Building energy loads prediction using bayesian-based metaheuristic optimized-explainable tree-based model

The study presents a sophisticated hybrid machine learning methodology tailored for predicting energy loads in occupied buildings. Leveraging eight pivotal input features—building compactness, surface area, wall area, roof area, overall height, orientation, glazing area, and glazing area distribution—we elucidate the intricate relationships between building characteristics and their corresponding heating load (HL) and cooling load (CL). We meticulously analyze these features across 12 diverse structural forms, each emblematic of unique architectural designs and building materials. Using a dataset encompassing 768 buildings, we demonstrate the prowess of our proposed models. Among the algorithms we employed, the extreme gradient boosting algorithm stands out, registering impressive accuracy metrics (HL: RSQ = 0.9986, RMSE = 0.3797, MAE = 0.2467 and MAPE = 1.1812; CL: RSQ = 0.9938, RMSE = 0.7578, MAE = 0.4546 and MAPE = 1.6365). We further integrate SHAP analysis, revealing that relative compactness positively influences both HL and CL the most, closely followed by surface area and glazing area. By merging an explainable extreme gradient boosting algorithm with a Bayesian-based metaheuristic optimization technique, we ensure both high predictive accuracy and interpretability. This study holds profound implications for enhancing building energy efficiency, curbing waste, and championing the shift to sustainable energy sources, aligning seamlessly with SDG 7.