Cross-scenario evaluation of explainable machine learning for non-invasive summer occupancy detection across five...

Accurate identification of indoor occupant presence is crucial for intelligent building energy management. Traditional monitoring methods are often invasive and lack standardized quantitative indicators. Therefore, this study proposes a non-invasive occupant presence state prediction method based on machine learning. Six machine learning algorithms-Logistic Regression, Decision Tree, k-Nearest Neighbor (KNN), Random Forest, CatBoost, and XGBoost-were evaluated across five building scenarios (hospitals, classrooms, dormitories, offices, and dwelling houses) using core environmental features including air temperature, relative humidity, sound pressure level, illuminance, CO₂ concentration, formaldehyde (HCHO), PM_2.5, PM_1.0, and PM₁₀. A temporally ordered walk-forward evaluation framework was adopted to prevent data leakage, with VIF screening and RFECV for feature selection. The best-performing models achieved mean accuracies ranging from 0.61 (dwelling) to 0.88 (office), with the optimal algorithm varying by scenario. SHAP-based interpretability analysis identified CO₂ concentration, illuminance as the most consistently influential predictors, with PM_2.5, temperature, and sound contributing in scenario-specific patterns. A leave-one-feature-out sensitivity analysis showed that removing CO₂ caused the largest performance drop in the hospital scenario. This study provides a comparative analysis of explainable machine-learning predictors across five single-room building scenarios under a temporally aligned protocol, providing comparative observations that may inform subsequent multi-room and closed-loop studies of smart building management.