In order to reduce various losses that may be caused by dense fog, the short-term prediction of dense fog’s trend has become a research hotspot in the field of meteorological short-term prediction. However, current researches focus on the temporal features of dense fog while ignoring its spatial features, so the prediction accuracy is still at a relatively low level. To this end, this paper proposes an algorithm of short-term prediction for dense fog’s trend based on fusion of spatio-temporal features. The algorithm takes multi-stations as nodes in the graph data. By introducing Graph Attention Network, we realize the extraction of spatial features. On this basis, using the Long-Short Term Memory network to further extract temporal features from the spatial features in order to realize feature-level fusion. Then we use the fully connected layer to output the predicted value of visibility. Further we can get to the predict result t based on the predicted value of visibility. We apply the algorithm on meteorological data released by National Centers for Environmental Information to carry out the experiment. The experimental results show that the F1-score and TS-score of the proposed algorithm take an improvement of 2%-12% on baseline models, which proves that the proposed algorithm has a great practical application value.