A temporal convolutional network model with attention mechanisms and quantile regression for state of health estimation of lithium batteries

Accurate and reliable estimation of the state of health (SOH) is essential for ensuring the safe and stable operation of lithium-ion batteries. However, due to the complexity of battery capacity degradation mechanisms and incompleteness of charge-discharge curves, achieving precise SOH estimation remains a significant challenge. To address this, this paper proposes an SOH estimation method based on a Time Convolutional Network with Attention (TCN-Attention) model. The approach first leverages TCN's strengths to extract spatiotemporal features and identify aging characteristics from constant-current charge curves. Next, an attention mechanism is employed to adaptively adjust feature weights, capturing subtle yet significant changes in aging features. Finally, the model is trained with a quantile regression loss function, enabling it to provide confidence intervals for its predictions. The experimental results show that the proposed method achieves accurate point estimates of SOH and reliable confidence intervals. By incorporating a piecewise estimation approach, the technique also ensures high estimation accuracy when only partial constant-current charging data are available. Notably, even when applied to multiple constant-current charging segments with voltage interval widths of 0.1 V or 0.2 V, the root-mean-square error and maximum absolute error of SOH estimation remain below 1.30% and 4.94%, respectively, demonstrating its robustness and practicality. • A TCN-Attention model is proposed to extract local and global aging features, improving SOH estimation accuracy. • Quantile regression is adopted to overcome deep learning point-estimate limits, providing 95% confidence intervals for SOH. • A piecewise estimation method is proposed to mitigate significant errors in specific charging segments.