In order to improve the accuracy of annual runoff prediction, the simultaneous heat transfer search (SHTS) algorithm was used to optimize the key parameters and the mixed weight coefficients of hybrid nuclear support vector machine (SVM). A mixed kernel SHTS-SVM annual runoff prediction model was proposed. The SHTS algorithm was verified by six standard test functions at different dimensions, and verified with the teaching optimization (TLBO) algorithm and gray wolf optimization (GWO) algorithm. Two of the annual runoff prediction examples were used to verify the mixed kernel SHTS-SVM model and compared with the prediction results of polynomial kernel SHTS-SVM, Gaussian kernel SHTS-SVM and SHTS-BP models. The results showed that the optimization accuracy of SHTS algorithm is better than the TLBO and GWO optimization algorithms, and it has better extreme value searching ability and robust performance. The mixed kernel SHTS-SVM model combines the advantages of the polynomial global kernel function and the Gaussian local kernel function. It is superior to the comparison model in terms of prediction accuracy, generalization ability and so on, and has good practical application value.