State of Health Estimation for Battery Packs Based on Ensemble Bayesian Neural Networks Integrating Uncertainty Quantification and Calibration

Many data-driven state of health (SOH) prediction models lack reliable uncertainty quantification (UQ) and calibration (UC), which are vital for practical decision-making. This paper integrates UQ and UC for battery pack SOH estimation by proposing an ensemble Bayesian neural network model. This model is constructed by combining a Bayesian neural network, an ensemble neural network, and a multi-layer perceptron using bagging and stacking. To address battery inconsistency, features are extracted from cell-level data. The model quantifies aleatoric and epistemic uncertainty through output variance and the mean squared error between predictions and actual values. Additionally, a novel temperature scaling method is employed to calibrate this output uncertainty. The proposed model, tested on electric vehicle battery pack data, shows an error below 1%. It also provides reliable UQ and UC, proving its accuracy and robustness.