基于物理信息神经网络的水库调度研究  

Research on Reservoir Operation Based on Physical Information Neural Networks

作　　者：王飞龙 胡挺 肖扬帆 WANG Fei-long;HU Ting;XIAO Yang-fan(China Three Gorges Corporation,Yichang 430010,China)

机构地区：[1]中国长江三峡集团有限公司,湖北宜昌430010

出　　处：《水电能源科学》2025年第4期213-216,共4页Water Resources and Power

基　　金：国家重点研发计划(2021YFC3200303);国家自然科学基金项目(U2340205)。

摘　　要：传统的水库调度智能模型多是以前馈神经网络或循环神经网络为代表的黑箱模型,模型的可解释性较差,缺乏内部规律的探索。因此,基于物理神经网络基本原理,结合水库调度方程,以出库流量误差及出力误差之和为模型总误差,建立了包含水库调度方程的物理机制神经网络,并以向家坝水库为例,对比了不含水库调度方程的神经网络,讨论了模型隐藏层数及隐藏层神经元个数对模型精度的影响。结果表明,包含水库调度方程的神经网络模型验证效果优于一般前馈神经网络,验证误差率为3%,隐藏层数和神经元个数太少会导致模拟效果较差,而太多的隐藏层数和神经元个数可能对模型精度提升不大,选择合适的隐藏层数和神经元个数是提高模型精度的手段之一。The traditional intelligent model of reservoir operation is mostly a black box model represented by feedforward neural network or recurrent neural network.The interpretability of the model is poor and lacks the exploration of internal laws.In view of this,based on the basic principles of physical neural networks,combined with the reservoir operation equation,this paper takes the sum of the outflow error and the output error as the total error of the model,and establishes a physical mechanism neural network containing the reservoir operation equation.Taking Xiangjiaba Reservoir as an example,the neural network without the reservoir operation equation is compared,and the impact of the number of hidden layers and the number of hidden layer neurons on the accuracy of the model is discussed.The results show that the verification effect of the neural network model containing the reservoir operation equation is better than that of the general feedforward neural network,and the verification error rate is 3%.Too few hidden layers and the number of neurons will lead to poor simulation results,while too many hidden layers and the number of neurons may not improve the accuracy of the model.Choosing the appropriate hidden layers and the number of neurons is one of the means to improve the accuracy of the model.

关 键 词：水库调度 物理约束 物理信息神经网络 前馈神经网络 PDE 

分 类 号：TV697.11[水利工程—水利水电工程]