基于JMI-CNN-LSTM耦合模型的梯级水电站间流量动态滞时关系  

Study on dynamic flow lag time between cascade hydropower stations based on CNN-LSTM coupling model

作　　者：闫孟婷 黄炜斌[1] 张天遥 马光文[1] 赵丽伟 YAN Mengting;HUANG Weibin;ZHANG Tianyao;MA Guangwen;ZHAO Liwei(College of Water Resource&Hydropower,Sichuan University,Chengdu 610065,Sichuan,China)

机构地区：[1]四川大学水利水电学院,四川成都610065

出　　处：《水利水电技术（中英文）》2023年第3期154-164,共11页Water Resources and Hydropower Engineering

基　　金：国家重点研发计划(2016YFC0402208);国家重点研发计划(2018YFB0905204)。

摘　　要：【目的】梯级水电站间存在的水力联系导致下游电站的运营模式往往受制于上游电站,以往的梯级水电站优化调度中通常选择忽略流量滞时或认为其是常数,少有考虑流量滞时与上游水电站出库流量及区间降雨等因素的动态关系。为提高下游电站预判流量精确度,将神经网络应用到梯级电站间流量动态滞时研究中。【方法】首先采用联合互信息理论选取下游电站入库流量的主要影响因素作为模型输入因子,其次根据卷积神经网络和长短时记忆神经网络的互补特性,建立上游出库流量与下游入库流量的JMI-CNN-LSTM深度学习网络模型,最后结合实际算例,将所建立模型的拟合结果与随机森林回归模型、固定滞时模型进行对比。【结果】结果显示:本文所建立的模型较相同条件下其他方法各类误差均存在不同程度的减少,其中MAE至少减少了14.6%。【结论】结果表明:相较其他方法,JMI-CNN-LSTM耦合模型预测精度更佳,能够更准确的体现梯级电站间流量滞时的动态关系。[Objective]The hydraulic connection between cascade hydropower stations leads to the operation mode of downstream hydropower stations is often subject to the upstream hydropower stations.the optimal operation of cascade hydropower stations in the past usually chooses to ignore the flow lag time or consider it as a constant,rarely consider the dynamic relationship between the flow delay and the upstream hydropower station outbound flow and interval rainfall,.In order to improve the accuracy of flow prediction for downstream power stations,the neural network is applied to the study of flow dynamic lag between cascade power stations.[Methods]Firstly,the joint mutual information theory was used to select the main influencing factors of the inflow of downstream power stations as the input factors of the model,then according to the complementary characteristics of the convolutional neural network and the long short-term memory neural network,the JMI-CNN-LSTM deep learning network model of upstream hydropower station outflow and downstream hydropower station inbound flow is established,finally the model fitting results are compared with the random forest regression model and the fixed lag time model base on the actual case study.[Results]The results show that compared with other methods under the same conditions,the errors of the model established in this paper are reduced to different degrees,and the MAE is reduced by at least 14.6%.[Conclusion]The results show that compared with other methods,JMI-CNN-LSTM coupling model has better prediction accuracy and can more accurately reflect the dynamic relationship of flow lag time between cascade power stations.

关 键 词：梯级水电站 动态滞时 联合互信息 卷积神经网络 长短期记忆网络 

分 类 号：TV697.1[水利工程—水利水电工程] TV76