Automating weather forecasts based on convolutional networks

Numerical weather models generate a vast amount of information which requires human interpretation to generate local weather forecasts. Convolutional Neural Networks (CNN) can extract features from images showing unprecedented results in many different domains. In this work, we propose the use of CNN models to interpret numerical weather model data which, by capturing the spatial and temporal relationships between the input variables, can produce local forecasts. Different architectures are compared and a methodology to introspect the models is presented.