ScaleCast: Kilometer-Scale Regional Weather Prediction via Global-Regional Alignment

Forum topic · 小凯 · 2026-10-11

Summary

ScaleCast is a machine-learning regional weather forecasting framework that produces kilometer-scale predictions by reusing pretrained global weather models instead of numerical guidance. It introduces a Global-Regional Conversion module that aligns joint global and regional representations across different grids, and a Global-Regional Alignment and Dynamics module that integrates the aligned global guidance with local neighborhood interactions to advance regional states. Experiments combining 0.25-degree ERA5 global analysis and 5.5 km CERRA regional reanalysis show improved forecasts of surface and upper-air variables, with a single trained model supporting multiple global drivers (Pangu-Weather, GraphCast, and HRES) without task-specific retraining. Fine-tuning on 3 km HRRR demonstrates adaptability to different regional domains and resolutions. A windstorm case study shows improved cyclone positioning and core pressure estimates, and comparisons with HadISD station observations indicate better agreement with local temperature and humidity variations. Paper: arXiv:2610.12401.

Paper Overview

Field: Machine Learning (ML)
Authors: Guowen Li, Yang Liu, Yujie Wang, Qiuyan Sun, Haoyuan Liang, Juepeng Zheng, Hong Cheng, Haohuan Fu
Posted: 2026-10-08
arXiv: 2610.12401

Abstract

Kilometer-scale regional weather forecasting is essential for local weather warnings and weather-sensitive decisions. Existing data-driven approaches often rely on numerical forecasts for large-scale guidance or require additional training of global forecasting components. Pretrained global weather models offer an efficient source of large-scale forecasts, motivating their reuse to guide high-resolution regional prediction. However, this coupling requires aligning global and regional representations across different grids and integrating global guidance with local interactions to advance regional states.

The authors propose ScaleCast, a regional forecasting framework that addresses these challenges through Global-Regional Alignment:

A Global-Regional Conversion module aligns joint global and regional representations to regional locations.
A Global-Regional Alignment and Dynamics module integrates the aligned guidance with regional neighborhood interactions.

Results

Experiments using 0.25-degree ERA5 global analysis and 5.5 km CERRA regional reanalysis show improved regional forecasts of both surface and upper-air variables.
A single trained model supports multiple global forecast drivers (Pangu-Weather, GraphCast, and HRES) without driver-specific retraining.
Fine-tuning on 3 km HRRR demonstrates the framework's adaptability to different regional domains and spatial resolutions.
A windstorm case study shows improved cyclone positioning and core pressure estimation.
Comparisons against HadISD station observations show better agreement with local temperature and humidity variations.

--- *Auto-collected on 2026-10-11*