Practice-Oriented System Identification Strategies for MPC of Building Thermal and HVAC Dynamics
System Identification, Building Thermal and HVAC Dynamics, Model Predictive Control
The increasingly competitive HVAC market has made it necessary to develop technologies that exploit the economic potential in such systems to reduce energy consumption. Smart HVAC operation through optimization-based control strategies, such as model predicative control (MPC), serves as one method for achieving this goal. MPC and Economic MPC have gained significant attention due to their ability to optimally operate HVAC system in order to minimize their energy consumption and/or energy cost while maintaining desired comfort temperatures. Several research work have attempted to use system identification in order to model building thermal and HVAC dynamics. Nonetheless, empirically modeling HVAC systems in a robust and scalable manner is very challenging due to the non-convexity of the system identification optimization problem and the existence of complex actuator saturation limits. Therefore, this work develops grey-box system identification strategies that attack these challenges to enable the development of accurate empirical models of HVAC systems in practice settings. Saturation refers to the usage of the maximum or minimum HVAC capacity to track a desired temperature set-point for buildings under heating and cooling modes. The existence of significant saturation in the data collected is a common problem that poses many challenges for identifying the dynamics of HVAC systems since they can affect the quality of the collected data and result in an inaccurate identified model. Classical approaches for dealing with saturation in system identification, such as using nonlinear functions to capture the saturation behavior, are not implementable due to the complex saturation behavior associated with HVAC systems. In addition, another challenge faced in identifying HVAC and building thermal dynamics is the existence of many roots in such non-convex system identification problems. Therefore, it is desirable in industry to avoid using initial guesses that lead to local optima which result in inaccurate models. In the first part of this work, we develop an algorithm that is capable of detecting and removing saturation data from system identification experiment input-output data. This is done to extract the useful data sections that represent the cited HVAC system dynamics which are necessary to identify reliable models of the HVAC dynamics. The second part of this work develops a strategy that avoids solving system identification problems all the way to local optimality using initial guesses that lead to local optima which result in poor models. The algorithm attempts to eliminate poor initial guesses and yield initial guesses that ultimately lead to great fits of the model to the data. The parameters of the grey-box models were identified via a two-step parameter estimation approach. In the first step, the model parameters were identified using simulation prediction error method. In the second step, the model was augmented with a disturbance model and the estimation gain (i.e., Kalman gain) was identified using the standard 1-step prediction error method. The proposed strategies were applied to data collected from real building HVAC systems and have shown to successfully work in practice. The results demonstrate accurate 1-step and multi-step predictions which are necessary for the implementation of MPC.