Abstract: The wheel wear of light rail trains is difficult to predict due to poor information and small data samples. However, the amount of wear gradually increases with the running mileage. The grey future prediction model is supposed to deal with this problem effectively. In this study, we propose an improved non-equidistant grey model GM(1,1) with background values optimized by a genetic algorithm (GA). While the grey model is not good enough to track data series with features of randomness and nonlinearity, the residual error series of the GA-GM(1,1) model is corrected through a back propagation neural network (BPNN). To further improve the performance of the GA-GM(1,1)-BPNN model, a particle swarm optimization algorithm (PSO) is implemented to train the weight and bias in the neural network. The traditional non-equidistant GM(1,1) model and the proposed GA-GM(1,1), GA-GM(1,1)-BPNN, and GA-GM(1,1)-PSO-BPNN models were used to predict the wheel diameter and wheel flange wear of the Changchun light rail train and their validity and rationality were verified. Benefitting from the optimization effects of the genetic algorithm, neural network, and particle swarm algorithm, the performance ranking of the four methods from highest to lowest was GA-GM(1,1)-PSO-BPNN > GA-GM(1,1)-BPNN > GA-GM(1,1) > GM(1,1) in both the fitting and prediction zones. The GA-GM(1,1)-PSO-BPNN model performed best, with the lowest fitting and forecasting maximum relative error, mean absolute error, mean absolute percentage error, and mean squared error of all four models. Therefore, it is the most effective and stable model in field application of light rail train wheel wear prediction.