Dynamic Equivalencing using Machine Learning

Global renewable capacity is projected to increase by nearly 5,500 GW (about 75%) between 2024 and 2030. While this rapid expansion of decentralized generation brings opportunities, it also introduces significant challenges, such as reduced system inertia, stochastic power output, time-varying operating conditions, and stability concerns. A couple of power blackouts in recent years have highlighted that the stability of system operation is a critical issue. Together, these factors reduce grid flexibility and impose heavy computational demands for accurate analysis, estimation, and control. This proposal, dynamic equivalencing using machine learning (DEML), aims to address these challenges from a data-driven and physics-guided perspective. The objective is to accelerate computationally intense electromagnetic transient simulations (EMT) using deep learning-enabled model order reduction (MOR) technologies. Unlike traditional methods, DEML reframes the classical dynamic equivalencing task as an unsupervised machine learning problem with physics-informed constraints, addressing key limitations of the state-of-the-art nonlinear MOR approaches. Specifically, DEML will deliver dynamic equivalents (DEs) of utility-scale wind farms with doubly-fed induction generators (DFIGs) or permanent magnet synchronous generators (PMSGs), leveraging the Koopman operator framework. Such DEs are expected to accelerate EMT simulations by a factor of 2-3, while maintaining <5% error in system response. Validation will be performed using real-time hardware-in-the-loop (HIL) setups. The project is novel because it is discovery-based and fundamental to the overarching goal of analyzing large-scale power systems (LSPS) in line with the EU’s decarbonization, digitization, and energy reliance objectives.