Development of a data‐driven ageing model for Lithium‐ion batteries: a nonparametric approach to learn from real operation data
Mattin Lucu Oyhagaray
- DIRECTORS: Haritza Camblong Ruiz y Egoitz Martinez Laserna
- UNIVERSITY: UPV / EHU
Lithium-ion (Li-ion) battery technology has gained a significant market share as the principal energy storage solution for many industrial applications, mainly due to its relatively high technological maturity, high energy efficiency and high specific energy and power. However, Li-ion batteries are still relatively expensive compared to other storage technologies, and their performance is known to decline over time and use, which threatens their competitiveness against more affordable solutions. In order to overcome such barriers and ensure the profitability of Li-ion based systems, the global research focusses on different paths. Some of them are the implementation of optimised sizing of the storage systems, second-life business strategies and the design of effective operation strategies, which allow the reduction of the total cost of ownership and make profitable the implementation of large-scale Li-ion energy storage systems. The latter points are strongly conditioned by the development of accurate Li-ion battery ageing models, able to relate the operating conditions of a battery system to their subsequent degradation.
A significant challenge for the development of conventional Li-ion battery ageing models is the amount of laboratory tests required to verify the accuracy of the model under realistic operating conditions. In order to reduce the number of laboratory tests and at the same time ensure the validity of the model under realistic operating profiles, the solution proposed in this thesis is the development of ageing models capable to continuously learn from streaming data. Following this approach, reduced laboratory tests could be used to develop a preliminary ageing model. Further, once the battery pack has been implemented and deployed, in-field data extracted by the data acquisition system could allow updating the preliminary ageing model. In this way, the ageing model would be continuously upgraded, improving prediction accuracy, extending the operating window of the model itself and providing useful information for predictive maintenance, adaptive energy management strategies or business case redefinition.
After an in-depth study of the state of the art, the Gaussian Process (GP) modelling framework was selected as the most suitable method to meet with the objective of the thesis. Compositional covariance functions were proposed in order to develop GP models tailored to the Li-ion battery ageing prediction application. A holistic ageing model was developed, composed of two separated models corresponding respectively to the cell degradation through calendar and cycle ageing. Both models were validated under a broad range of static, dynamic, and realistic operating conditions, using an extensive laboratory experimental dataset involving Li-ion cells tested during more than three years. A methodology was designed to validate the ability of the models to learn continuously from the data progressively observed. The research works carried out in this thesis bring the main following findings:
- Due to their nonparametric character, GP-based ageing models are capable to learn from progressively observed operating conditions: throughout the whole operating range of the Li-ion battery, the prediction accuracy of the model improves, and the confidence boundaries of the predictions are reduced, indicating an increased reliability of the models’ predictions.
- In this context, isotropic kernel components are suitable to host the features corresponding to the different stress-factors, in so far as the battery operates within the limited range of the recommended operating conditions.
- The sensitivity analysis based on the automatic relevance determination kernels allows to identify which stress-factors have highest influence on battery ageing, providing insightful inputs for the development of energy management strategies oriented to extend the lifetime of battery systems.
- There is a discrepancy between the capacity loss induced by static and dynamic profiles of the stress-factors, which could be related to the lower influence of the depth of discharge in dynamic operation. This highlights the increased interest of ageing models capable to evolve after the deployment and learn from the dynamic profiles observed in real applications.
The modelling approach proposed in this thesis comes aligned with the digitalisation trends observed in the recent years in different areas. In fact, there is significant interest from industry in the introduction of new data collection telemetry technology. This implies the forthcoming availability of a significant amount of real-world battery operation data. In this context, the development of ageing models able to learn from in-field battery operation data is an interesting solution to mitigate the need for exhaustive laboratory testing and reduce the development cost of ageing models. The findings presented in this work are therefore not only of technological but also of economic interest, and the proposed solution is particularly adapted to the industry trends upcoming.