Prefabricated partition walls are significant components in single-tube double-track large-diameter shield tunnels, while the large slenderness ratio will render them the vulnerable components under seismic reaction. Therefore, this study proposes a three-stage PINN-assisted framework for predicting the hysteresis behaviour of prefabricated partition walls, achieving both accuracy and physical consistency under small-sample conditions. Time-history parameters including displacement, velocity and acceleration, are taken as inputs of the proposed network. Guided by physics priors, a physics‑informed recurrent neural network (PI‑RNN) is developed and trained, and a residual boosting is employed to progressively correct prediction errors, ultimately predicting the hysteretic force. Firstly, initial parameters are identified utilizing classical hysteretic ordinary differential equation (ODE) models (Bouc–Wen and Baber–Noori) [1-2] coupled with numerical optimization, which provide an interpretable prior for subsequent learning. Secondly, the RNN including Bi-LSTM layers and dense layers are constructed to estimate the internal hysteretic variable, and the physics equation layer maps these quantities to the force prediction. Training employs a composite loss function comprising a data-misfit term, a physics-based ODE constraint and regularization terms [3-5]. Eventually, the residual is learned via a gradient boosting machine (GBM) and the final output is obtained, balancing physical consistency with improved predictive accuracy. By applying the proposed framework on the hysteresis performance of four test specimens, the coefficient of determination R² generally exceeds 0.95 (dropping to 0.88 for Specimen 3), the minimum RMSE is 6.98 kN, and the model exhibits no observable drift in energy dissipation or tangent-stiffness evolution, which reveals high accuracy and physical consistency across multiple specimens. The proposed framework is computationally efficient, enables rapid prediction and real-time hybrid-test assessment, offering a scalable physics-guided data-driven tool for designing, evaluating, and optimizing prefabricated partition walls.