In order to effectively improve the short-term water demand forecasting model of water distribution networks in terms of prediction accuracy and stability, a novel combined prediction modeling method which can predict water demand and residuals simultaneously is proposed in this paper. First, the Bayesian least squares support vector machine method (Bayesian-LSSVM) was used to establish the time series prediction model of user’s water demand (BL model) to obtain the initial water demand prediction values. Then, to predict the residual sequence of the initial water demand prediction values produced by the BL model, a chaotic time series prediction model (RM model) was constructed based on the Bayesian-LSSVM method. At the same time, the predicted residuals produced by the RM model were compensated to the water demand predictions by the BL model to correct the initial water demand prediction values. Results of the case study show that the RM model could accurately capture the change trend of the residual value of the BL model initial prediction values and the residual sequence of the BL model initial water demand prediction values. The combined forecast model consisting of the BL model and the RM model (BL+RM model) was superior to single BL model in the accuracy and stability of short-term water demand forecasting. BL+RM model was applicable for short-term water demand forecasting with different water demand characteristics such as small average water demand and large water volatility, and hence could effectively meet the needs of actual engineering.