Applications of Convolutional-Recurrent Neural Networks in Weather Forecasting

Abstract

In this thesis, we use convolutional long short-term memory based neural network archi- tectures to predict certain basic dynamics of weather variables. First, rainfall dynamics are studied across direct and iterative forecasting methodologies from India-restricted data set. Next, we develop models capable of predicting dynamics over a global cube- sphere grid and as a proof of concept study the dynamics of geopotential height while comparing them on a benchmark dataset namely WeatherBench. Experiments indicate that ConvLSTM2D-based models are advantageous for iterative forecasting than direct forecasting than conv2D-based models by slight margins and need further development. When applied to cube-sphere grid, CubeSphereConvLSTM did perform better than the benchmark CNN iterative model but did not beat the best operational systems.