Rainfall Forecasting with Variational Autoencoders and LSTMs

Abstract

In this paper we present a case study using a novel machine learning system for rainfall forecasting in a localized area over Colombia, South America. We explore a new forecasting approach inspired by established techniques used in computer vision and generative modeling to create a predictive model for precipitation maps. Using an ensemble made of a Variational Autoencoder and a stacked LSTM we were able to create a system which learns the spatial and temporal features of weather patterns in an integrated way, but also allows them to be measured and studied independently. Such a system introduces practical benefits in applications such as detection of rare weather patterns or analysis of anomalous events in addition to its regular forecasting capabilities.