Long-term multi-horizon weather prediction using geo-based physics-informed learning

Abstract Severe atmospheric conditions may trigger extreme weather events, leading to considerable economic losses and substantial human casualties. Weather prediction supports proactive mitigation strategies, but it remains a highly challenging process due to the nonlinear interactions between the motion dynamics and thermodynamics of the atmosphere. In this work, we introduce a geo-based physics-informed predictor (GeoPIP) for long-horizon weather prediction. By integrating geo-based atmospheric representations, transformer-based temporal modelling and physics-informed regularisation, GeoPIP is designed to address long-horizon uncertainty, geographical variability and physically plausible prediction behaviour. The proposed system can perform long-term multi-horizon dual estimations of air temperature and wind speed while preserving physical consistency, delivering a transparent interpretation of the spatiotemporal estimation behaviour and the contribution of the most influential atmospheric parameters. We evaluated the performance of our system by utilising real-world measurements collected across geographically distributed weather stations in Saudi Arabia. GeoPIP yielded a low prediction error of $$1.497^\circ $$ 1.497∘C for air temperature and 0.420 m/s for wind speed across large-scale meteorological environments. It has low computational complexity, and its architecture can be easily integrated into intelligent meteorological monitoring systems to realise next-generation long-horizon weather prediction. Graphic abstract