In this paper a comparison among six short-term water demand forecasting models is presented. The models differ in terms of forecasting technique, type of prevision (deterministic or probabilistic) and data requirement for calibration. In particular, the compared models are: an Artificial Neural Network based model, a model based on periodic patterns, both requiring a calibration over a year of historically observed water demands, two models that take into account the periodic behaviours using observed data only on a restricted time window preceding the time of forecast, a probabilistic model based on Markov chain and a Naïve model. All the models are evaluated applying them to seven real-life case studies, consisting in two-year time series of hourly water demands observed in districts/networks with a number of users variable from 120000 to 300. The comparison shows that all the models provide similar and medium-high forecasting accuracy for each case study but the models based on the moving-window technique are more robust and their performances do not worse moving from the calibration to the validation period as for all the other models considered.