Activated sludge process based on artificial neural network

Journal of Harbin Instiule of Technology (Ner Series)， Vol. 9, No. 4, 2002Activated sludge process based on artificial neural networkZHANG Wen-yi, CAI Jian-an张文艺，蔡建安( School of Chemical & Environmental Engineering , Anhui Lniversity of Tecbnology , Manshan 243002, (China)Abstract: Considering the difficulty of crealing waler quality model for activated sludge system,日lypicul BParificial neural network model haa been established to simulate the operation of a waste waler treatment facili-ties. The comparison of prediction results with the on-spot measurements shows the model, the model is accu-rate and this model can also be used to realize itelligentized on-line control of the wastewater processingprocess.Key words: artifeial neural nelwork ; back-propugation; activated sludge system; modelCLC number: X703Document code: AArticle ID: 1005-9113(2002 )04-0383-040 INTRODUCTIONprocess is similar t0 the thinking process of humanbrain. Since the biological treatment system is a multi-The design and circulation of wastewater microbio-variable and non-linear system, it is difficult to build alogical treatment has been mostly carried out accordingreasonable mathenatic model. The emerge of ANNto experience data for many years. Some scholars (i. e.model just reflects this denand in time.W. w. Eckenfelder, R. E. Mckinney, A. W. Lawrence,There are various type of ANN. The type chosenP. L. MeCarty and so on ) introduced reactor theory ofin this paper is the multiplayer feed forward networkchemical industry and the theory of microorganism bio-( back-propagation) which is very effective in functionchemistry to wastewater treatment, and had profoundlyoptimization modeling. Each neuron has a " state" or "studied the laws of removing organic and microbiologi-activity level" that is determined by the input receivedcal growth. They brought up the model of the dynamicfrom the other units in the network. In the hidden andof removing organic ( biodegrade) and the dynamic of output layer, the net input unit Iis of the formbiological growth. But these models were all simple.Both the parameter of solution and the process computa-S; =Ey,u。+b;tion are all relatively casual. It is dificult for these the-whereb; is the bias of unit, y; is the output from unitj,ories to be applied lo the computation of the biologicalwg is the weight vector of unit 1 and m is the number oftreatment process in accordance with actual data of aneurons in the layer preceding the layer including unitgiven sewage plant. Since 1990s artificial, neural net-I. This weighted sum s,, which is called the " incomingwork has become a leading academic subject ，whichsignal" of uniti, is then passed through a nonlinear ac-has been in active use and swiftly expanded in engi-tivation function to produce the outgoing signal: y;. Theneering fields. Artificial Neural Network has recentlymost common transfer function is the sigmoidal:become a hotspot of the mathematic model building. Ithis paper, we study a given sewage treatment plantF(s) =with this technology.The mainly computation method of BP neural net-.1 THE METHOD OF MODEL BUILDING OFwork is as follows:①Initialized variable. Before training， theARTIFICIAL NEURAL NETWORKweights w; and other variables are initialized with ran-As an overhang utensil of generalized function ,dom values in the range[0.1].②Select a couple of data (x‘,T*), presenls theANN ( artificial neural network ) , which is structuredinput vector into input layers (m = 0), as to all unit iand auto-adaptable, is suitable for the process of thethere isy!" = x,*, where k is the serious number of thedala which has no evident relation between the condi-data.tion of biological treatment and its outcome, and canestablish a definite relationship between them. This中国煤化工ANNMHCNMHGReceived 2001 -09- 19Sponsored by the Natural sicience Foundation of Educational Cormcitee of Anhui Province( Grant No.2002kj035 ).383●Journal of Hurbin Institute of Technology ( New Series), Vol. 9, No. 4, 2002y," = F(s;") = F( Ew"y"~I)concentration. Connection between nodes are showm bysolid lines: they are associated with synaptic weights④computation of the error value of the outputthat are adjusted during the training procedure. The sig-layers ofj nodemoid activation function are plotted within the node.8;=_{r-y/] =y"(1 -y,")[7-y"]HOand computation of the error of all forehead layers node87;~1= 52(478").⑤after the presentation of the (h + 1)th input ex-0il-(一-CODamplar ,each weight is computed according toOwT = rη8"yr-1 - aOw",-1V-cya一-(wm= w.od+ AW，η=0.01-1- NHI-NT-eya- -( twhere n is the learming rate (i. e. the fraction whichthe global error is minimized during each step) and a isa constant ( momentum term) that determines the effecton the current weight change.NH-N-( y⑥Going to step②for all the training data.⑦after training all the sample I一N (N is thesample number), a plausible measure of how poorlyFig. 1 Typical three-layered feed-forword artiflcial neuralthe network is performing with its current set of weightsnetworkis given where T, is the actual state of the output unit inresponse to the kth input exemplar and y, is desired2.2ANN Topology Construction and Modelingstate.The ANN uses a threelayered (8→4-→2) feed-forward network, which has eight input neurons forE=-Z[7-门]2coding the eight pollutant variables. The hidden layerhas 4 neurons ,determining as the optimal configurationE= ZE:and giving the lowest error in the lraining (0. 0001).Learning is thus reduced to a minimization proce-The output neuron computes the value of the dependantdure of the error measure whenE≤e(e = 0.001 is avariable ( COD and NH3-N). We thus have an ANNsmall preset constant ) the leaming is over, otherwiseconstruction (8 ,4 ,2).go to step②2.3 Network Training and Predicting Result Analysis⑧after the network training , a neural network2.3. I Network trainingmodel is obtained. When presented with an inpul pat-Put the data of Table 1 into the artificial neuronternx(k)，the network can produce an accurately de-network with column 2,5,12, 17 exception and trainsired output vector y( k).the network to produce a desired output vector, invol-ving systematical change of the weights until the net-ARTIFICLAL MODELING OF ACTIVATEDwork produce the desired output( within a given toler-SLUDGE PROCESS BASED ON NEURALance). The process is repeated over the entire trainingNETWORK CONSTRUCTION(step②-→+⑧, η = 0.29). After 12173 iterations ，Eis less than a given constant (e = 0. 001), then the2.1 Input Vector and Output VectorANN training is over. Table 2 ~ 4 show the ANNn the present study, we consider the contamina-weight value, valve value and valve values of hide lay-tion concentration of the input wastewater as input vec-ers and output layers respectively.tor : Phenol ( PHN) , COD, NH,-N, oil, vaporized cy-2.3.2 Analysis of the ANN model predicting resultanugens( V-CYA)，Total cyanogens( T-CYA), scn-,When we present testing data (Column, 2,5,12,Ph. COD, NH3-N concentration of output wastewater17 in Table 1) into the trained ANN model, a series ofwere used as outpul veclor ( as shown in Table 1). Thedesired output vectors ( C0D and NH3-N concentrationdata used in this study come from the coking plant ofof the wastewater) was achieved (Table 4). From thatBaoshan Steel Company , Shanghai，China.table, we can see that the eror between the predictedFig. 1 shows typical three _layered feed-forward ari-中国煤化工all less than 5% of theficial neural network. Eight input nodes cortespondingF conclude that the ANNnoto eight independent waslewater variables ,4 hidden lay-CN M H GNN prdieting vable iser nodes and 2 output nodes estimating COD & NHz-Naccurate( shown in Table 5). An activaled modelingbased on the ANN is achieved.,384●Journal of Harbin Institute of Technology (New Series), Vol. 9, No. 4, 2002Table 1 Sample of ANN/mg . L 1Inpul vectorOutput veclorSampleNo.PHN p ( COD)ilV-cya τ-cya SCN NH,-N pHIρ(COD) NH,-N72. 21 6287.11.958. 11242227.0 9. 0378.1 0. 8478.0 1 774 8.02.988.65219235.0 9. 2661.40. 7863.2 1 682 7.52.618. 43248218.0 8. 9492.7 0.4249.517108.53.848.32229199.0 9. 9581.5 0. 6539.413588.32.496.78221143.0 9.1698.1 0. 4039.7 I 6585.12.577.41232176.0 9.63115.0 0. 1635. 82 1984.51.317.98248148.0 9. 1782.6 0. 3:28. 51 614 6.60.910. 90300147.0 9. 54140.0 0. 4:32.9 1 406 4.0.37.42113.09. 2788.5 1. 01137.616507.11.949.07225156.0 9. 9484.6 0. 3845.2 1 8068.31.84 10. 80238134.0 9. 57122.0 3. 911248.01 5947.27251153. 09.10110.0 1. 001359.61 652.7 1.639. 83256142.09.08126.0 0. 4045.21 488.6 1.63 .8. 7096.19. 2288.74.521544.41 5124.01.628.4222894.79.2483.90.401656.07.211. 8029580.49.3172.4 0. 2241.915664.11.227. 9220893.99. 2680.6 0. 491832.4 .I468..1 1.03 .246117.0 9.2363.70.0648.81 5525.82. 239.5523089.79.4057.00.672(44.11 5604.81.119.9425888. 39.5162.80.5721 4942.21.46107.08. 8856.50.482239.31 5661.93.669.0I162.0 9.2752.80.4746.01 6101.90141.89.3086.30.80Table 2 ANN Hidden Layers Weight ValueSerial No. of input vectorLayers No. .570. 685.-0.391-8.9753. 017-1.7198. 031.8.535-2.06312. 9751. 990- 6.5652. 876-13.850 11. 622 .-6.913 -4. 1030.7290. 229.-0.609-0.177-0.6131. 0320.117-0.366Table3 ANN Output Layers Weight ValueSerial No. of hide vectorSerial No. of output vector中国煤化工4-3. 451-7. 4:YHCN M H G- 18.289- 12. 7796.336-19.198 .9. 821385●Journal of Harbin Institute of Technology (Neu Series), Vol. 9, No.4, 2002Table 4 Valve values of hidden layers and output layershide layensoutput layerslaver3Valve values3.761 031. 625427. 976 312.98755.-8.64121.783 8Table 5 The result of the ANN predictionTest NoEstimated valueObserved valueEmorErorρ (COD)ρ(C0D)NH,-NNH,- N61.261.40. 3300. 790.78- 1.2896.098.12.1400.390.402.5012115.2110. 0-4.7701.051.00-5. (X077.880. 63.4750.510.49-4.13function of activated sludge system. This tool could a3 CONCLUSIONSso be used in other areas of the environment protec-tion. ANN can be seen to be an effective prediction al-(1) In this paper, we have built a basic structureternative to traditional modeling techniques.of activated sludge system which is based on artificialReferences:neural network (8-+4- +2). As to this model, whenthe input neuruns is 8, hidden neurons is 4, the output[1] BREY T, JARRE-TEICHMANN A, BORICH 0. Anifi-neurons is 2, the learning ratio is 0. 29，the ineriaciul neural network vg multiple linear regression: predic-ing PB ratios from empirical data{ J]. Mar Ecol Progfactor is 0.5, it has the best quality.Ser, 1996 (140) :251-256.(2) By comparing the predicted result of the[2] LEK S, BELAUD A. Improved estimation, using neuralmodel with the on-spot data, the model is accurate andnetworks, of the food consumption of fish populationscan be used for wastewater process plants flexibly. This[J]. Mar Freshwaler Res, 1995 ,46 :1229-1236.model can also be used to realize intelligentized on-line[3] LEK Sovan . Predieting stream nitrogen concentration fromcontrol of the process.walershed features using neural networks[J]. Wat Res,(3) Finally, we put forward an effective tool for1999 ,33( 16) :3469-3478.the predietion of COD and NH3-N concentration as a中国煤化工MYHCNMHG.386●