基于机器学习的气温预报误差订正研究

Research on Machine Learning-Based Correction of Temperature Forecast Errors

摘要: 利用2023年12月~2024年11月青藏高原地区CLDAS实况融合逐日最高、最低气温资料，系统评估智能网格以及随机森林(RF)与梯度下降树(GBDT)两种机器学习后处理方案的订正效果。结果表明：(1) RF与GBDT对最高、最低气温的逐日预报均带来系统性正技巧，全年各季无一例外；其中GBDT准确率均优于RF。(2) 原始预报日最高气温和日最低气温RMSE分别为4.24℃与3.75℃，存在显著系统偏差。(3) RF与GBDT均能有效压缩误差，GBDT最优，日最高气温和日最低气温RMSE分别降至3.33℃与3.14℃，相对提升21%与16%。(4) 空间上，研究区域的西部地区改进幅度最大，准确率提升10%~25%。

Abstract: Using the CLDAS daily maximum and minimum temperature analyses for the Qinghai-Xizang Plateau from December 2023 to November 2024, we systematically evaluate the correction skill of the intelligent-grid forecast and two machine-learning post-processing schemes—Random Forest (RF) and Gradient Boosting Decision Tree (GBDT). The results show that: (1) Both RF and GBDT deliver systematic positive skill for daily maximum and minimum temperature forecasts in every season, with GBDT consistently outperforming RF. (2) The raw forecasts exhibit significant systematic biases, with RMSEs of 4.24˚C for daily maximum temperature and 3.75˚C for daily minimum temperature. (3) Both RF and GBDT effectively compress these errors; the best-performing GBDT reduces RMSEs to 3.33˚C and 3.14˚C, corresponding to relative improvements of 21% and 16%, respectively. (4) Spatially, the largest gains (10%~25% increase in accuracy) occur over the western part of the study domain.