Application of Bayesian Networks to the Forecasting of Daily Water Demand

In this paper, we investigate the application of Bayesian Networks (BN) for the 1-step ahead forecasting of dailywater demand. The water demand time series is associated with the series containing information for daily precipitation and mean temperature that play the role of the additional explanatory variables. To enable the application of the standard Bayesian network as a predictive model, all considered time series are discretized. The number of discretization intervals is assumed as a parameter of the following learn-and-test trials. To test forecasting accuracy, we propose a novel discrete type of mean absolute error measure. Then, the concept of growing window is used to learn and test several Bayesian networks. For comparative experiments, different algorithms for learning structure and parameters of the BNs are applied. The experiments revealed that a simple two-node BN outperformed all of the other complex models tested for the considered data.