基于多种机器学习算法的霜预报模型对比

Comparison of Frost Forecasting Models Based on Multiple Machine Learning Algorithms

摘要: 本研究针对机场霜预报问题，采用沈阳桃仙国际机场2019~2024年冬半年地面观测资料，基于霜形成的物理机制及皮尔逊相关系数筛选预报因子，构建了逻辑回归、SVM、随机森林、XGBoost、LightGBM和MLP共六种机器学习模型及两种集成模型。研究采用时间序列划分数据集，以TS评分最大化为目标进行分类阈值优化。结果表明SVM与随机森林表现最优，集成学习增益有限。阈值优化可大幅提升预报性能，各模型TS评分相对提升17.4%~51.9%。01时气温为核心预报因子，各因子重要性排序符合辐射冷却物理机制。本研究构建的随机森林模型综合性能最优，可为民航气象部门开展霜预报业务提供技术参考。

Abstract: Focusing on the airport frost forecasting task, this study utilizes ground meteorological observation data from Shenyang Taoxian International Airport collected during the winter half-years of 2019 to 2024. Based on the physical mechanism of frost formation and Pearson correlation coefficient analysis, representative forecasting predictors are screened out. Six machine learning models—logistic regression, support vector machine (SVM), random forest, XGBoost, LightGBM, and multilayer perceptron (MLP)—together with two ensemble models are established for frost prediction. The dataset is divided in chronological order to ensure evaluation objectivity, and classification threshold optimization is conducted with the goal of maximizing the Threat Score (TS). The results indicate that SVM and random forest deliver the optimal forecasting performance, whereas ensemble learning brings limited performance improvement. Threshold optimization can significantly enhance forecasting capability, with the relative TS improvement of each model ranging from 17.4% to 51.9%. The air temperature at 01:00 is recognized as the core predictor, and the importance ranking of all factors conforms to the physical mechanism of radiative cooling. Overall, the random forest model constructed in this study achieves the best comprehensive performance, and can provide technical reference for civil aviation meteorological departments in operational frost forecasting.