净水技术2025，Vol.44Issue(5)：57-68,186,13.DOI:10.15890/j.cnki.jsjs.2025.05.007
基于STL-MIKE-LSTM联合建模的陈行水库盐度快速预测
Rapid Prediction of Salinity for Chenhang Reservoir Based on Joint Modeling of STL-MIKE-LSTM
摘要
Abstract
[Objective]The study aims to probe into the feasibility of substituting neural network models for the mechanism models based on nonlinear hydrological and water quality processes for lakes and reservoirs,so that the output water quality under different water intake conditions can be calculated rapidly during salt tide.[Methods]The study proposed a novel method for predicting chloride output water concentration for lakes and reservoirs.In this method,seasonal-trend decomposition using LOESS(STL)was used first to extract the characteristics of intake water concentration data.Then,data augmentation based on Gaussian noise was applied to synthesize simulated data.After that,simulated operation conditions were constructed and calculated by MIKE 21 mechanism model.Finally,the calculation result were used to train an long-short term memory(LSTM)neural network model.[Results]The method proposed was applied to Chenhang Reservoir and it turned out that:(1)By comparing the result of STL decomposition under different parameters,it turned out that choosing periodicity parameter(np)=12 could lead to a better performance;(2)By comparing the prediction performances of the LSTM model under different numbers of hidden layer neurons and prediction time steps,it was found that the prediction performance went up at first and then went down as the number of neurons increased,and the performance continued to go down as the prediction time step increase.The result showed that selecting 128 neurons and a prediction time step of 24 hours had the best overall performance;(3)By comparing the prediction performances of neural networks with different structures,it was found that the LSTM performed the best in the prediction set[root mean square error(RMSE)was 0.13 mg/L,mean relative error(MRE)was 0.04,Nash-sutcliffe efficiency coefficient(NSE)was 0.96)];(4)According to the evaluation of established LSTM model using online monitoring and forecast data,it was demonstrated that the LSTM model indeed had high accuracy in predicting the outlet concentration(RMSE was 0.29 mg/L,MRE was 0.09,NSE was 0.58),and the computility and time required were far lower than those of the MIKE 21 mechanism model.[Conclusion]The method to predict the delivery water concentration of chloride that is proposed in this study has been verified to have both high computational accuracy and speed,and can replace the mechanism model to provide rapid decision-making support for reservoir managers in response to salt tide.关键词
咸潮/长短期记忆(LSTM)/周期趋势分解算法(STL)/数据增强/MIKE 21Key words
salt tide/long-short term memory(LSTM)/seasonal-trend decomposition using LOESS(STL)/data augmentation/MIKE 21分类
环境科学引用本文复制引用
吴畅,崔婧嫄,黄帆,宋辰煜,张晟,赵蓬勃,张海平..基于STL-MIKE-LSTM联合建模的陈行水库盐度快速预测[J].净水技术,2025,44(5):57-68,186,13.基金项目
上海城投水务(集团)有限公司科研项目:基于在线数据联动的陈行水源地水质预报及应对措施研究(KYYS220001) （集团）
上海市级科技重大专项:人工智能基础理论与关键核心技术(2021SHZDZX0100) （2021SHZDZX0100）
