Abstract
Quantifying building thermal performance under uncertainty remains a major computational challenge, particularly when traditional Monte Carlo (MC) simulations are applied to large-scale probabilistic analyses. To address this limitation, this study develops a scalable and generalizable data-driven stochastic framework that accelerates MC-based convergence analysis for building energy modeling with a focus on low-to-medium-sized commercial buildings in six cities across Europe to test the method. The proposed approach combines Latin Hypercube Sampling with an optimized tree-based meta-model ensemble tuned via Bayesian Optimization, replicating MC convergence behavior with up to 95–99% reduction in computation time while maintaining high accuracy (average R2 > 0.90). By using 700 MC baseline samples, more than 600,000 stochastic samples are generated to demonstrate the framework's ability to support high-resolution uncertainty and sensitivity analyses at scales impractical for direct MC simulations. Using a 0.1% Coefficient of Variation threshold, the study reveals that convergence characteristics of key output percentiles (5th, 50th, and 95th) vary systematically with building typology, climate, and scenario. Sensitivity analysis identifies the solar heat-gain coefficient of south-facing glazing as the dominant factor influencing future cooling demand, while persistent heating demand in cold climates highlights the need for balanced adaptation strategies. Overall, the proposed framework constitutes a robust and transferable methodology for efficient stochastic assessment of building thermal performance under climate uncertainty, offering a foundation for robust energy policy and design decisions.