Abstract: To reduce the prediction errors caused by the nonlinear, complex, and unstable characteristics of water surface evaporation, a VMD-LSTM water surface evaporation prediction model was introduced, employing a "decomposition-prediction-reconstruction" framework.This model integrated Long Short-Term Memory(LSTM) neural network with Variational Mode Decomposition(VMD).The VMD technique was utilized to decompose water surface evaporation and its primary influencing factors into an equivalent number of sub-mode components, thereby addressing issues of data non-stationarity.These sub-mode components served as inputs for the LSTM neural network, then we developed a hybrid deep neural network model named VMD-LSTM,which was applied to forecast monthly water surface evaporation at the Badong Station in the Three Gorges Reservoir.The results indicated that the VMD-LSTM model outperforms alternative models in terms of prediction accuracy and peak-valley fitting.Compared to a standalone LSTM model, the calibration period root mean square error(RMSE) and mean absolute percentage error(MAPE) was reduced by 54% and 48%,respectively, while the Nash Efficiency Coefficient(NSE) demonstrated an increase of 11%.Although the model's predictive accuracy diminishes with an extended forecast period, it maintains satisfactory performance.Specifically, as the forecast period extends from one month to seven months, the calibration period NSE decreases from 0.97 to 0.84.The research findings can provide valuable theoretical insights for the effective utilization and scientific management of water resources in the Three Gorges Reservoir.