Distributed Control
Distributed control coordinates many local controllers, each using limited local information, to achieve a global objective without a central authority.
Why decentralize
Large systems, power grids, sensor networks, multi-machine plants, are impractical to control from one central computer that gathers all measurements and commands all actuators: communication is limited, the central node is a single point of failure, and the computation may not scale. Distributed control assigns a local controller to each subsystem, each acting on its own and neighbors' information, coordinating through limited communication to reach a shared goal.
Structure and information
The core question is what each controller is allowed to know. Fully decentralized control uses only local measurements; distributed control adds communication with neighbors. The information structure, who knows what and when, determines what performance is achievable and whether the optimal controller is even tractable to compute. Some information structures make optimal distributed design convex; others, notably non-classical ones, make it hard.
Design approaches
Quadratic invariance is a condition on the plant and communication structure under which the optimal distributed controller can be found by convex optimization, a landmark result that identifies the tractable cases. Otherwise, distributed MPC, in which subsystem optimizers exchange predicted trajectories and iterate toward a coordinated solution, and dual-decomposition methods that split a global optimization across nodes, are widely used. Stability of the interconnection is checked with passivity or small-gain arguments.
- Local controllers with limited information, no central node
- Information structure decides achievable performance
- Quadratic invariance identifies convex-solvable cases
- Distributed MPC and dual decomposition coordinate iteratively
Scalability and resilience are the payoffs: adding a subsystem means adding a local controller and communication links, not redesigning a monolith, and failure of one node degrades performance locally rather than globally.
For a design-stage plant composed of many coupled subsystems, distributed control would let each subsystem regulate itself while cooperating with neighbors, evaluated in simulation. Its most cooperative form is multi-agent consensus.