Online action detection focuses on recognizing actions happening in the latest frames of streaming video. Given the strong correlation between human skeletons and actions, many researchers have attempted to use skeletons for online action detection. Recently, spatio-temporal graph convolutional methods achieve good action modeling effects, but they have limited capability for detecting actions in the latest frames. In this paper, we introduce a temporal enhancement technique to optimize the performance of skeleton-based online action detection, involving a temporal feature enhancement module and a motion difference module. The temporal feature enhancement module, modified based on Transformer, enhances the latest features temporally. The motion difference module introduces motion features into the network. Experimental results demonstrate that our method is competitive.