均方误差
平均绝对百分比误差
学习迁移
摄氏度
人工神经网络
稳健性(进化)
人工智能
计算机科学
健康状况
机器学习
电池(电)
统计
数学
化学
功率(物理)
生物化学
物理
量子力学
基因
热力学
作者
Yichun Li,Mina Maleki,Shadi Banitaan
标识
DOI:10.1016/j.est.2023.109185
摘要
Accurately estimating the state of health (SOH) of lithium-ion batteries in real-world scenarios, especially for electric vehicles (EVs) is challenging due to dynamic operating conditions and limited battery usage data. To address these challenges, transfer learning, as an effective approach that uses the knowledge gained from a source task to improve efficiency and accuracy on a similar target task, has been combined with deep neural networks (DNN) to determine the SOH of Li-ion batteries at various temperatures with a limited amount of target data. In this study, the SOH of Li-ion batteries has been estimated using one of the largest Electrochemical Impedance Spectroscopy (EIS) datasets collected at three different temperatures, 25, 35, and 45 degrees Celsius. The source task is to estimate the SOH using DNN models at one or two operating temperatures (25, 35, and 45 degrees Celsius), while the target task is to estimate the SOH at the other temperature. The results demonstrate that the average mean squared error (MSE) and mean absolute percentage error (MAPE) for estimating the SOH using transfer learning-based DNN (DNN-TL) at 35 degrees Celsius have been reduced up to 72.63% and 48.43%, respectively, compared with the DNN model without transfer learning. Similarly, the average MSE and MAPE for estimating the SOH at 45 degrees Celsius using DNN-TL have been reduced up to 79.51% and 55.11%, respectively. Also, our findings demonstrate the robustness of DNN-TL models, requiring target data with only 6% of the size of source data for retraining, leveraging previously learned knowledge from the source data. Furthermore, the exclusion of fixed layers in the retraining process enhances the performance of the DNN-TL model.
科研通智能强力驱动
Strongly Powered by AbleSci AI