Notes on finding black-box model of a large building

Finding a suitable dynamic, linear, time-invariant model is a major obstacle for the use of model-predictive control of buildings. While finding models based on physical properties of the system is time consuming, statistical models need the system to be excited, which is not always possible. This paper presents possibilities for finding a suitable model based on subspace identification methods for unexcited data and data containing a specially designed identification experiment, and presents practical experiences gained while finding suitable models for a large building. Finally, some other possibilities of finding a model by incorporating prior information to the identification process are discussed, and the performance of the models is evaluated with respect to their use as a part of an MPC controller.