SIGMAformer: a spatiotemporal Gaussian mixture correlation transformer for global weather forecasting

Abstract SpatIotemporal Gaussian Mixture correlAtion transformer (SIGMAformer) is a spatiotemporal forecasting architecture that integrates a Gaussian mixture pattern extractor (GMPE) with a dynamic spatiotemporal correlation (DSTC) mechanism. The DSTC module leverages GMPE to automatically compute spatiotemporal pattern-specific weights from the data. These weights are first used to calculate temporal correlations within each station and then integrated with global pattern weights to evaluate spatial correlations across stations. This nonlinear, dynamically adaptive modeling approach emphasizes critical spatiotemporal patterns while suppressing less relevant ones. Experiments on global weather datasets reveal that SIGMAformer consistently outperforms state-of-the-art forecasting models and significantly improves wind speed prediction. Removing DSTC increased the mean squared error values by up to 7.18% and 7.22% for wind speed and temperature predictions, respectively. These findings underscore SIGMAformer’s capacity to capture essential spatiotemporal patterns and establish a scalable methodology for intelligent sensor-network fusion in environmental forecasting.