Abstract: The pre-dam water level of the Three Gorges Reservoir is affected by many factors, the change mechanism is so complex to be accurately predicted by the traditional water balance method.Based on the historical operation data of the reservoir, this study analyzes the law of pre-dam water level change, and uses three machine learning methods, namely artificial neural network, support vector machine and random forest, to construct an hourly pre-dam water level prediction model, and evaluates the prediction effect of the model.The test results show that all three prediction models exhibit high prediction accuracy, among which the random forest model performs best in the accuracy evaluation.The mean square error of k-fold cross validation is 5.2,and the average value of R2 is 0.82.It has obvious advantages over the traditional water balance method in three typical test cases with different peak shaving amounts.The research results can provide technical support for the precise operation of short-term power generation in reservoirs.