The area of southern Xinjiang is a high occurrence area of dust weather, which has a serious impact on local residents’ life. To reveal the characteristics of local windsand weather variation and predict future trends, a wavelet decomposition method is used to decompose the time series of dust weather in the southern edge of the Taklimakan Desert from 2008 to 2016 into stationary fluctuation terms and nonlinear trend terms, according to the characteristics of the data. The autoregressive (AR) model and the least square support vector machine (LSSVM) are selected to predict the variation trend. Finally, the time series prediction of the number of dust weather days is achieved by the addition principle reconstruction. The results show that the dust weather is a typical spring and summer type, mainly concentrated in the period from March to September, and the peak value appears in May. The predicted value of the combined model is basically consistent with the measured value, and has a higher prediction accuracy (absolute error is 4 d, root mean square error is 3.764 d). Compared with the prediction results of AR model and LSSVM, the correlation coefficient of combined model increases 0.12 and 0.31 respectively), and has a better application prospect. Thus, it could provide scientific basis for preventing wind and sand disaster and guiding actual production and life in the research area.