基于BOA-BiLSTM模型的地表水水质预测  

Surface water quality prediction based on BOA-BiLSTM model

作　　者：章佩丽 赵文雅 许旭敏 包鑫磊 ZHANG Peili;ZHAO Wenya;XU Xumin;BAO Xinlei(Taizhou Pollution Control Technology Center Co.Ltd.,Taizhou 318000,Zhejiang Province,China;Key Laboratory of the Eco-environmental Big Data of Taizhou,Taizhou 318000,Zhejiang Province,China;Luqiao Branch of Taizhou Municipal Ecology and Environment Bureau,Taizhou 318000,Zhejiang Province,China;Taizhou Luke Testing Technology Co.,Ltd.,Taizhou 318000,Zhejiang Province,China)

机构地区：[1]台州市污染防治技术中心有限公司,浙江台州318000 [2]台州市生态环境大数据重点实验室,浙江台州318000 [3]台州市生态环境局路桥分局,浙江台州318000 [4]台州市绿科检测技术有限公司,浙江台州318000

出　　处：《浙江大学学报(理学版)》2025年第3期323-333,345,共12页Journal of Zhejiang University(Science Edition)

基　　金：浙江省生态环境科研和成果推广项目(2022HT0010);台州市科技计划项目(22gyb37)。

摘　　要：为准确评估监测条件有限的平原河网小流域河水水质演变趋势,预知水质变化情况,利用浙江省台州市南官河2021年6月至2023年6月的水质监测数据,基于贝叶斯优化算法(Bayesian optimization algorithm,BOA)和双向长短期记忆神经网络(bi-directional long short-term memory,BiLSTM)建立了地表水水质预测模型。利用箱线图和Spearman秩相关系数挖掘水质的时空分布规律,划定中间河段4个站点为重点研究区域,NH3—N和TP为治理重点。通过BOA和双向信息传递机制优化LSTM超参数和模型结构,结果显示,用BOA-BiLSTM模型预测,未来4 h NH_(3)—N浓度的均方根误差(root mean squared error,RMSE)分别为0.2132,0.3689,0.3327和0.3740;未来4 h TP浓度的RMSE分别为0.0246,0.0321,0.0422和0.0334。二者较基准LSTM模型的预测结果分别提升了15.8%,10.6%,10.6%,17.1%和22.6%,3.6%,14.8%,11.8%。以磨石桥NH_(3)—N浓度为例,对比了时序预测与加入上下游数据后的多变量预测结果,发现时序预测对监测参数较少的平原河网具有更强的适用性和更高的预测精度。同时结合研究区域现场勘查和地块分类情况,指出生活源、污水收集及处理设施不完善、雨污合流应为整治重点。当监测参数有限时,本文模型有助于提升对水质异常的监管水平,为环境执法、水环境治理提供数据支撑。To accurately assess the water quality evolution trend of small watersheds in plain river networks with limited monitoring conditions and predict the change of water quality in advance,based on the water quality monitoring data of Nanguan River in Taizhou,Zhejiang province from June 2021 to June 2023,a surface water quality prediction model was established by integrating a Bayesian optimization algorithm(BOA)with a bi-directional long short-term memory network(BiLSTM).The temporal and spatial distribution of water quality was discovered through the use of box plots and Spearman's rank correlation.Consequently,NH3—N and TP were identified as the primary control parameters,and four stations were designated as the primary research areas.The results indicate that the predicted RMSE of NH3—N by the BOA-BiLSTM model for the next four hours is respectively 0.2132,0.3689,0.3327 and 0.3740,the predicted RMSE of TP is respectively 0.0246,0.0321,0.0422 and 0.0334.Compared with the basic LSTM model,the predicted RMSE of NH_3—N in the next four hours is increased by 15.8%,10.6%,10.6%and 17.1%,respectively,the RMSE of forecast TP is increased by 22.6%,3.6%,14.8%and 11.8%respectively.Simultaneously,we take the NH_3—N of Moshi Bridge as an example,it is demonstrated that the time series prediction has higher prediction accuracy and stronger applicability for the plain river network with fewer monitoring parameters comparing with the multi-variable prediction results after adding upstream and downstream data.Finally,based on the on-site investigation and land classification of the study area,it is pointed out that the living source,imperfect sewage collection and treatment facilities,and the combined flow of rain and sewage are the key points of regulation.The established high-precision water quality prediction model and its associated findings provide reliable data to support environmental law enforcement and water environment governance,helping to enhance monitoring capabilities for abnormal water quality within limited

关 键 词：水质预测 平原河网 贝叶斯优化算法 双向长短期记忆神经网络 现场勘查 

分 类 号：X824[环境科学与工程—环境工程]