developing ann
TRANSCRIPT
-
7/27/2019 Developing ANN
1/8
6thEnvironmental Engineering Specialty Conference of the CSCE
& 2ndSpring Conference of the Geoenvironmentla Divisionof the Canadian Geotechnical Society
6econgrs spcial sur le gnie de l environmnement de la SCGCet 2econgrs de pr intemps de la secti on goenvi ronnementale
de la societcanadi enne de gotechnique
______________________________________________________
London, Ontario
7-10 June 2000 / 7 10 juin 2000
376
DEVELOPING ARTIFICIAL NEURAL NETWORK PROCESS MODELS: AGUIDE FOR DRINKING WATER UTILITIES
C.W. Baxter, S.J. Stanley, Q. Zhang, D.W. SmithDepartment of Civil and Environmental Engineering, University of Alberta, Canada
ABSTRACT: Due to the complex nature of drinking water treatment unit processes, utilities are often unable toquantify the interactions and relationships that exist between process inputs and process outputs. Processmodels, where they exist are site-specific and are unable to fully handle continuous and extreme variations in
more than one or two key process parameters. As such, process control strategies are based upon the results ofbench-scale and pilot-scale experiments rather than on process models. Presented is a framework fordeveloping artificial neural network (ANN) models of drinking water treatment processes. Then ANN modelling
technique is robust and can handle the dynamic nature of the treatment processes. The utility of ANN models isdemonstrated through a case study where a successful ANN model to predict filtration performance wasdeveloped.
1. INTRODUCTION
All unit processes in the drinking water treatment
industry are complex, involving many biological,physical, and chemical phenomena. While thenumerous parameters that directly influence the
performance of each process are known to plantoperators and researchers, the interactions andrelationships between process inputs and outputs is
often poorly understood and can not be easilyquantified. Process models, where they exist, areoften site-specific and developed using the results of
bench-scale and pilot-scale experiments where many
of the process parameters are held constant. Suchmodels are unable to cope with simultaneous
fluctuations in more than one or two key parameters.
The inability of water treatment utilities to quantify
process interactions and relationships can causegreat difficulty for water treatment process control. Ifeach unit process is considered to be a multiple-input
multiple-output (MIMO) non-linear optimizationproblem, effective control can only be achieved
through strategies that can accommodate thefluctuations of multiple input and output parameters ona real-time basis. In the absence of such strategies,
utilities resort to using bench-scale tests that provideonly a snapshot of the process conditions in order toachieve process control.
As utilities strive to meet more stringent customer-imposed and regulatory demands on finished water
quality, there is an urgent need for new tools toimprove process knowledge and provide effectivealternatives to current process control methodologies.
One such tool is artificial neural network (ANN)
modelling, a robust technique that allows for thedevelopment of multiple-variable, non-linear models
that can be integrated into real-time process controlstrategies.
Presented are guidelines for the development of ANNprocess models in the drinking water treatmentindustry. More specifically, the requirements and
applicability of the modelling technique are discussed,and the utility of the technique is demonstrated
-
7/27/2019 Developing ANN
2/8
377
through a sample process model and an associatedoffline process control application.
2. ARTIFICIAL NEURAL NETWORK PROCESS
MODELS
ANNs are categorized as an artificial intelligence
modelling technique due to their ability to recognizepatterns and relationships in historical data andsubsequently make inferences concerning new data.
ANNs can be used for two broad categories ofproblems: data classification and parameterprediction. For data classification problems, the ANN
uses a specified algorithm to analyze data cases orpatterns for similarities and then separates them into apre-defined number of classes. Credit agencies, for
example, often use ANN models to separate apotential list of clients into classes according to theirlevel of credit risk or buying patterns. For parameter
prediction problems, the ANN learns to accuratelypredict the value of an output parameter givensufficient input parameter information. The main
applications of the ANN technique in the watertreatment industry are in the development of processmodels and model-based process-control andautomation tools. These applications can be
categorized as parameter prediction problems, towhich the ensuing discussion is dedicated.
2.1 Advantages of ANN Modelling
The ANN modelling technique holds several
advantages over mechanistic modelling that make itparticularly suitable to process modelling in thedrinking water treatment industry. In developing
models of MIMO water treatment processes, a keyconsideration is the ability of the model to adapt to thedynamic nature of the process on a real time basis.
ANN models can handle non-linear relationships, andcan provide predictions of output parameters in realtime in response to simultaneous and independent
fluctuations of the values of model input parameters.ANN models are also fault-tolerant in model buildingand in end-use. Data patterns where the value of one
or more of the model inputs are missing can be
incorporated into model building if necessary,although model prediction will improve if only
complete data patterns are used. Similarly, whencompleted models are applied to new data patternswhere values are missing, the value of model outputs
can still be predicted. This feature is particularly usefulin process control applications where input data is fedto models in real time by instruments that are subject
to periodic failure. Finally, ANNs do not requirecomplicated programming, logical inference schemes,
or the development of complex algorithms to developa successful model. Many user-friendly ANN software
packages exist, offering the user a myriad ofmodelling options and allowing the user to customizethe modelling process to suit his or her knowledge of
modelling heuristics.
2.2 Challenges of ANN Modelling
Some aspects of the ANN modelling technique maypresent challenges to drinking water treatment utilities
that wish to develop successful process models. ANNmodels can only be developed where sufficienthistorical data for each of the process parameters
exists. Since the modelling technique requires that.While data requirements will be addressed further on,it is important to note at this point that not all utilities
have detailed historical records of influent waterquality, process control actions, and treated waterquality. As more utilities realize that historizing such
data is important for increasing process knowledgeand improving process efficiency, this challenge willbecome less influential with time. Perhaps the
greatest challenge that must be overcome is theperception of ANN models to be black-box modelsthat can not be understood by the end-user. Thisperception is fed by a new host of ANN software
packages that use proprietary algorithms to developquick and easy models with minimal user interference.From a public health standpoint, utility managers are
rightfully concerned about trusting plant operations tosuch a model. By using less sophisticated softwarepackages that feature well-understood training
algorithms, this challenge can be overcome.
2.3 Existing ANN Models
In recent years, water treatment utilities have come torealize the potential for the ANN modelling
techniques, resulting in a number of published modelapplications in water quality, treatment, anddistribution. With respect to water quality, models
have been developed for trihalomethane (THM)formation and speciation (Hutton, Sandhu and Chung1996), source water salinity forecasting (DeSilets et
al. 1992), and raw water colour forecasting (Zhang
and Stanley 1997). Process models have beendeveloped for alum and polymer dose forecasting in
coagulation (Mirsepassi, Cathers and Dharmappa1995), and turbidity and colour removal throughenhanced coagulation (Stanley, Baxter and Zhang
2000). Finally, for distribution systems, a model hasbeen developed for the prediction of residual chlorinein the distribution system (Rodriguez et al. 1997).
-
7/27/2019 Developing ANN
3/8
378
3. Building ANN Models of Drinking WaterTreatment Processes
3.1 Types of ANN Models
Each drinking water treatment unit process can beconsidered a non-linear MIMO process, as previouslydiscussed. The process inputs include influent water
parameters, such as pH and turbidity, as well asoperational parameters, such as chemical dosinglevels and flow rate. The process outputs are the
effluent water parameters. ANN models of drinkingwater treatment process can assume two distinctforms: process models and inverse process models.
In the former, the model predicts the value of one ormore process outputs, given the values of the processinput parameters. An example of this type of model is
the prediction of clarifier effluent turbidity usinginfluent water parameters and operational parameters.In the latter, the ANN model predicts the value of one
or more process inputs, given the values of theremaining process inputs and process output(s). Thistype of model is often used to predict the value of an
operational parameter required to reach a targeteffluent quality. An example of this type of model isthe prediction of alum dose required to maintain adesired value for clarifier effluent turbidity.
3.2 ANN Modelling Requirements
In order to build successful ANN process models,specific data, software, hardware, and personnelrequirements must be met. The key requirement of
the ANN modeling approach is the availability ofrelevant data to describe the process being modeled.Data must be available in a useable electronic format
for each of the process input and output parameters.The data used in model development must berepresentative of plant operations, spanning the range
of operating conditions that may be encounteredduring both routine operations and process upsetconditions. Model development requires the use of
appropriate ANN software. Many commercial softwarepackages are currently available; site-specificrestrictions on operating systems and data format, as
well as desired options will dictate the best software
choice for each utility. As previously indicated,software packages that use only proprietary
algorithms should be avoided in favour of those thathave a high degree of user input in model-building.With respect to hardware requirements, recent
advances in computing technology have made eventhe most modest new home computing systemcapable of performing the modelling calculations. To
ensure optimal performance however, a 500 MHzprocessor and 128 M of RAM should be considered to
be the minimum system requirements. With respect topersonnel requirements, the model developer should
have expert knowledge of the process being modelledand should understand basic modelling heuristics. Assuch, a plant operator or operations engineer with
some training in ANN model development would be asuitable candidate to develop process models.
3.3 Key Components of ANN Models
There are several key components to ANN models
that are collectively referred to as the ANNarchitecture. Processing units or neurons performprimitive operations such as scaling data, summing
weights, and amplifying or thresholding sums.Neurons are organized into layers with each layerperforming a specific function. The input layer serves
as an interface between the input parameter data andthe ANN model. Most models also contain one or twohidden layers, although more are possible. These
layers perform most of the iterative calculations withinthe network. The output layer serves as the interfacebetween the ANN model and the end-user,
transforming model information into an ANN-predictedvalue of the output parameter(s). Each neuron isconnected to every neuron in adjacent layers byweights; links that represent the strength of
connection between neurons. Each ANN model has apropagation rule that defines how the weightsconnected to a neuron are combined to produce a net
input. The propagation rule is generally a simplesummation of the weights. As discussed, the inputlayer serves as an interface between input parameter
data and the model. In this layer, a scaling function isused to scale data from their numeric range into arange that the network deals with efficiently, typically 0
to 1. The hidden and output layers contain anactivation function that defines how the net inputreceived by a neuron is combined with its current
state of activation to produce a new state of activation.The most common activation function used in processmodelling is the sigmoidal activation function. Finally,
each network has a learning rule that defines how theweights are modified in order to minimize predictionerror. As will be discussed, the backpropagation
algorithm is the most common learning rule employed
in process modelling.
3.4 ANN Learning
In developing ANN process models, a supervised
learning paradigm is employed. In supervisedlearning, historical data patterns, consisting of valuesfor each of the model input and output parameters,
are used to train the ANN. The goal of supervisedlearning is to minimize the error between the model-
-
7/27/2019 Developing ANN
4/8
379
predicted value and the actual value of the outputparameter(s). The error minimization takes place by
modifying the weights between neurons according toa learning rule, typically the backpropagationalgorithm. This algorithm increases or decreases the
value of the weights in order to minimize the squareddifference between the network-predicted value andthe actual value of the output parameter(s), summed
over all of the data patterns. ANN learning or trainingproceeds through repeated presentation of datapatterns to the ANN and subsequent weight
modification until the prediction error is sufficientlysmall, as defined by the user, or until a maximumnumber of iterations has been reached. A more
detailed description of the ANN learning process ispresented by Baxter, Stanley, and Zhang (1999).
3.5 Data Collection and Analysis
In order to develop successful ANN models of drinking
water treatment processes, careful attention must bepaid to the details of data collection and analysis. Incollecting data, several factors need to be considered.
First, the availability of the data must be ascertained.For data availability, the parameters for whichhistorical data exists, the time frame of historicalmeasurements, and the frequency of data
measurement must all be determined. The format andreliability of the data are also key considerations indata collection. Historical data can originate from of
grab-samples or real-time measurements, andmeasurements can be discrete or aggregated from anumber of samples. The reliability of the available
data should be ascertained through an examination ofquality assurance and quality control protocols.Finally, it is of considerable importance to note any
process changes that may have been implementedduring the time frame for which data are available.
With the above considerations in mind, it is possible todelineate a number of guidelines for selecting data tobe used in ANN process modelling. First and
foremost, data for each of the parameters known orsuspected to affect the process being modelled mustbe available. The quantity of data required to develop
a model is site specific, and is affected by seasonal
fluctuations in influent water quality, and the frequencyof process upsets. As such, it is more important to
ensure that the data are fully representative of therange of conditions that can be expected duringperiods of routine and upset operations. As a general
guideline, at least one full cycle of data must beavailable in order to ensure a representative data set.In temperate areas with four distinct seasons for
example, a data cycle would encompass a full year ofhistorical data. With respect to the format of the data,
the variability of the process as well as dataavailability will dictate whether to use hourly data,
daily averages, or some other frequency for each ofthe parameters. Successful process models can oftenbe made using the daily average or some daily
percentile value of each of the model parameters.With respect to the effect of major process changeson data selection, data collected prior to major
process changes should not be incorporated if theyare believed to differ significantly from current processdata. This guideline is necessary to ensure that the
data are representative of current process conditions.Finally, in order to maintain the integrity of the dataset, appropriate quality assurance and quality control
protocols for each model parameter must be in place.
Once an appropriate historical data set has been
selected, it should be fully characterized andsubjected to a comprehensive statistical analysis.Data characterization involves a qualitative
assessment of hourly, daily, and seasonal trends ofeach potential model parameter. The statisticalanalysis involves the determination of measures of
central tendency, measures of variation, percentileanalysis, and identification of outliers, erroneousentries, and non-entries for each data parameter. Incombination, the data characterization and statistical
analysis help to identify the boundaries of the studydomain as well as potential deficiencies in the dataset.
3.6 Model Building Protocol
Unfortunately, there is no widely accepted bestmethod of developing ANN models. When all of thepossible options in building an ANN architecture are
considered, an infinite number of distinct architecturesare possible. As such, each model developer mayuse a different protocol to reduce the number of
architectures that are evaluated. What follows is a six-step protocol that the authors have found to be usefulin developing drinking water treatment ANN process
models.
3.6.1 Selection of Model Inputs and Outputs
The first step involved in the selection of model inputsand outputs is the selection of the model output(s).
The output(s) are selected on the basis of operationalneeds, relevant literature, and data availability. Aspreviously discussed, drinking water treatment
processes can be considered MIMO processes. Assuch, models of such processes can have multipleoutput parameters. Since ANN models train by
minimizing the error between the predicted and actualvalues of model output parameters, the technique
-
7/27/2019 Developing ANN
5/8
-
7/27/2019 Developing ANN
6/8
381
being modelled however, model fine-tuning can makethe difference between acceptable and unacceptable
prediction performance.
3.6.6 Comparison of Candidate Models
Where more than one candidate model has beenidentified, the candidate models are compared and
the final model is selected. The selection processinvolves statistical and graphical analyses of themodelling results for each of the candidates. Plots of
actual versus predicted values of the model outputcan be useful in highlighting problem areas andprediction inconsistencies. Overall, the model that
best meets the goals of the modelling process isselected.
4. Case Study - ODP Plant
The Oxidation Demonstration Project (ODP) facility,owned and operated by the Metropolitan WaterDistrict of Southern California (MWD), is located on
the premises of the F.E. Weymouth Filtration Plant, inLa Verne, California. The facility began operation in1992 in order to evaluate the use of ozone andPEROXONE (a combination of ozone and hydrogen
peroxide) as primary disinfectants at the full-scaleMWD facilities. The facility has a capacity of 20.8ML/d and consists of three types of ozone generators,
three feed gas systems, an external mix ozonecontacting system, two ozone contactors, two types ofozone destruct systems, a hydrogen-peroxide
injection system, and conventional coagulation andfiltration unit processes. With respect to theconventional process, the ODP plant has the ability to
feed sulfuric acid, sodium hydroxide, metal coagulant(alum, ferric chloride, or polyaluminum chloride),polymer, filter aid, chlorine, and ammonia. The ODP
plant is operated as a stand-alone facility, although
treated water is currently pumped from the clearwell tothe head works of the Weymouth facility.
In addition to the evaluation of alternativedisinfectants, the ODP facility has been used to study
biological filtration, filter air binding, enhancedcoagulation, arsenic removal, and other filtrationissues. Each of these studies required different plant
configurations and had different data collectionrequirements. As such, while the plant has been incontinuous operation for over 7 years the data set was
often incomplete. A complete list of the parametersthat have been historically measured at the facility ispresented in Table 1.
Table 1. Data parameters historically measured atthe ODP plant
Parameter Classification
Source Water Plant Influent
Turbidity (NTU) Plant Influent
pH Plant Influent
Dissolved Oxygen (mg/L) Plant Influent
UV254(cm-1
) Plant Influent
Temperature (C) Plant Influent
Alkalinity (mg/L) Plant Influent
Bromide (mg/L) Plant Influent
Plant Flow (MGD) Operational
Coagulant Dose (mg/L) Operational
Polymer Dose (mg/L) Operational
Filter-aid Dose (mg/L) Operational
Total Ozone Dose (mg/L) OperationalFiltration Rate (gpm/ft
2) Operational
90th
Percentile Turbidity (NTU) Filter Effluent
50th
Percentile Turbidity (NTU) Filter Effluent
90th
Percentile Particle Counts (PC/mL) Filter Effluent
50th
Percentile Particle Counts (PC/mL) Filter Effluent
4.1 Model Development
The objective of this case study is to develop ANN
models to predict the 50
th
percentile filter effluentturbidity at the ODP using historical operational data.All percentile parameters are based on a 24-hour time
period. The six-step protocol previously defined wasused in model development. As per the protocol, themodel input and output parameters were selected
based on the results of a literature review oncoagulation filtration and data availability. A list of theinputs used in model development is presented in
Table 2
Table 2. 50thpercentile filter effluent model
inputs
Parameter Classification
Turbidity (NTU) Plant Influent
pH Plant Influent
Temperature (C) Plant Influent
Alkalinity (mg/L) Plant Influent
Coagulant Dose (mg/L) Operational
Polymer Dose (mg/L) Operational
Filter-aid Dose (mg/L) Operational
Total Ozone Dose (mg/L) Operational
Filtration Rate (gpm/ft2) Operational
-
7/27/2019 Developing ANN
7/8
382
0
0.02
0.04
0.06
0.08
0.1
0.12
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79
Pattern #
FilterEffluentTurbidity(NTU
)
Actual Model Prediction
Figure 1. Model results for 50th percentile filter effluent turbidity model
0.05
0.055
0.06
0.065
0.07
0.075
0.08
0.085
0.09
0.095
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
Polymer Dose (mg/L)
50thPercentileFilterEffluentTurbid
ity(NTU)
Optimized Polymer Dose
Figure 2. Effect of polymer dose on filter effluent turbidity for typical spring water conditions
Based on the protocol, the best candidate model was
a three-layer backpropagation network with 32 hiddenlayer neurons. The model predicted the 50
thpercentile
value of filter effluent turbidity with a mean absolute
error of 0.003 NTU and a coefficient of multiple
regression (R2) of 0.85, based on the production data
set. The model results are presented graphically inFigure 1.
-
7/27/2019 Developing ANN
8/8
383
While a detailed description of the numerous processcontrol and related applications of developed models
is beyond the scope of this paper, the trained modelcan be used as a virtual full-scale laboratory toprovide information about process operations at the
ODP facility. A graphical representation of the effect ofaltering the polymer dose on filter effluent turbidity fortypical spring water conditions is presented in Figure
2. The values of the filter effluent turbidity used tocreate this figure were generated from the trainedANN process model.
Based on typical raw water and operational conditionsexperienced in springtime at the ODP facility, a
polymer dose of approximately 2.4 mg/L wouldproduce a filter effluent with the lowest turbidity. Itshould be noted that this example of the virtual-lab
capability of the ANN models is extremelyrudimentary; more advanced virtual-lab systems canbe developed where the effects of simultaneously
changing the values of multiple input parameters canbe assessed.
5.0 CONCLUSION
While the challenges involved in developing ANN
process models may be prohibitive for some utilities,the current trend towards historizing process data, incombination with recent advances in computing
technology, have made the technology appropriate formany utilities. As demonstrated through the casestudy, The ANN modelling technique allows utilities to
improve process knowledge, and therefore facilitateprocess control, by allowing for the development ofmultiple-parameter, non-linear models of complex unit
processes. As customer and regulatory demands onfinished water quality continue to become morestringent, ANN models will continue offer drinking
water utilities a sophisticated process modelling andcontrol alternative to conventional methodologies.
6.0 ACKNOWLEDGEMENTS
The authors an indebted to the American Water
Works Association Research Foundation (AWWARF),the Metropolitan Water District of Southern California,
and EPCOR Water Services for their continuedlogistical and financial support of ANN research in thedrinking water treatment industry.
7.0 REFERENCES
Baxter, C.W., S.J. Stanley, and Q. Zhang.Development of a Full-Scale Artificial NeuralNetwork Model for the Removal of Natural Organic
Matter by Enhanced Coagulation. AQUA
48(4):129-136.
DeSilets, L., B. Golden, Q. Wang and R. Kumar.1992. Predicting Salinity in the Chesapeake BayUsing Backpropagation. Computers and
Operations Research19(3-4): 277-285.
Hutton, P.H., N. Sandhu, and F.I. Chung. 1996.
Predicting THM Formation With Artificial NeuralNetworks. In Proceedings of the North AmericanWater and Environment Conference, Anaheim,
CA.: ASCE.
Mirsepassi, A., B. Cathers and H. B. Dharmappa.
1995. Application of Artificial Neural Networks tothe Real Time Operation of Water TreatmentPlants. In IEEE International Conference on Neural
Networks: Proceedings. Perth, Australia: IEEE
Rodriguez, M. J., J. R. West, J. Powell and J. B.
Serodes. 1997. Application of Two Approaches toModel Chlorine Residuals in Severn Trent WaterLTD (STW) Distribution Systems. Water Science
and Technology36(5): 317-324.
Stanley, S.J., C.W. Baxter, and Q. Zhang. 2000.
Process Modeling and Control of Enhanced
Coagulation. Denver, Colo.: AWWARF andAWWA.
Zhang, Q. and S. J. Stanley. 1997. Forecasting Raw-water Quality Parameters for the North
Saskatchewan River by Neural Network Modeling.Water Research31(9): 2340-2350.