Neuro-Dynamic State Estimation for Networked Microgrids

The increasing integration of distributed energy resources (DERs) brings complicated dynamics in networked microgrids (NMs), calling for high-fidelity dynamic state estimation (DSE) of NMs. Traditional DSE, which requires accurate physical models of the entire NMs, is becoming increasingly unattainable. This paper devises neuro-dynamic state estimation (neuroDSE), a learning-based DSE algorithm to track the dynamics of inverter-interfaced NMs with unknown subsystems. The process and contributions include: 1) a data-driven neuroDSE algorithm is established for NMs with partially unidentified dynamic models by incorporating the neural-ordinary-differential-equations (ODE-Net) into Kalman filters; 2) a self-refined neuroDSEplus method is devised to tackle limited and noisy measurements. Specifically, Kalman filters are embedded into ODE-Net training for automatic filtering, augmenting, and correcting effects; 3) a neuroDSEknet algorithm is derived to relieve the model mismatch scenarios by integrating KalmanNet with neuroDSE. Numerical simulations carried out on typical four-microgrid NMs reveal that neuroDSE can track the dynamics under various control modes (e.g., droop/secondary controls) and components. Its variants increase the accuracy of neuroDSE under limited measurement and model mismatch scenarios.