ABSTRACT: In this paper, the average absolute error, standard deviation, correlation coefficient and accuracy of ECWMF, GRAPES and JAPAN models for wind speed prediction of 84 stations in Guizhou Province are tested by statistics. The average error of EC model is between - 1.9 and 1.7, GAPES model is between - 3.1 and 0, JAPAN model is between - 3 and 1.9, and the average absolute error of EC model is between 0.6 and 2.1. Among them, GAPES model is between 1.1 and 3.3, JAPAN model is between 0.6 and 3.1, EC model and JAPAN model are 0.5, GRAPES model is 0.4, the accuracy of wind speed prediction error absolute value is 2 m * S-1 is 83%, the lowest is 54% of weaving gold, GRAPES model is 56%, the lowest is 32% of Changshun, JAPAN model is 75%, the lowest is 29% of Changshun. By comparing the forecasting results of the three models in 9 prefectures, it is found that the forecasting effect of the EC model is better than that of the other two models. Among them, EC model has the best forecasting effect in Tongren area, southwest Guizhou is the worst, GRAPES model has the best forecasting effect in Zunyi area and Bijie area is the worst; JAPAN model has the best forecasting effect in Zunyi area and the worst forecasting effect in southwest Guizhou. Two of the three models have poor forecasting effect in southwestern Guizhou. The BP neural network is used to revise the wind speed prediction of the two models, and the results of the three models are improved obviously. The improvement of error and accuracy is obvious, and the improvement of correlation coefficient is small.