<p>Square guide rails, as typical ultra-long components, are prone to large deflection bending during machining. Therefore, accurate straightening is a prerequisite for ensuring assembly accuracy and in-service performance. Most existing straightening models are based on the plane-section assumption and are applicable to cases with small initial deflections. However, when straightening rails with large initial defects, the validity of the plane-section assumption declines significantly. Consequently, these models fail to accurately characterize the overall post-straightening morphology, and a dedicated evaluation framework and prediction model tailored to large-deflection scenarios are required. To address this challenge, this study proposes morphology descriptors for post-straightening profiles under large initial deflections, thereby overcoming the limitations of conventional indices that fail to comprehensively reflect the global rail shape. A finite element model of the pressure straightening process was developed using ABAQUS and validated against experimental measurements, demonstrating satisfactory reliability and accuracy. Subsequently, a parametric modeling strategy was adopted to build a dataset that covered combinations of initial deflection and straightening strokes. Finally, a deep neural network was developed to predict the straightening stroke. The application results indicate that the proposed model achieves favorable straightening performance and effectively improves the overall straightness of square guide rails with large initial defects.</p>