A Spatiotemporal Ensemble Approach to Rainfall Forecasting

Abstract

This paper proposes a new ensemble method built upon a deep neural network architecture. We use a set of meteorological models for rain forecast as base predictors. Each meteorological model is provided to a channel of the network and, through a convolution operator, the prediction models are weighted and combined. As a result, the predicted value produced by the ensemble depends on both the spatial neighborhood and the temporal pattern. We conduct some computational experiments in order to compare our approach to other ensemble methods widely used for daily rainfall prediction. The results show that our architecture based on ConvLSTM networks is a strong candidate to solve the problem of combining predictions in a spatiotemporal context.