Building energy prediction in a changing climate: an interpretable machine learning approach

Currently, the consideration of climate change and the interpretation of machine learning-based building energy prediction models are often limited to a specific building function. This study extends the interpretable machine learning approach to various building functions while considering the impacts of climate change. To generate a cooling energy use intensity dataset, Latin hypercube sampling, and physical method simulations are integrated, with future climate data morphed from an ensemble of Coupled Model Intercomparison Project Phase 6 results. Four machine learning models are developed for cooling energy prediction and subsequently interpreted using SHapley Additive exPlanations. The results from the Haikou case study demonstrate that machine learning models offer accurate and exponentially faster alternatives to computer dynamic simulation software for cooling energy prediction. Support vector regression exhibits the highest accuracy, achieving normalized mean absolute error, normalized root-mean-square error, mean absolute percentage error, and coefficient of determination of 6.29 %, 11.64 %, 6.52 %, and 95.03 %, respectively. Feature importance rankings vary while maintaining consistency across different models. The cooling coefficient of performance, lighting gain, compactness ratio, and window-towall ratio are essential features for a building’s cooling energy use intensity, while the U-values of the wall, roof, and window are among the least influential. All models perceive climate change as having a growing positive influence on buildings’ cooling energy use intensity. The interpretable machine learning models developed in this study provide valuable tools for building design and retrofit, empowering building professionals and the general public to pursue sustainable buildings for the future.