Federated Digital Twin-Empowered Online Control and Optimization for Cyber-Physical Systems

Digital twin (DT) technologies have the potential to revolutionize the online control and optimization methods for cyber-physical systems. However, the lack of the high-fidelity DT models that can accurately simulate the operation of the physical world is a key obstacle to their development. To address this issue, this paper proposes a federated DT (FedDT) architecture and modeling method to create a DT that can mimic the intrinsic dynamics and operational mechanisms of the physical world. Specifically, the FedDT architecture encompasses the internal components for specifying the assignment of functionalities and the external structure to leverage the collaboration of individual DTs. This model provides a digital representation of an online decision-making process for universal control and optimization problems, while considering DTs' collaboration to enrich their capabilities. Then, we design a federated self-learning algorithm to complete the DT modeling. By using the modeled FedDT, we are able to make purposeful planning and decisions in the digital space that are equivalent to those in the real world, without requiring knowledge of the dynamics of entities and environments. This enables achieving predictive evolution, precise estimation, and reliable decision for online control and optimization. The proposed FedDT offers practical advantages for engineering systems that require predictive and reliable decision-making with strong lookahead capabilities. It is particularly well-suited for applications such as autonomous driving, frequency control in smart grids, and real-time process control in Industrial Internet of Things (IIoT), where accurately modeling physical system dynamics is highly challenging. In real-world deployments, engineers only need to define the observations, actions, and reward signals. The proposed method then autonomously learns the underlying system dynamics and derives informed, data-driven optimization and control strategies. Finally, we demonstrate the effectiveness of our proposed FedDT model by applying it for the autonomous driving use case.