Abstract: The rapid development of artificial intelligence provides a customized solution for the free manipulation of electromagnetic waves by metasurfaces. This paper proposes a deep fully connected neural network model integrated with the idea of residual networks, which is used for the inverse prediction of structural parameters of broadband polarization-conversion metasurfaces from reflection coefficients. First, a three-layer metasurface unit structure is designed, and its 8 control parameters are determined. On this basis, by combining the refined parameter control idea of different metasurface structures with the efficient mapping capability of deep learning-based inverse design, an end-to-end mapping model from electromagnetic response to structural parameters is constructed. The residual connection mechanism is innovatively introduced, which effectively addresses the gradient vanishing problem in the training of deep networks. The paper focuses on elaborating the network architecture design integrated with residual connections, training strategies, and analyzes the impact of logarithmic transformation on prediction accuracy. Algorithm evaluation of the model shows that the coefficients of determination (R²) of the model's prediction results for all 8 structural parameters are greater than 0.9. The metasurface designed based on the predicted parameters maintains a polarization conversion ratio of over 90% across the entire frequency band of 8.8–24.4 GHz. Analysis indicates that this study provides an efficient and feasible method for the inverse design of metasurfaces, and this method can be further extended to the design of metasurfaces with more diverse functions.