A Deep Learning Multi-feature Based Fusion Model for Predicting The State of Health of Lithium-ion Batteries

Abstract

Lithium-ion batteries have become the preferred energy storage method with applications ranging from consumer electronics to electric vehicles. Utilization of the battery will eventually lead to degradation and capacity fade. Accurately predicting the state of health (SOH) of the cells holds significant importance in terms of reliability and safety of the cell during its operation. The battery degradation mechanism is strongly non-linear and the physics-based model have their inherent disadvantages. The machine learning method has become popular for estimating SOH due to its superior non-linear mapping, adaptive, and self-learning capabilities, made possible by advances in deep learning technologies. In this study parallel hybrid neural network is formulated for predicting the state of health of lithium-ion cell. Firstly, the factors that have an effect on the cell state were analysed. These factors are cell voltage, charging & discharging time and incremental capacity curve. The features were then processed for use as input to the model. Spearman correlation coefficient analysis shows that all the factors had a positive correlation with SOH. While charging time has a negative correlation with the other features. Next the deep learning models namely convolution neural network (CNN), temporal convolution network (TCN), long-short-term memory (LSTM) and bi-directional LSTM were used to make fusion models. The number of layers in CNN and TCN were also varied. The hyperparameters used in the models were optimized using Bayesian optimization algorithm. The models were validated through comparative experiments on the University of Maryland battery degradation dataset. The prediction accuracy with CNN 3-layer LSTM was found to be the best for the training and the test dataset. The overall R2 value, root mean squared error (RMSE) and mean absolute percentage error (MAPE) with the model was found to be 0.999646, 0.003807 and 0.3, respectively. The impact of the features on the model was also analysed by removing one feature and retraining the model with the other features. The effect of discharging time and the peak of the discharge incremental capacity curve was maximum. The analysis also reveals that either charging voltage or discharging voltage can be used. Further, the proposed model was also compared with the other studies. The comparison shows that the R2, RMSE and MAPE values of the proposed model was better.