基于Catboost和Stacking融合模型的 长江中下游短时临近降水预报研究

用中国自动站与CMORPH降水产品融合的逐小时降水量网格数据集、全球预报系统(Global Forecasting System, GFS)模式再分析资料,将机器学习特征算法筛选的特征变量作为模型输入数据,运用Catboost模型和以Catboost和随机森林为初级模型、径向基神经网络为次级模型的融合模型预测未来6 h累计降水等级,并应用公平TS评分(Equal Threat Score,ETS)、真实技巧评分(True Skill Statistic,TSS)、混淆矩阵、预报偏差(Bias值)、击中率(Probability of Detection,POD)对预报结果进行检验分析。结果表明:优化变量的输入有利于提高模型的准确率;Catboost模型和融合模型都可以在一定程度上辨别晴雨状况;仅非动力学变量参与的融合模型对雨区预报准确率最高,但容易将暴雨雨区预报得更加广泛。总体而言,融合模型具有更强、更稳定的预报性能,中到暴雨量级预报准确率还待进一步提高。

Short-range forecast of precipitation over the middle-lower reaches of the Yangtze River based on the Catboost and Stacking model

The Catboost model and the Stacking model which has Catboost and Random Forest(RF) as the primary predictive model, a secondary predictive model was developed using Radial Basis Function(RBF) to predict 6 h accumulated precipitation-grade over the middle-lower reaches of the Yangtze River based on data from the hourly precipitation data merged by China Automatic Weather Station(CAWS) and CMORPH precipitation products and the Global Forecasting System(GFS) reanalysis. Comprehensively considering ETS score, TSS score, confusion matrix, Bias and POD (Probability of Destection) were analyzed the precipitation-grade forecast ability of two models. The results showed that: optimized the input variables is beneficial to improve the accuracy of the models; both the Catboost model and the Stacking model can distinguish clear and rainy conditions to a certain extent; the Stacking model with only non-dynamic variables has the highest prediction accuracy for the rain areas, but it can overly predict the heavy rain areas more widely. Above all, the Stacking model has stronger and more stable forecasting performance, however the prediction accuracy of moderate to heavy rainfall needs to be further improved.