Few-Shot Machine Learning at the Grid Edge: Data-Driven Impedance Models for Model-Free Smart Inverters

Abstract The future electric grid will be supported by a vast number of smart inverters distributed at the edge of the grid. These inverters' dynamics are commonly characterized as impedances under small-signal perturbations and are critical for ensuring grid stability and resiliency. However, the operating conditions of these inverters can vary widely, resulting in various impedance patterns and complicating grid-inverter interaction behaviors. Current analytical impedance models necessitate a comprehensive and accurate comprehension of system parameters while also relying on multiple assumptions to simplify system complexities. They can hardly capture the complete electrical behaviors of physical systems when inverters are controlled with sophisticated algorithms or performing complex functions. Real-world impedance acquisitions across multiple operating points through simulations or measurements are expensive and impractical. Leveraging the recent advances in artificial intelligence and machine learning, we present InvNet, a few-shot machine learning framework capable of characterizing inverter impedance patterns across a wide operation range, even with limited impedance data for each inverter. The InvNet can extrapolate from physics-based models to real-world models and from one inverter to another. Comprehensive evaluations were conducted to verify the effectiveness of our proposed approach in various application scenarios, and all data and models were open-sourced. Our study demonstrates the effectiveness of machine learning and neural networks as powerful techniques for modeling the intricate black-box characteristics of grid-edge energy systems and analyzing the behaviors of such systems with larger-scales that cannot be accurately described using traditional analytical methods.