A fusion extended long short-term memory learning architecture for capacity trajectory prediction of Li-ion battery

Abstract

The capacity of lithium-ion batteries is a critical indicator for battery management systems(BMS), and accurate capacity prediction is essential for ensuring the reliable and safe operation of battery systems. Traditional deep learning methods often struggle to overcome the inherent limitations of time-series modeling, making it challenging to accurately identify phenomena such as capacity regeneration. To address this issue, this paper introduces a novel approach for predicting lithium battery capacity trajectories, leveraging data preprocessing and the fusion of extended Long Short-Term Memory(LSTM) networks. Initially, we apply the successive variational modal decomposition(SVMD) to preprocess the battery capacity aging curve, which is decomposed into intrinsic mode functions(IMFs) and residuals. A sliding window technique is then employed on these data. The varying IMF data is inputted into the extended LSTM network to facilitate accurate time-series modeling, while the smoothly decreasing residual data is fed into a faster Gated Recurrent Unit to enhance prediction efficiency. This dual-target learning approach effectively captures battery capacity trends, even in the presence of capacity regeneration phenomena and significant inflection points. Extensive experiments are conducted using the NASA and MIT datasets, and the results demonstrate that the proposed method successfully predicts battery capacity trajectories. Notably, compared to state-of-the-art methods, our approach achieves reductions in mean absolute errors(MAEs) and root mean square errors(RMSEs) by an average of 60.44% and 62.85%, respectively.