CNN framework enhanced with attention weighed and optimized features for state-of-health estimation of lithium-ion batteries

Accurate estimation of Lithium-ion battery State of Health (SoH) is crucial for battery energy storage system safety and reliability and remaining service life. Nonlinear degradation characteristics, capacity fading, and operating condition variations, however, hinder the accuracy of the conventional model based and deep learning methods. The major drawback of the existing hybrid methods is that they lack the ability to prioritize the degradation features and do not optimize the parameters optimally, which leads to poor generalization when the batteries are operated under different conditions. To solve these problems, a hybrid framework for SoH estimation is proposed, which is based on CNN-AM-APO. To overcome these issues, a hybrid framework, namely CNN-AM-APO is proposed for SoH estimation. To extract hierarchical degradation features from the battery discharge signals, a proposed framework uses one-dimensional CNN for extracting degradation features, and an adaptive channel attention mechanism is used to emphasize degradation-sensitive information and suppress all channels irrelevant to degradation. Artificial Protozoa Optimization (APO) is added to perform adaptive exploration/exploitation for optimizing the network parameters and hence stabilizing the convergence and minimizing the prediction error. The proposed framework is tested with the publicly available lithium-ion battery degradation dataset under the similar preprocessing, training and testing conditions. The proposed model outperforms CNN, CNN-LSTM, CNN-AM, and CNN-LSTM-APO models based on experimental results with an MSE value of 0.000426, RMSE value of 0.020639, MAE value of 0.017381, MAPE value of 0.023530, and R² value of 95.60%. The results show that the proposed framework can be used for intelligent battery management systems and real-time lithium-ion battery health monitoring, and that the integration of attention-guided feature refinement and the optimization of parameters through the APO can enhance prediction accuracy, convergence stability, and generalization capability.