Identification of controller models for model-based optimal control of large buildings

Model predictive control (MPC) is a control technique with a large potential to reduce the energy consumption at heat ventilation and air conditioning (HVAC) systems when comparing it with the classical Rule Based Control (RBC) strategies. The bottleneck of this technique is to find an adequate model to predict the thermal behavior of the building. This model should be simply enough to be introduced into an optimization problem that calculates the optimal heating inputs to consume as low energy as possible while maintaining the inside comfort temperature. At the same time, the model should be accurate enough when predicting the temperature in the optimization horizon. White-box identification techniques can reach high levels of accuracy, but the process is time-consuming and need of a large amount of information of the building. This paper studies the remaining grey and black- box techniques to develop models for MPC. Grey-box techniques need just a low quantity of information of the building to construct a simplified physical model which parameters are later adjusted with data. Black-box techniques are totally data- driven. A deep comparison has been developed for two levels of system complexity: a single-zone and a six-zones building. Four black-box identification procedures have been compared: Box-Jenkins (BJ), Neural Network (NN), Subspace (SUB) and the classical state space (SS) identification procedure. RC thermal networks (RC) have been used for grey-box identification. An innovative method has been developed to obtain RC Networks of complex systems automatically by just providing data. The results do not differ from the single to the multiple-zone identification. It has been demonstrated that black-box models are strongly sensitive to data, specially NN, which result unstable for closed loop simulations when identifying them with non-representative data, but behave the best for short prediction horizons (3 hours) and when using representative data to train them. BJ result also hardly sensitive, increasing their errors by a factor of 5 from representative to non-representative data. SUB scores the best with a RMSE lower than 0.7 degrees even when the data is non-representative. RC obtains RMSE always below 1.2 degrees. However, these models are characterized for their capability of predict longer prediction horizons without an appreciable increase on their errors. The same happens when some training data is hidden at the identification. It means that these models are the most robust.