Hybrid Attention Enhanced Deep Temporal Convolutional Network for Capacity Estimation of Lithium-Ion Battery Under Small Sample Conditions

Abstract

Deep learning has emerged as the mainstream method of lithium-ion battery (LIB) capacity forecasting owing to its expressive feature learning and nonlinear modelling ability. In high reliability application scenarios, point predictions of LIB capacity cannot support risk-based maintenance decisions. On the other hand, under insufficient sample sizes, prediction results also exhibit uncertainty. Therefore, uncertainty quantification of prediction outcomes is essential. This article proposes a statistical feature enhanced LIB capacity prediction framework for small sample conditions. Specifically, constructing a multidimensional high quality representation space through systematic statistical feature enhancement, coupled with an importance driven feature refinement mechanism. Then a novel Hybrid Attention Enhanced Deep Temporal Convolutional Network (HAE-DTCN) is proposed to reconcile convolutional locality with attention-based global dependencies, capturing multi-scale degradation patterns, in addition, a composite loss function is introduced to optimize point prediction and confidence intervals simultaneously. Finally, experiments on NASA, CALCE and XJTU public datasets demonstrate that the proposed framework achieves significant advantages in both accuracy and robustness, providing a feasible solution for LIB capacity health assessment under small sample conditions.

Keywords: Lithium-ion battery, Capacity prediction, Hybrid attention enhanced deep temporal convolutional network, Small sample conditions, Uncertainty quantification