Machine Learning-Based Lifetime Prediction of Lithium-Ion Cells

Abstract

Precise lifetime predictions for lithium-ion cells are crucial for efficient battery development and thus enable profitable electric vehicles and a sustainable transformation towards zero-emission mobility. However, limitations remain due to the complex degradation of lithium-ion cells, strongly influenced by cell design as well as operating and storage conditions. To overcome them, a machine learning framework is developed based on symbolic regression via genetic programming. This evolutionary algorithm is capable of inferring physically interpretable models from cell aging data without requiring domain knowledge. This novel approach is compared against established approaches in case studies, which represent common tasks of lifetime prediction based on cycle and calendar aging data of 104 automotive lithium-ion pouch-cells. On average, predictive accuracy for extrapolations over storage time and energy throughput is increased by 38% and 13%, respectively. For predictions over other stress factors, error reductions of up to 77% are achieved. Furthermore, the evolutionary generated aging models meet requirements regarding applicability, generalizability, and interpretability. This highlights the potential of evolutionary algorithms to enhance cell aging predictions as well as insights.

Keywords: aging models; evolutionary algorithms; lifetime prediction; lithium-ion cells; machine learning.