Learning to Leverage Centralized and Distributed Strategies in Moving Horizon Model Predictive Control

Distributed model predictive control (DMPC) offers a computationally efficient alternative to centralized model predictive control (CMPC) for enabling the optimal control of industrial process systems. Both strategies offer distinct advantages: DMPC tends to reduce the computational burden of the control problem, thereby producing solutions on a time scale more suitable for real-time process control, whereas CMPC tends to yield higher-quality control decisions. In this work, we demonstrate that problem parameters, such as state measurements and set points, significantly influence the computational costs or solution time for solving the nonlinear programming (NLP) problem associated with the centralized optimal control problem. We show that DMPC is preferable when computational costs are high, while CMPC is more effective when the costs are low. To navigate the trade-offs between these strategies in online control, we employ machine learning classifiers to identify when a problem instance favors a particular solution strategy. We augment the traditional feedback control loop with these classifiers to strategically choose between CMPC and DMPC during the moving horizon of the control. By doing so, we show that leveraging the individual strengths of each approach improves the overall quality of process control compared to a strictly centralized or distributed strategy. We demonstrate the efficacy of the classifier-informed algorithm selection across two case study process systems.