Abstract: In order to improve the forecast accuracy of 10 m wind, multiple models and verification methods are applied. Based on the ECMWF ensemble prediction system, the wind predict accuracy from direct model outputs at five stations over the Haituo mountain in Beijing is compared with that from four post-processing methods including one-variable regression, ridge regression, neural networks and particle swarm optimization-neural networks. The differences among these forecasts are discussed based on several verification methods. First, the verifications of systematic error and forecasting accuracy show that the prediction errors from regress and neural networks methods are much smaller than that from direct model outputs. Second, the wind frequency diagrams show that the forecast accuracy is improvement by regress methods in weak wind condition and by neural networks in the whole wind speed variance. The difference in the wind forecast varies greatly with wind direction for the different post-processing methods. The bias of wind direction forecast from direct model output in strong wind weather can be corrected by regress and neural networks methods. Finally, possible methods for improving wind forecast accuracy in complex terrain region are discussed.