Low-frequency oscillation identification is significant for the stability analysis of modern power systems under the dual-carbon objective. Although the classical identification methods represented by Prony and Hilbert-Huang Transform (HHT) algorithms have high identification accuracy, it is not suitable for real-time or online applications of the current power system. To solve this problem, this paper proposes Bi-LSTM-based low-frequency oscillation identification method. In which, the Prony algorithm is first utilized to pre-identify the oscillation mode of the historical data in order to construct a labeled dataset. And the dominant mode is selected as label to construct the complete dataset. Then, Bi-LSTM-based deep learning algorithm is constructed to identify dominant mode in a specific frequency range. Case studies in the 4-machine 2-area system show that the proposed method can identify dominant mode’s information accurately, including oscillation frequency and damping ratio. Meanwhile, compared with other recurrent neural network algorithms, the proposed method has a better discrimination effect and can effectively solves the problem of difficult on-line identification of low-frequency oscillation modes in power systems.