The statistical models of SVM, Kalman and multi dimensional linear regression are established for extreme temperature with the ECMWF (European Centre for Medium Range Weather Forecasts) grid data and the observation data of Qingyan from 2003 to 2009. The methods of average, weighting, and regression are used in forecast integration. The integrated forecasts indicate that the accuracies of minimum temperature are 0.8% to 24.2% higher than those of maximum temperature by means of SVM, multiple regression and integrated methods at various time from June to December 2010. The weighting method is the best, then the integration, but the weight is different for maximum and minimum temperature: the accuracy of maximum temperature is better when the SVM weight is greater, and the accuracy of minimum temperature is better when the weight of multiple regression is greater. The forecasting accuracy decreases with increasing led time for both single and consensus forecast methods. The month to month verification indicates that the accuracy of the single SVM method is relatively high and stable; that of the Kalman is relatively low; and that of the regression is unstably. The integrated results of three methods show improving, and the absolute errors of both maximum and minimum temperature after weighting integration are small, better than those of the average and regression methods.