On the basis of data sets of wind speed and wind power observed from 33 wind turbines in a wind farm in Shaanxi Province, China and with wind speeds at a height of 100 m forecast by the ECMWF model, the wind speed prediction was corrected using simple linear regression and introducing the optimal training period and forecast models for wind power business were comparatively studied using the perfect prognostic (PP) and model output statistics (MOS) methods. The optimal training period was defined as a sliding cycle, which should be given before the forecast date. For training days varying from 10 to 100 d, taking some past forecast days as an evaluation period to calculate the root-mean-square error (RMSE) of the forecast, the optimal training period was decided according to the variation of the RMSE of the wind speed or wind power forecast with the training days. A prediction equation of wind power was established using the PP method according to historical observations of wind speed and wind power in the training sample. Then, the numerical prediction product of wind speed was used to replace the observation data in the prediction equation, which was applied to a wind power prediction business. In contrast, the MOS method directly used the wind speed of the numerical weather prediction model to establish the prediction equation, which was quite different from the PP method. The result showed that wind speed predictions after correction and wind power were improved during the optimal training period, reducing the prediction error to the greatest extent. The optimal training days were determined by the RMSE of the forecast during the evaluation period of 10 d before the forecast date. The wind speed at a height of 100 m forecast by the ECMWF model was closer to that of the observation at the height of the fan hub, and the RMSE and average absolute error (ABE) respectively ranged from 2.5 to 4.1 m·s^-1 and 1.9 to 3.3 m·s^-1. The prediction error of the wind speed after correction became smaller, with the percentage of improvement in the RMSE of the wind speed forecast ranging from 33% to 46%, and the RMSE and ABE of the wind speed prediction respectively ranged from 1.5 to 2.4 m·s^-1 and 1.1 to 1.9 m·s^-1. Thus, wind speed predictions could be applied to wind power prediction businesses using both the PP and MOS methods. Although the wind power prediction error of the MOS method was nearly the same as that of the PP method when the wind speed prediction was corrected, the MOS method, which did not need to improve the quality of wind power prediction by correcting the wind speed prediction, was slightly better than the PP method in terms of complexity. The RMSE and ABE of the wind power forecasts of both the PP and MOS methods were 17%-25% and 11%-18%, respectively, within the forecast time of 3-72 h. Thus, both the PP and MOS methods could be applied to wind power prediction businesses.