Bayesian Informer for probabilistic remaining useful life forecasting of solid oxide fuel cells

Accurate forecasting of remaining useful life (RUL) for solid oxide fuel cells (SOFCs) is essential to improve operational reliability and support health-aware control and predictive maintenance. In this paper, a novel prognostics framework is proposed to forecast the state-of-health (SOH) trajectory using two Bayesian sequence models: Informer and long short-term memory (LSTM). Aleatoric uncertainty is modeled with a variance output head trained with Gaussian negative log-likelihood, and epistemic uncertainty is estimated via Monte Carlo (MC) dropout. The two uncertainty sources are combined to form calibrated confidence bands for SOH and derived RUL. Because the forecast horizon shrinks as the cell ages, these variable horizons are handled with a single model using an exponentially weighted zero-padding technique, ensuring uniform sequence length while softly enforcing SOH degradation toward end-of-life (EOL). Integrating impedance features with SOH as inputs enhances the capture of degradation dynamics relevant to control and supervisory decisions. Extensive degradation experiments under redox cycling are conducted, and the collected experimental data are used to evaluate the proposed method for RUL estimation. The Bayesian Informer achieves high forecasting accuracy with a mean absolute error (MAE) of 0.0087 and coverage of 93%, while producing credible RUL distributions from first-hitting times (FHT) of the MC trajectories.