Learning to Infer Voltage Stability Margin Using Transfer Learning

Preventing voltage collapse is critical for reliable operation of power systems. A challenging problem is that the voltage stability margin, i.e., the distance from a given power profile to the voltage stability boundary, is very computationally intensive to obtain. A novel machine learning based approach for real-time inference of voltage stability margin is developed, only needing a very small number of offline-computed voltage stability margin data. An accurate margin predictor is trained by first training a binary stability classifier and then transferring this pre-trained model to fine-tune on the small data set of margins. Numerical simulations demonstrate that the proposed method significantly outperforms Jacobian-based voltage stability margin estimation with even faster real-time computation.