Cooperation and Coordination in Decentralised Digital Twin Networks

The increasing complexity of modern cyber-physical systems, such as smart power grids, necessitates new governance paradigms. Digital Twins, which are live virtual models of physical systems, offer powerful simulation and decision support. However, traditional centralised Digital Twin architectures are ill-suited for large, multi-owner systems. They suffer from scalability bottlenecks, single points of failure, and critical privacy vulnerabilities, as stakeholders are often unwilling to share sensitive raw data with a central authority. This thesis addresses these limitations by delivering a novel architectural framework for decentralised collaboration and an algorithm for intelligent coordination.

The framework enables a federation of autonomous Digital Twins to collaborate by exchanging high-level, privacy-preserving "decision information", such as future intentions or predictions, rather than sensitive raw data. This is facilitated by a distributed middleware that integrates with independent Digital Twin instances. To ensure robustness in this open environment, the framework incorporates a trust mechanism that continuously verifies the historical accuracy of each peer's predictions against real-world outcomes, allowing agents to dynamically weigh the reliability of incoming information.

To enable intelligent, optimised coordination, this thesis further contributes the Digital Twin Assisted Multi-Agent Deep Deterministic Policy Gradient algorithm. This Multi-Agent Reinforcement Learning method introduces a "simulation-assisted critic" that leverages the shared decision-level information to construct a global simulation model for training. This unique approach achieves the global awareness needed for centralised training without requiring access to private agent data, thus resolving the privacy-performance paradox. The complete framework was validated in a complex Vehicle-to-Grid coordination scenario, where it learned sophisticated control policies and achieved performance comparable to fully centralised, non-private learning algorithms.