Based on the investigation of weather prediction models conducted in the previous fiscal year, we designed five prototype deep learning models specialized for weather forecasting around Japan. The architectures used were Swin-Unet, a type of Vision Transformer, and AR2U-Net, a type of CNN. All models were trained using 20 years of regional reanalysis data (RCDSJRA-55), with input variables including wind, temperature, humidity, and geopotential height at multiple pressure levels, as well as surface wind and precipitation. The models employed an autoregressive approach, predicting the next time step and using it as input for the following step.
The highest accuracy was achieved with the Swin-Unet–based model, which consisted of a Predictor that outputs future values and a Corrector that refines results using a diffusion model. The Corrector adds Gaussian noise to the predictions, then removes it through a diffusion process to restore high-frequency components, reducing excessive spatial smoothing. In the heavy rainfall event of October 25, 2019, the observed hourly maximum rainfall was 83.5 mm; the Predictor alone estimated 55.6 mm, while the combined use with the Corrector improved the estimate to 63.1 mm (Fig.1).

We also evaluated the impact of training methods and initialization settings on our improved ClimaX-based prediction model and compared it with publicly available data-driven models (Fig.2). The results revealed the limitations of autoregressive generation in improving prediction accuracy, as well as the significant influence of initialization settings.