Intelligent Online Health Estimation for Lithium-Ion Batteries Based on a Parallel Attention Network Combining Multivariate Time Series

With the development of cloud and edge computing, data-driven methods for estimating a Li-ion battery's state of health are becoming increasingly attractive. However, existing data-driven estimation methods have problems of low accuracy and weak robustness that need to be solved. Focusing on these points, this paper proposes a parallel attention network combining multivariate time series to extract the mapping relationship between the selected health features and the state of health. First, multivariate time series are extracted, which can describe battery aging characteristics at different scales. Then, a novel parallel learning framework is designed by integrating long short-term memory cells and an attention mechanism, which can make full use of the health features and help to solve the challenging issues of estimation accuracy and robustness. Experimental results show that the proposed model is able to obtain estimation results for different batteries with a mean absolute percentage error of less than 1%. Compared with existing methods, the maximum error of the proposed model is 38% lower on average. Furthermore, even under measurement noise injections of 50 dB, a preferable estimation result (maximum error of 3.36%) can still be achieved.