Abstract: A hot water consumption prediction model based on VMD-GWO-LSTM is proposed to solve the problem of poor stability and large error caused by traditional prediction methods that ignore the timing of water consumption for water storage electric water heaters. VMD decomposed the original time series data to obtain modal components, and GWO optimized the LSTM network parameters for each component to establish an LSTM prediction model. The predicted values of hot water consumption for a certain period in the future were obtained by superposing the results of each prediction component. The prediction results of three typical operating conditions show that the correlation coefficient (R) of the optimized VMD-GWO-LSTM prediction is stable at above 98.60%, and the RMSE decreases by at least 61.7% compared with the prediction of the unoptimized LSTM, and the MAE decreases by at least 51.4%. Compared with the prediction of BP, SVM, GWO-LSTM, and VMD-LSTM, the prediction error is smaller and the stability is better, and the energy loss caused by the deviation in the supply of hot water due to the prediction error is reduced.