Application of Convolutional Recurrent Neural Network in Precipitation Forecasting

Abstract

In this work, we use convolutional recurrent neural network-based architectures involving ConvLSTM and ConvGRU for precipitation forecasting over the Indian region. We first compare direct and iterative approach for forecasting and find the iterative approach better for our model. We use multiple variables such as specific humidity, orography, soil moisture and surface pressure as input features for better capturing the underlying dynamics. We also analyse the forecasts over different homogeneous rainfall regions. Finally, we compare ConvLSTM and ConvGRU based models. We find similar performance in both the models, ConvGRU being faster.