Bayesian adaptive elastic net for early prediction of battery cycle life

Abstract

Accurate prognosis of battery remaining useful life is essential for evaluating reliability and optimizing the utilization of energy storage systems. To address the challenge of multicollinearity among degradation features, we propose a Bayesian adaptive elastic net method that improves prediction accuracy while providing comprehensive uncertainty quantification. The proposed approach implements adaptive shrinkage, applying separate penalty parameters to each regression coefficient to shrink the non-significant variables while preserving the effects of important predictors. By representing Laplace priors as Gaussian-exponential mixtures through data augmentation techniques and approximating the normalizing constant numerically, we develop an efficient Metropolis-Hastings-within-Gibbs sampling framework for posterior inference. Validation using numerical simulations and battery cycle life data demonstrates that the proposed method delivers superior prediction accuracy and robust uncertainty quantification, even under severe multicollinearity, providing an effective solution for battery lifetime prediction.