Prediction of Mid-Span Prestress Loss Effects for Cable-Stayed Truss Bridges Using an Improved GA-BP Neural Network

In prestressed concrete cable-stayed truss bridges, the prestress losses in the upper chords, tensile web members, lower chords, and closure segment of the mid-span region have a significant influence on the stress distribution and node displacements of the truss structure. To rapidly and accurately reveal the structural response laws induced by multi-source prestress loss coupling, an improved genetic algorithm (GA) and backpropagation (BP) neural network hybrid model, referred to as the improved GA-BP algorithm, was developed based on orthogonal experimental design. The model was validated by comparing its performance with those of BP, GA-BP, PSO-BP, and SABO-BP models under the same data partitioning. The improved GA-BP model achieved R2 values of 0.9727, 0.9861, 0.9720, and 0.9524 for the maximum tensile stress, maximum compressive stress, maximum X-displacement, and maximum Z-displacement, respectively, with corresponding MAPE values of 2.28%, 0.86%, 0.43%, and 1.11%. Sensitivity analyses based on 800 sets of global stochastic predicted samples revealed that the prestress loss of the upper chord plays a dominant role in controlling node displacement and compressive stress, while that of the tensile web members serves as the core sensitive component inducing mid-span tensile stress exceedance. These findings provide practical guidance for structural health monitoring: the upper chord prestress loss should be prioritized for deformation control, while the tensile web members warrant close inspection to prevent tensile stress exceedance. The study demonstrates that the improved GA-BP neural network is suitable for predicting prestress loss effects and quantitatively evaluating structural responses for such complex bridges.