Based on the results of large-scale dynamic triaxial tests on coal gangue subgrade fillers, the grey correlation analysis method was used to analyze the influencing factors of cumulative deformation, and four characteristic parameters were determined, including confining pressure, compactness, gradation, and cyclic loading number. The particle swarm optimization (PSO) algorithm was used to globally optimize and assign the weights and thresholds of the back propagation (BP) neural network, and a PSO-BP neural network prediction model for cumulative deformation of coal gangue subgrade fillers was established. By comparing the traditional BP neural network with the established prediction model, the feasibility and superiority of the established prediction model were verified. The generalization ability of the model was further analyzed by comparing the prediction effects of the model under different learning levels, which proved the prediction potential of the model.