AI-Enabled Dynamic Model Discovery of Networked Microgrids

In real-world networked microgrids (NMs), the exhaustive dynamic model of each and every microgrid may not always be attainable, hindering the adoption of conventional model-based NM analytics. This chapter introduces learning-based dynamic model discovery approaches for NMs, which construct reliable and accurate dynamic models of microgrids from measurements. We first present a neural-ordinary-differential-equations (ODE-Net)-enabled method to learn continuous-time dynamic models of the unidentified subsystems of NMs under heterogeneous uncertainties. We further establish a physics-data-integrated approach to explicitly control the closed-loop accuracy of the learnt dynamic models during the training process. Experimental results in a typical NMs system and a large-scale distribution-microgrid-hybrid system (both with inverter-based resources) validate the accuracy, efficacy, and versatility of the proposed methods. The AI-enabled dynamic model discovery therefore lays a foundation for various model-free analytics of NMs.