Approach of using machine learning to optimize office building thermal and economy performance based on a high-resolution shading behavior model

Traditional building performance simulation-driven optimization methods face bottlenecks, including high computational costs and difficulties with large-scale iterations. This study develops and validates a building performance prediction model based on a high-resolution shading behavior model using machine learning techniques. By integrating the NSGA-II algorithm, we established a multi-objective optimization framework for office buildings that explicitly accounts for stochastic shading behavior. The introduction of this machine learning surrogate model significantly enhances optimization efficiency, enabling population evolution within a reasonable timeframe and ensuring thorough exploration and convergence of the Pareto front. The superiority of this approach is validated through comparisons with a static shading model and a comprehensive economic analysis. Results demonstrate that, compared to the baseline scenario, the stochastic-optimization method improves energy performance by 8.8% and achieves a 25% increase in optimization stability over static-model-based methods. Moreover, economic analysis confirms the framework’s viability, with a 9.5-year payback period and a 50-year net present value (NPV). Notably, while single-room uncertainties are high near the break-even point, they become negligible at larger building scales, ensuring robust economic feasibility. © 2026 Elsevier Ltd.