Short-term water demand forecasting is a required procedure for the optimal real-time control of water distribution systems when, for example, reducing operational costs associated with pumping. The forecasting literature presents a variety of methods ranging from linear seasonal ARIMA models to non-linear support vector regression to black box artificial neural networks. In general, hourly water demand time series are characterized by a strong double seasonal dependency and a non-linear temporal correlation. Therefore, in order to obtain the most suitable representation one needs to select a time series model that respects these features. This paper intends to evaluate the k-nearest neighbor (kNN) approach to forecast the short-term water demand time series. The kNN approach is a pattern recognition algorithm where the forecasted values are directly determined by the most similar past observations. To assess performance, robust evaluation criteria such as reliability and sharpness will be used to evaluate the kNN approach compared with the traditional seasonal ARMA model.