探究基于LightGBM与地形的黄冈站点温度订正模型

Exploring a Station-Scale Temperature Correction Model for Huanggang Based on LightGBM and Topography

摘要: 针对2 m气温预报中局地效应表征不足等问题，本文采用欧洲中心再分析资料ERA5高空数据、地面站点观测资料以及30 m分辨率地形数据，在黄冈地区构建了基于LightGBM算法的站点精细化温度订正模型。模型对比结果表明，LightGBM模型在预报精度和训练效率方面均表现最优，具有良好的业务化潜力。独立预报检验结果显示，4月(7月)模型的MAE和RMSE较ERA5分别降低16.02% (11.54%)和12.63% (10.78%)、准确率提升了9.71 (5.53)个百分点。虽然模型与ERA5的误差空间分布均表现出平原低、山区高的特征，但模型在山区的各项指标更高，改进效果更明显，且对温度场空间细节与局地中心的刻画能力明显优于ERA5，可为大别山区等复杂地形区域温度精细化预报提供技术支撑。

Abstract: To address issues such as the inadequate representation of local effects in 2-meter air temperature forecasts, this study uses upper-air data from ERA5 reanalysis, ground station observations, and 30 m resolution topographic data to construct a regional high-resolution temperature correction model based on the LightGBM algorithm in Huanggang. Model comparison results show that the LightGBM model performs best in both prediction accuracy and training efficiency, demonstrating strong potential for operational application. Independent forecast test results show that compared to ERA5, the model reduces MAE by 16.02% (11.54%) and RMSE by 12.63% (10.78%) in April (July), with the accuracy increased by 9.71 (5.53) percentage points. Although both the model and ERA5 exhibit a spatial error distribution characterized by lower errors in plains and higher errors in mountainous areas, the model achieves better performance metrics and more prominent improvement effects in mountainous regions. Furthermore, its ability to capture spatial details and local centers within temperature field is significantly superior to that of ERA5, providing technical support for high-resolution temperature forecasting in complex topographic regions such as the Dabie Mountains.