aspg-22-4426-1791483625 20261008182025 American Scientific Publishing Group admin@americaspg.com American Scientific Publishing Group International Journal of BIM and Engineering Science IJBES 2571-1075 2026 12 2 Anwar Shanwan Doctor in Mechanical Engineering, Researcher at University of Orléans, France Mariam Altaema PhD Candidate, Faculty of Natural Sciences and Engineering, Ankara Yıldırım Beyazıt University, Ankara, Türkiye Raphaël Omran Mechanical Engineering, University of Rennes, France Buildings account for nearly 40% of global final energy consumption, with heating, ventilation, and air conditioning systems responsible for the largest single share of that load. Conventional schedule-based HVAC controllers operate on fixed occupancy assumptions and are consequently unable to exploit the predictable but irregular occupancy patterns that characterise modern working environments. This paper proposes OccuTwin, an integrated framework that couples multi-step occupancy forecasting with a BIM-based digital twin to enable genuinely predictive HVAC optimisation. Four sequence models—Long Short-Term Memory, Gated Recurrent Unit, XGBoost with lag features, and a Temporal Fusion Transformer—are trained on the publicly available Mind Your Building occupancy dataset and the multi-building Building Data Genome Project benchmark to predict room-level occupancy at five-minute resolution. The best-performing Transformer model achieves 94.8% accuracy and an improved weighted F1- score on the 30% hold-out set, outperforming the LSTM baseline by 1.4 percentage points. An IFC-coupled co-simulation environment links real-time occupancy predictions to a virtual HVAC thermal model implemented in EnergyPlus, enabling zone-level set-point optimisation driven by predicted rather than sensed occupancy. Annual co-simulation across nine building zones documents 21.2% energy savings over a schedule-based rule controller while simultaneously improving the fraction of time within ASHRAE thermal comfort bounds from 92.4% to 97.2%. Ablation experiments identify temporal lag features and SMOTE-based class rebalancing as the two most critical preprocessing choices, and noise-injection tests confirm that the Transformer retains 95.1% accuracy under 15% sensor noise—a critical property for edge-deployed IoT environments. 2026 66 80 4426 https://creativecommons.org/licenses/by/4.0/ 10.54216/IJBES.120209 https://www.americaspg.com/journal/22/article/4426 https://www.americaspg.com/storage/articles/manuscripts/01KVKAABRGCRJ6W1FFM45BEY1X.pdf