Automatic Train Protection (ATP) system reliability significantly impacts rail safety and maintenance efficiency, but current strategies lack data-driven spare parts optimization. We propose a hybrid failure rate prediction model combining time-varying filtered empirical mode decomposition (TVFEMD) and machine learning. First, ATP operational data undergoes interval segmentation and zero-value preprocessing. Next, grey wolf optimization (GWO) adaptively tunes TVFEMD parameters to decompose failure rate series into intrinsic mode functions (IMFs). Each IMF is independently predicted via weighted least squares support vector machine (WLSSVM), with final outputs aggregated through superposition. Validated using real ATP system data, the model achieves 0.0028 MAE, 0.0066 RMSE, and 0.0139 MAPE for 10-minute interval predictions with zero-inflated data, surpassing baseline methods. Results confirm its effectiveness for ATP failure rate forecasting.