Improving the forecast accuracy of ECMWF 2-m air temperature using a historical dataset

The 2-m air temperature (T2m) is an important meteorological variable and has been the focus of meteorological forecasting. Although the numerical weather model is an important means of forecasting, it typically presents forecasting errors that cannot be eliminated by improving the ability of the numerical model to reproduce the processes. Thus, a statistical correction of the forecast results is required. In this study, we applied the local dynamical analog (LDA) method to correct the operational T2m forecast product obtained from the European Centre for Medium-Range Weather Forecasts with the lead time of 24–240 h. To our knowledge, for the first time, we used spatially adjacent grids from high-resolution grid data as potential analog pools to compensate for the short duration of historical data. The T2m of weather forecasts in East Asia for December 2018 was improved by LDA correction with a small sample condition. Compared with ERA5 and station observation data, the results show that the root mean square error can be reduced by 2%–4% and the correlation coefficient can be increased by 1%–5% for different lead times, with the most distinct improvement effect for the medium-term forecast time. The Qinghai Tibet Plateau, Mongolia Plateau, and other areas, where the raw prediction error is relatively high, presented better performance than other regions. For a cold-wave process, we also demonstrate that the corrected results based on analogs present better forecasting skill performance than raw forecast results. The analog correction with the LDA method, which combines statistical and model dynamical techniques, is proposed to be integrated with other advanced operational models. The forecast skill of T2m was improved by a historical dataset, which may contribute to energy management and the construction industry. • Numerical weather model forecasting can be improved by use of historical datasets. • Spatially adjacent grids compensate for short time duration of historical data. • Use of local dynamical analog method can reduce RMSE by up to 4%. • Temporal correlation coefficient is improved by up to 5%. • The forecast accuracy of a Cold wave process is improved by the correction.