This research proposes a data-driven model situated at the intersection of occupancy, sustainable design and building utilisation. The model is designed to assess the sustainability performance of the existing campus building stock while offering predictive insights for future developments. A university campus was selected as a controlled “micro-city” environment to address the difficulties of collecting consistent city-scale data. The main objective is to demonstrate how occupancy, operational and spatial information can be integrated through machine learning (ML) to support a gradual and human-centred urban transformation process.
An open-access Green Campus Dataset comprising 2,138 data points representing environmental, spatial and socioeconomic variables was analysed. Following exploratory correlation analysis and feature engineering, seven ML algorithms – linear regression, random forest, gradient boosting, multilayer perceptron (MLP), K-nearest neighbours (KNN), support vector regression (SVR) and XGBoost – were comparatively evaluated for predicting energy savings. Model performance was assessed using R2, RMSE, MAE and MAPE, and XGBoost was selected as the best-performing model. Explainable AI techniques, and sensitivity analysis, were subsequently applied to interpret the relationships between key variables and sustainability performance. The resulting framework was used to generate building-level sustainability scores and support a cloud-based decision-support prototype.
XGBoost achieved the highest predictive performance among the evaluated ML models. Feature importance analysis identified Beam Irradiance as the most influential predictor of energy savings, with an importance score of 74.45%, while explainability analyses further demonstrated the influence of energy efficiency, solar utilisation and occupancy-related variables on sustainability performance. The findings indicate that campus sustainability emerges from the interaction of environmental, operational and occupancy-related factors rather than from technical characteristics alone. The proposed framework demonstrates the potential of campus-scale ML to support human-centred, data-driven sustainability assessment and decision-making.
The study introduces a human-centred ML framework for campus-scale sustainability assessment, beginning with the comparative evaluation of seven supervised ML models to identify the most effective predictive approach. By integrating environmental, operational and occupancy-related variables with explainable AI techniques, the framework extends conventional performance-based assessment towards a more interpretable socio-technical approach. Treating the university campus as a bounded urban laboratory, the study demonstrates how machine learning and sustainability scoring can support data-driven decision-making and provide a scalable foundation for future smart and sustainable urban applications.
