This study quantitatively analyzes the dynamic acceleration performance of medium to-large agricultural tractors (60HP and 75HP classes) and validates the reliability of a system-level simulation model through real-vehicle field tests. As the development trend of agricultural machinery shifts toward higher performance and complexity, establishing a virtual verification-based design process is crucial for reducing development period and costs. A comprehensive time-domain dynamic model was constructed using Matlab/Simulink to simulate the entire power transmission path, encompassing the engine polynomial torque curves, sequential gear reduction ratios, and non-linear physical resistance factors. The consistency of this physical model was evaluated against empirical tests conducted strictly under ISO 5721-1 and OECD(Organisation for Economic Co-operation and Development) test codes [1,2]. The physical model demonstrated high reliability, with top speed and target reaching time errors of 1.1% and 3.7%, respectively, satisfying the general performance prediction tolerance of ±5%. A rigorous Design of Experiments (DOE) sensitivity analysis revealed that engine power and vehicle mass are the dominant factors, dictating over 83.4% of the acceleration variance. To minimize residual errors caused by unmodeled non-linearities such as tire slip and transmission delay, a Long Short Term Memory (LSTM) regression neural network was introduced [8]. The AI-based correction algorithm significantly improved prediction precision, reducing the Mean Absolute Percentage Error (MAPE) to within 0.44% and achieving a coefficient of determination (R2) of 0.993. This synergistic physics-data hybrid framework provides a highly robust foundation for digital twin-based performance verification in the off-road vehicle industry [35].