Uncertainty of ageing and sudden death behaviour in Lithium-ion cells: Can Deep Learning models help?

Lithium-ion batteries (LIBs) have become inevitable and growing as energy storage devices in today’s automotive world; accurate predictions of the states, i.e., State-of- Health State-of-Charge and state-of-function, are vital for its operation and system management. State-of-Health forecasting detects early anomalies in performance, facilitating preventive maintenance and supporting warranty management by predicting the Remaining Useful Life (RUL). Despite two decades of research, lithium-ion cell ageing largely remains hypothetical due to complex, poorly characterized side reactions, complicating the understanding of degradation mechanisms. Traditional electrochemical models like pseudo-two-dimensional (P2D) and single-particle models (SPM) lack accuracy due to a lack of mathematical description of ageing mechanisms interestingly Deep Learning (DL) models can take advantage of the experimental data to capture the nonlinearity and uncertainty of ageing and sudden death behaviour. In this work, authors have utilized the DL models like Long short-term memory (LSTM), Gated recurrent units (GRU), Temporal Convolutional Networks (TCN), and Transformer models to evaluate the fitness for ageing prediction.

The forecasting power of the DL models not only captures the capacity degradation trend but also the highly unpredictable sudden death behaviour. The authors used a publicly available dataset of 124 commercial lithium iron phosphate/graphite cells with widely different cycle lifetimes (ranging from 150 to 2,300 cycles) that were cycled under fast-charging conditions. Best practices for sequential data, including filtering, sampling, augmentation, and feature engineering, were carried out to enhance model performance. Given initial ageing data for 250 cycles., the developed models can forecast the entire trajectory till the SOH level of 0.8. Root Means Squared Error (RMSE), error percentage (%) and R2 are the metrics used to evaluate and compare models, cycle error is used to evaluate the sudden death behaviour. Through this work, it is evident that domain know-how, along with machine learning, can benefit by creating immense value in the automotive industry across systems.