This paper presents a deep-learning framework, Transformer LSTM/Bi-LSTM Conditional Generative Adversarial Network (cGAN), designed to address forced oscillation (FO) source localization challenges in power grids with high IBR penetration. Our approach offers dual contributions. Firstly, it synthesizes time series power grid status data during FO occurrences, by leveraging the fusion of transformer and Long Short-Term Memory (LSTM)/Bidirectional Long Short-Term Memory (Bi-LSTM) networks. Secondly, it employs the deep learning classifier to predict the oscillation source location. Leveraging extensive datasets enhances source identification accuracy. Evaluation results demonstrate the model's proficiency in generating synthetic oscillation data and enhancing FO source identification accuracy compared to exisiting OSL identification apporaches. Commendably, the Transformer Bi-LSTM cGAN architecture showcases exceptional performance, particularly in scenarios featuring intentional faults and mixed fault and fault-free conditions. These findings underscore the robustness of the proposed model in OSL identification, thereby showcasing its potential for practical deployment in power system applications.