Leveraging Convolutions in Recurrent Neural Networks for Doppler Weather Radar Echo Prediction

Precipitation forecasting for short duration is an important problem in weather prediction. In this work, we propose a deep learning based approach for precipitation forecasting using Doppler weather radar data. Our approach uses convolutions within recurrence structure in vanilla recurrent neural networks exploiting both spatial and temporal dependencies in the data. We show that this approach can be applied for fine grained precipitation forecast with similar accuracy as that of complex models while reducing the model size by 4 times. Results are presented on the task of echo state prediction and skill scores for rainfall estimates on the data from Seattle, WA, USA as well as from cross testing the model, trained on Seattle data, on unseen data from Albany, NY, USA.