SYNTHETIC OPERATIONAL DATA GENERATION FOR DEEP LEARNING APPLICATIONS IN POWER TRANSMISSION LINES

Deep learning (DL)-based protection algorithms for power transmission lines require large volumes of operational data for accurate training. However, such data is often complex to access due to confidentiality, restrictions, and proprietary limitations. This paper proposes a synthetic data generation method that combines principal component analysis (PCA) with a conditional tabular generative adversarial network (CTGAN). PCA reduces the dimensionality of high-frequency time-series data, allowing CTGAN to operate efficiently while retaining essential statistical characteristics. The generated synthetic data shows strong correlation with real data and effectively augments limited datasets. Validation using an LSTM-based fault classification model demonstrated an improvement from 50.93% to 86.07% accuracy. Additional validation using sub-synchronous oscillation data demonstrates broader applicability. The proposed method is scalable and supports DL training in data-scarce scenarios.