Evolutionary hybrid deep learning based on feature engineering and deep projection encoded echo-state network for lithium batteries state of health estimation

The state of health estimation for lithium batteries is crucial for optimizing their performance, extending their lifetime, ensuring safety, and reducing their maintenance cost. This study proposes a hybrid deep learning model for SOH estimation in lithium batteries. The model utilizes Random Forest (RF) and Variational Mode Decomposition (VMD) for feature processing, and then utilizes a deep projection-encoded echo-state network (DeePESN) for health state estimation. To improve the estimation accuracy, a logistic initialization method was used to optimize the Newton-Raphson-based optimizer (NRBO) algorithm. In this study, the LoNRBO algorithm is used to determine the hyperparameters of the model. To improve the generalization ability of the model, a Kernel Extreme Learning Machine Autoencoder (KELMAE) was used to reduce the dimensionality of the input data within the DeePESN model. After experimental verification, the hybrid deep learning model proposed in this paper exhibits excellent estimation performance under different working conditions: at a temperature of 24°C and a discharge current of 2A, at a temperature of 24°C and a discharge current of 4A, and at a temperature of 4 °C and a discharge current of 1A. The root-mean-square errors (RMSE) of SOH were 0.0768, 0.0998, 0.0734and 0.0979, respectively. In summary, the proposed hybrid deep learning method for the SOH estimation of lithium-ion batteries is feasible and can better fit lithium-ion battery data under different working conditions.