Abstract: This paper contributes to vehicle dynamics modeling by introducing a model architecture that addresses the state-space representations of nonlinear physical systems through data-driven techniques, thus bypassing the need for extensive parameterization. By effectively compressing time-series field test data, the model enables accurate long-term prediction of vehicle states and onboard sensor outputs, based on the sequential control inputs of driver assistance systems. This makes it suitable for both closed-loop simulation and offline analysis. Additionally, the model overcomes challenges such as sensor data unavailability and the effects of sensor noise in real-world scenarios. Utilizing a reduced-order approach within a neural state-space framework, the model comprises state and output networks configured as multi-layer perceptrons. The model is highly efficient in simulating vehicle dynamics during forward driving, its scalability is enhanced by a kinematics-based state limiter, which facilitates seamless transitions and improved adaptability across complex driving modes. The study also analyzes the impact of data down-sampling on model performance, which is crucial for the practical deployment of V2X. Validated through real-vehicle experiments under various driving conditions, the model demonstrates its capability to accurately replicate vehicle dynamics and its robustness in the face of sensor noise and down-sampled training data. This highlights the role of neural networks in refining both the credibility and practical utility of intelligent driving simulations. The source code used to train and evaluate the model is available at https://github.com/pansong/PyNSSM.