Abstract: The transmission distance and data capacity of coherent optical fiber communication systems are greatly limited by the Kerr effect and chromatic dispersion. To compensate the nonlinear impairments in optical fiber transmission, combined with a convolutional neural network layer, a bi-directional long short-term memory layer and an attention layer, a fiber nonlinearity compensation algorithm based on the CNN-BiLSTM-Attention model is proposed and simulated in a DP-16QAM 30Gbaud coherent optical system. The simulation results demonstrate that, compared to the CNN-BiLSTM model, the proposed algorithm achieves a reduction of 31.6% in the number of real multiplications required to equalize per symbol, but at the cost of lowering the Q-factor by 0.03 dB to 0.23 dB when transmitting over a distance of 1200 km. Additionally, the Q-factor of the algorithm is improved by 0.43 dB at the optimal launch power under similar complexity.