Different key factors are chosen and composed from the numerical forecast products of the models of ECMWF and T639 from 2013 to 2014, and the surface hourly temperature forecast are established by means of the nonlinear regressive methods—BP (Back Propagation) neural net， SVM (Support Vector Machine), and constructed nonlinear function methods. The results show that the error correction method can reduce forecast error more stably when the error is large; adjusted by the mean errors, all the three methods can forecast the hourly temperature satisfactorily, with mean absolute errors reduced generally by 0.5 ℃. The test of independent samples indicates that the mean absolute errors of hourly temperature forecast by utilizing the methods of BP, SVM, constructed nonlinear function, and error correction are less than 1.5 ℃, 1.7 ℃, 1.8 ℃, and 1.4 ℃, respectively. The constructed nonlinear function has a good performance in fitting and prediction in general.