A Novel Grid Impedance Estimation Method for Grid-Connected Inverter Systems Using LSTM Neural Networks

Accurate grid impedance estimation is critical for ensuring the stability and reliability of grid-connected inverters, especially under dynamic operating conditions. This paper proposes a data-driven approach based on Long Short-Term Memory (LSTM) neural networks, designed to estimate grid impedance directly from time-domain measurements-specifically voltages at the Point of Common Coupling (PCC), and currents on both grid and converter sides-without requiring frequency analysis or signal injection. The estimation is formulated as a 42-class classification problem, each class representing a unique resistive inductive impedance pair. Transient responses following a step change in active power reference are used as input sequences for the LSTM. Various LSTM architectures are tested, including single-and two-layer setups, showing classification accuracies up to 93%, with model sizes as small as 0.12 MB and prediction times below 0.1 ms. A benchmark with traditional machine learning classifiers confirms the superior trade-off of LSTM models in terms of performance, speed, and memory footprint. The proposed method offers a scalable and non-intrusive solution for real-time impedance monitoring in modern power electronic systems.