Abstract:

The capacity of lithium-ion batteries degrades after numerous charge-discharge cycles, posing a risk to energy storage systems. This study proposes a hybrid model for the early lifetime prediction of lithium-ion batteries considering their energy and temperature features. The proposed model addresses the insufficient analysis of temperature and energy features and the lack of research inon the significance of features extracted via deep learning. First, to fully mine the effective information from the temperature data, the voltage, current, and temperature data were employed to indirectly compute and extract the capacity, energy, and temperature signal energy curves of the battery. The first 100 cycles were selected to construct the corresponding two-dimensional features. Second, to address the inability of convolutional neural networks (CNNs) to filter extracted feature maps, a feature extraction architecture based on CNNs and a convolutional block attention mechanism was proposed, The attention mechanism identifies the importance of each feature map, facilitating mapping from features to early lifetime predictions. Experiments conducted on the MIT lithium-ion battery degradation dataset validated the effectiveness of the proposed features and methods. The results indicated that the proposed hybrid model outperformed the basic CNN, achieving superior prediction performance with an average root mean square error of 97.43. Furthermore, a series of experiments using different features as inputs revealed that the proposed temperature signal energy features provide superior prediction performance, whereas the multi-feature fusion technology can achieve better prediction performance. Finally, in scenarios with limited period data application, the model requires at least 70 cycles to maintain good prediction performance and high stability.

Key words: lithium-ion batteries, early lifetime prediction, deep learning, temperature feature, energy feature