High-Resolution Building Energy Model Bayesian Calibration Framework

This project aims to develop a framework that enables the Bayesian calibration of a building energy simulation model with hourly or sub-hourly historical building operational data.

Status: Current

Research themes: Energy and GHG emissions Health and comfort

Research areas: Intelligent/ Smart and Interactive Buildings; Applied AI and ML Technologies for buildings; HVAC control and component characterization and optimization

Project Objective

This project aims to develop a framework that enables the Bayesian calibration of a building energy simulation model with hourly or sub-hourly historical building operational data.

Approach

There have been efforts to automate the calibration process by applying data-driven techniques. However, previous methods focus on finding deterministic point estimates of model parameters, which fit the data best. Consequently, the methods often suffer from overfitting, and the estimates do not necessarily represent true building properties. To solve the problem, we develop the framework with efficient Bayesian inference methods that quantify uncertainty due to a lack of data instead. To improve computational efficiency, we utilize deep-learning-based surrogate modelling techniques.

Findings

We expect that the developed framework will ensure reliable and robust model calibration even under a significant level of uncertainty (e.g., unknown occupancy information) and even with the inherent imperfection of a simulation tool (e.g., limitation of 1-D heat diffusion model in most energy simulation tools). We plan to evaluate and demonstrate the framework through a series of simulation and experimental studies. We believe that this framework will be a step toward seamless knowledge-data integration that will enable the development of robust and scalable AI solutions for building operations.