Bayesian Networks for Probabilistic Weather Prediction

Several standard approaches have been introduced for meteorological time series prediction (analog techniques, neural networks, etc.). However, when dealing with multivariate spatially distributed time series (e.g., a network of meteorological stations over the Iberian peninsula) the above methods do not consider all the available information (they consider special independency assumptions to simplify the model). In this work, we introduce Bayesian Networks (BNs) in this framework to model the spatial and temporal dependencies among the different stations using a directed acyclic graph. This graph is learnt from the available databases and allows deriving a probabilistic model consistent with all the available information. Afterwards, the resulting model is combined with numerical atmospheric predictions which are given as evidence for the model. Efficient inference mechanisms provide the conditional distributions of the desired variables at a desired future time. We illustrate the efficiency of the proposed methodology by obtaining precipitation forecasts for 100 stations in the North basin of the Iberian peninsula during Winter 1999. We show how standard analog techniques are a special case of the proposed methodology when no spatial dependencies are considered in the model.