Upload
dtu
View
0
Download
0
Embed Size (px)
Citation preview
2199 © IWA Publishing 2011 Water Science & Technology | 63.10 | 2011
Application of the Morris method for screening the
influential parameters of fuzzy controllers applied to
wastewater treatment plants
M. V. Ruano, J. Ribes, J. Ferrer and G. Sin
ABSTRACT
In thispaper,weevaluate theapplicationofa sensitivity analysis to helpfine-tuninga fuzzycontroller for a
biological nitrogen and phosphorus removal (BNPR) plant. TheMorris Screeningmethod is proposedand
evaluated as a prior step to obtain the parameter significance ranking. First, an iterative procedure has
been performed in order to find out the proper repetition number of the elementary effects (r) of the
method. The optimal repetition number found in this study (r¼ 60) is in direct contrast to previous
applications of the Morris method, which usually use low repetition number, e.g. r¼ 10∼ 20. Working
with a non-proper repetition number (r) could lead to Type I error (identifying a not-important factor as
significant (false positive)) as well as Type II error (identifying an important factor as not significant (false
negative)), hence emphasizing the importance of finding the optimal repetition number for each study in
question. With the proper r found, the Morris Screening helped identify the parameter significance
ranking, thereby facilitating the calibrationof fuzzy controllers,whichcontainmanyparameters that need
to be adjusted for different wastewater treatment plant (WWTP) applications.
doi: 10.2166/wst.2011.442
M. V. Ruano (corresponding author)J. RibesDep. Enginyeria Química,Universitat de València. Dr. Moliner,50. 46100 – Burjassot, València,SpainE-mail: [email protected]
J. FerrerInstituto de Ingeniería del Agua y Medio Ambiente,Universidad Politécnica de Valencia,Camino de Vera s/n. 46022,Valencia, Spain
G. SinCAPEC-Department of Chemical and Biochemical
Engineering,Technical University of Denmark,Building 229, DK-2800 Kgs. Lyngby,Denmark
Key words | fine-tuning, fuzzy controllers, parameter significance ranking, screening method,
sensitivity analysis
INTRODUCTION
Wastewater treatment plant (WWTP) models are used formany applications/purposes including plant design, optim-isation and control. It is generally accepted that the
modelling and simulation of WWTPs represents a powerfultool for control systems design and tuning. However, themodel predictions are not free from uncertainty, as these
models are an approximation of reality (abstraction), andare typically built on a considerable number of assumptions.In this regard, sensitivity analysis provides useful infor-mation for the modellers as this technique attempts to
quantify how a change in the input model parameters affectsthe model outputs. Different strategies have been applied inthe literature (see, for instance, Saltelli et al. ), which
are typically classified in two main categories: global sensi-tivity analysis, where a sampling method is taken and theuncertainty range given in the input reflects the uncertainty
in the output variables (Monte Carlo analysis; FourierAmplitude Sensitivity Test (FAST), Morris Screening(1991)); and local sensitivity analysis, which is based on
the local effect of the parameters in the output variables(Weijers & Vanrolleghem ; Brun et al. ).
Fuzzy logic based controllers have been successfully
applied to wastewater treatment processes (see e.g. Ferreret al. ; Serralta et al. ), since fuzzy sets theoryoffers an effective tool for the development of intelligent
control systems (Zhu et al. ). Fuzzy control algorithmscan be used to create transparent controllers that are easyto modify and extend because the fuzzy-rules are writtenin the language of process experts and operators (Yong
et al. ). Although these control systems have beenshown to be more robust than classical controllers (Manesiset al. ; Traoré et al. ), they usually contain quite a
number of parameters, which complicates their calibration.So far, these control systems have been tuned by trial anderror methods, based on technical knowledge on the process
and controller performance (Chanona et al. ). Whateveroptimisation method is applied, the fine-tuning of these con-trollers requires a previous selection of the most important
2200 M. V. Ruano et al. | Application of Morris method on fuzzy controllers to WWTPs Water Science & Technology | 63.10 | 2011
parameters to be adjusted in each particular application.
Sensitivity analysis can be used to find the proper par-ameters for fine tuning of a fuzzy controller applied toWWTP plant operation. Ruano et al. () proposed a sys-
tematic approach for fine tuning of fuzzy controllers basedon model simulations that employs three statistical methods:(i) Monte-Carlo procedure: to find proper initial conditions,(ii) Identifiability analysis: to find an identifiable parameter
subset of the fuzzy controller based on local sensitivityanalysis and (iii) minimization algorithm. However, thismethodology is based on local sensitivity analysis, and
then requires an iterative procedure to confirm that theidentifiable parameter subset does not depend on the localpoint in the parameter space where the identifiability study
has been carried out. Global sensitivity analyses are pro-posed to overcome the problem of selecting the properinitial point in parameter space.
In this work, a global sensitivity analysis is applied aimed
at screening the most influential parameters of the fuzzy-con-trol systems to be used in the fine-tuning procedure. TheMorris method of Elementary Effects (EEi) (Morris )
was used as sensitivity analysismethod. In addition, the impli-cations of using this methodology at the usual low repetitionnumber of elementary effects (r¼ 10∼ 20) is assessed. This
screening approach is proposed as a prior step to obtain theparameter significance ranking and the best initial point tocarry out the aforementioned identifiability study.
MATERIAL AND METHODS
WWTP and control system description
The fuzzy logic-based control system was implemented
to control the aeration in a nutrient-removing WWTPbased on a modified University of Cape Town (UCT)scheme (see Figure 1).
This control system was previously developed by theresearch group and it has been applied in several full scaleWWTPs (Ribes et al. ). The developed control system
consists of two hierarchical control layers: the supervisorycontrol and the process control layer. In the process controllayer, several independent fuzzy logic-based control loopswere developed for two process variables: dissolved
oxygen in each one of the aerobic zones and blowers thatsupply air and maintain pressure in the air pipelines.The first controller manipulates the air valve opening
according to dissolved oxygen (DO) concentrations, whilethe rotational speed of the blower is manipulated by a
frequency converter to reach the blowers’ pressure set
point. Since each control valve is governed by an indepen-dent DO controller, the air pressure controller isimplemented in order to enhance the overall control
system performance, especially when there is more thanone control valve in the same air pipeline system. This con-troller aims at obtaining the ideal situation where themovement of one valve, which is governed by its own DO
controller, does not affect the air flow rate through theother valves in the same air pipeline system. In the supervi-sory control layer, the optimum set point of the blowers’
pressure is established according to the control valves’ open-ing and the DO concentrations. In this way, the dischargepressure level remains at the minimum level required to
maintain the DO concentration at the set points in all theaerobic zones. Thus, this control system allows differentDO set points to be fixed for different reactors, which makesit possible, for instance, to achieve simultaneous nitrifica-
tion-denitrification, as was found in Denia WWTP (Ribeset al. ). However, it should be mentioned that only oneaerobic zone has been considered in this case study (see
Figure 1) in order to reduce the computational cost of the sen-sitivity study. Thus, the aeration control system is simplified toone DO controller and one air pressure controller.
For the DO controller, the input variables are the oxygenerror (OE) and the accumulated oxygen error (AOE) and theoutput variable is the increment/decrement of the air valve
opening (IV). For the air pressure controller, the input vari-ables are the pressure error (PE) and the accumulatedpressure error (APE) and the output variable is the incre-ment/decrement of the rotational speed of the blower (IB),which is governed by a frequency converter. Both controllersare fuzzy logic-based controllers, which consist of five stages.Figures 2(a) and (b) show these five stages of the dissolved
oxygen controller and the air pressure controller, respectively.The first stage and the last stage are the same for
both controllers. Stage 1 is the input stage (the measured
or calculated variables are input to the controller) andstage 5 is the output stage. In stage 2, so-called fuzzification,the data collected from on-line sensors are converted
into linguistic variables (fuzzy set), represented by member-ship functions (Gaussian shape in this study). In stage 3, aset of rules (so-called inference engine) are applied tothe fuzzy set obtained in stage 2. The output linguistic
variables are obtained in this stage by the Max-Prod oper-ator, following the Larsen’s fuzzy inference method(Larsen ). These linguistic variables are converted
into numerical control actions in stage 4, which iscalled defuzzification. In order to obtain a single output
Figure 2 | Fuzzy control stages in the dissolved oxygen controller (a) and in the air
pressure controller (b).
Figure 1 | Flow diagram of the control system applied to a modified UCT process.
2201 M. V. Ruano et al. | Application of Morris method on fuzzy controllers to WWTPs Water Science & Technology | 63.10 | 2011
value from our fuzzy linguistic set, the Height Defuzzifiermethod was employed (Mendel ), which only uses thecentre of the Gaussian defuzzification membership func-
tions. The total number of parameters of both controllerscomes from the different stages of the fuzzy logic-basedcontrollers, mainly derived from the defuzzification and
fuzzification steps (Ruano et al. ). Both controllersuse three Gaussian membership functions to fuzzify eachinput (“High Negative”, HN; “Low Negative”, LN; and
“Low Positive”, LP), which gives a total of 12 membershipfunctions from both fuzzification stages; and four to defuz-zify each output (“High Negative”, HN; “Low Negative”,LN; “Low Positive”, LP; and “High Positive”, HP),which gives a total of eight membership functions from
both defuzzification stages. As each Gaussian curve is
defined by two parameters (centre, c, and amplitude, a),the control system has a total of 24 parameters corre-sponding to the fuzzification stages. In contrast, for the
defuzzification stages, only the centres are used, giving atotal of eight parameters. Including the response time ofthe control system (RT ), i.e., time interval between two
control actions, a total number of 33 parameters need tobe adjusted. In order to identify the parameters of thiscontrol system, acronyms for each parameter have beenused. These acronyms are constructed as follows: “ab-
breviation of input variable”þ “c/a”þ “fuzzification/defuzzification membership function abbreviation”. Forinstance, the acronym OEaHN means the amplitude of
the High Negative membership function for the input vari-able Oxygen Error; and the acronym IVcLN means thecentre of the Low Negative membership function for the
output variable Increment air Valve opening.In order to decrease the computational demand in this
case study, we simplified the control system to 17 par-
ameters, assuming symmetric behaviour of themembership functions defined for each fuzzy variable.This symmetric behaviour involves that the amplitude forthe three Gaussian membership functions is the same and
their centres are equidistant. This control structure is thesimplest one that can be implemented in a WWTP, as theoxygen concentration is controlled in only one reactor.
When more aerobic reactors are to be controlled, thenumber of parameters will increase proportionally. As anexample, this control system has recently been implemented
in Denia WWTP (Denia, Spain) (Ribes et al. ), with ninedifferent oxygen variables and one pressure (as the samegroup of blowers is used to aerate the whole system).Hence, reducing the number of tuning parameters in this
kind of controller is essential.
2202 M. V. Ruano et al. | Application of Morris method on fuzzy controllers to WWTPs Water Science & Technology | 63.10 | 2011
The fuzzy controller and the WWTP model were
implemented and simulated using the WWTP simulationsoftware DESASS (Ferrer et al. ). This software includesthe plant-wide model Biological Nutrient Removal Model
no 1 (BNRM1, Seco et al. ). The simulation strategyconsisted of a steady-state simulation to obtain proper initialconditions followed by 28 days of dynamic simulations. Thelast 14 days were considered to evaluate the performance of
the control system. The standardised influent file fordry weather proposed by Copp () was used in thisstudy. The Integral Absolute Error (IAE, integral of the
absolute value of the time dependent error function) foreach controller (Oxygen and Pressure) was selected asoutput measure. So, in this study IAEO and IAEP are the
output variables.
Sensitivity analysis
Morris screening
The method of Morris () evaluates the so-called distri-bution of Elementary Effects (EE) of each input factor tomodel outputs, from which basic statistics are computed to
derive sensitivity information. This distribution function isdenoted as Fjk, which stands for the distribution of theeffects of the jth input parameter on the kth output. The
EEjk attributable to each input parameter is obtained fromthe following differentiation of model output, syk, withrespect to the input, θj:
EEjk ¼sykðu1;u2;uj þD; . . . ; uMÞ � sykðu1;u2;uj; . . . ; uMÞ
Dð1Þ
whereΔ is a predetermined perturbation factorof θj, syk(θ1, θ2,θ3, θj,… , θM) is the scalarmodel output evaluated at input par-ameters (θ1, θ2, θ3, θj,…, θM), while syk(θ1, θ2, θ3, θjþ Δ,… , θM) is the scalar model output corresponding to a Δ
change in θj. Each input parameter, θj, can only take valuescorresponding to (a predefined set of) p levels within itsrange. The calculation of the elementary effects,EEjk, is repli-
cated r times, at randomly sampled points in the input space,leading to a distribution of EEjk used to infer a global sensi-tivity measure. To do that, an effective one-factor-at-a-time(OAT) design has been developed (Morris ). In this case
study, the input uncertainty for each control parameter wasfixed at 50% of the default value.
To compare the sensitivity measures obtained for differ-
ent model outputs, the scaled elementary effects SEEjk
proposed by Sin et al. () were applied. The resulting
elementary effects of each output variable (IAEO and
IAEP) show the sensitivity of each controller separately(oxygen and pressure controller, respectively). However,both controllers must be tuned as a global MIMO (Multiple
Input Multiple Output) control system. Thus, sensitivityanalysis aiming at selecting the most influential parametersin this MIMO system was carried out giving an equalweight contribution of the scaled elementary effects
obtained from both output variables. The scaled elementaryeffect of the jth input parameter on the weightedcontribution of both output variables was calculated as
follows:
SEEj ¼EEj;IAEO
sIAEOþ EEj;IAEP
sIAEPð2Þ
where σIAEO and σIAEP are the standard deviations of thecorresponding output variables. Once the distribution ofthe scaled elementary effects is obtained, the sensitivity
measures mean (μ) and standard deviation (σ) of each Fjkare determined. In order to identify the influential par-ameters, these sensitivity measures were then interpretedusing the graphical approach proposed by Morris. In this
approach, the values of μ and σ obtained for all the Fjk dis-tributions are displayed together with two linescorresponding to μi ¼ +2SEMi, where the SEMi represents
the standard error of the mean that can be estimated asSEMi¼ σi/
ffiffi
rp
. Parameters with low μ and low σ aredeemed as non-influential (Morris ).
One issue of particular interest is the selection of the res-olution, p, and number of trajectories, r. Cropp & Braddock() pointed out that a good choice of the repetitionnumber (r) is more critical for obtaining a good estimate of
the effects than the resolution (p). In this case study, an opti-mal setting of r was searched with a constant resolution ofp¼ 8. To this end, the number of repetitions of elementary
effects calculations (r) for each distribution Fjk wasincreased until the influential parameters remained moreor less stable, i.e., the Type II error was minimised (type II
error: identifying an important factor as insignificant).Once the ropt was found, the graphical Morris approachwas used to find the significant parameters.
RESULTS AND DISCUSSION
The Morris method was applied to a different number of
elementary effects calculations, r, until the sensitivity of par-ameters remained more or less stable. Table 1 shows the
Table 1 | Sensitivity measures of control parameters at the different r evaluated
r¼ 5 r¼ 10 r¼ 15 r¼ 30
Parameter μ σ Parameter μ σ Parameter μ σ Parameter μ σ
OEaHN 1.728 2.383 OEaHN �1.268 3.621 APEaHN �2.049 4.918 IVcHN �0.980 2.516
APEaHN �1.651 1.831 IVcHN �1.256 3.452 IBcLN �1.150 4.356 APEaHN �0.665 2.930
IBcHN �1.461 1.494 PEcLN 0.859 2.118 IBcHN 1.019 2.601 RT 0.586 1.287
APEcHN 1.371 3.676 IBcLN 0.600 1.798 IVcHN �0.904 2.132 PEaHN �0.445 1.925
AOEcHN �1.158 2.313 APEaHN �0.371 2.960 PEcLN �0.736 2.443 OEaHN �0.437 2.103
OEcHN 0.838 2.005 OEcHN 0.322 2.700 RT 0.662 1.370 OEcHN 0.259 1.806
PEcHN �0.562 1.156 RT 0.303 0.911 OEcHN 0.578 1.233 APEcHN �0.259 1.247
IBcLN 0.300 1.989 PEcHN 0.302 2.784 AOEcHN �0.243 0.500 IBcLN 0.154 1.648
IVcHN 0.229 0.325 IBcHN 0.300 2.971 PEcHN 0.206 3.832 PEcHN 0.147 2.986
APEcLN 0.184 0.408 APEcHN �0.292 1.966 APEcHN �0.108 0.363 IBcHN �0.072 3.733
PEaHN �0.161 3.278 AOEaHN �0.219 1.283 PEaHN �0.090 1.608 AOEaHN �0.067 0.774
RT �0.134 0.496 PEaHN �0.206 1.404 APEcLN 0.084 0.294 AOEcHN 0.058 0.216
PEcLN �0.072 3.598 AOEcHN 0.054 0.094 AOEcLN 0.082 0.412 AOEcLN 0.037 0.123
AOEaHN 0.043 0.062 IVcLN 0.040 0.435 OEaHN 0.050 0.378 IVcLN 0.028 0.241
OEcLN 0.041 0.090 OEcLN �0.033 0.113 AOEaHN 0.025 0.177 PEcLN 0.015 1.896
IVcLN 0.028 0.022 AOEcLN �0.024 0.072 IVcLN �0.011 0.053 APEcLN �0.006 0.126
AOEcLN 0.006 0.013 APEcLN 0.017 0.090 OEcLN �0.010 0.142 OEcLN �0.005 0.126
r¼ 40 r¼ 50 r¼ 60
Parameter μ σ Parameter μ σ Parameter μ σ
IVcHN �0.955 2.149 IVcHN �0.731 1.963 APEaHN �0.757 2.676
APEaHN �0.559 2.592 PEcHN �0.525 2.841 IBcHN �0.591 2.685
RT 0.423 1.276 PEaHN �0.507 2.069 RT 0.562 1.224
PEaHN �0.371 1.698 APEaHN �0.470 2.674 IVcHN �0.557 1.970
OEaHN �0.338 1.881 IBcHN �0.441 3.108 PEaHN �0.514 2.153
IBcHN �0.207 3.255 RT 0.438 1.281 OEcHN 0.422 1.508
OEcHN 0.185 1.849 OEaHN �0.310 1.977 PEcLN 0.344 1.871
AOEaHN �0.137 0.644 OEcHN 0.255 1.724 OEaHN �0.306 2.086
PEcLN 0.119 1.774 PEcLN 0.174 1.876 PEcHN �0.301 2.780
PEcHN 0.119 2.628 APEcHN �0.138 1.588 IBcLN �0.262 1.454
AOEcLN 0.062 0.260 AOEcLN 0.094 0.303 APEcHN �0.145 1.741
APEcHN �0.035 1.654 IBcLN �0.071 1.384 AOEcLN 0.033 0.625
AOEcHN 0.021 0.187 AOEaHN �0.051 0.668 APEcLN �0.029 0.229
OEcLN �0.020 0.155 APEcLN �0.043 0.281 OEcLN �0.028 0.450
IBcLN 0.014 1.493 AOEcHN 0.028 0.190 AOEaHN �0.013 0.942
APEcLN 0.009 0.116 OEcLN �0.017 0.138 IVcLN �0.010 0.305
IVcLN 0.009 0.186 IVcLN 0.001 0.164 AOEcHN �0.007 0.309
2203 M. V. Ruano et al. | Application of Morris method on fuzzy controllers to WWTPs Water Science & Technology | 63.10 | 2011
resulting sensitivity measures (μ and σ) for the differentnumber of elementary effects calculated.
As can be seen in this table, increasing the number ofruns, due to a higher r, clearly demonstrates a closer
similarity between the sensitivity measures of the par-ameters. However, increasing the number of elementary
effects from 50 to 60 does not show significant differencesbetween the significance ranking of the parameters
2204 M. V. Ruano et al. | Application of Morris method on fuzzy controllers to WWTPs Water Science & Technology | 63.10 | 2011
considered influential (see below for the identification of
influential parameters). As a result, r¼ 60 was selected asthe optimal number of repetitions for this case study. Theoverall model evaluation costs were therefore 1,080 simu-
lations (¼ r*(kþ 1), k¼ 17) (computational cost of onesimulation was around 10 min obtained using a PC with4 GHz Intel® Pentium® processor). These results are in con-trast to previous applications of the Morris method, since
most of these studies used a low repetition number, e.g.,r¼ 10∼ 20 (Campolongo et al. ). As our results indicate,in agreement with Cropp & Braddock (), r has a signifi-
cant effect on the identified parameter sensitivity,particularly for a low value of r, (lower than r¼ 30, in thiscase study). This fact implies the necessity of finding out
the optimal repetition number for SEEjk calculations (r). Anon-optimal selection of r would lead to Type II error(false negative), failing in the identification of a parameterof considerable influence in the model and Type I error
(false positive), as well, considering a factor as significantwhen it is not. For instance, in this case study, for a rep-etition number of r¼ 10, the parameter RT (response time)
was characterised by a low mean and low standard devi-ation and, for a repetition number of r¼ 15, the parameterOEaHN (amplitude of the High Negative membership func-
tion of the Oxygen Error) was also characterised by a lowmean and low standard deviation. In contrast, the resultsfor the optimal repetition number showed that these
parameters present a considerable effect in the output vari-ables, as was expected from practical experience. On the
Figure 3 | μ versus σ, for ropt¼ 60. Lines correspond to μi¼+2SEMi. (a) IAEO output variable; (
other hand, for a repetition number of r¼ 40 the parameter
AOEaHN (amplitude of the High Negative membershipfunction of the accumulated oxygen error) was consideredas influential compared to the results for the optimal rep-
etition number, where it is non-influential.Figure 3 shows the graphical Morris approach for the
optimal number of repetitions obtained for the two outputvariables (Figure 3(a) IAEO and Figure 3(b) IAEP). The
resulting influential parameters from both output variablesare similar (only one parameter is different). Figure 3(c)shows the same graph for the weighted contribution of the
elementary effects obtained from both output variables(Equation (2)). This figure was used to screen out the non-influential parameters of the control system (i.e. the six par-
ameters that are not labelled in Figure 3(c)). From the 11influential parameters, RT, OEcHN (High Negative centreof the Oxygen Error), APEaHN (High Negative amplitudeof the Accumulated Pressure Error) and IVcHN (High Nega-
tive centre of the increment of the air valve opening)presented high mean and low deviation, lying outside ofthe wedge formed by the two lines corresponding to μi ¼+2SEMi. Thus, the effect of these parameters on theoutput variables is expected to be linear and additive,which is desirable for parameter estimation based on optim-
isation algorithms, since they are characterised by a highermean than the rest of the parameters and not too high stan-dard deviation.
These results are in agreement with the experience-based knowledge. RT is one of the most important
b) IAEP output variable; (c) for the weighted contribution of IAEO and IAEP output variables.
2205 M. V. Ruano et al. | Application of Morris method on fuzzy controllers to WWTPs Water Science & Technology | 63.10 | 2011
parameters, which is in agreement with its importance in the
controller actions. High RT values would give a slowresponse to changes in the process variables and, inversely,low values could cause too fast actions, which would lead to
instabilities when controlling a process with either a lowresponse time or a high time delay. IVcHN is also importantin the control system because it determines the change inthe manipulated variable, i.e. the air valve opening. The
magnitude of the oxygen control actions will basically bedetermined by this parameter and the frequency of eachcontrol action, given by RT. In contrast, the rest of the par-
ameters characterised by low mean and high standarddeviation present an effect that is dependent on the valuesof the other parameters through interaction effects or a non-
linear effect.Compared to the results obtained in the previous work
(Ruano et al. ), in which local sensitivity analysis wasused to find the significant parameters, the following could
be said: (i) the significant parameters identified in the localsensitivity analysis agreed mostly with those identified bythe method of Morris, e.g., OEcHN and RT parameters;
(ii) only one parameter identified as influential in the localsensitivity analysis is found to be non-influential accordingto the Morris screening study (AOEaHN, i.e. High Negative
amplitude of the accumulated oxygen error). Hence, theseresults demonstrate that global methods, in particular theMorris Screening, are more resilient to Type II error, i.e.,
identifying non-influential parameters as influential.Hence, it is recommended for use in identifiability problems.
It is worth pointing out that the significant reduction ofthe important controller parameters could suggest a simpli-
fication of the control system design, i.e., to eliminatesome redundant membership functions. However, it mustbe highlighted that one characteristic of fuzzy control is
that, when the knowledge-based rule system is formulated,some redundant membership functions are defined inorder to represent in a general way all possible system beha-
viours. In this way, the fuzzy control systems become moreflexible at describing system dynamics, but at the expense ofa large number of parameters (around 17 in this case). This
is certainly different from proportional-integral-derivative(PID)-type controllers, which use much fewer numbers ofparameters (around three per control loop) to describeinput-output dynamics of the system; hence there are com-
paratively fewer parameters to fine-tune. However, whenthe system to be controlled has strong non-linear responseto input dynamics, as the one presented in this work, fine-
tuning PID controllers becomes a challenging (if proble-matic) issue since the tuning procedure is done for a
specific working point (thereby ignoring other operating
points in the system). From this point of view, fuzzy controlemerges as an interesting alternative to PID controllers,which otherwise may require gain-scheduling-type adaptive
controller strategies.Finally, one caveat has to be mentioned when using the
Morris screening method, which has to do with the selectionof a proper repetition number of EEi calculations (ropt).Probably, a high value of the parameter r will be neededwhen either a highly nonlinear model is used, or a largeinput uncertainty is defined (lower and upper values in uni-
form distributions). It, thus, comes with a slightly highercomputational cost, which is increasingly feasible to do.
As a guideline to find the proper repetition number, the
Morris method must be applied to a different number ofelementary effects calculations (at least for two r, startingwith e.g. r¼ 10). In general, we proposed that the parametersignificance ranking (as defined by the graphical approach
of Morris) obtained for two different “r” must be similar.The r is optimal when the following holds: (i) those par-ameters, which are found non-influential, are the same in
both iterations with different r, and (ii) those parameters,which are influential, maintain similar ranking. Since thetwo measures must be evaluated at the same time, the
graphical Morris approach facilitates this analysis.We recommend the Morris screening as an efficient and
promising method for sensitivity analysis in the wastewater
modelling community for applications ranging from modelcalibration to controller fine-tuning applications.
CONCLUSIONS
Fuzzy controllers, in contrast to PID controllers, contain a
large number of parameters to describe system dynamicsin all possible operating conditions. This is an issue fortheir application, since fine-tuning of fuzzy controller par-
ameters becomes challenging. Hence, the Morris screeningapproach has been applied to help with fine-tuning of afuzzy logic based control system of a WWTP. Firstly, an
iterative procedure has been executed in order to find outthe proper repetition number of EEi calculations (ropt). Theoptimal repetition number found in this study is in directcontrast to previous applications of the Morris method,
which usually use a low repetition number, e.g. r¼ 10∼ 20.Working with a non-proper repetition number of elementaryeffects (r) could lead to Type I error as well as Type II error.
This approach is proposed as a prior step to obtain theparameter significance ranking for the identifiability
2206 M. V. Ruano et al. | Application of Morris method on fuzzy controllers to WWTPs Water Science & Technology | 63.10 | 2011
methodology used in the previous work (Ruano et al. ).As regards parameter significance ranking, compared to thelocal sensitivity analysis, one can conclude that both resultsare in fairly good agreement with each other with the excep-
tion of one parameter identified falsely by the local method.Overall, the Morris approach provides a good approxi-mation of a global sensitivity measure, thereby overcomingthe drawback of the differential analysis method whose
results are only locally valid. The downside of the Morrismethod is the computational cost, which is increasingly feas-ible to do. Overall, the results supplemented with process
engineering knowledge provide valuable insights into theengineering problems at hand, thereby helping engineersin doing their day-to-day work, e.g., helping to fine-tune a
controller, checking the robustness of plant design, amongothers.
REFERENCES
Brun, R., Kühni, M., Siegrist, H., Gujer, W. & Reichert, P. Practical identifiability of ASM2d parameters-systematicselection and tuning of parameter subsets. Water Research 36(16), 4113–4127.
Campolongo, F., Cariboni, J. & Saltelli, A. An effectivescreening design for sensitivity analysis of large models.Environ. Modell. Softw. 22, 1509–1518.
Chanona, J., Pastor, L., Borras, L. & Seco, A. Fuzzyalgorithm for pH control in a struvite crystallization reactor.Water Sci. Technol. 53 (12), 161–168.
Copp, J. The COST Simulation Benchmark – Descriptionand Simulator Manual COST. European Cooperation in thefield of Scientific and Technical Research, Brussels, Belgium.
Cropp, R. & Braddock, R. The new Morris method: anefficient second-order screening method. Reliab. Eng. Syst.Safe. 78 (1), 77–83.
Ferrer, J., Rodrigo, M. A., Seco, A. & Penya-Roja, J. M. Energysaving in the aeration process by fuzzy logic control. WaterSci. Technol. 38 (3), 209–217.
Ferrer, J., Seco, A., Serralta, J., Ribes, J., Manga, J., Asensi, E.,Morenilla, J. J. & Llavador, F. DESASS: a software toolfor designing, simulating and optimising WWTPs. Environ.Modell. Softw. 23, 19–26.
Larsen, P. M. Industrial application of fuzzy logic control.Int. J. Man Mach. Stud. 12 (1), 3–10.
Manesis, S. A., Sapidis, D. J. & King, R. E. Intelligentcontrol of wastewater treatment plants. Artif. Intell. Eng. 12(3), 275–281.
Mendel, J. M. Fuzzy logic systems for engineering: a tutorial.Proceedings of the IEEE 83 (3), 345–376.
Morris, M. D. Factorial sampling plans forpreliminary computational experiments. Technometrics33, 161–174.
Ribes, J., Serralta, J., Seco, A., Ferrer, J., Lloret, R., Bernácer, I. &Morenilla, J. J. Implementation of a supervisoryfuzzy logic based control system in the Denia WWTP. InProceedings 10th IWA Specialised Conference on Design,Operation and Economics of Large Wastewater TreatmentPlants (LWWTP07), Vienna, Austria, September 9–13, 2007(on CD-ROM).
Ruano, M. V., Ribes, J., Sin, G., Seco, A. & Ferrer, J. A systematic approach for fine-tuning of fuzzycontrollers applied to WWTPs. Environ. Modell. Softw. 25(5), 670–676.
Saltelli, A., Scott, E. M. & Chen, K. (eds). SensitivityAnalysis. Chichester, Wiley.
Seco, A., Ribes, J., Serralta, J. & Ferrer, J. Biologicalnutrient removal model no. 1. Water Sci. Technol. 50 (6),69–78.
Serralta, J., Ribes, J., Seco, A. & Ferrer, J. A supervisorycontrol system for optimising nitrogen removal and aerationenergy consumption in wastewater treatment plants. WaterSci. Technol. 45 (4/5), 309–316.
Sin, G., Eliasson Lantz, A. & Gernaey, K. V. Good modellingpractice (GMoP) for PAT applications: propagation of inputuncertainty and sensitivity analysis. Biotechnol. Prog. 25,1043–1053.
Traoré, A., Grieu, S., Corominas, L., Thiery, F., Polit, M. &Colprim, J. Fuzzy control of dissolved oxygen in asequencing batch reactor pilot plant. Chem. Eng. Jour. 111(1), 13–19.
Weijers, S. R. & Vanrolleghem, P. A. A procedure forselecting best identifiable parameters in calibrating activatedsludge model no. 1 to full-scale plant data. Wat. Sci. Tech. 36(5), 69–79.
Yong, M., Yong-zhen, P., Xiao-lian, W. & Shu-ying, W. Intelligent control aeration and external carbon addition forimproving nitrogen removal. Environ. Model. Softw. 21, 821–828.
Zhu, G., Peng, Y., Ma, B., Wang, Y. & Yin, C. Optimizationof anoxic/oxic step feeding activated sludge process withfuzzy control model for improving nitrogen removal.Chemical Engineering Journal 151 (1–3), 195–201.