Hybrid CNN-BiLSTM-Attention Conditional GAN for power system forced oscillation localization

Data-driven methods for Forced Oscillation (FO) sources localization in power systems face challenges including limited dataset size, class imbalance, and high data acquisition costs, which limit the ability to train large and sophisticated deep learning models. To address these issues, in this paper, we propose a hybrid Conditional Generative Adversarial Network (CGAN) to generate high-fidelity synthetic data, which synergistically integrates (1) Convolutional Neural Network (CNN) for spatial correlation extraction, (2) Bidirectional Long Short-Term Memory (BiSTM) for bidirectional temporal dynamics, and (3) attention mechanisms for adaptive spatiotemporal feature weighting. Furthermore, to ensure training stability while maintaining time–frequency consistency between synthetic and real data, we modify the training strategy by incorporating three key methodologies, i.e., gradient-penalized Wasserstein distance metric, generative–discriminative dynamic equilibrium, and multi-objective optimization. Finally, extensive validation on IEEE 39-bus and WECC 179-bus test systems demonstrates our CNN-BiLSTM-Attention CGAN model’s robustness and effectiveness across challenging scenarios: multi-source coupled oscillations (91.3% F1-Score), severe data class imbalance (30.9% F1 improvement), high noise levels (40 dB noise interference), and extreme Phasor Measurement Unit (PMU) data loss (80% loss rate). Remarkably, in the WECC 179-bus system with adequate PMU coverage, the proposed model can achieve an F1-Score exceeding 96% even when the random loss rate amounts to 80%. • Hybrid CGAN integrates CNN, BiLSTM and attention mechanisms for FO localization. • Novel training strategy with Wasserstein metric and multi-objective optimization. • Achieves 91.30% F1-Score for multi-source coupled oscillations. • Handles 80% PMU data loss with 96% F1-Score in WECC 179-bus system. • 30.9% F1 improvement under severe data imbalance conditions.