A Parallel Self-Attention Transformer for Predicting the Remaining Useful Life of Lithium-Ion Batteries

• Published in: IEEE journal of emerging and selected topics in industrial electronics (Print) pp. 1 - 11
• Main Authors: Zhang, Shurong, Cui, Zhongrui, Zhang, Qin, Li, Yueyang
• Format: Journal Article
• Language: English
• Published: IEEE 08.07.2025
• Subjects: Batteries; Computer architecture; Data mining; Degradation; Encoding; Feature extraction; Forecasting; Lithium-ion battery; remaining useful life prediction; Time series analysis; time step variable block; transformer; Transformers; Vectors
• ISSN: 2687-9735; 2687-9743
• DOI: 10.1109/JESTIE.2025.3585988

Precisely forecasting the lifetime of lithium-ion batteries is crucial for addressing consumer worries regarding their safety and dependability. However, existing research predominantly focuses on individual degradation characteristics of batteries, neglecting their multiple degradation features or their interactions. To address these issues, this paper proposes a time series method based on Transformer with a parallel self-attention mechanism to forecast the remaining useful life of lithium-ion battery. First, the method processes lithium-ion battery data through a sampling layer and integrates a time step variable block, where the latter combines a time step encoding layer and a variable encoding layer to capture degradation information from both the time and feature dimensions. The time step encoding layer learns long-term dependencies through the self-attention mechanism, while the variable encoding layer focuses on the local degradation features from different sensors. The time step encoding layer and the variable encoding layer operate in parallel to extract both temporal data and sensor degradation features. Then, these two layers focus on the different aspects within the feature vector, capturing the correlations between these features through a multi-head self-attention mechanism, and determining the relative significance of each feature in forecasting the current time step. These correlations and the weighted features are fused into a new feature vector. Finally, the new feature vector is passed into the decoder to compute the prediction result. Experimental results on two classical lithium battery datasets show that our approach surpasses the existing methods in predicting battery remaining useful life.