Predicting voltage degradation at a fixed operating load from proton exchange membrane fuel cells (PEMFCs) subjected to dynamic load cycling is important for durability assessment and maintenance scheduling. This study develops a hierarchical residual forecasting scheme that combines exponential smoothing (ETS), autoregressive integrated moving average (ARIMA), and iTransformer. For every repeated load cycle, the voltage measured at the same predefined load plateau and within-cycle phase is extracted to form a condition-consistent degradation sequence. Pearson-correlation screening retains four synchronous operating variables. ETS first estimates the slowly changing degradation path, ARIMA represents the autocorrelated component of the first residual, and iTransformer forecasts the nonlinear secondary residual with the selected operating information. The three forecasts are then added to obtain the final voltage estimate. The framework is evaluated using a 1000 h dynamic durability dataset from a single PEMFC. Over the held-out interval from 720 to 1000 h, rolling one-step-ahead forecasts yield an RMSE of 1.51 mV, an MAE of 1.12 mV, and an R2 of 0.9759. Relative to the ablated configurations and the compared deep-learning models, the framework follows recovery behavior and nonstationary fluctuations more closely and produces a tighter absolute-error distribution.