Improving 2 m temperature forecasts of numerical weather prediction through a machine learning-based Bayesian model

Abstract

Temperature is a fundamental meteorological factor significantly impacting human life and socio-economic development. This study applies a multi-model fusion technique, integrating three artificial intelligence (AI) methods, to improve temperature forecast accuracy by addressing systematic errors and biases in the European Centre for Medium-Range Weather Forecasts (ECMWF) 2 m temperature predictions for Xiong’an New Area and its upstream regions. Using ECMWF forecast data from January 1, 2018, to December 31, 2021, along with ERA5 reanalysis data, we optimized a Bayesian model averaging (BMA_OP) approach, combining linear regression, LightGBM, and UNet to revise the 2 m temperature forecast. BMA_OP demonstrated improved performance, achieving an overall root-mean-square error (RMSE) of 1.15 °C, an average prediction accuracy of 73% for the ECMWF model, and an accuracy of over 91% for BMA_OP, marking a 24.7% improvement. To further assess generalization, we tested the model using full-year 2022 data, where BMA_OP outperformed the ECMWF model with an RMSE, mean absolute error (MAE), and accuracy of 1.31 °C, 1.03 °C, and 87%, respectively—exceeding the ECMWF model’s results by 16%, 13%, and 6%. These findings confirm the effectiveness of BMA_OP-based multi-model fusion technology for temperature correction.