Federated Learning Framework for Adaptive Spatio-Temporal Graph Neural Networks in Weather Forecasting

Accurate weather forecasting is vital for sectors such as agriculture and disaster management, where timely and precise predictions can significantly impact outcomes. Conventional forecasting methods often struggle in accurately capturing the complex spatio-temporal dependencies inherent in meteorological data. The proliferation of modern sensor-equipped devices, such as smartphones, enables decentralized and real-time data collection, providing opportunities for large-scale predictive modelling. However, the sensitive nature of user data requires privacy-preserving algorithms. In this paper, we propose a federated spatio-temporal adaptive graph neural network (FedSTAGNN) model, which combines attention-based graph convolutional network layers and temporal convolutional layers in a federated learning framework to capture spatial and temporal patterns. The proposed model employs hierarchical aggregation at both city and nationwide levels to ensure scalability across diverse regions. Experimental evaluation on real-world weather dataset demonstrates that FedSTAGNN outperforms state-of-the-art models in terms of predictive accuracy, communication efficiency, and data privacy, highlighting its potential for practical deployment in large-scale, privacy-sensitive applications.