CNCast: Leveraging 3D Swin Transformer and DiT for Enhanced Regional Weather Forecasting

TL;DR

CNCast addresses the need for accurate hourly regional weather forecasts by combining a boundary-conditioned Swin Transformer 3D backbone for multi-variable predictions with a latent-diffusion-based DiT approach for high-resolution precipitation diagnosis. The framework leverages ERA5 regional data and CMPAS precipitation, using VAEs to compress inputs for efficient diffusion modeling and employing an ensemble approach to improve precipitation outputs. Key contributions include enhanced boundary treatments to stabilize longer lead forecasts and a two-stage system that delivers competitive, and often superior, performance to the global Pangu model, particularly in wind forecasts. The work demonstrates practical impact by enabling finer-scale precipitation analysis and providing a scalable pathway to improve regional forecast reliability in data-rich environments like China.

Abstract

This study introduces a cutting-edge regional weather forecasting model based on the SwinTransformer 3D architecture. This model is specifically designed to deliver precise hourly weather predictions ranging from 1 hour to 5 days, significantly improving the reliability and practicality of short-term weather forecasts. Our model has demonstrated generally superior performance when compared to Pangu, a well-established global model. The evaluation indicates that our model excels in predicting most weather variables, highlighting its potential as a more effective alternative in the field of limited area modeling. A noteworthy feature of this model is the integration of enhanced boundary conditions, inspired by traditional numerical weather prediction (NWP) techniques. This integration has substantially improved the model's predictive accuracy. Additionally, the model includes an innovative approach for diagnosing hourly total precipitation at a high spatial resolution of approximately 5 kilometers. This is achieved through a latent diffusion model, offering an alternative method for generating high-resolution precipitation data.

Figures (9)

• Figure 1: Swin Transformer 3D architecture.
• Figure 2: Patch embedding with boundary condition. The boundary is taken from target with 4 outmost pixels and then flattened to be fed into patch embedding layer. The embedded boundary is then split to 4 parts and concatenated back to the embedded feature.
• Figure 3: Precipitation diagnosis with DiT based on latents.
• Figure 4: The comparison of forecast accuracy in terms of latitude-weighted RMSE and ACC of 4 surface variables.
• Figure 5: Visualization of 12 hour weather forecast and spatial MAE of mslp(top) and u-component of wind(bottom) produced by CNCast, Pangu and the ERA5 ground-truth. The initial time point (i.e. the forecast is performed on) is 2021072001 UTC.