基于多模型集成学习的对流云覆盖率预测研究

Predictive Research of Convective Cloud Cover Rate Using Multi-Model Ensemble Learning

摘要: 基于2016~2024年5~9月ERA5再分析资料的逐小时97个大气物理特征变量数据，以实况的组合反射率大于35 dBZ的终端区覆盖率为目标变量，利用随机森林(RandomForest)、极端随机树(ExtraTrees)、梯度提升决策树(GBDT)、XGBoost、CatBoost和LightGBM共6种机器学习模型。采用Pearson相关性分析进行特征筛选，利用MinMax标准化进行数据预处理，通过3折交叉验证评估模型泛化能力，并引入SHAP (SHapley Additive exPlanations)方法对在所测模型中表现最优模型进行全局和局部可解释性分析。结果表明：ExtraTrees模型预测性能表现最好，测试集决定系数(R²)为0.1761，平均绝对误差(MAE)为0.0242，均方根误差(RMSE)为0.0627，交叉验证R²为0.1837；大气可降水量(pw)、降水总量(tp)和850 hPa露点温度(td_850)是预测云覆盖率最重要的三个特征，模型学习到的特征影响模式与气象学物理机制高度一致；学习曲线分析表明模型尚未达到性能平台期，增加训练数据量有望进一步提升预测精度。

Abstract: Based on hourly atmospheric physical characteristic variables from ERA5 reanalysis data from May to September 2016 to 2024, the target variable was the terminal area coverage rate with a combined reflectance greater than 35 dBZ. Six machine learning models were used: Random Forest, ExtraTrees, Gradient Boosting Decision Tree (GBDT), XGBoost, CatBoost, and LightGBM. Pearson correlation analysis was employed for feature screening, MinMax normalization for data preprocessing, and 3-fold cross-validation for model generalization assessment. The SHAP (SHapley Additive exPlanations) method was introduced for global and local interpretability analysis of the optimal model. The main conclusions are as follows: The ExtraTrees model achieved the best prediction performance with a test R² of 0.1761, MAE of 0.0242, RMSE of 0.0627, and cross-validation R² of 0.1837. SHAP feature importance analysis identified precipitable water (pw), total precipitation (tp), and 850 hPa dewpoint temperature (td_850) as the three most important predictors of cloud cover rate, with the learned feature impact patterns being highly consistent with meteorological physical mechanisms. Learning curve analysis indicates that the model has not yet reached a performance plateau, suggesting that increasing training data volume could further improve prediction accuracy.