Abstract: The maximum and minimum water level is an important constraint to be considered in the calculation of cascade reservoir operation and economic operation of hydropower station. The commonly used iteration for multi period prediction results has low credibility. This article selects a long-short term memory model (LSTM) that has good processing effects on time series problems and improves its parameter calibration ability. It is applied to the prediction of the maximum and minimum water levels of the Three Gorges Reservoir in the next four days. The traditional forecast model is constructed according to the water balance forecast framework. Two kinds of deep learning models were built based on LSTM model with different characteristic variables. The calculation results show that the deep learning model prediction considering the propagation law of water surface profiles in the Three Gorges Reservoir area has accurate and stable prediction effects, Make the absolute error of 99% prediction less than 40cm.

Key words: economic operation in hydropower station, water level prediction, LSTM, deep learning, neural network