Aiming at the problems of anti-noise, anti-high resistance and complex threshold setting of traditional pole selection methods, a fault selection method of flexible DC distribution line based on Res-BiLSTM network was proposed. Firstly, the original fault signal was subjected to complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and then the reconstructed signal was obtained by using the correlation coefficient and Shannon entropy for reconstruction. Secondly, the Res-BiLSTM network model was constructed for the pole selection. In order to improve the network accuracy and the convergence speed, the channel attention module was introduced into the split-attention network. The reconstructed signal features were extracted using the convolutional bidirectional long short-term memory and the improved split-attention network at the same time. The extracted features were fused using the attention feature fusion module, and the fused features are classified. Finally, PSCAD/EMTDC was employed to construct the model and to verify the proposed methodology. The simulation results show that the proposed pole selection method is highly accurate, anti-interference, and independent of fault distance.