Water resource shortage is a realistic problem faced by water supply systems in river basins. Forecasting water demand is crucial for sustainable management of water supply systems. In this paper, the ARMA—DNN model is established to predict the water demand of the basin by combining the deep neural network (DNN) and the autoregressed-moving average (ARMA) mixed model. Taking the economic growth water demand and the actual social water demand as the main prediction targets, a mixed prediction model based on 14 statistical indicators is built. The model uses data from 2010 to 2020 to forecast water demand in the Minjiang River Basin. The results show that :(a) the model can accurately predict the future water consumption of the basin under the condition of actual water consumption changes; (b) The forecast of future water consumption has a significant impact on agricultural grain yield, industrial economic output value and domestic water satisfaction. In each region of the basin, agricultural grain yield and industrial economic output value and domestic water satisfaction are mutually restricted; (c) When climate conditions deteriorate and water shortages become severe, effective water demand forecasting can alleviate water demand contradictions to some extent. In a word, watershed managers need to make industrial water allocation schemes in different regions based on the forecast results of future water consumption, so as to balance the relationship among agricultural and food output, economic output value and domestic water satisfaction.