Data-driven Battery SOC Estimation: A CNN-ResNet-BiLSTM Framework

With the advancement in artificial intelligence technology, a series of hybrid algorithms have been applied to battery state-of-charge (SOC) estimation. However, existing approaches face challenges in extracting deep features and suffer from gradient vanishing, which prevents the construction of deeper networks. A novel data-driven fusion algorithm, the CNN-ResNet-BiLSTM framework, is proposed to further improve the SOC estimation accuracy. The proposed method incorporates a residual network (ResNet) to resolve gradient vanishing in deep convolutional neural networks (CNNs), thereby enabling deeper integration between the CNN and bidirectional long short-term memory (BiLSTM). Furthermore, this study employs grey relation analysis for feature construction to enhance the quality of the derivative and historical category data. Experimental results show that the proposed framework achieves 0.348% root mean square error and 0.255% mean absolute error during a driving cycle, indicating exceptional performance in SOC estimation. This demonstrates the outstanding generalization capability of the proposed framework through cross-condition validation, which provides a theoretical foundation for practical SOC estimation applications.