基于深度学习的排污口污染物浓度预测研究

Predicting concentrations of water contaminates from sewage outlet based on deep learning algorithms

入河排污口是污染物进入生态环境的最后一道关口,预防超标排放是改善流域生态环境质量的基础.为实现排口超标排放事前预警,本研究以长江泰州段两类典型排口(污水处理设施排口和工业企业清净下水排口)为例,利用排口污染物监测数据与气象数据,基于长短期记忆神经网络(LSTM)、门控循环单元(GRU)、卷积循环神经网络(CRNN)等深度学习算法构建多污染因子(总氮、氨氮、COD、总磷)浓度预测模型,并结合SHAP分析结果识别影响排口水质预测的重要因素.结果表明:(1)单层与双层GRU模型在排污口未来6 h污染物浓度预测中表现较好,R^(2)可达0.67~0.81;(2)自相关变量的累积重要性绝对值占比超80%,对排口污染物浓度预测的影响显著大于其他输入变量.该方法有潜力拓展应用至其它排污口类型及其它污染因子的浓度预测,为排口污染预警和全链条管理提供技术支撑.

The sewage outlet is the final gate before pollutants entering the ecological environment.To prevent water pollution and achieve early warning of emissions exceeding the standard,it is necessary to develop models to predict concentrations of water contaminates from sewage outlet.This study took two typical types of sewage outlets in the Taizhou section of the Yangtze River as cases and developed a multi-factors model to predict concentrations of water contaminates from these two sewage outlets based on three deep learning algorithms,including long short-term memory neural network (LSTM),gated recycling unit (GRU),and convolutional recurrent neural network (CRNN),using water contaminate monitoring data and meteorological data as input features.Then the importance of different input features was analyzed based on the method of Shapley Additive Explanation.Results showed that:(1)Both single-layer and double-layer GRU models performed well in predicting pollutant concentrations for the next6 hours,with R^(2) reaching 0.67~0.81;(2)The importance of autocorrelated variables,significantly higher than other input variables,accounted for more than 80%among all variables.This method also shows high potential to be applied to other types of sewage outlets and pollutant factors,thereby supporting early warning of pollution from sewage outlets.