Optimizing zone temperature setpoint excitation to minimize training data for data-driven dynamic building models

Poorly designed excitation signals could lead to inaccurate or, even worse, highly correlated parameter estimates in a data-driven model, so it is critical to have an informative training data set in order to obtain an accurate model in a cost effective manner. This paper investigates a sequential optimal design of experiments (DOE) approach to generate an optimal training data set for varying zone temperature setpoints that maximizes the accuracy of parameter estimates for an intended building model structure. This method was applied to a whole building case study in which a simplified thermal network modeling approach was adopted and the thermal parameters were estimated. The obtained optimal trajectory was always under a bang-bang type which enables an exhaustive search scheme to start the sequential design procedure. The designed excitation signals led to significant improvements in model accuracy compared to night-setup/back control strategies that are typically used to vary setpoints.