Based on the observation data from 126 national meteorological stations in Yunnan Province from 2016 to 2025 and and ERA5 reanalysis data for the same period, the spatial distribution characteristics of regional heavy fog in Yunnan were diagnosed and analyzed using Rotated Empirical Orthogonal Function (REOF), and fog zoning was carried out combined with topographic factors. Furthermore, machine learning methods such as random forest were employed to establish zonal heavy fog forecasting models, and the results were compared and verified with SCMOC and ECMWF numerical forecast products. The results show that: 1)Heavy fog in Yunnan Province exhibits four dominant spatial modes: an out-of-phase pattern between the southwest and the rest of the region, southeastern Yunnan, northeastern Yunnan, and northwestern Yunnan. Combined with topography, the province is finally divided into four fog zones: Zone 1 (western and northwestern Yunnan), Zone 2 (northeastern and central Yunnan), Zone 3 (southwestern and southern Yunnan), and Zone 4 (southeastern Yunnan). 2)The random forest model achieves an accuracy of 0.82~0.84 for modeling samples and 0.80~0.83 for test samples in each zone, which is significantly superior to decision tree, K-nearest neighbor, logistic regression, and Bayesian methods. 3)There are regional differences in the importance of forecast factors in each fog zone. High-altitude areas (Zone 1 and Zone 2) are strongly dependent on water vapor conditions at 600—700 hPa. Regions significantly affected by eastern industrial activities (Zone 2 and Zone 4) are largely contributed by initial visibility. In relatively low-altitude areas (Zone 3 and Zone 4), the importance of sea level pressure and temperature factors within the boundary layer is more prominent. 4)The comprehensive performance of the short-range heavy fog forecast model established based on random forest is superior to those of two mainstream forecast products, SCMOC and ECMWF. The improvement in forecast effect is more significant in autumn and winter, and Regions 3 and 4 exhibit higher forecast accuracy.. For a regional heavy fog event that lasted two consecutive days, the TS scores of the random forest model are 0.38 and 0.41, which are significantly higher than those of SCMOC (both < 0.1) and ECMWF (0.26 and 0.11, respectively).