长短期记忆网络在中长期径流预报中的应用

Application of long-short term memory network in medium-long term runoff forecast

长短期记忆神经网络(long-short term memory neural network,LSTM)是一种适合模拟与预测时间序列长程变化的神经网络,可用于研究中长期水文预报问题。以黄河源区唐乃亥流域为研究区域,根据2000-2009年的日降雨量、蒸发量和流量资料,构建了未来30 d的逐日流量LSTM预报模型,并以2010-2012年的资料进行验证。采用相对误差绝对值平均、确定性系数为评价指标,将LSTM与传统的误差反向传播(back propagation,BP)算法模型进行对比。结果表明,LSTM模型具有精度高、预报效果稳定等特点,特别是随着预见期的增长其精度衰减较BP模型明显变缓。研究成果可供研究区径流中长期预测提供参考。

The long-short term memory neural network(LSTM)is a kind of neural network suitable for simulating and predicting the long-range variation of time series,so it can be used to study the medium-long term hydrological forecasting.Based on the daily rainfall,evaporation and discharge data from 2000 to 2009,a daily discharge LSTM prediction model for the next 30 days is constructed in Tangnaihai basin,the source area of the Yellow River,and verified by the data from 2010 to 2012.The mean absolute value of relative error and the coefficients of determination are used as evaluation indexes to compare LSTM with traditional back propagation(BP)model.The results show that LSTM model has the characteristics of high precision and stable prediction effect,especially with the increase of prediction period,the precision attenuation of LSTM model becomes slower than that of BP model.The research results can provide a reference for the medium-long term runoff prediction in the study area.