Abstract: Currently the application of real-time hybrid simulation is limited by the low computational efficiency of numerical substructures, thus cannot be extended to the seismic tests of large and complex structural models. In this case, a real-time hybrid simulation method based on parameter identification and neural network surrogate model is proposed. The method first adopts a neural network to establish a surrogate model for the numerical substructure, with the constitutive parameters of the experimental substructure as the inputs of the neural network; then these constitutive parameters are identified online during the test, and the surrogate model is used to predict the boundary displacement of the numerical substructure. The core of this method is that the inputs of the neural network are the constitutive parameters of the experimental substructure instead of the restoring forces, so that even if there is an error in the predicted displacement, it will not affect the identification of the constitutive parameters, thus the prediction error will not be passed to the next time step. This ensures the accuracy of the surrogate model for the numerical substructure. Virtual real-time hybrid simulation on an 8-story reinforced concrete frame structure with buckling-restrained braces was conducted using the proposed method, in which a damper at the bottom story was taken as the experimental substructure. The results show that the proposed surrogate model has good accuracy and computational efficiency.