Abstract
Model- and optimization-based control (MBC) strategies are increasingly regarded as promising technologies in the building and energy engineering domains, due to their ability to respond to price signals and enable buildings to contribute flexibility to the energy system. A key barrier to broader adoption is the identification of suitable control-oriented building models, including characterization of system noise and tuning of state estimators. In this work, we propose a system identification algorithm comprising several novel elements aimed at reducing implementation effort in commercial buildings. The method features decomposed identification of physical and statistical parameters, continuous re-identification of both parameter sets, and supports a time-varying modeling paradigm to capture unobservable inputs and varying uncertainty. Performance verification is carried out on two datasets from a medium-sized office building equipped with radiator-based heating and dedicated energy metering. For each dataset, models are initially identified using two weeks of normal operation data, followed by continuous re-identification over an additional two-week period. The resulting models exhibit (i) good day-ahead simulation performance on test data and (ii) white noise residuals on training data in most experiments. Compared to a reference model identified using pseudo-random binary sequence (PRBS) excitation, the proposed method improves normalized root mean squared error (NRMSE) by up to 200%. These findings suggest that the proposed method and modeling approach can support the deployment of room-level, model-based controllers without the need for intrusive excitation experiments, and may be suitable for scalable implementation in well-instrumented non-domestic buildings.