A capacity fade reliability model for lithium-ion battery packs based on real-vehicle data

Degradation characteristics of lithium-ion battery pack system (LIBPs) cannot be well described directly by the existing life model of cell, such as the interference imposed by stochastic uncertainty and coupling effect of multiple cells. In this article, we devise a battery capacity estimation and prediction algorithm leveraging deep learning and least squares regression, and develop a stochastic model of capacity degradation obeying a lognormal distribution and a reliability evaluation methodology to reveal the internal evolutionary rules and health status of LIBPs with multi-state logic. We generate a comprehensive dataset consisting of 150 cells and 3 battery packs derived from the cloud platform for cross-validation and take real-world vehicle daily operating charging and discharging data as an input for the proposed framework. The validation results indicate that our best model achieves 0.33 % mean absolute percentage error and 0.44 % root mean squared percent error for state of health estimation and 0.317 mean absolute error in remaining lifetime prediction. Further, remaining capacity probability distribution curves are established with estimated and predicted results, and reliability fade tendencies are meticulously profiled based on the series-parallel structure, which enables accurate electric vehicle service time evaluation under different conditions of uncertainty factors introduced by battery inconsistency and randomness of driving behaviors. This work highlights the promise of deep learning to supplant resource- and time-consuming traditional approaches, and emphasizes the potential of employing reliability theory to rapidly design new-generation battery safety management frameworks.