Abstract: Due to the low data dependency, to the high accuracy, and to the universality, physics-information neural networks (PINNs) have become a useful method for solving differential equations. However, existing studies on PINN for solving structural dynamic response equations mostly base upon simple assumptions about parameters to analyze their accuracy and feasibility, which do not consider issues such as large differences in the physical parameter scales and high vibration frequencies in real structures, leading to poor approximation performance of PINN under real parameters. Therefore, this study constructs a parameter-balanced Fourier-feature PINN model (PBFF-PINN) suitable for practical engineering structures. Firstly, the model performs parameter scales transformation on the structural vibration equations, solving the problem of learning imbalance and poor stability in PINN caused by large differences in the parameter scales of real structures. Then, it introduces Fourier feature mapping to transform the traditional time-domain input into time-frequency domain input to enhance the expression capability and solution accuracy of PINN in practical structural vibration problems. Ablation experiments on vibration systems with practical parameter scales validate the necessity of these two processes for solving real structural vibration equations and the applicability of PBFF-PINN.