Towards Smart-Building Digital Twin: Data Integration, and Probabilistic Frameworks for Reliable Virtual Sensing, and Continuous Model Calibration
Abstract
Buildings responsible for around 40% of global energy use and a third of emissions are increasingly scrutinized for their potential to lower energy consumption and emissions. Consequently, adopting advanced technologies such as IoT and AI in the building sector becomes crucial for more efficient and sustainable operations. Despite ongoing efforts, the rising trend emphasizes the need for more advanced, holistic technologies. In light of this need, smart-building digital twins have emerged, marking a significant shift toward building digitization.
Smart-building digital twins employ data, information, and models to mirror static and dynamic aspects of buildings virtually. This approach establishes a bidirectional connection between physical buildings and their digital counterparts, enhancing performance with real-time monitoring, autonomous control, and visualization.
While gaining momentum, adopting smart-building digital twins faces several challenges, such as achieving data interoperability, overcoming limited sensing environments, and continuously calibrating physics-based models with real-time sensor measurements. Therefore, this thesis tackles these three significant challenges to advance research in smart-building digital twins.
Data interoperability challenges in smart buildings stem from heterogeneous, siloed data sources, impeding unified access for smart-building digital twin applications. This research presents a novel multi-layer architecture and method for integrating BIM and IoT data using domain ontologies, achieving unified access and semantic interoperability. Evaluations using actual building data demonstrated the effectiveness of these approaches in providing a robust data backbone for digital twin applications.
Sensing limitations in buildings arise from the absence or malfunction of sensors and the difficulty of measuring certain variables. This thesis presents a novel probabilistic virtual sensor framework to estimate unobservable variables using existing sensor data while offering confidence measures in these estimates. Evaluation with simulated building data proved these models' accuracy, efficiency, and reliability in addressing building sensing limitations.
The challenge in continuously calibrating physics-based models is the unobservability of influential model inputs. This thesis introduces a novel framework and calibrator model for continuously calibrating physics-based models while quantifying uncertainty and enabling multi-variable calibration under missing or noisy sensor data. The evaluation results confirmed that the proposed calibrator model accurately synchronized the outputs of physics-based models, meeting standard error thresholds for calibration accuracy.