Some paradigms for coordinating feedback control with applications to urban traffic control and smart grids

This lecture treats the problem of designing local control agents for cooperating components in a network of interacting dynamic systems. Each local control agent must ensure that all local specifications are met, but at the same time must ensure that the different components help each other in achieving good global behavior as well as good local behavior. This problem will be illustrated by using urban traffic control and smart electric power grids as examples. Centralized or hierarchical control approaches are not robust against failures in communication networks, and require unrealistic assumptions on the knowledge of each agent about the overall model. A completely decentralized approach, where each local control agent selfishly tries to achieve its local specifications only, runs a high risk of global interactions that may destabilize the system, making it impossible to achieve the specifications. This talk proposes two paradigms for distributed feedback control that require very little information exchange and very little global model knowledge. The leader/follower control paradigm is illustrated for urban traffic control: heavily loaded leader agents send messages to their follower neighbors requesting that these followers give green only to platoons of vehicles travelling towards the leader intersection at those times when this will be optimal for the performance of the leader. Another coordination paradigm is called the coordinating model predictive control (CMPC). Consider a power transmission network that has been partitioned in interacting regions, where CMPC is used in order to prevent the spread of the disturbances following incidents like line or machine failure. CMPC tries to resolve this by having each local control agent apply a model predictive controller, using as on-line available information not only the local voltage and current measurements, but also information on the planned sequence of future control actions of neighboring agents, communi- ated to it from time to time. This talk will discuss some of the minimal requirements for modeling, communication and control agent set-up in order to robustly achieve specifications using distributed control.