摘要
长短时记忆神经网络(LSTM)具有很强的时间序列关系拟合能力,非常适用于模拟及预报流域产汇流这一复杂的时间序列过程。基于LSTM针对不同预见期的径流预报分别建立了流域降雨径流模型,以探讨LSTM在水文预报当中的应用。模型采用流域降雨、气象及水文数据作为输入,不同预见期后的径流过程作为输出,率定期为14a,验证期为2a。结果显示,在预见期为0~2d时LSTM预报精度很高,在预见期为3d时预报精度较差,但仍优于新安江模型。隐藏层神经元数量作为神经网络复杂程度的代表既会影响模型预报精度,也会影响模型训练速度。而输入数据长度则仅会在预见期为0的条件下影响模型预报效果。
The Long Short-Term Memory(LSTM)is suitable for rainfall-runoff modelling and forecasting since it has a strong ability in fitting time series.In this study,the LSTM was employed in predicting runoff in different foresight periods,in order to assess the capability of the LSTM in rainfall-runoff modelling and forecasting.The historical precipitation,meteorological and hydrological data were used as input data,runoff at after different foresight periods were selected as model output.The calibration period is 14 years and the validation period is 2 years.As expected,the proposed model shows a great ability to predict runoff 0-2 days ahead.With 3 days of foresight period,the LSTM performs relatively poor but still better than the Xinanjiang model.The number of hidden nodes has a primary impact on the prediction accuracy and training efficiency.While the length of input data has an impact on model performance only when the foresight period is 0 day.
作者
殷兆凯
廖卫红
王若佳
雷晓辉
YIN Zhaokai;LIAO Weihong;WANG Ruojia;LEI Xiaohui(State Key Laboratory of Hydraulic Engineering Simulation and Safety,Tianjim University,Tianjin 300072,China;State Key Laboratory of Simulation and Regulation of Water Cycle in River Basin,China Institute of Water Resources and Hydroporwer Research,Bejjing 100038,China;Department of Information Management,Peking University,Beijing 100871,China;Institute of Ocean Research,Peking University,Beijing 100871,China)
出处
《南水北调与水利科技》
CAS
北大核心
2019年第6期1-9,27,共10页
South-to-North Water Transfers and Water Science & Technology
基金
“十三五”国家重点研发计划(2017YFB0203104)
国家自然科学基金(51709273)
广东省水利科技创新项目(2017-06)~~
关键词
降雨径流模拟
水文预报
机器学习
深度学习
长短时记忆
Rainfall-runoff modelling
Hydrological forecast
Machine learning
Deep learning
LSTM
