Extensive development of a Bayesian calibration approach for building energy models using an innovative case study: a shipping container building.

The building sector is responsible for almost a third of global energy consumption and a quarter of CO2 emissions. Innovative architectural designs that promote the reusability of raw materials, such as shipping container architecture, can help to reduce construction’s environmental impact. By creating physical models to analyze energy consumption, we can develop practical tools to explore other ways to minimize the environmental impact throughout the building’s life cycle. Calibrating these models is essential since some parameters remain uncertain.
Bayesian calibration [1] is an effective automatic method that naturally considers parameter uncertainty and prior knowledge. However, it still needs to be improved with new case studies to optimize energy consumption. This study introduces an innovative real case study: a modular scholar building based on shipping containers (also called Smart Building) with over 150 sensors and actuators, enabling the collection of a wide array of detailed data [2]. Based on the data describing the physical behavior and the operating conditions, we developed an energy model based on EnergyPlus software. Despite our comprehensive understanding of this building, the physical
model exhibits uncertain parameters.
This work provides a statistical framework to enhance existing methods for Bayesian calibration of building energy models. We aim to expand the current calibration approach by introducing
new mathematical tools and thoroughly exploring different configurations of this approach in both simulated and real case studies. The simulated case study involves an energy model of an artificial reference building provided by the Department of Energy (USA). This synthetic case study offers a fully manageable testbed to evaluate the capabilities of the approach. The real case study, in turn,
relies on data collected from the Smart Building in 2023 (for the training dataset) and 2024 (for the validation dataset) and permits the demonstration of the usability of the approach on original data.
One of the initial steps in this calibration approach involves parameter screening or sensitivity analysis (SA) to reduce the problem’s dimensionality. This way, only the most influential parameters
will be calibrated. We compared three different SA methods (Sobol, Morris, EASI-RDB-FAST) and evaluated their convergence and computational cost. Our findings indicate that EASI-RDBFAST
[3], previously unused in this field, is the most suitable, as it provides quantitative information while maintaining reasonable computational time, unlike the other methods.
The Bayesian framework requires an extensive number of model evaluations. Simulating the physical model remains costly even with a reduced set of uncertain parameters. Calibrating the building
energy model involves replacing the physical model with a less costly mathematical one or meta-model [4]. The most common meta-model is the nonparametric Gaussian Process (GP) emulator,
which performs well but at a specific cost. In contrast, the Multiple Linear Regression (MLR) model is less accurate but cost-effective, which might be suitable for continuous calibration. By
fixing the uncertain parameters in the simulated case study, we can demonstrate that the GP model is better at estimating the parameters than the MLR, even if both provide accurate predictions.
Predictions are also very satisfying in the real case study for both meta-models, even though the discrepancy between the estimates of each meta-model is open to discussion. We also emphasize
that using MLR is not straightforward as it requires some statistical conditions before performing the Bayesian inference on real data. We demonstrated that certain transformations of the MLR model can address this issue when requirements are initially not met. This issue is rarely discussed in current literature.
Finally, we are improving the current approach by considering a different time scale for calibration.
After conducting a monthly sensitivity analysis, we found that the most influential parameters change throughout the year due to seasonal variations. Our preliminary findings show that we can
at least discriminate two periods over a year: from October to March (Winter) and from April to September (Summer). This outcome calls into question the relevance of calibrating only one annual energy model that doesn’t fit the seasonal dynamics. In our ongoing exploratory work, we suggest an alternative to yearly calibration. Our goal is to develop multiple models based on the outcomes of sensitivity analysis and to train these models using more targeted datasets.

References:
[1] Adrian Chong et al. “Bayesian Calibration of Building Energy Models with Large Datasets”.
In: Energy and Buildings 154 (Nov. 2017), pp. 343–355. issn: 0378-7788. doi: 10.1016/j.
enbuild.2017.08.069.
[2] Pierre-Antoine Cormier et al. “Dataset of an operating education modular building for simulation
and Artificial Intelligence.” In: Data in Brief (2024), p. 110889.
[3] Jeanne Goffart and Monika Woloszyn. “EASI RBD-FAST: An Efficient Method of Global
Sensitivity Analysis for Present and Future Challenges in Building Performance Simulation”.
In: Journal of Building Engineering 43 (Nov. 2021), p. 103129. issn: 23527102. doi: 10.1016/
j.jobe.2021.103129.
[4] Hyunwoo Lim and Zhiqiang John Zhai. “Comprehensive Evaluation of the Influence of Meta-
Models on Bayesian Calibration”. In: Energy and Buildings 155 (Nov. 2017), pp. 66–75. issn:
0378-7788. doi: 10.1016/j.enbuild.2017.09.009.