Impedance Identification and Stability Assessment for Grid-Connected Converters Based on Sensitivity-Driven Dataset Optimization

This paper presents an impedance identification model and stability assessment method for grid-connected converter systems using K-medoids clustering for small datasets. The sensitivity of input features is evaluated via K-medoids clustering to optimize the dataset, improving model training accuracy. The impedance identification model is built with a BP neural network and combined with the generalized Nyquist criterion for stability analysis. The study tackles the issue of limited measured data in online impedance identification, which often impacts stability analysis accuracy. K-medoids clustering determines the sensitivity ranking of variables, guiding dataset acquisition by selecting more samples from high-sensitivity variables and fewer from low-sensitivity ones. This approach enhances dataset quality and converter impedance identification accuracy. Results show that the proposed method accurately predicts converter admittance characteristics under various conditions and effectively assesses system stability. The well-trained BP neural network can learn from limited data and predict admittance over a wide frequency range. Combined with the generalized Nyquist criterion, it provides an effective stability evaluation for grid-connected systems. This research offers a novel approach for impedance identification and stability assessment in scenarios with limited data.