In this paper, a prediction approach for abnormal behavior of substation staff is proposed by combining global and local information using generative adversarial networks based on video scenes. In substations, this method can be used to issue timely warnings to operating and maintenance personnel who may trigger dangerous actions in the process of operation, which provides an important guarantee for the safety of the workers. The prediction task of human behavior aims to predict future action video frames based on given of human action video frames. Considering that human action videos contain both background scenes which are relatively time-invariant among frames, and alseo human actions information which are time-varying components in videos, a global generative adversarial network is used to model the time-invariant background and coarse human profiles; Then a local generative adversarial network is utilized to furtherly refine the time-varying human body details in the video. Compared with existing methods which only obtain pixel-level action predictions, these experiments indicate that our global and local combined approach can capture both spatial appearances and temporal dynamics of videos simultaneously.