[Objective] To address the limited adaptability of traditional numerical methods for water hammer under complex boundary and uncertain conditions， as well as the constrained convergence efficiency and prediction accuracy of Physics-Informed Neural Networks （PINNs） in strongly nonlinear transient flows， this study proposes a PINNs-based framework for transient pressure prediction in pipelines aiming to enhance the prediction accuracy of pressure variations in pipeline hydraulic transients. [Methods] In this framework， the continuity and momentum equations are embedded into the neural network loss function， and automatic differentiation is employed to solve the governing partial differential equations， enabling high-accuracy pressure simulation under limited labeled data conditions. Furthermore， an improved Locally Adaptive Activation Function （LAAF） is introduced by incorporating a trainable offset parameter β， while a mini-batch gradient descent strategy is adopted to enhance the model’s training stability and prediction robustness. [Results] Numerical case studies demonstrate that the proposed model effectively captures the transient characteristics of water hammer pressure. Compared with the Baseline-PINNs model and LAAF2-PINNs model， it achieves significantly higher predictive accuracy， with the relative L₂ error reduced by 37.32% and 23.83%， respectively. In addition， sensitivity analysis of wave velocity， verification under multiple initial flow velocity conditions and generalization ability analysis demonstrate that the improved model has superior generalization ability and robustness. [Conclusions] The proposed M-LAAF2-PINNs model provides a novel and effective approach for achieving high-precision water hammer transient analysis under limited observational data and uncertain boundary conditions， offering superior technical support for pressure monitoring， parameter inversion， and intelligent scheduling in long-distance pipeline systems.