Answers are grounded in this paper and cite their sources.
Accurate remaining useful life prediction of lithium-ion batteries plays a critical role in ensuring the safety of energy storage. Data-driven time-series forecasting methods have been widely adopted due to their powerful ability to extract temporal degradation features. However, most methods are limited in the extraction of multi-timescale degradation features. Moreover, the process by which these methods learn battery degradation features often suffers from poor intuitiveness. To address these issues, an exponential smoothing gate (ESG) network based on multi-timescale feature modulation (MTM) is proposed. First, a lightweight spatiotemporal fusion attention layer is designed to enhance the feature representation of battery capacity degradation through a non-parametric computational structure. Then, the MTM layer is developed. It combines varying sizes of convolutional kernel and dilation factor to extract feature representations along the temporal dimension. Subsequently, an interpretable ESG layer is constructed, where smoothing parameters are utilized to adaptively control the information flow between gating units. These smoothing parameters explicitly reveal the process of information transfer, thus improving the intuitiveness of internal information propagation within the network. Finally, a regression layer is introduced to establish the nonlinear mapping relationship between battery capacity degradation features and remaining useful life. Extensive prediction experiments are conducted on the CALCE and NASA battery degradation datasets. Experimental results demonstrate both the prediction accuracy and intuitiveness of the proposed model, confirming its effectiveness and strong generalization capability across different battery types and operating conditions.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: