Application of Deep Learning Bias Correction Method to Temperature Grid Forecast of 7-15 Days

Abstract

The forecast error of numerical weather forecasting is inevitable, and there are still difficulties in temperature forecast of 7-15 days. To improve forecast accuracy and timeliness, the deviation correction technique is often used in operation. In recent years, deep learning methods have shown great potential in statistical post-processing of model forecasts. To improve the accuracy of Global Ensemble Prediction System of China Meteorological Administration (CMA-GEPS) for 7-15 days, error characteristics of 2 m temperature and 10 m wind products of CMA-GEPS control forecast provided by TIGGE data center from 25 December 2018 to 5 July 2022, and ERA5 data provided by ECMWF are analyzed. The U-Net model and residual connection model is used to conduct 2 m temperature lattice forecast error revision experiment for the lead time of 168-360 h in the region 15.75°-55.25°N, 73°-136.5°E. The experiments are designed with various data features to explore differences of the deep learning methods for longer lead time with different sample characteristics and model parameters, and performances of models are examined by comparing the bias, mean absolute bias and root mean square error. The results show that 2 m temperature forecast errors of 7-15 days become larger as the lead time increases. The model forecast skill gradually decreases, and in the target area, performance in eastern and southern marine and offshore areas is better than in western and northern the plateaus and mountains. The differences in the spatial distribution of errors are more prominent. Among the revised models, the effect of the U-Net model is better than that of the U-Net residual connection model, and adding the initial 2 m temperature data of ERA5 can greatly improves the performance, but the effect of adding CMA-GEPS control forecast 10 m wind product of CMA-GEPS control forecast is not apparent. For 9 lead times, the revised root mean square errors are reduced by 10%-25%, and the model can effectively reduce the large forecast errors for the northern Mongolian Plateau and the western Tibetan Plateau, and some mountainous areas in the target area.

Keywords

• temperature grid forecast of 7-15 days
• bias correction
• deep learning method