Quentin MAYEMBA defended his PhD on jan. 13th, 2025 at 2PM.
Place : Amphi BU Sciences La Doua, 20 avenue Gaston Berger à Villeurbanne
Jury :
Rapporteurs :
M. Tedjani MESBAHI, Maître de Conférences HDR, INSA Strasbourg
M. Nassim RIZOUG, Maître de Conférences HDR, ESTACA
Examinateurs :
Mme. Ouafaé EL GANAOUI-MOURLAN, Maître de Conférences, IFP School,
Mme. Marie-Cécile PERA, Professeur des Universités, Université Bourgogne Franche-Comté
M. Ali SARI, Professeur des Universités, Université Claude Bernard Lyon 1
Encadrement :
M. Pascal VENET, Professeur des Universités, Université Claude Bernard Lyon 1, Directeur de thèse
M. Gabriel DUCRET, Ingénieur de recherche, IFP Energies Nouvelles
M. An LI, Responsable produit, Siemens Digital Industries Software
M. Rémy MINGANT, Ingénieur de recherche, IFP Energies Nouvelles
Abstract :
The lithium-ion battery market is experiencing significant growth thanks to its booming applications such as portable electronics, electric mobility, and stationary storage. However, the storable energy and available power decrease over time and with the number of charge and discharge cycles the battery undergoes. Parasitic reactions occur inside of each cell of the storage systems contributing to battery aging. This results in increased internal resistance and decreased available capacity. Foreseeing these effects requires to model the performance degradation of batteries. Developing robust battery aging models is essential to address this issue. In this context, machine learning techniques appear promising.
This research project focuses on the development of machine learning models for lithium-ion battery aging, specifically at the cell level. First, a state-of-the-art review of lithium-ion battery aging and the use of machine learning to model it is presented. As these models are data-driven, a study of the available databases is conducted. This study led to the selection of three databases that stand out due to their quality, and the range of use-cases they encompass: one focuses on calendar aging while the other two include profiles of electric vehicles and aircrafts. The database focused on calendar aging is used to compare different machine learning algorithms on the same data. Results indicate that XGBoost demonstrates the best performances. Subsequently, this work discusses several approaches employed to train and test models across various aging scenarios. These general approaches have been applied to the three selected databases. The performances obtained were compared to those of empirical models (a commercial one and another one, considered basic). The models developed in this study were either superior to or competitive with the commercial model, depending on the database studied. Overall, the proposed approach performed better, yielding a RMSE that is half that of the commercial model. Finally, complementary approaches were explored, such as using the results of empirical models as input for machine learning ones. This work paves the way for perspectives that could lead to better optimization of lithium-ion battery performances, with potential applications in fields such as energy’s predictive management, or the enhancement of energy storage systems’ dependability.
Keywords:
Batteries, Degradation, Machine Learning, Data, Ageing, lithium-ion