Abstract:
Digital Twins have significant potential in managing building energy systems, reducing operational risks, improving energy efficiency, and achieving decarbonisation goals. Driven by data and models, digital twins can perform monitoring, simulation, prediction, and optimisation. At the core of a digital twin lies its modelling capability, which precisely replicates the physical entity, enabling it to deliver functional services and meet application requirements. Present modelling approaches often depend on physics-based models, such as those found in the TRNSYS and Modelica IBPSA library, which can be complex, computationally intensive, and require substantial physical knowledge. In contrast, purely data-driven black-box models often lack physical interpretability and may not effectively capture the underlying dynamics. This work reconstructs the deep Gaussian process emulators for digital twinning in building energy systems. This method connects individual data-driven component models based on their physical associations, incorporating uncertainty propagation to quantify predictive uncertainties. Results demonstrate that it significantly reduces computational costs by refining emulators of individual submodels based on their varying functional complexities. The results underscore the effectiveness of advancing digital twin models for risk-aware decision-making within building energy systems, thus providing a viable pathway to improve overall energy performance and achieve building decarbonisation goals.