<p>The vanilla physics-informed neural networks (PINNs) approach struggles to solve partial differential equations (PDEs) with strong nonlinearity and high-order derivatives, and doesn’t respect some inherent physical properties of the governing PDEs, such as causality, energy and mass conservation. In this article, the solitary wave solutions of the generalized Rosenau-type equations with power-law nonlinearity are investigated employing a structure-preserving PINNs (spPINNs) method. An intermediate variable is introduced to alleviate the difficulties posed by fourth and fifth order spatial derivatives. To consider the energy and mass conservation properties of the governing equations, the Gauss-Legendre numerical integration formula is utilized to calculate energy and mass integrations, and the training points are selected according to the quadrature nodes through an invertible transformation. In addition, a shallow neural network is imported as an auxiliary tool, and a new structure loss is attached to the loss function as a penalty term. Numerical examples successfully simulate the motion of bright and dark soliton solutions, as well as the interaction between two solitons, verifying the effectiveness and robustness of the proposed spPINNs algorithm. Moreover, comparative experiments with the conventional PINNs manner and some improved PINNs approaches highlight the advantages of the presented spPINNs method both in improving accuracy and respecting intrinsic physical properties.</p>