Development of adaptive model-based digital twins for evolving power systems

Adaptation is a critical function that guarantees a digital twin (DT) remains accurate and provides reliable and trustworthy information about its behaviour. Lack of a suitable adaptation function compromises the efficiency and reliability of DT-based applications. Data-driven DTs that rely on machine learning (ML) adapt to new datasets acquired from the physical twin (PT) by employing transfer learning techniques or by undergoing complete reconstruction. However, these approaches cannot be implemented on model-based DTs that are based on known laws and principles of physics, since their accuracy is dependent on specific parameters that may be neither estimated from the measurements nor derived from existing datasets. This paper proposes a generic workflow to adapt model-based DTs to minimise their deviation and introduced errors from their corresponding PTs. The two-step method uses ML to first identify the deviating components within an interval of confidence and then estimate new parameters based on information from the PT. A relational database is adopted for flexible and efficient data access and storage. To verify the effectiveness and feasibility of the proposed method, a DT testbed is used for a power system digital twin (PSDT) based on real-time simulations. The PSDT can be successfully adapted for specific changes, including system configurations and component parameters. The proposed adaptation workflow provides an opportunity to adapt system-level model-based PSDTs across wider scenarios and/or conditions.