Abstract: To address the issue of noise interference in sampling signals associated with power system subsynchronous oscillation (SSO), a composite model for SSO modal parameter identification is proposed based on convolutional neural networks (CNN), bidirectional gated recurrent units (BiGRU), and self-attention mechanism. By abandoning the serial structure of traditional models and the empirical selection of hyperparameters, the proposed model adopts a parallel structure of CNN and BiGRU and utilizes the sparrow search algorithm (SSA) to determine the model’s hyperparameters. The model is trained on noise-free SSO simulation datasets and evaluated through noisy simulation signals and measured torsional vibration signals. Experimental results demonstrate that, compared to traditional serial structures and other deep learning models, the proposed model significantly improves the accuracy of modal damping and frequency identification. This model exhibits enhanced robustness against noise interference and frequency drift, making it suitable for SSO modal parameter identification in actual power systems characterized by low signal-to-noise ratios and frequency drift.