Health management of power batteries in low temperatures based on Adaptive Transfer Enformer framework

Abstract

Accurate State of Charge (SOC) estimation is essential for extending battery life and improving the safety of battery management systems. However, many existing methods face challenges, including a lack of sufficient samples in specific driving modes, overlooking hidden factors such as low temperatures, and experiencing negative transfer in transfer learning. This paper introduces the Adaptive Transfer Enformer (ATE) Framework, which integrates an Enhanced Transformer (Enformer) model with Adaptive Transfer Learning (ATL). The Enformer incorporates Multilevel Residual Attention (MRA) and Pattern Dynamic Decomposition (PDD), forming the backbone of the pre-trained model. MRA addresses gradient vanishing issues due to limited samples and captures the underlying relationships at each time point. PDD dynamically learns temporal trends, hidden factors, and their interactions. ATL provides an effective feature learning strategy to promote positive transfer in SOC estimation. Experimental results on two public datasets with added noise show that the proposed method improves average accuracy compared to state-of-the-art methods. Additionally, results from nine transfer scenarios demonstrate the strong generalization and noise resistance capabilities of the ATE Framework.