Cycle life prediction of lithium-ion batteries based on data-driven methods

Abstract

Predicting the cycle life of lithium-ion batteries (LIBs) is crucial for their applications in electric vehicles. Traditional predicting methods are limited by the complex and nonlinear behavior of the LIBs, whose degradation mechanisms have not been fully understood. Recently, machine learning techniques attract increasing attention for empirically learning and predicting battery behavior. Herein, we first generate a comprehensive dataset with 104 commercial LiNi 0.8 Co 0.15 Al 0.05 O 2 /graphite 18650-series batteries under a wide range of demanding test conditions by systematically varying the ambient temperature, charge/discharge current, and cut-off voltages. Next, a linear regression model, a conventional neural network (NN) model, and a convolutional NN model are built to predict the cycle life. Models based on experts-extracted features achieve around 35% of the training error and the test error. In comparison, the model based on algorithm-extracted features can achieve 9.28% of the training error and 22.73% of the test error. These algorithm-extracted hidden features show higher covariance (0.88) with the cycle life of LIBs than experts-extracted features (0.71). This study demonstrates the capability of machine learning techniques to capture hidden features for complex, nonlinear systems, which can be used to accurately predict their cycle life. • An extensive cycle life dataset with 104 commercial 18650 lithium-ion batteries (LIBs) is generated. • Data-driven methods are applied to predict the cycle life of LIBs based on their initial information. • Machine learning algorithms can capture hidden features better than human experts. • A convolutional neural network shows the best prediction performance.