A Catboost-Based Machine Learning Approach for Fault Detection in VSC-Based Multi-Terminal HVDC Grid

High-voltage direct current (HVDC) transmission systems face protection challenges due to the absence of natural current zero-crossings and the rapid rise of fault currents. This paper proposes a hierarchical fault detection and classification framework for multi-terminal HVDC networks that combines transient feature extraction with gradient boosting decision trees. The proposed method utilizes traveling wave transients distinguish between internal and external faults across a range of fault resistances and noise levels. The three-stage architecture employs asymmetric loss weighting to maintain system security alongside dependability for high-impedance fault detection. When evaluated on a four-terminal VSC-HVDC benchmark network, the protection scheme achieved zero false trips against normal operational fluctuations and remote pole-to-ground disturbances, alongside 92.24% overall dependability across fault resistances up to 500 Ω and noise levels down to 20 dB SNR. For cumulative fault resistances up to 150 Ω, the dependability exceeded 99%. Furthermore, the framework demonstrates generalization across measurement noise and provides interpretability through feature importance analysis. The proposed framework offers a data-driven approach for HVDC protection that balances the requirements of relay security and dependability.