Spatio-Temporal Weather Prediction with Graph Neural Networks

Abstract

Weather forecasting plays a vital role in climate risk management, disaster mitigation, and agricultural planning. Traditional forecasting models often rely on sequential methods or spatially coarse datasets, limiting their ability to capture fine-grained interactions across geographically distributed locations. This research proposes a Graph Neural Network (GNN)-based approach for spatio-temporal weather prediction, utilizing ERA5 reanalysis data from 1979 to 2020 for 1,359 locations in Nepal. Key meteorological features, including precipitation, relative humidity, and temperature, are incorporated into the model. The proposed framework constructs a graph representation, where each node corresponds to a geographic location and edges represent spatial adjacency or environmental similarity. The GNN architecture integrates graph convolutional layers to capture spatial dependencies and a Gated Recurrent Unit (GRU) to model temporal patterns. Performance evaluation against historical weather data demonstrates that the model achieves lower Mean Squared Error (MSE) than traditional sequential baselines, while maintaining computational efficiency. Results highlight the model’s ability to generalize across diverse climate zones, making it a promising tool for large-scale weather monitoring. Future enhancements could incorporate real-time sensor feedback and probabilistic uncertainty quantification to develop a more robust forecasting pipeline. This study underscores the potential of GNNs in enhancing weather prediction accuracy by effectively modeling spatial dependencies—an aspect often overlooked in conventional approaches. While the model achieved strong accuracy for temperature and humidity, precipitation predictions exhibited modest visual deviations. These differences are largely attributable to the bursty, sparse nature of rainfall and vertical scale exaggeration in plotted values. Nonetheless, the predictions remained temporally aligned with actual events and yielded low MSE, underscoring the model’s validity.