Daily travel plans and early warnings for extreme weather all rely on traditional numerical weather prediction. However, both traditional numerical weather prediction and AI forecasting large models have long suffered from systematic biases, which compromise forecast accuracy.
To address this challenge, the research group led by Prof. Xiaomeng Huang from Tsinghua University, China, in collaboration with the National Climate Centre, China, has developed an AI bias correction framework based on spatiotemporal correlation deep learning. This framework accurately corrects forecast biases, achieving a maximum 20% reduction in the root-mean-square error of 7-day 2-meter air temperature forecasts. In addition, it can support the bias correction of oceanic variables, making forecasts in meteorological and oceanic scenarios more accurate. The findings were recently published in Atmospheric and Oceanic Science Letters .
The research team systematically integrated three key innovations into their model design: dynamic climatological normalization, ConvLSTM with temporal causality constraints, and residual self-attention mechanisms, enabling systematic bias correction of European Centre for Medium-Range Weather Forecasts (ECMWF) numerical forecasts.
The model was trained and validated using 41 years (1981–2021) of global atmospheric data, with ERA5 (fifth generation ECMWF atmospheric reanalysis) data serving as the ground truth. A decadal stratified sampling strategy, i.e., five non-consecutive years (1981, 1991, 2001, 2011, 2021) selected at 10-year intervals as a testing set, was employed to ensure the model's generalization capability across distinct climate phases.
Results show that the model boasts outstanding generalization capability. After being trained on the air temperature variable, it only takes 20 minutes to perform cross-variable correction for wind fields and air pressure, cutting the retraining time by 85%. Integrated as a plug-in into existing AI forecasting models, it further improves the forecast skill by 10%. Moreover, the corrected atmospheric data can significantly enhance the prediction performance of ocean models, enabling cross-domain empowerment from meteorology to oceanography.
This research was supported by the National Natural Science Foundation of China. The model code has been made publicly available, providing a reproducible technical solution for meteorological AI research.
