Calibration of Building Models for Supervisory Control of Commercial Buildings

Past research of predictive optimal control of active and passive building thermal storage inventory has confirmed the importance of accuracy in the employed building model. In a subsequent investigation of modelfree learning control for the same application, a hybrid model-based/model-free control scheme based on simulated reinforcement learning has been proposed. Experimentation validated this approach, yet the experiment data analysis also revealed that the accuracy of the training model for the learning controller can significantly affect the knowledge base of the controller and its resultant performance. As a result, procedures of model calibration to improve the model accuracy are needed to improve the control quality in both model-based predictive optimal control and hybrid control schemes. This paper describes a methodology for model calibration that is based on system identification and has been successfully applied in both approaches. A numerical analysis demonstrates that by carrying out the model calibration the performance of the controllers has been improved.