Physics-guided Gaussian process for HVAC system performance prognosis

Prognosis is crucial for tracking and predicting a system’s performance and provides the basis for predictive maintenance. Prognosis techniques are broadly classified into model-based and data-driven methods, which rely on physical knowledge and data learning, respectively. While data-driven methods alleviate limitations associated with model-based methods due to assumptions needed to simplify system complexity, they generally require a large amount of training data to properly capture the system behavior and make the models generalizable. This study addresses these issues by presenting a physics-guided Gaussian process (PGGP) that integrates physical knowledge with data learning in three aspects: (1) analytical equations, which model the physical degradation trend, are embedded as the mean function of Gaussian process (GP) to guide the prognosis; (2) kernel functions of GP are designed to model the time-varying variation in degradation; (3) the parameters of physical mean function and kernel functions are jointly optimized through data learning to capture the information not embedded in existing physical knowledge. Evaluated using a dataset of heating, ventilation, and air conditioning system, PGGP has consistently achieved higher prediction accuracy as compared to data-driven methods without embedded physics, including standard GP, Support Vector Machine (SVM), and Recurrent Neural Network (RNN), by at least 11.4%, 31.7% and 5.0%, respectively. In addition, PGGP has demonstrated improved data efficiency, outperforming standard GP, SVM and RNN by up to 59.9%, 75.5% and 68.4%, respectively, in the scenario of the least amount of training data, which is critical to ensuring early forecast of system degradation and support predictive maintenance.