Low-order building model identification in presence of unmeasured disturbance for predictive control strategies

Predictive control strategies for building heating and cooling systems have been proposed as an energy efficient alternative to traditional strategies. The performance of such strategies is highly dependent on the underlying system models used. In an effective strategy, these models used are required to be accurate enough for informative predictions to be made yet simple enough to be used within a numerical optimization problem. Identification of such models from measured data may not be trivial in the presence of a large amount of unmeasured disturbance. In this paper, methods for deriving low-order zone models in the presence of unknown disturbances are considered. A high-order RC-network representing the complexity of a building is used to generate data for the identification process. An estimate of the disturbance affecting each zone of the network is first developed using Kalman filtering. Disturbances common to several zones are isolated by a spatial filtering process using principal component analysis. The new disturbance estimates are then included in the model identification formulation. The models and disturbance estimates are refined through several iterations of the process. Significantly improved prediction accuracy is shown to result when the disturbance estimates are incorporated.