Predicting how electric vehicle batteries lose capacity over time remains one of the most pressing challenges in EV reliability. Researchers conducted a systematic experimental study comparing three classes of predictive algorithms—machine learning, statistical learning, and reliability-based approaches—to forecast battery degradation at both the individual cell and complete pack level.
The team tested a 26 ampere-hour two-wheeler battery pack alongside its 18650-type cells under identical thermal and electrical conditions. This dual-scale approach revealed how degradation patterns differ between single cells and integrated packs, a critical distinction often overlooked in battery modeling. Five specific techniques were implemented and evaluated: support vector regression with radial basis function kernels, Gaussian process regression with Bayesian hyperparameter optimization, artificial neural networks with 15 hidden neurons, polynomial regression with ridge regularization, and Weibull distribution fitting.
Results demonstrated that support vector regression achieved the highest accuracy, reaching R² values of 0.9995 at pack level and 0.9988 at cell level. Gaussian process regression and artificial neural networks performed nearly as well, while traditional reliability models showed notably reduced accuracy across both scales. The researchers evaluated performance using standardized metrics including root mean square error, mean absolute error, and coefficient of determination.
These findings carry significant implications for real-time battery management systems in production vehicles. Accurate health monitoring reduces the conservative range estimates that manufacturers typically apply to account for uncertainty, directly improving customer experience and EV competitiveness. The data-driven framework accommodates the nonlinear aging behavior and operational variability that plague physics-based models, while remaining computationally feasible for embedded automotive systems.
Next steps involve validating these algorithms across additional battery chemistries, operating profiles, and vehicle platforms. Integration into production battery management systems could enable predictive maintenance alerts and adaptive charging strategies that further extend pack longevity.
