基于深度学习模型的中试纳滤系统膜污染预测研究

Prediction of Membrane Fouling in Pilot Nanofiltration System based on Deep Learning Model

膜污染的预测或模拟对于阐明膜污染机理和制定有效的污染控制措施具有重要意义。建立了一种门控循环单元(GRU)模型,用于预测中试纳滤系统的过滤性能,并探究膜的污染机理。采用跨膜压差(TMP)作为输出变量,考察以TMP自身为输入变量的单输入时间序列模型的预测效果,并探讨了9种水质参数:pH、水温、电导率、总溶解性固体(TDS)、总硬度(TH)、浊度、高锰酸盐指数(CODMn)、溶解性有机碳(DOC)以及UV254的独立和耦合效应对膜污染及预测结果的影响,筛选出最优模型与膜污染预测中常用的机器学习算法——随机森林(RF)和时间序列预测中广泛使用的长短时记忆网络(LSTM)进行比较。结果表明:以TMP自身数据作为输入的LSTM以及GRU单输入时间序列预测模型的R^(2)分别达到了0.9613和0.9861;水质参数对TMP的相关性顺序依次为:温度>电导率>TDS>总硬度>高锰酸盐指数>DOC>浊度>UV254>pH;以温度和电导率为输入的多变量GRU模型预测效果最佳(R^(2)=0.8344),预测精度优于相同时间步长的单输入GRU模型(R^(2)=0.4555)以及相同输入参数的多变量LSTM(R^(2)=0.6428)和RF模型(R^(2)=-4.1894),在此基础上增加或减少水质参数作为模型输入,模型的预测精度均有所下降。时间序列模型在膜污染预测方面展现出了较高的可靠性,GRU模型预测精度更高,在膜污染预测中具备更高的应用潜力。输入变量的特征选择对高效预测膜污染具有重要意义,对输入水质数据进行特征选择,可以显著提升模型的预测性能。此外,预测结果反向验证了水温和无机离子耦合是造成冬季纳滤膜污染的主要原因,冬季纳滤运行过程中要注意温度和无机离子污染对纳滤膜运行稳定性的影响。

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,agated recurrent unit(GRU)model was established to predict the filtration performance of a pilot-scale nanofiltration system and explore the fouling mechanism of membranes.Trans membrane pressure(TMP)was adopted as the output variable,and the predictive effectiveness of a single-input time series model using TMP itself as the input variable was examined.Furthermore,the independent and coupled effects of nine water quality parameters:pH,water temperature,conductivity,total dissolved solids(TDS),total hardness(TH),turbidity,permanganate index,dissolved organic carbon(DOC),and UV254on 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 0.9613and 0.9861,respectively.The order of correlation between water quality parameters and TMP was as follows:temperature>conductivity>TDS>TH>CODMn>DOC>turbidity>UV254>pH.The multivariate GRU model with temperature and conductivity as inputs exhibited the best prediction performance(R2=0.8344),with higher accuracy than the single-input GRU model with the same time step(R2=0.4555)and the multivariate LSTM(R2=0.6428)and RF models(R2=-4.1894)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 prediction 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.