In order to objectively describe the prediction performance of numerical forecast visibility products in complex terrain areas, this paper defined the low visibility process and selected typical processes in 2016 autumn and winter in Guizhou mountains to test visibility forecast of ECMWF global numerical model and south China mesoscale numerical model. Study found that the visibility level’s prediction performance of radiation fog was significantly worse than frontal fog process, and the mesoscale model of south China was significantly worse than that of ECMWF model. Fog’s prediction skills of each model were very low, while the prediction skills of light fog were obviously higher. In particular, ECMWF model had certain reference for light fog in frontal fog process, while prediction skills of each model for radiation fog were particularly insufficient. In frontal fog process, ECMWF model could predict fog in the western plateau of Guizhou and light fog in the east, but the prediction field of each time and time and method is not good in the central region of guizhou.