Multi-scale Feature-driven Temperature Estimation for Lithium-ion Battery under Driving Scenarios Using Physical-enhanced Transfer Learning

Acquiring accurate temperature data for Lithium-ion batteries (LiBs) under dynamic driving conditions is crucial for ensuring their safety, performance, and extended lifespan. However, precise temperature estimation presents significant challenges due to environmental variations, limitations in sensor deployment, and complex driving cycles. To address this issue, this paper proposes a multi-scale feature-driven temperature estimation method for LiBs under driving scenarios using physical-enhanced transfer learning. First, a physical-constrained data enhancement (PCDE) approach is designed for driving conditions. By considering frequency matching and battery electrothermal characteristics, enhanced data that comply with physical laws are generated. Then, a TimesNet temperature estimation model is constructed to leverage its multi-period decomposition capability, enabling the extraction of multi-scale and multi-period features from current and voltage signals under dynamic conditions. Finally, a transfer learning strategy is developed by integrating the PCDE approach with the TimesNet model. Experiments are conducted across various driving scenarios using two datasets. The results demonstrate that the proposed method maintains excellent generalization performance under diverse and complex driving scenarios, with RMSE consistently below 0.4°C. The proposed method provides an effective technical solution for improving the temperature estimation accuracy of battery management systems in electric vehicles under complex driving scenarios.