Hybridization of semi-empirical models and machine learning models to assess the State-of-Healt[...]

PowerUp; Les Ulis; Sur place; EUR 30 000 - 50 000; Plein temps

Résumé du poste

Une entreprise innovante, spécialisée en énergie, recherche un expert en machine learning pour développer un cadre d'estimation de l'état de santé et de la durée de vie des batteries lithium-ion. Ce poste, essentiel pour le soutien à la transition énergétique, impliquera la comparaison entre approches ML et semi-empiriques, tout en optimisant la performance des modèles tests en milieu réel.

Qualifications

• Compétences avancées en machine learning.
• Expérience en gestion de données de batteries.
• Connaissance en modélisation de systèmes.

Responsabilités

• Développer un cadre ML pour l'estimation de l'état de santé des batteries.
• Comparer les résultats avec des méthodes semi-empiriques.
• Tester les algorithmes sur une plateforme cloud.

Connaissances

Machine Learning

Analyse de données

Modélisation

Formation

Master ou Doctorat en informatique ou ingénierie

Description du poste

Social network you want to login/join with:

Hybridization of semi-empirical models and machine learning models to assess the State-of-Health and predict the Remaining Useful Lifetime of lithium-ion batteries, , ile-de-france, fr

col-narrow-left

Client:

PowerUp

Location:

Job Category:

Other

-

EU work permit required:

Yes

col-narrow-right

Job Reference:

e8f53a80acfb

Job Views:

4

Posted:

21.06.2025

Expiry Date:

05.08.2025

col-wide

Job Description:

Lithium-ion batteries play a crucial role in the current energy transition, whether to support the massive introduction of renewable and intermittent energies, such as solar or wind, or to support the electrification of mobility. However, to sustain this growing demand, the complex behavior of Li-ion batteries need to be more predictable in terms of aging, since their lifetime performance have a direct impact on the return on investment (ROI) of all projects involving batteries, especially in the case of large or heavy-duty batteries (e.g. of hundreds of kWh needed for e-buses, and tens to hundreds of MWh of utility-scale energy storage systems).

To solve this issue, relevant methods have already been developed to predict lifetime of batteries according to dynamic usage profiles, based on the characterization of cells under various aging conditions, leading to so-called semi-empirical models. However, this approach takes long time periods and requires huge resources, usually based on an experimental plan carried out on 40 to 50 cells over 18 months up to 2 years. Furthermore, this approach also suffers from the fact that it is built on the behavior of individual cells, whereas in real usage the cells are then used in modules and packs (corresponding to the association of individual cells in series and parallel, to obtain the expected voltage and capacity, respectively). This can alter their aging, mainly due to the difference in thermal behavior, but also to the possible heterogeneity of the conditions of use of the cells in the modules and the packs.

On another hand, ML-based methods are nowadays used to predict the SOH and RUL of Li-ion batteries. These methods are receiving huge popularity and a great deal of interest since they can exploit the already available large amounts of battery data, extracting relevant health indicators without the need of comprehending the battery physics. However, most of these ML-based methods share serious limitations, since they are developed using laboratory data generated on separated cells of slight capacity (e.g., 1 .0 Ah). Additionally, these data are generated assuming standard and idealistic loading conditions that would never be encountered during the real usage of the battery. As a result, the performance of the existing ML-based methods could be inefficient when applied in real operation of packs.

The objective of this thesis is to develop a novel ML-based framework for estimating the SOH and predicting the RUL of Li-ion batteries using their monitoring data acquired during in-field operation. The results obtained will then be compared with the classical semi-empirical methods. Finally, the hybridization of the two types of approaches will be studied, to take advantage of the best prediction of each one. Furthermore, all the algorithms developed will be tested on a data processing platform hosted on the Cloud to qualify/quantify their performance and cost from an industrial point of view.

Topic description

Lithium-ion batteries play a crucial role in the current energy transition, whether to support the massive introduction of renewable and intermittent energies, such as solar or wind, or to support the electrification of mobility. However, to sustain this growing demand, the complex behavior of Li-ion batteries need to be more predictable in terms of aging, since their lifetime performance have a direct impact on the return on investment (ROI) of all projects involving batteries, especially in the case of large or heavy-duty batteries (e.g. of hundreds of kWh needed for e-buses, and tens to hundreds of MWh of utility-scale energy storage systems).

To solve this issue, relevant methods have already been developed to predict lifetime of batteries according to dynamic usage profiles, based on the characterization of cells under various aging conditions, leading to so-called semi-empirical models. However, this approach takes long time periods and requires huge resources, usually based on an experimental plan carried out on 40 to 50 cells over 18 months up to 2 years. Furthermore, this approach also suffers from the fact that it is built on the behavior of individual cells, whereas in real usage the cells are then used in modules and packs (corresponding to the association of individual cells in series and parallel, to obtain the expected voltage and capacity, respectively). This can alter their aging, mainly due to the difference in thermal behavior, but also to the possible heterogeneity of the conditions of use of the cells in the modules and the packs.

On another hand, ML-based methods are nowadays used to predict the SOH and RUL of Li-ion batteries. These methods are receiving huge popularity and a great deal of interest since they can exploit the already available large amounts of battery data, extracting relevant health indicators without the need of comprehending the battery physics. However, most of these ML-based methods share serious limitations, since they are developed using laboratory data generated on separated cells of slight capacity (e.g., 1 .0 Ah). Additionally, these data are generated assuming standard and idealistic loading conditions that would never be encountered during the real usage of the battery. As a result, the performance of the existing ML-based methods could be inefficient when applied in real operation of packs.

The objective of this thesis is to develop a novel ML-based framework for estimating the SOH and predicting the RUL of Li-ion batteries using their monitoring data acquired during in-field operation. The results obtained will then be compared with the classical semi-empirical methods. Finally, the hybridization of the two types of approaches will be studied, to take advantage of the best prediction of each one. Furthermore, all the algorithms developed will be tested on a data processing platform hosted on the Cloud to qualify/quantify their performance and cost from an industrial point of view.

Starting date

-10-02

Funding category

Partial or full private funding (CIFRE agreement, foundation, association)

Funding further details