Machine learning‐based approach for useful capacity prediction of second‐life batteries employing appropriate input selection

• Published in: International journal of energy research Vol. 45; no. 15; pp. 21023 - 21049
• Main Authors: Bhatt, Ankit, Ongsakul, Weerakorn, Madhu, Nimal, Singh, Jai Govind
• Format: Journal Article
• Language: English
• Published: Chichester, UK John Wiley & Sons, Inc 01.12.2021; Hindawi Limited
• Subjects: Ageing; Aging; aging and regeneration; Batteries; charging and discharging profile; Comparative analysis; Cost analysis; Discharge; discharging capacity prediction; Distributed generation; Electric vehicles; Energy management; Energy storage; Errors; Forecasting; Learning algorithms; Long short-term memory; Machine learning; machine learning model; Mathematical models; Regeneration; second life battery; Storage capacity; Storage conditions; Training
• ISSN: 0363-907X; 1099-114X
• DOI: 10.1002/er.7160

Summary Electric vehicle‐discarded second‐life batteries still contain 80% of usable capacity and can serve as a low‐cost alternative for microgrid storage applications where the battery storage capacity is flexible against transport applications. By accurately predicting the remaining useful capacity or state of health of these batteries, using the data from their first life operation, their cost‐effectiveness for microgrid energy management can be analyzed. For this purpose, three machine learning models are proposed here. The input parameters for the models are selected from the charging and discharging profiles of batteries, considering both the aging and regeneration phenomenon. Eight different input cases with and without K‐fold (K = 10) cross‐validation are used for training the proposed models. Based on the comparative analysis it is found that all the models trained with K‐fold cross‐validation, show minimum error as compared to without K‐fold. The forecasting results from multiple approaches showed that the long short term memory model trained with battery discharging profile outperformed other models, quantified by multiple error indices, including root mean square error (0.009147), mean absolute error (0.005841), and R2 (0.9713). The robustness of the model is validated with multiple battery datasets. Further, the study illustrates the importance of activation functions, in machine learning models used for forecasting.