基于灰色关联-长短时记忆网络的水质预测研究  

Study on water quality prediction based on Grey Relation-Long and Short-Term Memory network

作　　者：方国华[1] 张钰[1] 袁婷 廖涛 丁紫玉 FANG Guohua;ZHANG Yu;YUAN Ting;LIAO Tao;DING Ziyu(College of Water Conservancy and Hydropower Engineering,Hohai University,Nanjing 210098,China;Jiangsu Surveying and Design Institute of Water Resources Co.,Ltd.,Yangzhou 225127,Jiangsu,China)

机构地区：[1]河海大学水利水电学院,南京210098 [2]江苏省水利勘测设计研究院有限公司,江苏扬州225127

出　　处：《安全与环境学报》2023年第12期4557-4568,共12页Journal of Safety and Environment

基　　金：江苏省水利科技项目(2018033)。

摘　　要：为了有效预测河流水质变化趋势,充分利用水质序列的时序性与多元相关性信息,构建基于灰色关联-长短时记忆网络(Grey Relational Anlysis-Long and Short-Term Memory Network,GRA-LSTM)水质预测模型,以改善水环境质量。选择长江南京段河流型水源地进行实例研究,结果显示:当滑动时间窗口(d)为2,最大训练次数(MaxEpochs)为220,隐含层神经元个数(numHiddenUnits)分别为80与100时,总磷与溶解氧预测效果最佳。将长短时记忆网络(Long and Short-Term Memory Network,LSTM)预测结果与误差反向传播神经网络(Back Propagation,BP)、极限学习机(Extreme Learning Machine,ELM)、支持向量机(Support Vector Machine,SVM)模型预测结果进行对比分析,显示LSTM网络在水质预测研究中具备较强的适用性。运用灰色关联分析选择多元特征输入变量,实现了关键水质指标影响因子重要性定量化排序与冗余信息的消除,相较于单一特征输入的LSTM网络,GRA-LSTM网络能够进一步降低模型预测误差,其中总磷与溶解氧质量浓度预测均方根误差分别降低了11.6%与12.4%。The paper intends to establish a water quality prediction model based on the grey relational method and long and short-term memory network(GRA-LSTM),making full use of the time sequence and multivariate correlation information of water quality series,to predict the changing trend of river water quality and improve the water environment quality.The river-type water sources in the Nanjing section of the Yangtze River are selected for the case study.The results show that the prediction performances of total phosphorus and dissolved oxygen are the best when the input size is 2,the maximum number of training(MaxEpochs) is 220,and the number of neurons in the hidden layer(numHiddenUnits) is 80 and 100,respectively.The prediction results are then compared with those of error back propagation neural network(BP),extreme learning machine(ELM),and support vector machine(SVM) models,which verifies that LSTM has strong applicability in water quality prediction.The type of data distribution and the appropriate method for data normalization are identified,which is necessary to eliminate the dimension impact of different water quality indicators.The common parameters are calibrated based on the built-in adaptive moment estimation(ADAM) optimization algorithm,while the super parameters are obtained through sensitivity analysis combined with trial and accumulative experience.Further,the grey relational method is used to select multiple characteristic input variables to achieve the quantitative ranking of the impact factors concerning the key water quality index and the elimination of redundant information,and then the water quality prediction model based on the GRA-LSTM network is obtained through repeated training.Compared with the LSTM network with single characteristic input,the GRA-LSTM network can further improve the model prediction performance,and the root mean square errors of total phosphorus and dissolved oxygen prediction reduce by 11.6% and 12.4%,and the correlation coefficients increase by 4.39% and 1.05%,respectively

关 键 词：环境工程学 水质预测 总磷 溶解氧 长短时记忆网络 灰色关联 长江南京段 

分 类 号：X832[环境科学与工程—环境工程]