以ERA5通用数据集构建的气象大模型普遍存在输入变量多、计算资源消耗大和可解释性差的问题，且针对小概率极端气象事件如冻雨的建模研究较少。针对以上技术难题，以冻雨多发的南方地区为研究对象，基于XGBoost（extreme gradient boosting）模型和SHAP（Shapley additive explanations）方法构建冻雨最优模型，并探讨冻雨的预报机制。结果表明：基于冻雨形成机制的研究成果，在ERA5数据基础上新构建了48个与冻雨密切相关的气象因子，并通过相关性分析筛选出16个影响或潜在影响冻雨的因子作为输入集，相较通用数据集中的变量数量减少了约92%。通过因子重要性分析并结合贝叶斯优化，构建XGBoost最优冻雨预测模型。其在预测冻雨方面表现优于现有基于物理机制与统计相结合的冻雨参数化方案，且选取的7个关键因子具有明确的物理机制。模型因子的平均重要性依次为2 m气温、海拔高度、大气廓线类型、固态水成物比例、次冻结层最低温度、暖层顶部高度和600 hPa相对湿度。SHAP值揭示了各因子在预测过程中的贡献大小与模型的预报机制，但由于因子间的交互作用，SHAP值在不同时间和地点有所不同，表明冻雨预报非常复杂，不同时间和地点的预报机制均不同。研究成果通过以上步骤研制了仅包含7个因子、预报能力强大且预报机制清楚的XGBoost冻雨预测模型，可为小概率极端气象事件的预报提供借鉴，具有较高的科学价值和应用价值。

Large weather models(LWMs) constructed using the ERA5 general dataset often face challenges such as numerous input features, high computational demands, and poor interpretability. Moreover, research on modeling rare extreme weather events, such as freezing rain, remains limited. To address these issues, taking the southern region with frequent freezing rain as the research subject, an optimal freezing rain model was developed using the extreme gradient boosting (XGBoost) algorithm and the Shapley additive explanations(SHAP) method to explore the forecasting mechanism of freezing rain. The results indicate that, based on research into freezing rain formation mechanisms, 48 meteorological features closely related to freezing rain are newly constructed from ERA5 data. Correlation analysis identifies 16 features as the key inputs, reducing the feature count by nearly 92% compared to the original dataset. Using feature importance analysis in combination with Bayesian optimization, an XGBoost-based optimal freezing rain prediction model is developed. This model outperforms existing freezing rain parameterization schemes, which combine physical mechanisms and statistics, with the seven selected key features showing clear physical relevance. The average importance of the model features is ranked as follows: 2-meter air temperature, altitude, atmospheric profile, proportion of solid precipitation particles, minimum temperature of the subfreezing layer, height of the top of the warm layer, and relative humidity at 600 hPa. SHAP values reveal the contribution of each feature to the prediction process and the model's forecasting mechanism. However, due to interactions between features, SHAP values vary across different times and locations, highlighting the complexity of freezing rain forecasting and the differing mechanisms at play in various temporal and spatial contexts. By incorporating only seven features, research results successfully develop an XGBoost-based freezing rain prediction model with strong forecasting capabilities and clear prediction mechanisms. This model provides valuable insights into forecasting low-probability extreme weather events, offering both scientific and practical value.