A hybrid modeling approach integrating first-principles knowledge with statistical methods for fault detection in HVAC systems

Abstract

This work presents a hybrid modeling technique that combines first-principles knowledge with principal component analysis (PCA) to detect faults in heating, ventilation, and air conditioning (HVAC) systems. Residuals, defined as the discrepancies between expected and observed behaviors, along with temperature measurements are used to develop multiple hybrid PCA models. Each model describes the normal behavior of the system in a particular operating state of air-handling units (AHUs). Hotelling’s T2 and square prediction error (SPE) statistics corresponding to the new observations are calculated using the hybrid PCA model in order to monitor the process and detect deviations from the expected behavior. The efficacy of the proposed approach to detect faults is evaluated and compared with two benchmark approaches: (1) residual analysis (based on first-principles models) and (2) a data-driven method (based on PCA) applied to raw temperature measurements. The superior performance of the proposed methodology, over the benchmarks, is shown via simulation tests with commonly occurring fault scenarios.