Data-Driven Control: Overview and Perspectives
Wentao Tang, Pródromos Daoutidis
Abstract
Process systems are characterized by nonlinearity, uncertainty, large scales, and also the need of pursuing both safety and economic optimality in operations. As a result they are difficult to control effectively. Data-driven techniques such as machine learning algorithms can provide complementary tools and insights to classical model-based control by enhancing the capability of modeling the dynamics of complex systems and the maintenance of control performance. Moreover, by learning the behavior of plants and controllers as black boxes, data-driven techniques can enable a completely model-free control paradigm. Hence, data-driven process control has the potential to mitigate the challenges of state-of-the-art control technology and yield generic, adaptive, and scalable strategies. This paper aims at providing an overview and conceptual classification of the main approaches in this emerging and promising field, and identifying current limitations and future directions.