Data-Driven ICA-Bi-LSTM-Combined Lithium Battery SOH Estimation

• Published in: Mathematical problems in engineering Vol. 2022; pp. 1 - 8
• Main Authors: Sun, Hanlei, Sun, Jianrui, Zhao, Kun, Wang, Licheng, Wang, Kai
• Format: Journal Article
• Language: English
• Published: New York Hindawi 28.03.2022; John Wiley & Sons, Inc
• Subjects: Accuracy; Aging; Artificial neural networks; Back propagation; Back propagation networks; Battery chargers; Charging; Correlation coefficients; Deep learning; Electric potential; Energy; Engineering; Lithium; Lithium batteries; Lithium-ion batteries; Machine learning; Mathematical models; Methods; Neural networks; Parameters; Power; Random variables; Rechargeable batteries; Recurrent neural networks; Standard deviation; Support vector machines; Voltage
• ISSN: 1024-123X; 1563-5147
• DOI: 10.1155/2022/9645892

Lithium battery state of health (SOH) is a key parameter to characterize the actual battery life. SOH cannot be directly measured. In order to further improve the accuracy of SOH estimation of lithium batteries, a model combining incremental capacity analysis (ICA) and bidirectional long- and short-term memory (Bi-LSTM) neural networks based on health characteristic parameters is proposed to predict the SOH of lithium-ion batteries. First, the health characteristic parameters are initially selected from the lithium battery charging curve, and the health characteristics are extracted by the Pearson correlation coefficient, including the charging time of constant current, charging time of constant voltage, voltage change rate from 300 s to 1000 s, 200s of voltage per cycle at a time. Second, ICA was used to deeply mine the deep associations related to SOH and the peaks of IC curves and their corresponding voltages were extracted as additional inputs to the model. Then, Bi-LSTM is used to form a combined SOH estimation model through adaptive weighting factors. Finally, the verification is based on the 5th battery parameters of the NASA lithium battery data set. The experimental results show that the proposed combined model reduces the mean square error by 55.17%, 49.28%, and 41.47%, respectively, compared with single models such as BP neural network (BPNN), LSTM, and gated recurrent neural network (GRU).