Unlocking calibration periods: energy clustering with white-box models

Building performance simulation is crucial for achieving building energy efficiency. These simulation tools allow stakeholders to evaluate the effectiveness of Energy Conservation Measures (ECMs) and implement effective energy-saving strategies. However, considerable studies highlight a discrepancy between simulated performance and real-world results. This gap, known as the performance gap, has been addressed by several researchers, leading to model calibration methods to bridge the gap between simulation and reality.

One approach involves calibrating white-box models. These models incorporate building physics, offering insights into the building’s energy performance. However, calibrating white-box models can be complex due to the numerous parameters involved. A significant contributor to the performance gap in white-box models is the inaccuracy of the building envelope definition. This inaccuracy can stem from poor or lack of detailed material information.

This gap can be addressed through envelope-focused calibration during free oscillation periods. However, identifying these periods can be challenging in buildings without a fixed heating, ventilation and air conditioning (HVAC) operation schedule, as occupant needs dictate energy use patterns. Several studies have attempted to predict occupant behaviour through algorithms to determine indoor and outdoor conditions that influence specific behaviours.

This research, in contrast, explores using a Building Energy Model (BEM) to analyse monitored indoor temperature data. The hypothesis is that the energy required by the model to maintain a specific indoor temperature can be clustered to identify the HVAC system’s operation schedule. This methodology proves to be cost-effective as it relies on indoor temperature sensors and a non-calibrated model, requiring minimal sensor deployment and a simpler model.

Despite relying on a non-calibrated model, this methodology maintains a 90% accuracy or higher in identifying free oscillation periods compared to measured data. However, it may not capture all available free oscillation timesteps, which translates to an average 25% reduction in calibration space. Nonetheless, this methodology offers valuable insights for calibration, paving the way for further studies with a fully calibrated model.