Remaining Useful Life Prediction of Photovoltaic Modules Based on Kalman Filter and Variational Inference

Long‐term field degradation of photovoltaic (PV) modules is nonstationary and uncertain. This makes RUL prediction difficult when only limited observations are available. This study proposes a rolling probabilistic RUL framework based on Kalman filtering and variational inference (VI). Each sliding window represents the current degradation stage. Within each window, Kalman filtering and Rauch–Tung–Striebel smoothing estimate the latent degradation state. A drifted Wiener process describes local degradation. VI then jointly estimates the posterior distributions of the drift, diffusion variance, and observation‐noise variance. These posteriors are used in a first‐passage‐time model to obtain the RUL distribution. The method is tested with a 12‐year field dataset and four additional module datasets. At the LAST stage, relative errors range from 0.18% to 0.44%, with a mean of 0.279%. The method also shows lower errors than the compared stochastic models for these datasets. Sensitivity results show stable performance under different prior settings and initial‐state covariances. The rolling framework supports RUL prediction under changing PV degradation conditions.