Research On ANN Models Stimulating Water Quality Dynamic Parameters In Carrousel Oxidation Ditch System

Stimulating the correlation of water quality parameters correctly is an important topic in the real-time control on-line of water treatment system. Artificial neural networks(ANNs) have ieen widely used as a promising alternative approach in system stimulation and prediction because of their distinguishing features- data-driven, self-organized, self-adaptive, error-contained, parallel. The paper,at the beginning,summarizes the current application of ANNs in the field of environmental science and analyzes the effectiveness and flexibility of ANNs in pattern recognition,dependent variable approach,association and recollection, etc.Then,the carrousel oxidation ditch system in Luohe Center of wastewater treatment ,as a case being difficult to control on-line because the influent characteristics are complex and vary significantly,is presented which makes it necessary to stimulate the dynamic behavior of the oxidation ditch system.To resolve the problem,two kinds of ANNs,error back-propagation(BP) neural network(NN) and radius basis function(RBF) NN are employed and three models are developed to predict effluent SS,COD,total nitrogen(TN) and total phosphrate(TP).The prediction of effluent SS and COD in the oxidation system are based on BPNN respectively,while the effluent TN and TP are predicted simultaneously based on RBFNN.The three models have good performance in test and are capable of adapting to various situations.Especially in the validation ,they can recognize high concentration of effluent with accuracy. Meanwhile.the test of stimulation shows that the models' performance strongly depends on ANN's architecture and the quality and size of training and test sets.Finally,the three models,which can be used to predict effluent SS, COD,TN and TP in the Carrousel oxidation system and stimulate the relationship between input variables including influent parameters(SS,COD,TN,TP,NH4+-N),control parameters(water temperature,MLSS, MLVSS,SV30) and output variables including effluent parameters(SS,COD,TN, TP),provide an easy method to optimize the system real-time control on-line and prevent the system from accidents.In comprehensive comparison,RBFNN model predicting effluent TN and TP is superior to BPNN models predicting effluent SS and COD. BPNN model predicting effluent COD has a relative narrow input range of influent COD.Focusing on negligence and mistakes existing in some researches,the paper states strict and integrated modeling steps, including data collection andtransformation,parameter qualitative analyse,condition assumption,the selection of ANN's architecture and algorithm,NN training and validation, stimulation test and so on. According to the dynamic mechanism of activated slude system,which acts as an auxiliary criterion to evaluate candidate models' performance,three satisfactory models are achieved .To improve models' adaptability and ability to generalize, some measures are put forward,such as three rules of data trans--formation,how to cope with the contradiction between precision and accuracy,the model's extent analyse and the restriction from dynamic mechanism of activated sludge system and so on.Furthermore,the author discusses the three models' limitation in application and why it is.The last but not the least is that the paper accumulates modeling experience which is helpful to the study of intelligent control of water treatment system.lt is a stepstone to deepen the research on consequent feed-back control on-line of the oxidation ditch system .The author hope all he has done brings related researches valuable assistance.