Ensemble learning of dynamical systems with multiple operating conditions via statistical process control

This paper addresses the adaptation and performance monitoring of ensemble data-based models over time. Once a model of a system is identified using a specific training dataset, it may fail to accurately represent system dynamics under varying operating conditions not included in the original training dataset. To continuously adapt a data-based model to evolving operating conditions, we propose an ensemble learning framework characterized by (i) a combination rule that weights different models based on the statistical proximity of their training dataset to the current operating condition, and (ii) a monitoring algorithm leveraging statistical control charts to supervise the ensemble model's reliability and trigger the identification and integration of a new model when a new operating condition is encountered. The proposed methodology is tested on an energy system referenced in the literature, which exhibits multiple operating conditions, showing promising results from both adaptation and monitoring perspectives.