For the poor adaptability of multi-fidelity methods for single-surrogate models, a multi-fidelity optimization method to build a hybrid surrogate model is proposed. This method combines three regular surrogate models, including PRS, Kriging, and RBF. According to the characteristics and prediction accuracy of different surrogate models, a hybrid improved surrogate model is constructed. A ship to shore gantry crane is as the research object. High and low fidelity finite element models of this crane are established. Latin hypercube sampling is used to sample load cases of the crane, and stresses are calculated via the finite element model and extracted into a stress database. On this basis, a low-fidelity surrogate model and a residual surrogate model for the stress points of the crane's key parts are established. Furthermore, these models are combined into a multi-fidelity surrogate model. This model can rapidly predict stress values according to new inputs. The accuracy of the predicted outputs is compared with the actual finite element outputs, which verifies the performance of the proposed model.