基于注意力机制的LSTM长江汛期水位预测方法研究  

An Attention Mechanism-Based LSTM Method for Predicting Water Level of the Yangtze River in Flood Season

作　　者：王迎飞 黄应平[2,3,4] 肖敏[1,2,4] 熊彪[1,2] 周爽爽[1,2] 靳专 WANG Yingfei;HUANG Yingping;XIAO Min;XIONG Biao;ZHOU Shuangshuang;JIN Zhuan(College of Computer&Information Technology,China Three Gorges Univ.,Yichang 443002,China;Hubei Engineering Technology Research Center for Farmland Environment Monitoring,China Three Gorges Univ.,Yichang 443002,China;College of Hydraulic&Environmental Engineering,China Three Gorges Univ.,Yichang 443002,China;Engineering Research Center of Eco-Environment in Three Gorges Reservoir Region of Ministry of Education,China Three Gorges Univ.,Yichang,443002,China)

机构地区：[1]三峡大学计算机与信息学院,湖北宜昌443002 [2]湖北省农田环境监测工程技术研究中心(三峡大学),湖北宜昌443002 [3]三峡大学水利与环境学院,湖北宜昌443002 [4]三峡库区生态环境教育部工程研究中心(三峡大学),湖北宜昌443002

出　　处：《三峡大学学报（自然科学版）》2022年第3期13-19,共7页Journal of China Three Gorges University：Natural Sciences

基　　金：国家自然科学基金(22136003);湖北省引智项目(2019BJH004);长江水环境教育部重点实验室(同济大学)开放基金项目(YRWEF202103);宜昌市自然科学研究项目(A22-3-005)。

摘　　要：为提高长短期记忆网络(Long Short-Term Memory,LSTM)在水位预测任务中的准确性,以及提高LSTM对数据中时空信息的利用率,本文提出了一种基于Softmax函数的注意力模块,并将其应用在LSTM的输入前,使模型可以根据输入数据中的时间和空间信息,自主地生成带有权重的词义向量,并赋予输入序列时间和空间注意力权重值.注意力机制增强了LSTM模型对时间序列的处理能力,同时使LSTM模型能够对数据空间信息具备一定的应对能力.最后在长江汛期的水位数据上进行预测试验,以E_(MAE)、E_(RMSE)、E_(MAPE)和R^(2)作为评价指标,将所提出的模型与原始LSTM模型进对比,同时分析了权重分布情况.结果表明,该模型对水位预测的精度有明显提升,应用了时空注意力机制的LSTM(AT-LSTM)相比于原始LSTM,E_(MAE)、E_(RMSE)和E_(MAPE)分别降低了21.77%、31.15%和17.4%,R^(2)提高了9.69%.To improve the accuracy of water level prediction in long short-term memory network(LSTM),this paper proposes an attention module based on Softmax function and applies it before the input of LSTM,to give higher weight to the data in time and space dimensions for predicting more valuable information.The attention mechanism enhances the time series processing capability of the LSTM model,and it enables the LSTM model to have some response capability to the data space information.Finally,the prediction test is carried out on the water level data of the Yangtze River in flood season.E_(MAE),E_(RMSE),E MAPE and R^(2) are used as evaluation indicators,the proposed model is compared with the original LSTM model,and the weight distribution is analyzed.The results show that,the model has a significant improvement in the accuracy of water level prediction.Compared with the original LSTM,the LSTM with the spatio-temporal attention mechanism(AT-LSTM)applied reduces the E_(MAE),E_(RMSE)and E MAPE by 21.77%,31.15%and 17.4%,respectively,and improves the R^(2)by 9.69%.

关 键 词：注意力机制 水位预测 长短期记忆网络 长江 

分 类 号：TV697.2[水利工程—水利水电工程]