In order to improve the prediction accuracy of the detonation parameters, a new method based on Support Vector Machine(SVM) theory and Hessian Local Linearly Embedding algorithm (HLLE) was proposed to predict 16 traditional composite explosives. The original data after dimension reduction with HLLE, was input to regressively predict the heat and velocity of composite explosion by SVM. The best kernel function parameter and penalty factor are selected by Genetic Algorithm(GA). The calculated results of the explosives almost agree with those of the literature, and the relative error is within ±3%. Using the method, the values of detonation velocity for 2, 4-diamino-3, 5-dinitro pyrazine-1-oxide/1-methyl-3, 5-dinitro-1, 2, 4-triazole and 2, 4-diamino-3, 5-dinitro pyrazine-1-oxide/1-methyl-4, 5-dinitroimidazole were predicted and compared with experimental and the relative errors are 2.91% and 3.72%, respectively, showing that the proposed method is comparatively accurate.