An Efficient Algorithm for Weather Forecasting Using Causal Graph Neural Network

The rapid accumulation of large-scale, long-term meteorological data presents unprecedented opportunities for data-driven weather modeling and high-resolution numerical weather prediction. While various deep learning techniques—such as Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNNs), and Graph Neural Networks (GNNs)—have been explored for weather forecasting, the complex spatial dependencies within historical meteorological data, particularly dynamic spatial correlations, remain insufficiently addressed. To tackle this challenge, we propose a Dynamic Spatio-Temporal Fusion Graph Network (DSTFGN), a novel module that integrates multivariate time-series analysis with graph-based causal inference to capture intricate and time-varying interdependencies among weather variables. The DSTFGN module fuses real-time inputs (e.g., sensor data, live weather feeds, external events) with historical records to model the propagation of disruptions—such as accidents or road closures—through the meteorological network. By effectively capturing dynamic spatial-temporal interactions, our approach significantly enhances forecasting accuracy and supports adaptive weather management strategies. Experimental evaluations on two real-world datasets demonstrate that DSTFGN consistently outperforms existing baseline models across short, medium, and long-term forecasting horizons.