Predicting the operational remaining life of lithium-ion battery with dynamic attention transformer

The objective of this study is to introduce and examine an approach to improve and forecast the operational remaining life of batteries, specifically lithium-ion, with the aid of a dynamic attention transformer. This transformer will be aided by Bayesian change point detection. With cleaner energy sources such as lithium-ion batteries making inroads into various sectors ranging from transportation to storage, determining the useful life that is left after these batteries have been in operation for some time is of utmost importance: it increases efficiency, reduces downtime, and improves the cost of maintaining energy systems. The proposed method is based on the use of singular spectrum analysis, Bayesian change point detection, and a dynamic attention transformer. This method aims at capturing the drastic degradation pattern of batteries and learning the correlation of health indicators at those points. This dynamic attention transformer approach serves as an improvement over neural network models and vanilla transformer approaches. The results achieved by monitoring battery health indicators at drastic degradation points help in both the prediction of operational remaining life and its extension by changing the charging pattern. The performance of the algorithm was analyzed with the aid of error metrics, e.g., the Mean Average Error (MAE) is 0.0189, translating into an accuracy improvement of 74.9% over neural network methods and 7.47% over other vanilla transformer-based methods. In addition, the operational remaining life improved by 15%. These outcomes offer valuable insights into the learning capabilities of DAT and present an efficient, cost-effective method for accurately estimating battery lifespan, outperforming traditional learning approaches.