资料长度对深度学习方法日径流预报效率的影响

Influences of data length on efficiency of daily runoff forecast by deep learning method

长短期记忆(Long Short-Term Memory,LSTM)神经网络深度学习方法具有显著的时序分析能力,在径流预报方面有其独特的优势,但该模型预报的最优输入输出长度组合尚不太明确,探析不同输入输出长度对LSTM日径流预报效率的影响对相关应用具有实际意义。以四川省西部大渡河、雅砻江、岷江支流以及嘉陵江上游等流域为研究区,选取了20个子/区间流域,试验不同长度的前期输入资料预报不同预见期下的径流,研究了不同资料长度下LSTM模型的日径流预报效率,分析了该方法在不同流域的适用性与最优输入输出长度的特征。结果表明:①以前期降水、气温以及径流作为输入,前期资料长度对预报结果影响不太明显,但预报准确性会随预见期延长而下降,采用该种输入方案的预见期不宜超过7 d;②仅以前期降水、气温资料作为输入,预报准确性会随前期资料长度增加而提高,也会随预见期的延长而下降,建议该种方案的资料长度大于7 d、预见期最好为1 d,不宜超过3 d;③径流变异性是显著影响预报效率和最优输入输出长度组合的重要因子,变异性强,则预报效果较差,对输入输出长度的敏感性偏弱。研究成果可为提高深度学习径流预报效率提供参考,有助于结合流域特性确定适用的输入输出长度组合方案。

Long Short-Term Memory(LSTM)neural network model,a deep learning method with strong capability of temporal series analysis,has unique advantages in runoff prediction.However,the optimal scheme of input and output lengths in this model is still not clear,so it is of practical significance to explore the influence of different input and output lengths on daily runoff prediction efficiency by LSTM.Taking Dadu River,Yalong River,tributaries of Minjiang River and upper reaches of Jialing River in western Sichuan Province as the study area,20 sub-catchments were selected to test the daily runoff in different forecast periods with input data of different lengths and the daily runoff prediction efficiency of LSTM model under different data lengths was studied.The applicability of this method in different river basins and the characteristics of optimal input and output length were analyzed.The results show that:①When the previous precipitation,temperature and runoff are taken as inputs,the input length has little effect on daily runoff prediction,but the accuracy will decrease with the extension of forecast period.Therefore,the forecast period should be set within 7 days to guarantee forecast accuracy.②When only previous precipitation and temperature are taken as inputs,the forecasting accuracy will increase with the extension of previous data and decrease with the extension of forecast period.Therefore,the length of preliminary data should be beyond 7 days,and the forecast period is preferably 1 day,and should not exceed 3 days.③Runoff variability is an important indicator affecting the prediction efficiency and the optimal combination of input and output lengths.The prediction results in catchments with strong variability show low accuracy and weak sensitivity to input and output lengths.The research results can provide a reference for improving the runoff prediction efficiency by deep learning method,and help to determine the suitable input and output length combination scheme in consideration of watershed characteristics.