This paper proposes a general Multiple-Model Predictor frame-work for forecasting time series that presents multiple patterns (behaviours). The proposed approach allows off-line identifying the different time-series behaviours, training a model for each behaviour and identifing on-line which is the pattern and associated model to use at each time instant. A particular case based on K-means and Radial Basis Functions (RBF) is proposed to exemplify a possible implementation. To illustrate the use and performance of the proposed approach, an application of the short-term water demand forecast is presented using demand data from the Barcelona drinking water network.