Application of Bayesian neural network to prediction of urban short-term water consumption

Under the background of implementation of the most stringent management regulations on water resources, the prediction of short-term water consumption is playing an increasingly significant role in urban water supply system scheduling. Based on the analysis of the temporal evolution pattern and random factors of short-term water consumption, a Bayesian neural network prediction model for urban short-term water consumption was built with the daily maximum temperature, daily water consumption of the previous 7 days, ratio of water consumption of the current month to the annual amount, daily precipitation, and holidays as predictors of short-term water consumption. Meanwhile, Bayesian regularization was used to optimize BP neural network. Both BP network model and the optimized model were applied to a running-water company in Guangzhou City for tesing. The results indicated that the mean absolute percentage error of the Bayesian neural network prediction model was 0.87%, while that of the BP neural network prediction model was 1.85%. Compared to the BP neural network prediction model, the optimized model has stronger generalization ability, with accuracy improved by about 0.98%. Thus, it fits better with the high-precision requirement of urban short-term water prediction.

Project Supported:

Clc Number: