Supervisory system identification for bilinear systems with application to thermal dynamics in buildings

When system identification is performed online and predictions of the system response are requested often (as in model predictive control formulations), the identification model with the best performance may not be fixed with time. Besides, more accurate models may require larger training times compared to low-order linear models. This is particularly evident in thermal dynamics in buildings where operating conditions may change throughout the year. To this end, this paper introduces a supervisory identification process, tailored specifically for input-output stable bilinear systems, where two parallel decision processes run periodically. The first one is concerned with the selection of the appropriate partition of the input(s) domain, while the second one is concerned with the selection of the identification model for each one of the resulting partition sets. The overall identification model constitutes a switched system. We show analytically that the proposed scheme is adaptive and robust to changes in the performance of the identification models, while convergence is attained (in probability) to the best model.