Multi-Model Fusion Methods with Strong Generalization Capability for Online SOH Estimation of Lithium-Ion Batteries

Abstract

Lithium-ion cells are essential in daily life, but accurately assessing their state of health (SOH) is challenging because deep learning models need large, consistent cycling datasets that are often impractical to obtain. Moreover, the real-world data can be incomplete or inconsistent. In this study, the SOH was predicted online utilizing fusion models of convolutional neural network (CNN), long short-term memory (LSTM) convolution block attention mechanism (CBAM), and multi-head attention mechanism (MHA) based architecture. Datasets prepared by the reaserchers at Massachussets Institute of Technology (MIT), and Sandia National Laboratory (SNL) were combined for training various models. Bayesian optimization algorithm was used to optimize the hyperparameters. The results reveal that the CNN-LSTM fusion model enhances the accuracy of capacity estimation. Oxford University dataset was used to test the efficacy of the models. Highest accuracy in prediction was found with the CNN-LSTM model having lowest root mean squared error (RMSE) of 0.0136. Ablation experiments were carried out and the performance of the fusion models was found better than the base models. The trained models were also retrained for another voltage range. CALCE dataset, provided by the University of Maryland, was utilized for the experiments. 2 and 7 cells dataset were separately used for the training. The CNNCBAM framework with 5 CNN layers was found to be the best model. The models generated in this study can be used for predicting the SOH online for cells having different kind of form, chemistry, charged/discharged at different rates and varying temperatures demonstrating its practicality and generalization ability.