Scenario reduction is an essential process for solving large-scale two-stage stochastic optimization models, as the number of realizations directly affects complexity and computational burden. Traditional methods often focus primarily on statistical distributions between the original and reduced sets, or struggle with the trade-off between solution quality and computation time. This study proposes a novel scenario reduction approach termed K-means clustering with Nearest Economic Selection (K-NES). The methodology combines data space partitioning with a recourse evaluation phase, utilizing the expected value problem as a proxy to estimate costs and identify optimal economic medoids. Furthermore, a progressive iteration framework is presented to systematically scale the subset size. The proposed algorithm was implemented in Python using the Gurobi solver and tested on two benchmark cases: Aircraft Allocation (GBD) and Electricity Planning (LandS). Experimental results show that K-NES outperforms conventional techniques, such as Fast Forward Selection, K-means clustering, and Monte Carlo sampling, in both computational efficiency and solution quality. Even when scenarios were reduced to 1-4% of the sample size, the algorithm achieved relative gaps of 0.014% and 0.001% in GBD and LandS, respectively. Our approach requires approximately one second for the reduction step. Under the progressive iteration framework, K-NES met the stopping criterion of a relative gap of less than or equal to 0.001% in 95% and 90% of the evaluations for GBD and LandS, respectively. These findings highlight the effectiveness of K-NES as a practical and scalable tool for addressing uncertainty in large-scale operations.

Keywords

Two-stage stochastic optimization, Scenario reduction, K-means clustering, Cost-driven heuristic.