结合退水曲线和长短期记忆网络的中长期低流量预报  

A low flow forecasting model based on recession curve and long short-term memory(LSTM)network

作　　者：周俸嘉 杨汉波[1] 董宁澎 ZHOU Fengjia;YANG Hanbo;DONG Ningpeng(Key Laboratory of Hydrosphere Sciences of the Ministry of Water Resources,Department of Hydraulic Engineering,Tsinghua University,Beijing 100084,China;State Key Laboratory of Simulation and Regulation of Water Cycle in River Basin,China Institute of Water Resources and Hydropower Research,Beijing 100038,China)

机构地区：[1]水利部水圈科学重点实验室清华大学水利水电工程系,北京100084 [2]中国水利水电科学研究院流域水循环模拟与调控国家重点实验室,北京100038

出　　处：《水利水电技术（中英文）》2024年第9期99-107,共9页Water Resources and Hydropower Engineering

基　　金：国家重点研发计划(2021YFC3000202)。

摘　　要：【目的】长短期记忆网络(LSTM)在水文预报研究中显示出较强的预报能力,但通常依赖于大量数据的训练。为使LSTM模型更好的适用于数据量较少的流域。【方法】研究采用退水曲线,对LSTM模型施加物理约束,提出了适用于低流量预报的混合模型。【结果】在中国西南不同地区3个流域的应用表明:(1)随着预见期的增长,混合模型预报结果的合格率有轻微下降,预见期10 d以内准确率可达到90%以上;(2)混合模型的预报精度显著高于LSTM,且能够显著降低误差累积效应的影响;(3)混合模型在减少训练样本数和减少预报因子维数的情况下均优于LSTM模型。【结论】结果表明,引入退水曲线可以降低混合模型对训练数据量的要求,有效延长预见期,对深度学习预报低流量提供了新的改进思路,并可以为抗旱方案设计等提供技术支持。[Objective]Long Short-Term Memory(LSTM)networks have shown strong forecasting capabilities in hydrological research,but they typically rely on a large amount of training data.In order to better adapt LSTM models to watersheds with limited data and introduce some physical mechanisms during the forecasting process,[Methods]this study applies recession curves to impose physical constraints on LSTM models,proposes a hybrid model for low flow prediction.[Results]The hybrid model is tested in three different watersheds in southwest China.The results are as follows:(1)as the forecast horizon increases,the accuracy of the hybrid model slightly decreases,but the accuracy can exceed 90%for a forecast horizon within 10 days;(2)the hybrid model significantly outperforms LSTM in terms of prediction accuracy and mitigates the effects of error accumulation;(3)the hybrid model performs better than LSTM when reducing the number of training samples and the dimensions of prediction factors.[Conclusion]The results indicate that the introduction of recession curves can reduce the training data requirement of the hybrid model,extend the forecast horizon,which can provide a new approach for deep learning in low flow prediction,and offer technical support for drought mitigation planning and other related fields.

关 键 词：基流 退水曲线 LSTM 低流量预报 神经网络 水资源 径流 

分 类 号：P338.3[天文地球—水文科学]