Simultaneous extraction of intra- and inter-cycle features for predicting lithium-ion battery's knees using convolutional and recurrent neural networks

The timely and accurate prediction of degradation patterns of lithium-ion batteries (LIBs) is crucial for investigating their lifespan. Although charging-discharging cycling data are common for this purpose, there has been limited exploration into simultaneously extracting features related to both intra-and inter-cycle behaviors, despite their significant relevance. In this study, machine learning methods are proposed to enable the simultaneous extraction of intra-and inter-cycle features from voltage-current-temperature (VIT) cycling datasets for predicting the knees of LIBs. To standardize the length of each cycling data, the time-series data are realigned based on voltage or employing zero-padding, depending on the dataset type. The cycling dataset is then organized as an array along the time and cycle axes before applying convolutional neural networks (CNNs) and/or recurrent neural networks (RNNs) for feature extraction. Three approaches to utilizing these nonlinear regression tools are explored: 2-dimensional CNN (2D CNN), RNN + 1-dimensional CNN (1D CNN), and RNN + 2D CNN. The resulting knees prediction models are evaluated using a benchmark dataset. Our results underscore the advantage of explicitly considering cycle-to-cycle behavior in conjunction with intracycle temporal behavior when constructing a data-driven prediction model. We also investigate reducing the input data requirement to facilitate early knee prediction. The findings demonstrate that the input size can be effectively reduced from 100 to the first 60 cycles without compromising prediction performance.

Automated summary

This study proposes a machine learning approach to predict the knees of lithium-ion batteries (LIBs) by extracting features from voltage-current-temperature (VIT) cycling datasets. The findings highlight the advantage of considering cycle-to-cycle behavior in conjunction with intracycle temporal behavior when constructing a data-driven prediction model. Additionally, the study demonstrates that the input size can be reduced to facilitate early knee prediction.