ConvScaleCast: a physics-informed scale-bridging framework for 0–24 h urban convective precipitation forecasting

Mesoscale convective systems generate heavy-tailed extreme rainfall that conventional numerical weather prediction and standard deep learning architectures consistently underestimate, creating persistent forecasting bottlenecks over coastal urban zones. This study develops ConvScaleCast, a physics-informed scale-bridging framework generating 0–24 h precipitation predictions through integrated multi-scale atmospheric information fusion. The architecture unifies adaptive multi-band radar synthesis, the vertically constrained GVIP network for short-range radar extrapolation, and a novel PODA autoencoder paired with focal ordinal regression to correct systematic NWP biases. Unlike MSE-optimized models that oversmooth intense rainfall signatures, PODA prioritizes hard-to-characterize extreme samples to reconstruct the full skewed statistical distribution of heavy precipitation. Validation against multi-year observational records from Shanghai shows that the framework outperforms ECMWF operational forecasts and mainstream baseline algorithms, resolving the critical 2–6 h prediction gap while retaining coherent convective spatial structures. This multi-scale fusion architecture delivers a potential technical solution for skillful prediction of high-impact precipitation events.