Improving building stock energy models: a bayesian calibration approach with multi-resolution data

Building Stock Energy Models (BSEMs) are essential for evaluating large-scale energy strategies, but face challenges in achieving accurate real-world predictions. Although Bayesian calibration effectively manages uncertainties in individual building models, its application to BSEMs is limited by the observational data limitations and multi-temporal scale uncertainties.

Traditional approaches that focus on calibrating only a few sensitive parameters and fail to capture all the uncertainties in complex building stocks. Using a Japanese commercial building archetype with open-source multi-temporal data, we demonstrate that while traditional calibration improves prediction accuracy, it forces unrealistic parameter values and creates inconsistent posterior distributions across different temporal scales.

Additionally, our results reveal fundamental limitations in current calibration frameworks, showing parameter identification instability when different numbers of parameters are included. These findings highlight the need for improved calibration frameworks that holistically address parameter interdependencies and temporal variations in BSEMs.