Research On Some Technologies Of Soft Sensors For Wastewater Treatment Process

With the ever-increasing development,more and more attentions have paid to the sustainable development of environment and resources.Thus,the requirements of wastewater treatment are constrained by more and more rigid regulations.Wastewater treatment process is a complex biochemical system with multivariable coupling,time-varying parameters,strong nonlinearity and great uncertainty.There are many important variables which cannot be measured online by traditional hardware sensors,while these variables relate to the effective monitoring,optimization and quality of the process.To solve this problem,the soft sensors can measure the difficult-to-measure process variables by establishing the mathematical model of the easy-to-measure process variables and difficult-to-measure process variables.However,to establish a soft sensor model by the raw data collected from the industry directly may result in poor prediction performance,even leading to the overall performance degradation of a wastewater treatment system.Besides,the model may suffer from performance degradation or be no longer suitable for the current working point with time goes on.Against the background of wastewater treatment process,this paper focus on the data processing,modeling,model maintenance and interpretability of soft sensor.The main research contents of this paper are as follows:1.To address the problems of data noise,data drift and other abnormality existing in wastewater treatment,two data preprocessing methods,i.e.,the zero mean normalization and time difference algorithm,are introduced to the modeling process.To deal with the problems of multivariable coupling,strong correlation and the redundancy in original sample data set,four variable selection methods are presented,including variable importance in projection(VIP),genetic algorithm(GA),moving window method(MW),minimum absolute contraction and selection operator(LASSO).The performance to predict the effluent BOD of the model is used to verified the presented four variable selection methods.Results show that variable selection can significantly improve the prediction accuracy of the model while reducing the data dimension and model complexity.2.To solve the on-line measurement of BOD,extreme learning machine(ELM)and gaussian process regression model(GPR)are studied.The parameter selection,advantages and disadvantages of the two model are discussed.To enhance the performance of the models,the aforementioned four variable selection methods are combined with the two model.The performance of the models is compared with the partial least squares model which is commonly used in wastewater treatment.Meanwhile,in order to address the problem of model performance degradation,a TD-JIT adaptive framework is proposed,which extends the model from the traditional offline training and online measurement to the adaptive model to ensure the overall performance of the model even in the case of system changes.3.The interpretability of soft sensor model based on multivariate Granger causality is studied.Due to its simpleness and effectiveness,data-driven modeling is widely used in the soft sensor modeling for wastewater treatment.The data-driven modeling method take the wastewater treatment process as a black box,only use the input and output variable to establish the model to predict the target variables which leading to poor interpretability.To combine the moving window and multivariate Granger causality analysis(MW-MVGC),the dynamic causality sequence of input and output variables of the model can be obtained.Thus,study the interpretability of the model using the Granger causality,and multi-step prediction of the variable Granger causality by ARMA model can capture the evolution trend of the variables of the wastewater treatment process and adjust the model in real-time.