Automatic power line segmentation is an important prerequisite for the safe operation of intelligent inspection platforms. However, power line segmentation is a small target segmentation problem in complex backgrounds and multiple climatic environments, which is highly prone to encounter false or missed detections. In order to improve the robustness and accuracy of power line segmentation, an end-to-end segmentation model based on hierarchical attention fusion is proposed in combination with an encoder-decoder framework. The model proposes a reduced-dimensional residual convolution unit that increases the network depth while significantly reducing the network parameters, making it easier to deploy in embedded devices, enabling the model to capture global information and emphasize the target regions of powerlines, a chain-based hierarchical attention fusion module is designed for multi-scale feature fusion to address the category imbalance problem. To improve the model's attention to the unique line prior features of power lines, the line prior loss function is combined with the Focal loss function and Dice loss function to form a joint loss function to further improve the accuracy of power line segmentation. The experimental results show that the depth of the proposed model network increases to about 2.8 times that of the base network, while the number of parameters is only about 1/3 of the original one. Robust segmentation of power lines can be achieved for both regular weather and foggy weather aerial images. The proposed model can be applied to the field of power inspection, making the inspection more intelligent and efficient.