Ageing prediction models that generalise well across different types of second-life batteries
Batteries are referred to as “second-life” after they have served their initial purpose, in for example an electric vehicle, and then are reused for different applications. Even though these batteries have reached the manufacturer’s rated end of life, which usually means 70 to 80 per cent of the initial capacity, they can still store a lot of energy. However, if not handled properly the usage will not be optimised, and the deterioration process will be unnecessarily rapid.
Benedick Allan Strugnell-Lees, PhD student at Chalmers University of Technology, is developing battery ageing prediction models optimised for second-life usage within the SESBC project “Optimal usage and properties of battery storage units using 2nd life batteries”.
Benedick, what can second-life batteries be used for?
| “Typically, they are used in stationary charging applications for frequency modulation, peak-shaving, back-ups, and things of that sort of nature. However, one must consider that second-life batteries have a lower nominal capacity, but they still work quite well in applications with lower immediate loads and yields. Extending the lifetime in this way is a good option since recycling lithium-ion batteries can be costly and technically difficult, and outright discarding them is obviously a waste of resources.” |  |
What is your research about?
“Currently, my research focuses on modelling, predicting, and controlling second-life behaviour. For context, most battery ageing prediction models are developed for specific batteries, or chemistries, for a specific model from a specific manufacturer, and they are mainly focussed on capacities between 100 and 70 percent. My approach is to build models that can generalise across different manufacturers, cells, and chemistry types, without requirement for extensive datasets, as well as model capacities well beneath 70 percent. For this, I use physics-based machine learning, with the major contribution being the physics-based feature calculations and applications, used with very light machine learning models.”
You have recently published an article showing good results.
“Yes, in the article ‘An entropy-based, self-adaptive predictive algorithm for battery degradation’ we show that, by using the right physics based features, simple machine learning models can make a very accurate series of predictions within the acceptable range of error consistent with the current state of the art prediction models. What I tried to address was that by using physics-based features that are present in all battery energy storage systems you can use rather simple machine learning models to get accurate results across the board.”
Link to the article: https://www.sciencedirect.com/science/article/abs/pii/S0378775325017562
Why is this important?
“Most second-life battery farms or plants will probably be populated with batteries from many manufacturers and, as I have mentioned, the majority of today’s models are mainly developed for specific use/prediction. The two problems that this presents are that it is impractical to attempt to tie dozens of different predictive models together, and hope that they will work well for data that they have not been trained for, and that many manufactures will not store or share data from a batteries first life, meaning that one will not know anything about a given battery entering into second-life service, other than its nominal capacity. That’s why we need predictive models that can generalise well over many different types of battery, without extensive data on their performance, while maintaining low error margins across the board. If we can stay below one percent, the current industry “gold standard”, I think it’s worth the effort. Without accurate predictive ageing models batteries will not be used optimally in second-life and will likely degrade rapidly.”
What is next?
“Now I’m working with larger data-set of batteries that have more severely degraded, it is my hope that by harnessing the features that I have identified, self-adaptive feedback control, and transfer learning, it will be possible to generate a ready-made model that will rapidly adapt to accurately predict the behaviour of different types of battery with little to no retraining necessary. The aim is to have a sort of “plug and play” model, if you could call it this, that will accurately predict battery behaviour throughout second-life when very little is known about the battery when it is received.”
Can you tell us a little bit about yourself, Benedick
“I’m originally from the United States, but I did my undergraduate and master’s within a program called integrated mechanical and electrical engineering at the University of Bath in Great Britain. At first my plan was to get involved in the medical field, but I discovered that the time-lag from development to implementation was far too long for me. Since I wanted to see faster adaptation in the real world, and at the same time do something beneficial for people, following this logic, and my own interests, the next pertinent issue would be in environmental and sustainability based engineering. So, I started to apply for PhDs, and Sweden had some good and sensible programs in feilds that I was interested in. Having been in Sweden for just over two years now, I can definitely say that I like living here, in Sweden and Göteborg; I enjoy the nature, the culture, the mentality, and I find it very peaceful. I also like the structure of the academic system and the work life balance is great. The only real downside I would say is that I, like many Swedish people, am not a big fan of the long, cold, dark winter here though, especially as I’m coming from California originally.”
Contact:
Benedick Allan Strugnell-Lees
benedick@chalmers.se