Modern Multi Terminal HVDC (MT HVDC) grids are expanding rapidly with large scale renewable integration, yet disturbance detection still faces challenges in achieving the required speed, adaptability, and selectivity. Recent advances highlight a shift toward data driven approaches that learn disturbance signatures directly from measurements, reducing reliance on handcrafted features, threshold tuning, and complex signal processing pipelines.
This study proposes a hybrid AI based disturbance detection method that integrates a lightweight LSTM network with an XGBoost classifier for real time operation in MT HVDC systems. The method relies solely on voltage measurements sampled at 1 kHz, significantly reducing data rate requirements while preserving the temporal resolution necessary for fast decision making. Voltage waveforms are segmented into 1 ms windows, from which integrated voltage and rate of change features are extracted and encoded into compact temporal embeddings by the LSTM. These embeddings are subsequently classified by XGBoost to provide fast, interpretable, and probabilistic disturbance decisions.
A detailed MT‑HVDC test system was developed in PSCAD to generate a wide range of fault scenarios and operating conditions. Comprehensive benchmarking and multi‑objective hyperparameter optimization identified an efficient single‑layer LSTM with hidden size 8 and batch size 32 as the optimal balance between accuracy and computational efficiency. Results demonstrate strong generalization capability and robust performance under non‑ideal conditions.
Overall, the proposed method offers a scalable, data centric and deployment ready solution that enhances disturbance detection speed, improves fault clearance verification, and contributes to reducing Total Fault Clearance Time in next generation HVDC protection architectures.