基于复合神经网络的多元水质指标预测模型

Multivariate water quality parameter prediction model based on hybrid neural network

长江流域在我国水资源配置体系中具有重要地位,对其进行水质预测尤为重要。基于现有研究结果,结合循环神经网络(recurrent neural network,RNN)中的门控循环单元(gate recurrent unit,GRU)模型与全连接神经网络(fully connected neural network,FCNN),提出了改进的多元水质指标预测(MWQPP)模型,并用其预测长江流域水体的pH、溶解氧(DO)、高锰酸盐指数(CODMn)、氨氮(NH_(3)-N)。基于长江流域2011—2018年23个水质监测点7 566条原始数据,经对比实验,证明了用MWQPP模型预测得到的均方根误差(RMSE)、平均绝对误差(MAE)、平均绝对百分比误差(MAPE)和决定系数(R^(2))均优于传统水质预测模型,有效提升了水质预测的精度,具有较好的鲁棒性,为水质预测和流域管理提供了科学支撑。

The Yangtze River basin plays an important role in Chinese water resources allocation.What proves common knowledge is that it is particularly important to predict the water quality in the Yangtze River basin.Based on the existing research,the recurrent neural network(RNN) model with gate recurrent unit(GRU) and fully connected neural network(FCNN) are combined in this study to improve a multiple water quality parameter prediction(MWQPP) model.It is proposed to predict the four water quality parameters,such as pH,dissolved oxygen(DO),permanganate index(CODMn) and ammonia nitrogen(NH_(3)-N) in the Yangtze River basin.Based on 7 566 raw data of 23 water quality monitoring points in the Yangtze River basin from 2011 to 2018,the comparative experiments show that the root mean square error(RMSE),mean absolute error(MAE),mean absolute percentage error(MAPE) and coefficient of determination(R^(2)) obtained from the MWQPP model’s prediction results are better than traditional models,such as the multiple linear regression model,the random forest model,FCNN model and LSTM model,and the MWQPP model also has better robustness than these traditional water quality prediction models.As we can say,the MWQPP model can provide scientific,reasonable and effective support for water quality assurance and water management in Yangtze River basin.