Enhancing Convective Wind Prediction: Two Machine Learning Approach with Multi-Regime Flow Analysis and Adaptive Model Integration

This study explores the initiation mechanisms of convective wind events, emphasizing their variability across different atmospheric circulation patterns. Historically, the inadequate feature categorization within multi-faceted forecast models has led to suboptimal forecast efficacy, particularly for events in dynamically weak forcing conditions during the warm season. To improve the prediction accuracy of convective wind events, this research introduces a novel approach that combines machine learning techniques to identify varying meteorological flow regimes. Convective winds (CWs) are defined as wind speeds reaching or exceeding 17.2 m s^–1 and severe convective winds (SCWs) as speeds surpassing 24.5 m s^–1. This study examines the spatial and temporal distribution of CW and SCW events from 2013 to 2021 and their circulation dynamics associated with three primary flow regimes: cold air advection, warm air advection, and quasi-barotropic conditions. Key circulation features are used as input variables to construct an effective weather system pattern recognition model. This model employs an Adaptive Boosting (AdaBoost) algorithm combined with Random Under-Sampling (RUS) to address the class imbalance issue, achieving a recognition accuracy of 90.9%. Furthermore, utilizing factor analysis and Support Vector Machine (SVM) techniques, three specialized and independent probabilistic prediction models are developed based on the variance in predictor distributions across different flow regimes. By integrating the type of identification model with these prediction models, an enhanced comprehensive model is constructed. This advanced model autonomously identifies flow types and accordingly selects the most appropriate prediction model. Over a three-year validation period, this improved model outperformed the initially unclassified model in terms of prediction accuracy. Notably, for CWs and SCWs, the maximum Peirce Skill Score (PSS) increased from 0.530 and 0.702 to 0.628 and 0.726, respectively, and the corresponding maximum Threat Score (TS) improved from 0.087 and 0.024 to 0.120 and 0.026. These improvements were significant across all samples, with the cold air advection type showing the greatest enhancement due to the significant spatial variability of each factor. Additionally, the model improved forecast precision by prioritizing thermal factors, which played a key role in modulating false alarm rates in warm air advection and quasi-barotropic flow regimes. The results confirm the critical contribution of circulation feature recognition and segmented modeling to enhancing the adaptability and predictive accuracy of weather forecast models.