Leveraging Trend-Aware Attention in Transformers for Lithium-Ion Battery Capacity Prediction

Abstract

The prediction of lithium-ion battery capacity plays an essential role in ensuring the reliability and safety of modern electronic devices. To effectively capture the local trend information inherent in lithium-ion batteries and enhance the accuracy of capacity forecasts, this letter presents an innovative Transformer model that incorporates a specialized trend-aware attention mechanism. This novel model synergistically combines the strengths of trend-aware attention and the Transformer encoder. It introduces 1-D convolution within the trend-aware attention framework, thereby replacing the traditional linear projections of queries and keys found in conventional self-attention mechanisms. This strategic enhancement enables the model to more adeptly and efficiently capture both local trends and global features, surpassing the performance of standard self-attention approaches. Extensive validation using the NASA and CALCE lithium-ion battery datasets reveals that the proposed model significantly outperforms existing state-of-the-art models across a variety of evaluative metrics. This noteworthy performance underscores the model's advantages in effectively managing the complexities of time-series data for accurate battery capacity prediction.