Large scale systems

Model-Based Predictive Control (MPC) has proven to be one of the widely accepted advanced control techniques for large-scale system control. Applied to different large-scale infrastructures such as drinking water networks, sewage networks, open channel networks, or electrical networks, this technique confirms the applicability of this technique. The main reason is that once the dynamic model of the plant is obtained, the MPC design consists only of expressing the desired performance specifications through different control objectives (for example, weights on tracking errors and actuation efforts as in classical linear quadratic regulation) and constraints on system variables (for example, min/max of selected process variables and/or their rates of change) that are necessary to ensure process safety and asset health.

The rest of the MPC design is automatic: the given model, constraints, and weights define an optimal control problem over a finite time horizon in the future (this is why the approach is called predictive). This translates into an equivalent optimization problem and is solved online to obtain an optimal sequence of future control moves. Only the first of these moves is applied to the process, as in the next step it solves a new optimal control problem to exploit information from new measurements. In this way, an open-loop design methodology (optimal control) is transformed into a feedback one.

However, the main obstacle for MPC control, like any other control technique, when applied to large-scale systems in a centralized manner, is scalability. The reason is that a huge control model is needed, which is difficult to maintain or update and has to be rebuilt at every change in the system configuration, such as when some part of the system has to be interrupted due to maintenance actions or malfunctions. Later, a model change would require re-tuning the centralized controller. Clearly, the cost of establishing and maintaining the monolithic solution to the control problem is prohibitive. One way to circumvent these problems could be to investigate decentralized MPC (DMPC) or distributed MPC techniques, where local networked MPC controllers are responsible for controlling part of the entire system. The main difference between decentralized and distributed MPC is that the latter requires negotiations and recomputations of local control actions within the sampling period to increase the level of cooperation.