Enhanced distributed learning-based coordination of multiple approximate MPC for large-scale systems

In the field of process control, there exists a category of large-scale systems composed of multiple interrelated subsystems. Model Predictive Control (MPC) provides an effective control framework for these systems; however, it still faces challenges in handling larger-scale systems and reducing online computation costs. To address these issues, this paper innovatively proposes a distributed learning and approximate control scheme. Firstly, for each subsystem, we obtain a neural network-based approximate MPC controller by utilizing local information for offline training. Subsequently, we design a distributed reinforcement learning method, where the local approximate controllers make coordinated decisions by sharing information with their neighbors. This strategy not only focuses on the performance of the subsystems themselves but also takes into account the performance of neighboring subsystems, thereby effectively enhancing the overall performance of the approximate controllers. Additionally, we adopt a strategy that combines online learning and offline training to cope with changes in system characteristics during the process. The proposed scheme demonstrates good tracking performance, robustness and adaptability, illustrated through an example of a typical chemical process.