Abstract: The large-scale integration of new energy sources and non-linear loads has led to increasingly complex voltage flicker issues in distribution networks. To address the limitations of traditional mechanistic analysis methods and purely data-driven approaches in scenarios involving multi-source disturbances, low signal-to-noise ratios, and high-penetration distributed photovoltaic integration, this paper proposes a voltage flicker type identification method that integrates elastic net regularized variational mode decomposition with an Inception-ResNet deep feature learning framework incorporating the Convolutional Block Attention Module (CBAM). First, the mechanisms of typical voltage flicker phenomena in distribution networks are modeled and analyzed. Second, ER-VMD is used to adaptively decompose flicker signals, extracting multi-scale intrinsic mode components to construct feature representations. Subsequently, the processed features are fed into a CBAM-enhanced Inception-ResNet network to achieve automatic identification of multiple flicker types. Finally, simulation results demonstrate that the proposed method can effectively identify various types of voltage flicker under complex new energy disturbance conditions, thereby accomplishing the task of distinguishing different voltage flicker types with an accuracy rate of 98.9%.