Prediction of Membrane Fouling in Pilot Nanofiltration System based on Deep Learning Model
The prediction or simulation of membrane fouling is of great significance for elucidating the mechanism of membrane fouling and devising effective pollution control measures.In this study,a gated recurrent unit(GRU)model was established to predict the filtration performance of a pilot-scale nanofiltration system and explore the foul-ing mechanism of membranes.Trans membrane pressure(TMP)was adopted as the output variable,and the predic-tive effectiveness of a single-input time series model using TMP itself as the input variable was examined.Further-more,the independent and coupled effects of nine water quality parameters:pH,water temperature,conductivity,to-tal dissolved solids(TDS),total hardness(TH),turbidity,permanganate index,dissolved organic carbon(DOC),and UV254 on membrane fouling and prediction results were investigated.The optimal model was selected and compared with commonly used machine learning algorithms in membrane fouling prediction,namely random forest(RF)and long short-term memory(LSTM)networks widely used in time series prediction.The results indicated that the R2 values of the LSTM and GRU single-input time series prediction models using TMP itself as input reached 0.961 3 and 0.986 1,respec-tively.The order of correlation between water quality parameters and TMP was as follows:temperature＞conductivi-ty＞TDS＞TH＞CODMn＞DOC＞turbidity＞UV254＞pH I.The multivariate GRU model with temperature and con-ductivity as inputs exhibited the best prediction performance(R2=0.834 4),with higher accuracy than the single-in-put GRU model with the same time step(R2=0.455 5)and the multivariate LSTM(R2=0.642 8)and RF models(R2=-4.189 4)with the same input parameters.Therefore,increasing or decreasing water quality parameters as model inputs will result in a decrease in the prediction accuracy of the model.Time series models have demonstrated high reliability in predicting membrane fouling,with the GRU model showing higher prediction accuracy and greater potential for application in membrane fouling prediction.Feature selection of input variables is crucial for efficiently predicting membrane fouling,and selecting features from input water quality data can significantly enhance model predictive performance.Additionally,the pre-diction results confirmed that the coupling of water temperature and inorganic ions is the main cause of winter nanofiltration membrane fouling.Therefore,attention should be paid to the influence of low temperature and inorganic ion pollution on membrane operation stability during winter membrane operation processes.
membrane fouling predictiondeep learningGRUnanofiltrationpilot test
万方数据
维普
