An evolving learning method —growing Gaussian mixture regression—for modeling passive chilled beam systems in buildings

Abstract

Although chilled beam systems are considered as one of the best performing HVAC systems, few studies have looked into data-driven-based modeling for chilled beam systems. Furthermore, no learning-based data-driven methods have been applied to chilled beam systems in buildings. In this study, we used an evolving learning method, growing Gaussian mixture regression (GGMR), to predict cooling rates for passive chilled beam (PCB) systems where the training, evolution, and validation were carried out using data from real system measurements and from building energy simulation. GGMR updates key parameters such as weight coefficients, means, and covariance matrices of Gaussian components to adapt to changes in system operation beyond training data. This case study demonstrated that GGMR is an effective evolving learning-based data-driven method for accurately predicting cooling rates of PCB systems. The selection of key performance parameters of GGMR models including the number of components, training data size, and the learning rate was discussed in this paper. It is recommended that GGMR models could be further explored for predicting the performance of other complex HVAC systems such as radiant slab or mixed-mode ventilation systems.