Downstream Water Level Prediction of Reservoir based on Convolutional Neural Network and Long Short-Term Memory Network

Abstract

Accurate calculation of power generation output is crucial to the operation and management of reservoir. The calculation of power generation output is related to the downstream water level, which usually is obtained by interpolation of discharge flow. However, the interpolation method has a large error and adversely affects the output calculation, especially for medium and low water head reservoirs. This study explored the relevant factors of the downstream water level and accurately predicted it from historical operational data. The maximal information coefficient and feature combination were used to select feature inputs, and a deep neural network was designed based on a convolutional neural network and a long short-term memory network to predict the downstream water level of a reservoir. To verify the performance of designed model, it was compared with the interpolation method and 4 state-of-the-art prediction methods using 12 validation sets of Gezhouba Reservoir. The experimental results showed that downstream water level obtained by the designed model was closer to the actual water level than was the interpolated water level. Compared with four state-of-the-art prediction methods, the designed method also was very competitive. Finally, the influence of CNNLSTM on power generation output is compared with traditional interpolation method. The comparison results showed that the convolutional neural network–long short-term memory network method reduced the influence of the interpolation method by 92.74% on average.