The rapid digitalization of the Architecture, Engineering, Construction, and Operations (AECO) industry has introduced Digital Twin (DT) technology as a transformative approach for life cycle management. However, bridging the gap between static Building Information Models (BIM) and dynamic operational data, particularly occupant behaviour, remains a challenge in public infrastructure. This thesis proposes an AI-enhanced DT framework designed to integrate energy performance with behavioural analytics. The methodology utilizes a systematic data-driven approach, leveraging the Building Data Genome Project 2 (BDG v2) dataset. Four machine learning algorithms, XGBoost, LightGBM, Random Forest (RF), and Multi-Layer Perceptron (MLP), were implemented and compared to predict energy consumption patterns. A significant contribution of this research is the transition from a Global Model to a localized Autoregressive approach, mitigating the "memorization" effect of non-physical features and enhancing prediction accuracy for individual buildings. Furthermore, eXplainable AI (XAI) via SHAP values was employed to interpret the impact of non-human and human-related factors on energy loads. The framework was operationalized by integrating the trained models into Autodesk Tandem, providing a 3D visual interface for predictive energy management. The results demonstrate that the proposed DT framework enables facility managers to shift from reactive to proactive strategies, potentially reducing energy waste in educational facilities. This research underscores the necessity of human-centric data in DT environments to achieve meaningful sustainability goals.
La rapida digitalizzazione del settore dell'Architettura, dell'Ingegneria, delle Costruzioni e delle Operazioni (AECO) ha introdotto la tecnologia del Digital Twin (DT) come approccio trasformativo per la gestione del ciclo di vita. Tuttavia, colmare il divario tra i Modelli Informativi dell'Edificio (BIM) statici e i dati operativi dinamici, in particolare il comportamento degli occupanti, rimane una sfida nelle infrastrutture pubbliche. Questa tesi propone un framework DT potenziato dall'intelligenza artificiale, progettato per integrare le prestazioni energetiche con l'analisi comportamentale. La metodologia utilizza un approccio sistematico guidato dai dati, sfruttando il dataset Building Data Genome Project 2 (BDG v2). Sono stati implementati e confrontati quattro algoritmi di machine learning, XGBoost, LightGBM, Random Forest (RF) e Multi-Layer Perceptron (MLP), per prevedere i modelli di consumo energetico. Un contributo significativo di questa ricerca è la transizione da un Modello Globale a un approccio Autoregressivo localizzato, mitigando l'effetto di 'memorizzazione' delle caratteristiche non fisiche e migliorando l'accuratezza della previsione per i singoli edifici. Inoltre, è stata impiegata l'eXplainable AI (XAI) tramite i valori SHAP per interpretare l'impatto dei fattori umani e non umani sui carichi energetici. Il framework è stato reso operativo integrando i modelli addestrati in Autodesk Tandem, fornendo un'interfaccia visiva 3D per la gestione predittiva dell'energia. I risultati dimostrano che il framework DT proposto consente ai facility manager di passare da strategie reattive a proattive, riducendo potenzialmente gli sprechi energetici nelle strutture scolastiche. Questa ricerca sottolinea la necessità di dati antropocentrici negli ambienti DT per raggiungere significativi obiettivi di sostenibilità.
Optimizing campus energy efficiency: a low-cost AI-driven digital twin framework demonstrating the primacy of human behavior
Torabian, Negar
2025/2026
Abstract
The rapid digitalization of the Architecture, Engineering, Construction, and Operations (AECO) industry has introduced Digital Twin (DT) technology as a transformative approach for life cycle management. However, bridging the gap between static Building Information Models (BIM) and dynamic operational data, particularly occupant behaviour, remains a challenge in public infrastructure. This thesis proposes an AI-enhanced DT framework designed to integrate energy performance with behavioural analytics. The methodology utilizes a systematic data-driven approach, leveraging the Building Data Genome Project 2 (BDG v2) dataset. Four machine learning algorithms, XGBoost, LightGBM, Random Forest (RF), and Multi-Layer Perceptron (MLP), were implemented and compared to predict energy consumption patterns. A significant contribution of this research is the transition from a Global Model to a localized Autoregressive approach, mitigating the "memorization" effect of non-physical features and enhancing prediction accuracy for individual buildings. Furthermore, eXplainable AI (XAI) via SHAP values was employed to interpret the impact of non-human and human-related factors on energy loads. The framework was operationalized by integrating the trained models into Autodesk Tandem, providing a 3D visual interface for predictive energy management. The results demonstrate that the proposed DT framework enables facility managers to shift from reactive to proactive strategies, potentially reducing energy waste in educational facilities. This research underscores the necessity of human-centric data in DT environments to achieve meaningful sustainability goals.| File | Dimensione | Formato | |
|---|---|---|---|
|
Final thesis_Negar Torabian_271075.pdf
non accessibile
Descrizione: Final thesis document
Dimensione
3.58 MB
Formato
Adobe PDF
|
3.58 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/259719