CH-24-C071 - Calibration of Chiller Component Model with Limited Operational Data: The Use of Manufacturer’s Catalog Data via Bayesian Principle

This research study introduces a new method for calibrating a chiller component model using limited operational data and the manufacturer's catalog data through the Bayesian principle. Calibration of component models, such as electric chillers, plays a crucial role in the development of model-based applications, including predictive control or retrofit analysis for HVAC systems. Traditional approaches rely on the manufacturer's catalog data or operational data to construct a performance map due to the complex and dynamic nature of HVAC components. However, these approaches have limitations when the HVAC component deviates from the catalog performance due to aging, or faults, or when the component model is used for predictions beyond the training data range for predictive control applications. In this study, we propose a new calibration approach that combines the manufacturer's catalog data and operational data. Operational data is primarily used to calibrate model parameters, while the catalog data provides prior information for the calibration process using the Bayesian principle. This method was tested using air-cooled chiller data obtained from a laboratory environment, considering the effect of the amount of operational data. The proposed approach yields improved and more reliable results, reducing the root mean square error (RMSE) from 0.27 kW to 0.03 kW and the coefficient of variation of the RMSE (CV(RMSE)) from 9.9% to 1.3% compared to the traditional approach in power prediction, with data from only 7 different operating conditions.