Applied Research on the Correction Method of CMA Climate Model Prediction Products Based on Convolutional Neural Network

Applied Research on the Correction Method of CMA Climate Model Prediction Products Based on Convolutional Neural Network Abstract This study applies the Convolutional Neural Network (CNN) approach to conduct post-processing correction on key operational prediction products of the third generation climate operational prediction system of the China Meteorological Administration (CMA), namely CMA-CPSv3. The targeted products include monthly 2-meter air temperature, precipitation over China, and the El Ni?o-Southern Oscillation (ENSO) index during the period 2001–2023. Using reanalysis data from the National Centers for Environmental Prediction (NCEP) as the observa-tional benchmark, a dedicated correction model was developed through deep learning training of a multi-layer CNN architecture. After model construction, changes in prediction performance before and after correction were evaluated during an independent test period. Results indicate that the CNN model significantly improves the prediction accuracy of climate model products. For temperature and precipitation predictions over China, the correlation coefficient for forecasts with a lead time of 1–7 months increases by 0.1–0.5. Specifically, the Root-Mean-Square Error (RMSE) of temperature decreases by 0.5–1°C (a reduction rate of 20%–30%), while the RMSE of precipitation declines by 0.1–1 mm/day (a reduction rate of 3%–30%). For the ENSO index, the correlation skill for forecasts with a lead time of 1–7 months is enhanced by 5%–7%, and the RMSE at a lead time of 7 months is reduced by 50%, effectively addressing the issue of excessive oscillation amplitude of the ENSO index in the original CMA-CPSv3 model. Furthermore, this study explicitly identifies a limitation of the CNN model, i.e., excessive intensity smoothing, when applied to the correction of extreme climate events, and proposes mul-ti-dimensional directions for future optimization. It thus provides a technical solution that integrates scientific rigor and practical applicability for the operational post-processing of CMA"s climate models. Keywords: Convolutional Neural Network, Climate Model, Post-processing Correction, Deep Learning