Hybrid LSTM-CNN Generative Adversarial Network for Power System Data Generation Under Unconventional Events

Recent years have witnessed a growing frequency of unconventional events in power systems. Due to their low-probability, high-impact nature, pre-event operational data for early warning remains scarce. However, deep learning models demand substantial high-quality training data, while conventional data generation methods often fail to adequately represent the complex characteristics of power system data, including nonlinearity, multivariate coupling, and abrupt variations. To address this challenge, this paper proposes a novel LSTM-CNN Residual Wasserstein Generative Adversarial Network (LCR-WGAN), which integrates LSTM networks for capturing long-term temporal dependencies, CNN for extracting localized feature variations, and residual connections to enhance feature fusion and gradient flow. The model is validated using real-world SCADA measurements and fault recordings from power systems, with performance evaluated through metrics including Root Mean Squared Error , Dynamic Time Warping and Cosine Similarity. Experimental results demonstrate that LCR- WGAN consistently outperforms baseline models (WGAN, DCGAN, and SeqGAN) across all metrics for both single-phase and two-phase fault scenarios, confirming its capability to generate high-fidelity and diverse synthetic data even under limited training samples. This work establishes a promising technical pathway for early warning of unconventional power system events.