<p>Soliton systems play a pivotal role in nonlinear optics and communications, where accurate identification of their physical parameters is crucial for understanding their dynamic behavior. Although Physics-Informed Neural Networks (PINNs) have been widely applied to inverse modeling of partial differential equations, challenges such as unstable convergence and insufficient accuracy remain when dealing with high-order complex-valued systems. To address these issues, this study proposes CSD-PINN, a Complex-Symmetric, Dynamic-Weighted PINN framework that jointly optimizes the network architecture and training strategies. Structurally, a symmetric dual-network architecture is adopted, where two independent yet structurally identical subnetworks are employed to separately model the real and imaginary parts of complex solutions, with multiple trainable physical parameters explicitly embedded into the residual terms. On the training side, dynamic loss weighting, gradient clipping, and learning rate scheduling are incorporated to enhance convergence and stability. Numerical experiments on five soliton cases governed by two types of high-order nonlinear Schrödinger equations (HNLSEs) demonstrate that the proposed method reduces the average relative error of all parameters to within 1%, significantly outperforming traditional PINN models. Notably, in scenarios where the traditional model fails to converge, the CSD-PINN achieves over 90% reduction in the relative error of some parameters and over 80% reduction in the overall relative error range. Additional ablation studies further validate the model’s strong robustness and generalization ability under varying activation functions, neuron configurations, noise levels, etc. In summary, this study provides an effective and reliable tool for modeling and parameter identification in nonlinear wave systems, with broad application prospects in optical physics, soliton-based communications, and light–matter interactions.</p>