Five standard test functions were used to verify the multi-group teaching optimization(MGTLO)algorithm, and the simulation results were compared with those of the basic teaching optimization(TLBO)algorithm, shuffled frog leaping algorithm(SFLA), differential evolution(DE)algorithm and particle swarm optimization(PSO)algorithm. MGTLO was used to search for the optimum model parameters and the weight coefficients of the combined model based on the generalized regression neural network(GRNN), the radial basis function neural network(RBF)and the support vector machine(SVM)model elements. Four combined prediction models, including MGTLO-GRNN-RBF, MGTLO-GRNN-SVM, MGTLO-RBF-SVM, and MGTLO-GRNN-RBF-SVM were proposed and case studies of the runoff prediction were performed at the Yamadu Hydrological Station of the Yili River in Xinjiang and a hydrological station in Yunnan Province. The predicted results were compared with the following six single models, MGTLO-GRNN, MGTLO-RBF, MGTLO-SVM, GRNN, RBF and SVM. The results show that the optimization accuracy of MGTLO algorithm is better than that of TLBO, SFLA, DE and PSO, with good convergence speed and global optimization ability. The combined model merges the advantages of MGTLO algorithm, GRNN, RBF, and SVM model elements. It is superior to single models in terms of prediction accuracy and generalization ability. MGTLO algorithm can effectively optimize the parameters and weight coefficients of the combined models and the MGTLO-GRNN-RBF-SVM model has the highest prediction accuracy.