Abstract: Q195 steel is widely utilized across various industries, including construction, mechanical manufacturing, and automotive production, owing to its favorable plasticity, excellent weldability, and low cost. However, during the actual hot rolling process, thin plates of this grade are highly susceptible to shape defects such as edge waves, primarily caused by non-uniform distribution of rolling forces. These defects significantly compromise the dimensional accuracy of strip materials and the quality of final products. Therefore, accurately characterizing the rheological behavior of Q195 steel under high-temperature conditions and establishing a high-precision model for predicting its mechanical response during rolling are crucial for optimizing the rolling process and improving product quality. In this study, stress–strain data under different temperatures and strain rates were obtained through thermal compression experiments conducted on a Gleeble-3800 thermo-mechanical simulator. A hyperbolic sine-type Arrhenius constitutive model was developed using linear regression analysis. The results indicate that the flow stress of Q195 steel decreases with increasing deformation temperature and increases with rising strain rate. The hot deformation activation energy (Q) was determined to be approximately 258.2234 kJ·mol^-1. To further enhance prediction accuracy, a BP neural network model was developed based on the discrepancies between the constitutive model predictions and experimental data, enabling effective stress compensation. The integrated model achieved an average relative error of 0.729% and a correlation coefficient of 0.99983 when compared with experimental results. Subsequently, a VUMAT user-defined material subroutine compatible with ABAQUS was developed based on the improved constitutive model. Using actual industrial rolling parameters, a single-pass hot rolling finite element model was established. Comparison between simulation results and plant-measured data revealed that the predicted rolling forces agreed well with the measured values, yielding an average error of 1.239%. The prediction accuracy surpasses those obtained using the original Arrhenius and Johnson–Cook constitutive models, thereby validating the reliability and engineering applicability of the proposed modeling approach.