Comparison of machine learning algorithms for building performance evaluation

Predicting building performance is crucial for achieving energy efficiency and occupant comfort. Traditional building performance simulation, while comprehensive, is often computationally intensive and time-consuming. Machine learning (ML) offers a potential solution by accelerating prediction processes and enabling rapid design iterations. This study evaluates the efficacy of various ML algorithms in predicting critical building performance metrics: Energy Use Intensity (EUI), Percentage of Comfort Hours (PCH), Spatial Daylight Availability (sDA300,50%), and Annual Sunlight Exposure (ASE1000,250h).

A comparative analysis was conducted on Random Forest, Epsilon-Support Vector Machine, K-nearest Neighbors, Gradient Boosting, and Extreme Gradient Boosting. The dataset comprised performance data from typical Indonesian office buildings, based on 18 passive design variables and four distinct climate types within the country. Models were trained and tested to accurately predict the aforementioned performance metrics. Evaluation metrics including R² and computational time assessed surrogate model performance.

Results indicate that Extreme Gradient Boosting outperformed other algorithms in all considered performance metrics, achieving an R² score exceeding 0.8 while reducing calculation time by 25 times compared to traditional methods. These findings highlight the potential of ML as a powerful tool for enhancing building design and operation through rapid and reliable performance estimates.