Graph-regularized data-driven control: Simultaneous optimization across multiple operating conditions

This study proposes a framework that integrates knowledge into data-driven control methods to simultaneously optimize control parameters for multiple operating conditions. The method automatically identifies similarities among different datasets from the viewpoint of the controller and constructs a graph structure based on parameter transferability. This graph structure is utilized in unified optimization, incorporating prior knowledge as a regularization term to maintain connectivity between parameters. The proposed approach is validated using data obtained from a hot rolling simulator. The results show that a graph structure implicit in the simulation conditions can be estimated through the controller even if the relationships among datasets are unknown and that the regularization strength enables flexible controller design from a specialized to a generalized solution.