Building energy model calibration is commonly performed using dynamical fidelity criteria, such as statistical goodness-of-fit metrics, to assess agreement between simulated and measured data. However, the extent to which such criteria ensure model re...
Building energy model calibration is commonly performed using dynamical fidelity criteria, such as statistical goodness-of-fit metrics, to assess agreement between simulated and measured data. However, the extent to which such criteria ensure model reliability under hypothetical decision-making contexts remains unclear. This thesis examines the limitations and potential unreliability of traditional calibration approaches that focus primarily on dynamical fidelity.
An EnergyPlus model of a complex office building was developed using a highly automated, parametric modeling workflow implemented through Rhino, Grasshopper, and Python. Model calibration was conducted using a sampling-based approach, enabling efficient exploration of high-dimensional parameter spaces and repeatable sample selection through parallel computation. Conventional dynamical fidelity metrics, including daily and hourly CVRMSE and MAE, were used to select calibrated samples on a weekly basis, which were then employed to generate short-term forecasts.
To evaluate the robustness of these calibrated models, five fit-for-purpose what-if scenarios were designed to expose the models to varying contextual conditions representative of realistic operational and decision-making environments. The results demonstrate that marginal differences in calibration metrics can lead to substantially divergent predictions under different conditions, indicating that satisfaction of dynamical fidelity criteria alone does not guarantee representational fidelity. Even samples identified as “overall best” through retrospective evaluation exhibited unreliable performance across several what-if scenarios.
It can be said that conventional error tolerance-based calibration secures dynamical fidelity but not representational fidelity, and dynamical fidelity is a necessary but fundamentally insufficient condition for model reliability. Evaluating models not only under as-is conditions but also under counterfactual, purpose-driven what-if scenarios is essential for establishing the structural adequacy and reliability for decision-making objectives.