RainODE: Continuous-Time Precipitation Forecasting with Latent Neural ODEs

Traditional precipitation forecasting models are typically trained on discrete temporal snapshots (e.g., 5-minute intervals), which leads to error accumulation in autoregressive rollouts or high computational costs in many-to-many architectures. This paper asks whether reformulating precipitation forecasting as a continuous-time dynamical process can improve prediction stability and allow for flexible, arbitrary-time inference without the constraints of fixed temporal grids.

To address these challenges, the authors propose RainODE, a framework that models the evolution of precipitation in a latent space using a Neural ODE. By learning derivative-consistent dynamics, the model captures the large-scale advective motion of precipitation systems. To overcome the tendency of deterministic ODEs to produce over-smoothed results, the authors introduce a Stochastic Source Modeling (SSM) module based on a Brownian Bridge formulation. This module injects controlled stochasticity to refine high-frequency intensity variations and localized structures, such as growth and decay, while maintaining the underlying advective consistency learned by the ODE.

The authors evaluate RainODE on the established SEVIR dataset and a newly introduced benchmark, the Radar-based Precipitation Integrated Dataset (RAPID), which supports evaluation across 10-, 30-, and 60-minute intervals. RainODE consistently outperforms state-of-the-art discrete-time baselines across multiple temporal regimes. Quantitative results show that RainODE maintains superior structural coherence and intensity accuracy, particularly at high-intensity thresholds. Qualitative analysis demonstrates that while discrete baselines suffer from intensity dissipation or spatial misalignment over long horizons, RainODE produces coherent trajectories that remain stable even when evaluated at temporal resolutions different from those used during training.