Physics-Informed spatiotemporal deep learning for multivariate atmospheric forecasting

With the rapid development of the low-altitude economy, emerging applications such as general aviation, drone delivery, and urban air mobility are placing higher demands on the accurate sensing and forecasting of the low-altitude atmospheric environment. Current research faces two major challenges: (1) traditional numerical weather prediction (NWP) methods have limited responsiveness to rapid disturbances in the lower atmosphere, making them inadequate for precision requirements in low-altitude flight; and (2) data-driven deep learning methods often lack physical consistency constraints, leading to poor performance in forecasting short-term rapid fluctuations and compromising flight safety. To address these challenges, this paper proposes an innovative short-term forecasting framework for low-altitude meteorological variables, which integrates physical knowledge with short sequence modeling. First, a temporal sequence branch mechanism is introduced to precisely model the multi-scale dynamic variations in the low-altitude atmospheric field, significantly enhancing the model’s ability to detect and predict both rapid disturbances and slow-varying trends, thereby improving flight safety assurance. Second, an adaptive physics-guided loss is designed to incorporate conservation constraints, effectively improving the physical consistency of predictions and mitigating unrealistic forecasts in complex low-altitude environments. The experimental results show that the proposed method outperforms the mainstream baseline model in several metrics. Meanwhile, the model maintains stable generalization performance in different climate regions, reflecting good regional adaptability. In the distribution consistency assessment of the validation variables, such as relative humidity and wind speed, the KL divergence of the model is as low as 0.004 and 0.008, respectively, which demonstrates its good fitting ability to the actual distribution. Overall, the method in this paper achieves a good balance between prediction accuracy, physical consistency, and timeliness, and provides an effective solution for real-time weather prediction oriented to low-altitude traffic and environment sensing.