<p>Due to the complexity of the winding structure, it is challenging to establish an accurate prediction model for the frequency response function. A parallel hybrid model HWVM (Hybrid Winding Vibration Model) in guided mode is innovatively proposed based on the hybrid model, which organically combines the mechanism and data. Firstly, the error model dataset is constructed with the mechanism model as the basis; at the same time, the mechanistic vibration mode error is innovatively introduced as a bootstrap (guided) layer to provide a priori guidance of physical information for the network learning. Gaussian process regression is used for its construction, DNN (Deep Neural Networks) is used for the main network, and DDEM (Data-driven Error Model) is finally constructed. Further, physical information-guided layer has been verified on different networks (on different machine algorithms). Meanwhile, the overall performance of the fourth-order key-modes prediction model is improved under the concurrent hybrid model, HWVM (higher ability to inhibit overfitting, improving the accuracy and reducing the average number of iterations of the model training). The model has both interpretability and generalization ability with high accuracy, and is capable of extrapolating the response prediction in both spatial and frequency domains. The method proposed can quickly and accurately predict the displacement response of the winding structure under forced vibration, which can be used to optimize the design in the development of intelligent turbine.</p>