Aiming at the practical problem that substations are prone to encounter unmanned aerial vehicle （UAV） intrusion, an improved anti UAV target detection method is proposed based on YOLOv5. Firstly, a four scale features fusion structure is proposed by improving the original model structure of YOLOv5 to enhance the detection capability of small-scale objects. Secondly, the C3 module in the original model is introduced into the Transformer encoder to improve the learning ability of small target feature information. Finally, the convolution channel attention module is integrated into the network, focusing on the learning of the target area to improve the representation ability of the model for features. The test results show that the overall recognition rate of the improved model is 90.2%, the average accuracy is 89.5%, and the forward reasoning speed is 160 frames per second. In addition, compared with other existing frontier algorithms, the overall performance of this method is better, and it can better meet the real-time detection requirements of anti UAV in substations.