The water demand data of various water use industries in Zhejiang Province from 2000 to 2020 were collected, and the main driving factors affecting the water demand of each industry were screened using the main driving factor screening method based on Spearman rank correlation analysis. An improved long short-term memory (LSTM) neural network water demand prediction model was constructed to make dynamic rolling forecasts of the water demand of each industry, and the prediction results of the improved LSTM model were compared with those of the traditional univariate LSTM prediction model, convolutional neural network (CNN) model, and support vector regression (SVR) model. The results show that the LSTM model improved by the principal driving factor screening method can predict the annual water demand of each industry in real time and dynamically, and the prediction accuracy of the improved model is higher than that of the other three models.