基于GAT-BILSTM-Res的水质预测模型  

Water Quality Prediction Model Based on GAT-BILSTM-Res

作　　者：杨振舰 庞瑛 YANG Zhenjian;PANG Ying(School of Computer and Information Engineering,TCU,Tianjin 300384,China)

机构地区：[1]天津城建大学计算机与信息工程学院,天津300384

出　　处：《天津城建大学学报》2024年第1期60-65,共6页Journal of Tianjin Chengjian University

基　　金：天津市科技计划项目(22YDTPJC00840)。

摘　　要：针对水质数据在时间维度的依赖关系以及水质监测站点在空间维度的依赖关系,基于海河流域天津段实际监测的历史水质数据,设计了有效提取时空特征的方法,提出一种融合图注意力网络(GAT)、双向长短期记忆网络(Bi-LSTM)以及残差块(ResBlock)的时空水质预测模型(GAT-BILSTM-Res).该模型首先通过GAT捕获水质监测站点之间的拓扑关系,建立空间相关性模型;同时通过Bi-LSTM捕捉水质监测数据的动态变化,并对时间相关性进行建模;然后将时空特征融合,输入残差块;最后使用全连接层对预测结果进行输出.实验结果表明,相较于基线模型,该模型能够实现6.6%~25.2%的性能提升.For the dependence of water quality data in the time dimension and the dependence of water quality monitoring stations in the spatial dimension,this paper is based on the actual monitoring of historical water quality data in the Tianjin section of the Haihe River basin.It designs a method to effectively extract spatio-temporal characteristics,and proposes a spatio-temporal water quality model(GAT-BILSTM-Res)that combines graph attention network(GAT),bi-directional long and short-term memory network(Bi-LSTM)and residual block(ResBlock).The model first captures the topological relationship between water quality monitoring stations through GAT and establishes a spatial correlation model;at the same time,the dynamic changes in water quality monitoring data are captured through Bi-LSTM,and the temporal correlation is modeled.Then the spatio-temporal features are fused and input into the residual block.Finally,the prediction results are output by using the fully connected layer.The experimental results show that the model is able to achieve a performance improvement of 6.6%~25.2%compared with the baseline model.

关 键 词：水质预测 图注意力网络 双向长短时记忆网络 残差块 

分 类 号：X52[环境科学与工程—环境工程]