Mechanical and Civil Engineering Seminar

Abstract:  The control of a network of interacting dynamical systems is a central challenge for addressing a range of emerging application problems; examples include energy systems balancing a network of generation, load and storage devices, or robotic systems comprising a large number of components or agents. Utilizing the connectivity and interactions in the network by exploiting advances in communication and computation technologies offers the potential for pushing these systems to higher performance while increasing efficiency of operation, which will reduce system over-design and associated costs. However, safety requirements and high system complexity represent key limiting factors for leveraging these new opportunities.

This talk will present some of our recent work that brings high-performance control with hard guarantees on system safety to distributed systems, offering a scalable and modular approach that exploits interconnection effects and flexibly adjusts to network changes. A new framework for plug and play distributed predictive control will be introduced and we will discuss essential theoretical and practical aspects for certifying distributed decision-making based on an optimization-in-the-loop paradigm. We will show how the proposed scheme ensures the fundamental properties of stability and constraint satisfaction of the global system without recourse to any centralized coordination and even in the presence of online network changes, while allowing the control systems to optimize for performance. Application examples in area generation control and grid-aware electric vehicle charging will demonstrate the capabilities of the proposed theory. Lastly, we will address the computational aspects of the framework and present new results for certifying optimization with limited-precision computation or communication. The talk will conclude with future research directions in high-performance control of safety-critical systems, focusing on safe learning-based techniques that enable controllers to leverage large-scale online data for performance.