The conventional wind power prediction method is primarily generated by numerical weather prediction based on historical scada build statistical model or physical model to get the total power prediction. As the new wind farm could not collect the historical scada data, and the accuracy of the wind farm power prediction relies on the accuracy of short-term wind speed prediction. Therefore, in order to improve the accuracy of short-term wind speed prediction for new wind farm. First of all, numerical weather model exported the meteorological element on turbine hub height layers; secondly, the output data are corrected by the establishment of neural network model and multiple linear regression models; finally, the sources of the error are classified and analyzed. The wind farm test in Jiangsu province results indicate that: this method can improve the accuracy of wind speed prediction significantly and eliminate the amplitude deviation of numerical weather prediction compared with traditional methods, but the phase deviation is still the main source of error.