<![CDATA[Operationally used early warning systems for significant weather generally rely on conventional weighting-scoring schemes, namely, fixed thresholds set based on empirical experience or expert meteorological consensus. The fundamental limitation of conventional threshold-based methods and supervised classification models lies in their reliance on labeled data, while extreme weather events are naturally rare and imbalanced in historical station observation records. Such approaches are prone to generating false alarms. Therefore, a multilayered unsupervised learning approach is needed to objectively classify upper-air thermodynamic patterns without relying on subjective extreme-event labels. Although widely used, the K-Means algorithm has a fundamental weakness, namely its high sensitivity to the initial centroid determination. This study applies hybrid metaheuristic techniques with K-Means, namely GA-KMeans, PSO-KMeans, and GWO-KMeans, within a consistent evaluation framework, using the same cluster validity index, and applied to the operational meteorological domain. The test results show that the best fitness values are produced by the PSO-KMeans model with a WCSS value of 250.6012, followed by GWO-KMeans with a WCSS value of 304.7890, and finally the GA-KMeans model with a WCSS value of 330.0701. The third model used also consistently produces higher Silhouette Score values compared to the classical K-Means baseline and shows that metaheuristic hybridization is proven to be effective in improving the quality of cluster structures. Specifically within the field of computer science, the results demonstrate that employing appropriate cluster center optimization techniques can improve both clustering quality and resource efficiency when grouping various types of data.]]>
