基于卷积神经网络和长短期记忆网络的坝上水位精细化建模方法

A Refined Modeling Method for Forebay Water Level Based on Convolutional Neural Network and Long Short-Term Memory Network

摘要: 坝上水位是水电站调度运行的重要依据，然而受调峰非恒定流的影响，传统插值计算的水电站坝上水位与实际值存在较大的误差，不利于水库水位的精细控制和实际调度。本研究采用最大互信息系数探索水电站坝上水位变化的关联因素，并提出一种基于深度学习的CNN-LSTM模型计算方法，实现了准确计算受调峰非恒定流影响的水电站坝上水位。为验证本文所提模型的有效性，将其与传统法在三种评价准则进行对比，结果表明，所提的CNN-LSTM模型在汛期和枯水期的各种评价准则下均优于传统法，模型计算结果更接近实际坝上水位。本文所提模型在水电运行时可有效避免计算水位不准确带来的控制风险，降低水电站运行风险。

Abstract: The forebay water level is an important basis for the scheduling and operation of hydropower plants. However, due to the influence of unsteady flow generated by peak shaving, the forebay water level calcu-lated by the traditional interpolation method has a large error with the actual value, which is not condu-cive to the precise control of the water level and actual scheduling of hydropower plants. The maximum mutual information coefficient was used to explore the relevant factors of the forebay water level, and a deep neural network was designed to accurately calculate the forebay water level affected by peak shaving unsteady flow. To verify the effectiveness of the designed model, it was compared with traditional methods in three evaluation metrics. The results show that the proposed CNN-LSTM model outperforms traditional methods during flood and dry seasons, and the forebay water level obtained by CNN-LSTM is closer to the actual water level than the traditional method. The model proposed in this paper can effectively avoid the control risk caused by inaccurate calculation of the forebay water level and reduce the operation risk of the hydropower plant.