ST-GNNs for Weather Prediction in South Africa

Abstract

Spatial-temporal graph neural networks (ST-GNN) have been shown to be highly effective for flow prediction in dynamic systems, but are under explored for weather prediction applications. We compare and evaluate Graph WaveNet (GWN) and the Low Rank Weighted Graph Neural Network (WGN) for weather prediction in South Africa. We compare these results to two basic temporal deep neural networks architectures, i.e. the Long Short-Term Memory (LSTM) and the Temporal Convolutional Neural Network (TCN), for maximum temperature prediction across 21 weather stations in South Africa. We also perform rigorous experiments to evaluate the stability and robustness of both ST-GNNs. The results show that the GWN model outperforms the other models across different prediction horizons with an average SMAPE score of 8.30%. We also analyse and compare learnt adjacency matrices of the two ST-GNNs to gain insights into the prominent spatial-temporal dependencies between weather stations.