Modern power systems’ growing complexity demands precise load-curve clustering for effective demand response and dispatch. However, ISODATA’s reliance on random initialization and Euclidean distance undermines both stability and clustering accuracy. This paper proposes an improved ISODATA framework that adopts a farthest-point probabilistic sampling strategy for initial center selection, leverages a composite kernel function to embed load curves into a high-dimensional space capturing both global and local features, and incorporates adaptive split–merge operations guided by kernel-based distances. Experimental results indicate that, compared with traditional clustering methods, the improved ISODATA algorithm achieves significantly superior performance on both the Davies–Bouldin and Dunn indices, demonstrating higher clustering quality and delivering more reliable, interpretable classifications for power-load pattern recognition.