To quickly and accurately predict the critical flashover voltage of composite insulators in catenary systems and reduce the workload of artificial pollution tests, a prediction model for composite insulator pollution flashover voltage is proposed. First, the performance of the back propagation (BP) neural network is enhanced using the subtraction average based optimizer (SABO) algorithm improved by the golden sine algorithm (GSA) and piecewise linear chaotic map (PWLCM). Second, artificial pollution tests are conducted to obtain the flashover voltage of 10 different composite insulators, and relevant test parameters are collected. Third, the Obenaus model is used to analyze the pollution flashover behavior of composite insulators, and the Spearman correlation coefficient method is employed to select 4 parameters closely related to the critical flashover voltage of composite insulators as input features for the prediction model. Finally, the prediction model is comprehensively evaluated using five-fold cross-validation and compared with prediction results from commonly used intelligent optimization algorithms. The results show that the GSABO-BP model predicts the flashover voltage of composite insulators with an average absolute error of 1.244 kV, an average absolute percentage error of 2.25%, and a coefficient of determination consistently above 0.98. Compared to the original SABO-BP model, the average prediction error is reduced by 67.80%. The GSABO-BP model demonstrates high prediction accuracy for the flashover voltage of composite insulators, which is significant for the anti-pollution protection of electrified railway power supply systems.