settler. Correct
operationof the processrequires, among other things, the
following three items. First, the influent sewageentering
the processhas to be mixed intimately with the recycled sludge;
in principle there exist rangesof desirableproportions in
which substrate(sewage) and organism (sludge) should be mixed
(cashion, Keinath, and Schuk, 1977). Second, air is blown
into the Mixed liquor through diffusers placed along the base
of the tank; this gives the required agitation of the mixed
liquor, provides the necessaryaerobic environment for growth
of the sludge organisms, and can be a key factor in operational
control (Olsson and Andrews, 197C). Third, the settler must
effect good separationbetween the biological floc (sludge)
-3-
and the clarified effluent; excessbiological solids in the
ASP セ。ケ be manipulatedby the removal of waste sludge. Recent
work on the application of automation and control to the ASP
is surveyedby Olsson (1977).
Local control action is taken at the Norwich ASP to help
achieve these aims. Thus, as shown in Figure 1, recycle sludge
ratio (defined as the ratio of influent flow rate, QI' to
recycled sludge flow rate, QR) is regulatedby feedforward
control of QR using measurementsof QI. Dissolved oxygen
(DO) in the aeration tank is regulatedby feedback control of
the airblowers using a measurementof DO. v]aste sludge flow,
Qw' is set by the plant manager.
An ASP that is performing as required will be producing
an effluent that meets some standard. In Britain, this is
simply a イ・」ッセセ・ョ、。エゥッョ that the total (S-day) biochemical
oxygen demand exertedby the effluent should be less than
20 gm-3 and that the amount of suspendedsolid material in the
-3effluent should be less than 30 gm . Whilst these are
hard constraintson the process,a plant managercan choose
to operateas close to them as he feels is practical. In
a real sense, therefore, there is Borne fuzziness associated
with thesevalues. There are secondarygoals, but these
will differ from installation to installation and will depend
primarily on the quality and type of sewage that the process
receives. However, one of the most important of these is
that ammonia in the effluent is kept at acceptablelevels.
The major disturbancesto the processare in the form of
fluctuations in the composition and flow of the influent.
-4-
There are short term diurnal variations as well as long term
trends which together can easily produce changesof up to
50% in the averagequality of the influent.
Our control problem is thus simply stated. How can we
manipulate recycle ratio set point (RRSP), dissolved oxygen
set point (DOSP) and waste sludge flow (SWR) so that we main-
tain effluent quality despite these large variations in the
influent?
3. The Fuzzy Controller
At the core of the controller is a fuzzy algorithm for
determining the appropriatecontrol actions given the current
state of the process. Becausethe algorithm expects fuzzy
sets as inputs, the non-fuzzy processmeasurementshave to be
converted in some way. \1e have adopted the conventional
technique of representationby fuzzy singletons. Similarly,
since the processrespondsonly to non-fuzzy control actions,
the fuzzy control sets generatedby the algorithm have to be
de-fuzzified. We do this by selecting that control value
which divides the area under the membershipfunction in half.
The closed loop configuration is thus as shown in Figure 2.
The basic design problem is to construct the fuzzy algo-
rithm. In doing this we have relied on the considerableprac-
tical experienceof one of us (AL) in the day-to-daymanagement
of the ASP. The first task is to determine the fuzzy input
(measurement)variables for the algorithm, the fuzzy output
(control) variables and the primary fuzzy sets associatedwith
-5-
each of these. The set definitions have not been included here
becauseof space limitations; however, the input and output
variables are listed in Table 1.
We have experimentedwith several algorithms but will re-
strict our discussionto one that has several interesting
features. It ·is shown in Table 2 and consistsof 20 rules.
Symbols such as S, セ l and SP are mnemonics for fuzzy sets
which in this case are "small," "not large" and "small positive."
Each rule in the algorithm is interpretedas a fuzzy statement
of the form
WHEN:t.. DO セ
where セ is a fuzzy proposition about processconditions in terms
of the measurementsand whpre Q is a fuzzy proposition about
appropriatecontrol actions. The propositionsare interpreted
as multi-dimensional fuzzy sets and the rule itself is defined
as their cartesianproduct. Individual rules are cOIDbined
using the union H a 。 ク ゥ セ オ m I operation to forn the overall con-
troller relation.
The reasoningbehind the algorithm is briefly as follows
(for a more detailed descriptionof the role of similar rules
see Beck, Latten, and Tong, 1978). Rules 1-3 are resetting
rules in that, if the process is in a satisfactorystatebut
DOSP and/or SWR are at abnormal levels, then DOSP and/or SWR
are adjustedaccordinqly. Rules 4-7 deal with high effluent
suspendedsolids causedby a rising or bulking sludge (these
terms are briefly defined in Table 1). Rules 8-11 deal with
high NII 3-N levels in the effluent. Rules 12-13 cater for high
-6-
effluent solids causedby factors other than FIL or DNIT and
rules 14-18 describethe required control action if MLSS is
outside its normal range. Rules 19-20 deal with the problem
of a high effluent BOD.
Notice that most rules are concernedwith changesto SWR.
This reflects the fact that in practice waste sludge flow is
used most often to correct for effluent quality variations.
Notice too, that the rule set does not, by any means, exhaust
all the possibleprocessstates. It may be thought of as a
"sparsealgorithm." To test this algorithr'l we ran a simula-
tion of the ASP using a non-linear differential equation
model (with 14 state variables) and applying a disturbance
sequencederived from correspondingobservationsrecordedat
the Norwich plant. Conceptualaspectsof the model are described
in Beck, Latten, and Tong (1978); some accompanyingidentifi-
cation results are reported in Beck (1979).
Figure 3 shows thus the open loop responseof ESS and
ETBD on an hourly basis for 600 hours (i.e., 25 days). Clearly,
the processis not functioning properly. There are large
excursionsin both ESS and ETBD in the early and late parts
of the simulatjon (causedin fact by a bulking sludge condi-
tion). Also, NE 3-N levels in the effluent are high except for
the first 100 hours.
The controlled responsesto the saMe disturbancesequence
are shown in Figures 4-6. The controller sampling period is
set at 4 hours in this run (i.e., six possiblechangesin con-
trol action in each 24 hours). Figure l セ shows the hourly ESS
and ETBD values; Figure 5 shows the hourly NH 3-N and MLSS
-7-
values and Figure 6 shows the correspondingvalues of RRSP,
DOSP and SWR. Notice, straight away, that the early and late
bulking sludge conditions have been suppressed. Notice too
that the ETBD and ESS levels are well within the 20:30 limits,
except for three occasionsbetween 400 hours and 475 hours.
These three occasions,which representa significant loss of
solids over the clarifier weir, are precipitatedby problems
of a rising and a bulking sludge, with both problems being
partly a complex function of over-aeration(see DOSP in Figure
6). There is generally good nitrification throughout the
period with n h セ M n rarely being above 15 gm-3J
Our preliminary conclusion must be that the controller
works rather wAll. However, it does have some defects and in
exploring these we shall make use 0f an analytical tool which
we call a "rule activity chart." Figure 7 shows the rule
activity for 6DOSP (top two traces) and for 6RRSP (lower four
traces). Figure B shows the rule activity for 6SWR. The
horizontal axis is time and the vertical axis is the degree
to which the input propositionL is satisfied. Thus these
charts tell us which rules are contributing to the non-fuzzy
control actions applied to the process.
Since in this paper we are primarily concernedwith the
operationalaspectsof the fuzzy algorithm, rather than a
detailed analysis of the ASP's responses,we shall limit
ourselvesto a discussionof just two featuresof the closed
loop behavior. Despite our assertionthat in practice SWR
is the most often used control variable, we see from Figure 6
-8-
that RRSP is frequently changed. The activity charts of
Figure 7 indicate エ ィ セ エ this is primarily due to rules 12 and
17. But notice that these rules are often activatedat very
low levels. This suggeststhat \>ore ュゥセィエ introduce some kind
of threshold for rule activation.
A way of doing this that is consistentwith the fuzzy
set theory is to employ the concept of "truth qualification"
(see Zadeh.. 1978). Thus we can modify the rules in our
algorithm so エ ィ セ エ they have thp. form
where T is a fuzzy truth set which modifies the proposition
t . +:t. 0 glve 1.. • +Following Zadeh, Y is defined by a membership-function such that
セ +(y)Y-
= セtHセ (y))Y-
A suitable choice for T will achieve the desiredeffect (e.g.,
セ (t) = t if t > threshold, セ (t) = 0 if t < threshold).T - T
We believe that this techniquecould have been applied to
all the fuzzy controllers that have been reported. Because
it allows us to weight the importanceof individual rules,
it is a very flexible and useful design tool.
The secondpoint we should like to highlight is the
behavior of the control variable SWR. A comparisonof the two
responsesshows that SWR is highly correlatedwith MLSS but
lags it by approximately 20 hours (see Figures 5 and 6). There
are two main reasonsfor this. Firstly, becauseof the
-9-
incremental form of the rules for SNR it takes tiMe for SWR to
achieve the necessarylevels demandedby the processconditions.
Then secondly, becausethe rules do not take into account
rate-of-change of MLSS they cannot 、 ゥ セ エ ゥ ョ ァ オ ゥ ウ ィ between a
condition in which action is required (e.g., MLSS low and
decreasing)and one in which it is probably not (e.g., MLSS
low but increasing). Thus SWR is being changedlong after
such changesare required. Consequently,it is felt that in
general an incremental fuzzy algorithm should take account
of both the level and rate-of-changeof the appropriate
measuredvariables. We note that many of the pUblished algo-
rithms do exactly this.
Obviously, there are many other featuresof these responses
which are of interest. However, they require a detailed under-
standing of the ASP and are outside the scope of this paper.
4. Conclusions.
Our aim in this brief paper has not been to presenta
final solution to the ASP control problem. Rather it has been
to show that a fuzzy algorithm basedon practical experience
can be made to work on this difficult process. In doing so,
we have made some general comments about the form and struc-
ture that fuzzy algorithms should take.
Our results must clearly be qualified by the fact that
evaluationof the controller has been undertakenwith a
processsimulation. The present level of accuracy for such
models for biological waste treatmentprocessesis but little
advancedbeyond the primitive stage. Nevertheless,we do not
-10-
hesitate in assertingthat a fuzzy algorithm would be a
useful and practical way of regulating the activatedsludge
process. Indeed, a recent ウ オ セ L ・ ケ of factors limiting waste-
water treatmentplant performanceby Hegg, Raknessand
Schultz (1978) lends substantialsupport to our argument.
They observed, in particular, that:
"The highest ranking factor contributing toroor plant performancewas operatorapplica-tion of conceptsand testing to processcon-trol."
" •.• presentplant personnelare an untappedsource for achieving improved performance."
TABLE 1.
Input Variable
ETBD
ESS
MLSS
RASS
DNIT
DOSP
SWR
Output variable
DOSP
RRSP
-11-
Description
the rotal BOD exertedby the effluent
the suspendedsolids in the effluent
the suspendedsolids in the sludge
leaving the aeration tanks (the mixed
liquor)
the suspendedsolids in the recycled
sludge
the ammonia-N in the effluent
a measureof a condition in the clari-
fier called "bulking sludge"; this is
causedby the presenceof filamentous
bacteriawhich prevent settling.
a measureof a condition in the clari-
fier called "rising sludge"; this is
causedby denitrification whereby
nitrogen gas is fermed and then rises
to the surface of the clarifier 「 イ ゥ ョ セ ゥ ョ ァ
sludge with it.
the DO set point in the aeration tank
the waste sludge flow rate
Description
change in DOSP; i.e., DOSP(t)=DOSP(t-l)
+llDOSP(t)
change in recycle ratio set point; i.e.,
RRSP(t)=k+llRRSP(t) where k is a constant
SWR
-12-
change in SWR: i.e., s w r H エ I セ s w r H エ M ャ I
K l | ウ セ ュ (t)
TABLE 2.
-13-
0Z 0-, P-t P4'-
P4 U) U) p::(lJ Cl U) U) I E-i p:: 0 p:: セr-l o::l U) U) U) M H H U) ..-::l E-i U) H セ
::r: H Z 0 :?: Cl p:: U)
p:: セ セ セ Z セ Cl Cl U) <J <J <J
1 S S M M S L LN
2 S S M H S S SP
3 S S M M S L SN
4 M 1 SP
5 L 1 LP
6 M 1 SN
7 L 'I LN
8 S M SP
9 S M SN
10 S L LP
11 S L LN
12 .L M SP
13 L L LP
14 L LP
15 S SN
16 VS LN
17 VS S SP
18 L L SN
19 M S S SN
20 L S S LN
REFERENCES
Beck, M.B. (1979) On-line Estimation of Nitrification Dynamics.(Submitted to Proc. Am. Soc. Civil Engrs., J. Env.Engng. Div. for possibJepublication).
Beck, セ Q N b N L A. Latten and R.M. Tong (1978) Modelling andOperationalControl of the Activated Sludge ProcessinWastewaterTreatment. PP-78-10. Laxenburg, Austria:International Institute for Applied SystemsAnalysis.
Cashion, B.S., T.M. Keinath, and W.W. Schuk (1977) Evaluationof instantaneousF/M control strategiesfor the activatedsludge process. Progr. Nat. Technol., 9, Nos. 5/6.
Hegg, R.A., K.L. Rakness,.andセ N r N Schultz (1978) Evaluatiopof operation and maintenancefactors limiting municipalwastewatertreatmentplant performance. J. Wat. PolluteContr. Fedn., 50, 419.
Olsson, G. (1977) State of the art in sewage treatmentplantcontrol. Am. Inst. Chemical Engrs. SymposiumSeries,No. 159, 72, pp 52-76.
Olsson G. and J.F. Andrews (1978) The dissolvedoxygen profile- A valuable tool for the control of the activatedsludgeprocess. Wat. Res., 12, 985.
Tong, r N ャ セ N (1977) A control engineeringreview of fuzzy systems.Automatica, 13, 559.
Zadeh, L.A. (1978) PRUF - a meaning representationlanguagefornatural languages. Int. J. Man-Machine Studies, 10, 395.
-14-
Air
10- DOSPDO
ControllerrIIIL_ - __- __, ,
- ..,III
,..----------
RR
Influent? : セK,
IIIII
Aeration TankMixed
liquorClarifier
Effluent
ControllerI
! I Recycled sludge I セ __JL _Waste sludge
•
RRSP
OwController
SWR
Figure 1. Schematicof the Norwich ASP.
Disturbance
Non-fuzzy controls.--- ---1.1'-
7 Non - fu zzy measurements....
ASP
FuzzificationAlgorithm r'J"r-----1
III fuzzificat ion "-......r----...,III II IIL セ
r----- --- -- ---- -------- - -------- -- 1-----...,I I1 セ 7 II IIDe- vエセ⦅MMMT Fuzzy セ B B M M M M B G B G B B B B G ャ :
IIII
Figure 2. Closed loop configuration.
150·0
100.0
50.0
50.0
1.0.0
30.0
20·0
10.0
30·0
20·0
10·0
100·0 200.0 300·0 1.00·0 500·0
Time (hrs)
Figure 3. Open loop responses.
150·0
100·0
50·0
50·0
40·0
30·0
20·0
10·0
100·0 200·0 300·0 400·0 500·0
Time (hrs)
Figure 4. Closed loop responses: I.
r----r-----,----- --I ---,--------,,------;
15.0
10·0
5·0
3500·0
3000·0
2500·0
2000·0
'--__L-__LI__-1. I -L.__......J
100·0 200·0 300·0 400·0 500·0
Time (hrsl
Figure 5. Closed loop responses: II.
30·0
20.0
10·0
RRSP
'·0
o·gI
0·8
1.0
o·g
0·8
0·7
0·6
l
100·0 200·0 300·0 400·0 500·0
Time (hrs)
Figure 6. Closed loop control actions.
I:: 1Rule7
,::("le8
100·0IJ
200·0iii
300·0 400·0 500·0
Time (hrs)
Rule 121.0
·5
r500·0
11r
400·0I
300·0
Time thrsl
,200·0
': 1Rule 13
'.: ャjヲQKMMMMMNMMdGMMMMMNLNNNNMMMMMMMMMイMMMセjNNN⦅⦅L _Oil
':1Rule'B
I
100·0
Figure 7. Rule activity chart: I.
Rule 2
1::l C c\ セ lilli
101Rule 3
·5
セ oセI
101Rule 5
U·5
i
10lRule14
)0·5
Cl o nni
1:(IJCD dセ J[I
1:lJf1p セ OILi
'.: (1019Lrb no セ , oWn Do
1·01Rule 20
·5n IJ
i I , J i
1000 200·0 300·0 400·0 500·0
Time (hrs)
Figure 8. Rule activity chart: II.