Abstract: Based on the model identification and an analogue forecasting, a new approach based on Self-Organizing feature Map (SOM) and cross validation is constructed, which is called K-nearest neighbor nonparametric estimation bootstrap model (KNN). 500 hPa geopotential height and 700 hPa u, v wind field over Northwest China are analyzed by the model clusterings at first, then the optimal K combination is sought using cross validation aiming at past samples under different weather patterns. Forecasting identification value of each synoptic pattern is determined by K-data, according to historical record. When forecasting in real time, what kind of synoptic pattern is to be known first, then K-data of different time is used to compute the nearest neighbor of real forecasting predictor to historical material predictor. Finally forecasting conclusion is obtained by using the standard of forecasting identification value. In order to validate the effect on cluster synoptic pattern to KNN, T213 NWP products from 2003 to 2006 in winter half year and the data of daily maximum velocity in Ningxia are used to construct prediction models of daily maximum velocity≥6 m/s pattern in Ningxia under synoptic and non-synoptic patterns at one time, data from Jan to May in 2007 is used for forecast experiments. The forecast evaluation results show that although the probability of original sample is reduced when adding the Self-Organizing feature Map of KNN, more false alarms in forecasting are avoided, so that the effect of forecasting is improved in general, especially the forecasting effects of some synoptic patterns compared with those that aren't patterned. The result is that the forecasting information of Ningxia high wind can be reflected by improved KNN. What's worth pointing out is that, the number of synoptic patterns is reduced when patterned, so the forecasting will be effected to some extent. It has a good effect for meteorological observing station which has more original samples, but it is not good for the ones that have less original samples. Therefore if there are more historical data which can reflect the wide range of system changing, the forecast accuracy will be improved significantly and it has a great value for operational usage. Classification analogue prediction thinking can be expanded by these results.