Using the WRF mesoscale numerical model and WRF threedimensional variational data assimilation (3DVar), a combined scheme of cumulus process, microphysical processes and radiation processes for local area is selected based on comparison experiments and the NCEP/GFS data are chosen as background field of the model; the background error covariance over the region centered in Yunnan is computed statistically and replaces the primary background error covariance in 3DVar; At the same time, considering the elevation difference between model bottom layer and surface observing station, the topography in surface data is corrected. Through the above study the localized WRF mesoscale numerical forecast system has been established, which could be better to characterize the dynamic and thermodynamic conditions in the local underlying surface, and has a significant improvement in forecast capability.