Abstract: Using the observed data and three numerical weather prediction products of T7 online (T7),ECMWF and T639,the air temperature accuracy rates and prediction errors were tested and analyzed from January of 2014 to December of 2015 in Benxi city.With the BP neural network algorithm,a air temperature prediction error correcting model was established based on the error analyses of the models.The results show that,for the annual test,the accuracy rate of the minimum air temperature is higher than that of the maximum air temperature for the three models.The forecasting effects of the minimum air temperatures in summer and autumn are better than those in winter and spring.While for the maximum air temperature predictions,the performance of T7 is superior to those of EC and T639.The predication accuracy rate of these models decreases under a large air temperature disturbance.The forecasting errors of average minimum air temperatures for these models are less than 2.00℃.There is a big difference in the maximum air temperature forecasting among these models.T7 has the smallest error in the maximum air temperature forecasting.The systematic deviation of T639 is higher,with -1.34℃ and -2.87℃ for the minimum and maximum air temperatures,respectively.After correcting,the average absolute error decreases from 2.40℃ to 1.40℃,the systematic deviation and root-mean-square error are significantly reduced,and the forecasting accuracy rate obviously is improved from 50% to 80%.It indicates that this approach is valuable to be used in operation.

Key words: BP neural network, Air temperature, Error, Test