New AI model brings breakthroughs in five-day regional weather forecasting

Institute of Atmospheric Physics, Chinese Academy of Sciences – Researchers at Northwestern Polytechnical University in China proposed a novel deep learning–based framework that dramatically improves the accuracy of forecasts, even when data is limited.

To address challenges in regional forecasting, the team introduced a new method that integrates three major innovations:

• the use of semantic segmentation models originally designed for medical image analysis;
• a learnable Gaussian noise mechanism that improves the model’s robustness;
• and a cascade prediction strategy that breaks the forecasting task into manageable stages.

“Our goal was to make regional forecasting smarter, faster, and more reliable, even in data-limited scenarios,” said Prof. Congqi Cao, one of the study’s authors. “This is especially valuable for areas where a dense network of meteorological observations is not available.”

The method was tested on the East China Regional AI Medium Range Weather Forecasting Competition dataset, which included 10 years of reanalysis data from ERA5. The task involved using past atmospheric variables to predict five key surface weather indicators — including temperature, wind, and precipitation — every six hours for the next five days.

The model achieved significant improvements in prediction performance, outperforming many mainstream global AI forecasting models. Specifically, the method reduced temperature forecast errors by 9.3 per cent, improved the precipitation F1-score by 6.8 per cent, and lowered wind speed errors by 12.5 per cent.

“This is the first time semantic segmentation and learnable noise mechanisms have been used together for regional weather forecasting,” explained Cao. “It opens up new possibilities for accurate forecasting in other data-scarce regions.”

Looking ahead, the team plans to extend their method to real-time systems and apply it to more regions across China. They hope their work will eventually serve public safety, agriculture, and disaster prevention needs—delivering smarter, faster, and more local forecasts when they matter most.