In order to improve large-scale water consumption prediction models for mainland China and provide a technical support for the scientific utilization of water resources, the annual total water consumption prediction models of provincial administrative divisions, river basins and geographical regions were established based on the principles of ARMA, gray GM (1,1) and BP neural network model. The optimized results of the models were statistically analyzed, and the selected optimal models were used to predict the total water consumption from 2021 to 2025. The results show that for the provincial administrative division scale, the optimal prediction model of total annual water consumption is ARMA model in nine administrative divisions, gray GM (1,1) model in six administrative divisions, and neural network model in 16 administrative divisions. At the river basin scale, the optimal prediction model is ARMA model in five basins, gray GM (1,1) model in three basins, and neural network model in the Yangtze Basin. For the scale of large geographical regions, the optimal model of the six northern regions is the neural network model, and that of the four southern regions is the gray GM (1,1) model. The water consumption in the next five years will be generally stable. These findings are expected to provide some statistical reference for total water consumption management in China.