A novel spatio-temporal feature fusion attention model for lithium-ion battery capacity estimation using graph convolutional network

Precise estimation of battery capacity is crucial for ensuring the sustainability, safety and reliability of lithium-ion batteries. This study proposes a novel framework to estimate battery capacity, which integrates spatio-temporal graph convolutional networks (GCNs) and a bidirectional long short-term memory network (BiLSTM), both enhanced with attention mechanisms (AMs). Firstly, to evaluate the effectiveness of the proposed method, three diverse battery datasets are utilized: a custom-designed experimental dataset, the NASA dataset, and the CALCE dataset, covering battery pack data, small single-cell data, and large single-cell data. Subsequently, the battery health indicators (HIs) are extracted from battery operational data, and to avoid redundancy, temporal and spatial graph structures are created by analyzing the correlations between HIs and battery capacity using Pearson correlation coefficients to select high-quality HIs. Additionally, temporal and spatial features of the selected HIs are captured by GCNs respectively, then these features are dynamically fused using sigmoid AMs, and the output is fed into a BiLSTM for further learning of the temporal correlation. Finally, a softmax AM is applied to the output of the BiLSTM, assigning varying weights to each HI to enhance the precision of the capacity estimation. Comparative experiments demonstrate that the proposed model achieves superior accuracy and exhibits robust adaptability across diverse datasets.