This study leverages the Openpose system to capture skeletal key points of electric power operators, simplifying network complexity by sharing convolutional layers during the ReLU activation phase. We introduce a graph convolutional network (GCN) to model these skeletal sequences, creating a spatio-temporal deep learning approach for behavior recognition. Tested on a relevant dataset, our Openpose-GCN network demonstrates stability with a training loss of 0.11 after 700 iterations, achieves over 90% accuracy in recognizing operator actions and behaviors, and maintains a recognition error below 0.003 for operations with varying risk levels. These findings underscore the potential of our approach to enhance electric power operation safety through real-time risk warning and control.