Aiming at the problem that the current deep learning network model for substation meter detection has too many parameters and is difficult to be deployed in mobile devices and embedded devices with limited computing resources, we propose a lightweight substation meter detection algorithm with improved YOLOv5. Based on the YOLOv5 network, the improved algorithm introduces the SE fusion attention mechanism module, and adaptively learns the relationship between feature channels to improve the model’s ability to extract important features from the instrument. Meanwhile, TensorRT technology is used to reconstruct and optimize the improved model, which can reduce the number of model parameters, improve the detection speed and ensure the accuracy of the model detection. Experimental results demonstrate that compared with YOLOv5 on the embedded device Jetson Nano, the improved algorithm proposed in this paper presents significant advantages, which increase by 1.5% and 2.3% respectively on mAP@.5 and mAP@.5:.95, and the detection frame per second increases by 130%, reaching 23FPS. It can realize real-time instrument detection in substation scene, and has practical application significance.