Federated learning-based distributed model predictive control

This work develops a federated learning-based distributed model predictive control (FL-DMPC) method for nonlinear networked systems to improve data privacy in the development of machine learning models. Specifically, a novel framework of FL method with personalized optimization (Fedpo) is proposed to obtain the global FL model for the entire networked system by updating and aggregating the personalized models for subsystems. This new FL framework significantly reduces the complexity of the learning algorithm and improves computational efficiency compared to existing FL methods. Additionally, it addresses system heterogeneity due to various dynamics of subsystems. Subsequently, the convergence of the Fedpo framework is proved by deriving an upper bound for its generalization and personalization errors, followed by theoretical analysis of the closed-loop stability of nonlinear networked systems under FL-DMPC. Finally, a chemical process network is adopted to demonstrate the effectiveness of the proposed Fedpo modeling and FL-DMPC method.