An Improved Two-Stage Dimensionality Reduction and Clustering Framework for Characterizing Renewable Energy Output

Renewable energy scenarios are widely used for power system stochastic optimization and risk evaluation, yet massive redundant scenarios boost computational complexity, waste resources and destabilize results, necessitating scenario reduction. This paper proposes an improved two-stage clustering method to overcome the manual parameter tuning defects of conventional clustering-based reduction. Specifically, the Snow Ablation Optimizer is embedded into DBSCAN to auto-adjust core hyperparameters, realizing efficient initial scenario reduction and suppressing the interference of abnormal data. Afterwards, K-means optimized via the Calinski–Harabasz index is adopted for secondary reduction to adaptively identify optimal cluster numbers and enhance the representativeness of reserved scenarios. Basic comparative simulations validate its technical superiority: handling 5000 raw scenarios only takes around 2 min, with the Wasserstein distance decreased by 7.73% and 17.79% versus backward reduction and forward selection, while the silhouette coefficient rises by 12.73% and Davies–Bouldin index drops by 7.95% compared with classic K-means. Further stochastic unit commitment tests quantify tangible economic and low-carbon gains in day-ahead scheduling, and empirical coefficient-based scaling analysis extends these benefits to long-term grid operation and policy deployment for TSOs/DSOs. The integrated results confirm the method’s technical, economic and sustainable merits, delivering actionable quantitative support for high-renewable power system low-carbon planning and energy policy formulation.