Generalizing Climate Forecast to Tremendous-grained Temporal Scales through Physics-AI Hybrid Modeling
Authors: Wanghan Xu, Fenghua Ling, Wenlong Zhang, Tao Han, Hao Chen, Wanli Ouyang, Lei Bai
Summary: Knowledge-driven synthetic intelligence (AI) fashions have made important developments in climate forecasting, notably in medium-range and nowcasting. Nevertheless, most data-driven climate forecasting fashions are black-box programs that target studying information mapping reasonably than fine-grained bodily evolution within the time dimension. Consequently, the restrictions within the temporal scale of datasets stop these fashions from forecasting at finer time scales. This paper proposes a physics-AI hybrid mannequin (i.e., WeatherGFT) which Generalizes climate forecasts to Finer-grained Temporal scales past coaching dataset. Particularly, we make use of a fastidiously designed PDE kernel to simulate bodily evolution on a small time scale (e.g., 300 seconds) and use a parallel neural networks with a learnable router for bias correction. Moreover, we introduce a lead time-aware coaching framework to advertise the generalization of the mannequin at totally different lead occasions. The burden evaluation of physics-AI modules signifies that physics conducts main evolution whereas AI performs corrections adaptively. Intensive experiments present that WeatherGFT skilled on an hourly dataset, achieves state-of-the-art efficiency throughout a number of lead occasions and displays the potential to generalize 30-minute forecast
