State-of-health estimation and classification of series-connected batteries by using deep learning based hybrid decision approach.

In rechargeable battery control and operation, one of the primary obstacles is safety concerns where the battery degradation poses a significant factor. Therefore, in recent years, state-of-health assessment of lithium-ion batteries has become a noteworthy issue. On the other hand, it is challenging to ensure robustness and generalization because most state-of-health assessment techniques are implemented for a specific characteristic, operating situation, and battery material system. In most studies, health status of single cell batteries is assessed by using analytical or computer-aided deep learning methods. But, the state-of-health characteristics of series-connected battery systems should be also focused with advances of technology and usage, especially electric vehicles. This study presents a data-driven, deep learning-based hybrid decision approach for predicting the state-of-health of series-connected lithium-ion batteries with different characteristics. The paper consists of generating series-connected battery degradation dataset by using of some mostly used datasets. Also, by employing deep learning-based networks along with hybrid-classification aided by performance metrics, it is shown that estimating and predicting the state-of-health can be achieved not only by using sole deep-learning algorithms but also hybrid-classification techniques. The results demonstrate the high accuracy and simplicity of the proposed novel approach on datasets from Oxford University and Calce battery group. The best estimated mean squared error, root mean square error and mean-absolute percentage error values are not more than 0.0500, 0.2236 and 0.7065, respectively which shows the efficiency not only by accuracy but also error indicators. The results show that the proposed approach can be implemented in offline or online systems with best average accuracy of 98.33 % and classification time of 58 ms per sample.