The grid data of numerical weather prediction (NWP) are used to forecast the 2min wind velocity and direction at fixed time for the station affected by typhoon. Integrating and analyzing the station data of MICAPS, we select the relatively complete observational materials of more than 400 stations in coastal regions and islands as the raw data for prediction. Meanwhile， the nine suitable forecast factors are fixed by the correlation analysis with the NCEP reanalysis grid data. Based on the back propagation (BP) network, latitudinal and longitudinal artificial neural network models are developed for each station respectively. The absolute error of fitting wind velocity is 1.3 m·s-1. The test of independent samples shows that the absolute error of wind velocity is less than 2.0 m·s-1. This is a statistical interpretation of NWP, and it can pre-estimate the wind velocity effectively in landfall typhoon-affected areas.