基于EOF_AP聚类模型的京津冀冬季强降水中期预报技术研究

Application of EOF_AP Clustering Model in Medium-Range Forecasting of Winter Heavy Precipitation Events in Beijing-Tianjin-Hebei

摘要:

基于欧洲中期天气预报中心（ECMWF）集合预报系统（IFS）的降水预报资料，采用EOF分解和近邻传播聚类（Affinity Propagation Clustering，简称AP）方法，构建了EOF_AP聚类模型，针对2019—2021年冬季京津冀地区的三次冬季强降水天气过程进行了中期预报特征的识别及归类研究，研究结果表明：一是EOF_AP聚类模型的聚类效果良好，能够准确识别并归类京津冀地区冬季强降水过程的集合预报累计降水量预报的主要类型及其发生概率，尤其能提取低发生概率的极端降水预报类型，聚类结果具有较高可信度；二是ECMWF集合预报系统对冬季极端强降水具有一定的预报能力，在提前168h的时效下，已有少量成员能够稳定预报出接近实况的极端降水预报类型，而在提前72h时效左右，模式预报出现转折，能够识别极端降水的成员数显著增加；三是集合预报对于降水落区的预报能力要高于降水强度的预报，表现在预报效果好的成员数更多，且预报时效更长。

Abstract:

Based on precipitation forecast data from the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble forecasting system (IFS), this study employs Empirical Orthogonal Function (EOF) and Affinity Propagation Clustering (AP) methods to construct the EOF_AP clustering model. The model is applied to analyze and classify three major heavy precipitation events in the Beijing-Tianjin-Hebei region during the winters of 2019–2021, focusing on medium-range forecast features. The results shows: firstly, the EOF_AP clustering model demonstrates excellent performance, effectively identifying and classifying the primary types of cumulative precipitation forecasts in ensemble models and their occurrence probabilities. It is particularly adept at extracting low-probability extreme precipitation forecast types with high clustering credibility. Then, the ECMWF ensemble forecasting system exhibits a certain capability for predicting extreme heavy precipitation in winter. Around 168 hours in advance, a small number of ensemble members start to stably predict extreme precipitation types close to actual observations. At approximately 72 hours in advance, a turning point occurs, and the number of members capable of identifying extreme precipitation increases significantly. Finally, the ensemble model shows greater predictability for precipitation location than for precipitation intensity, with more ensemble members providing accurate forecasts for longer lead times.