A Bayesian optimized LSboost framework for energy usage intensity prediction in a university office building

Accurately predicting Energy Use Intensity (EUI) has become increasingly important in efforts to enhance energy efficiency and sustainability in buildings. This study aims at comparing the performance of three machine learning approaches, namely, Baseline Ensemble, Auto Hyperparameter Optimized Ensemble and Bayesian Optimized Ensemble, using real world sensor data collected from four zones of a university building in Thailand via a Building Energy Management System. The models are tested on both un-normalized and min-max normalized datasets to examine how data preprocessing influences prediction accuracy, error reduction, and training efficiency. The results of the study demonstrate that normalization improves prediction precision by significantly reducing mean absolute error and mean squared error values, although it has a limited effect on R 2 values. Among the three approaches, the Bayesian Optimized model trained on normalized data provides the most accurate and stable results while maintaining reasonable training times. These results highlight the practical value of integrating normalization and automated tuning when designing building energy models. The proposed Least Squares Boosting (LSBoost) – Bayesian Optimization model in the study offers a reliable and adaptable tool for forecasting Energy Use Intensity, with potential applications in real-time control, diagnostics, and long-term energy planning. Practical application This study presents a robust, data-driven framework for accurately forecasting Energy Use Intensity (EUI) in real-world building operations using Bayesian-optimized LSBoost models. By integrating indoor sensor data, external weather variables, and advanced machine learning, the proposed method supports energy managers, building operators, and HVAC control engineers in enhancing predictive maintenance, operational efficiency, and real-time energy management. The approach is particularly suited for smart building systems and retrofitting strategies that require scalable, accurate, and resource-efficient energy modeling under variable occupancy and environmental conditions.