📝 SummarXiv

Abstract

Accurate hourly weather forecasting is critical for numerous applications. Recent deep learning models have demonstrated strong capability on 6-hour intervals, yet achieving accurate and stable hourly predictions remains a critical challenge. This is primarily due to the rapid accumulation of errors in autoregressive rollouts and temporal discontinuities within the ERA5 data's 12-hour assimilation cycle. To address these issues, we propose FlowCast-ODE, a framework that models atmospheric state evolution as a continuous flow. FlowCast-ODE learns the conditional flow path directly from the previous state, an approach that aligns more naturally with physical dynamic systems and enables efficient computation. A coarse-to-fine strategy is introduced to train the model on 6-hour data using dynamic flow matching and then refined on hourly data that incorporates an Ordinary Differential Equation (ODE) solver to achieve temporally coherent forecasts. In addition, a lightweight low-rank AdaLN-Zero modulation mechanism is proposed and reduces model size by 15% without compromising accuracy. Experiments demonstrate that FlowCast-ODE outperforms strong baselines, yielding lower root mean square error (RMSE) and better energy conservation, which reduces blurring and preserves more fine-scale spatial details. It also shows comparable performance to the state-of-the-art model in forecasting extreme events like typhoons. Furthermore, the model alleviates temporal discontinuities associated with assimilation cycle transitions.