Water distribution network models are widely used by water companies. Consumer demands are one of the main uncertainties in these models, but their calibration is not feasible due to the low number of sensors available in most real networks. However, the behaviour of these individual demands can be also calibrated if some a priori information is available. A methodology for calibrating demand patterns based on singular value decomposition (SVD) is presented. Demand stochastic nature is overcome by using multiple data samples. The methodology is applied to two water distribution systems: an academic network and a real network with synthetic data.