Voltage source converters (VSCs), particularly modular multilevel converters (MMCs), play a pivotal role in modern power grids due to their application in interconnected systems and renewable energy integration. However, their complex internal dynamics and control systems can cause stability issues under a variety of operating conditions. Impedance-based stability analysis is critical for understanding these interactions and ensuring grid reliability. Traditional small-signal models based on physics and circuit analysis often neglect nonlinearities and practical implementation details, limiting their accuracy. In contrast, black-box impedance measurement techniques necessitate significant computational resources and human effort, particularly when analysing both AC and DC side impedances.

This thesis presents a transfer learning framework for accurate impedance characterization of MMCs. The framework uses a combination of linear time-invariant (LTI) modeling and black-box impedance measurements to estimate impedances based on system-level parameters such as AC and DC side voltages, active power, and reactive power. The methodology involves pre-training artificial neural network (ANN) models on a diverse dataset derived from LTI models, followed by fine-tuning with real-time electromagnetic transient (EMT) simulations. Unlike traditional approaches, which use a single ANN model across the entire frequency range, this study discusses the use of multiple smaller ANN models tailored to specific frequency ranges and operating points, which improves prediction accuracy.

The effectiveness of the proposed transfer learning framework is demonstrated with a practical MMC converter from the CIGRE B4 DC grid test system. The results show that impedance estimation outperforms traditional ANN methods even in scenarios involving confidential converter parameters. This approach allows for precise impedance characterization while reducing computational complexity and data requirements.