Abstract: The time series of water demand can be decomposed into several stationary detailed time series and a tendency time series according to the algorithm of this multi-scales in this paper.Decomposed time series are forecasted with BP neural network to obtain the prediction series.Then the forecasting results are reconstructed by wavelet theory.So，the forecasting result is gained.An example of water demand of Shihezi in Xinjiang Province is used to testify the feasibility of the new method.The results show that the method of coupling multi-scale decomposition and BP neural network has advantages over the traditional BP neural network in predicted qualification-rate，and has feasibility in forecasting of time series.