A high-speed recurrent state network with noise reduction for multi-temperature state of energy estimation of electric vehicles lithium-ion batteries

The state of energy (SOE) of lithium-ion batteries is a critical metric for evaluating the remaining driving range of electric vehicles. It is also an important parameter monitored by the battery management system. However, current machine learning-based methods for SOE estimation suffer from issues such as high complexity, long training times, and output fluctuations, which make them unreliable and inefficient. To address these challenges, this study proposes a high-speed and accurate SOE estimation method under realistic automotive working conditions. Specifically, a double reservoir deterministic jump recurrent state network with noise reduction is constructed. The double reservoir structure is designed for feature mapping, while the deterministic jump connection mechanism is used to reduce complexity and enhance dynamic performance. Additionally, the denoising module is used to improve prediction accuracy and reduce output noise. Finally, the proposed method is validated and analyzed using experimental data from various real-world automotive conditions at multi-temperatures. The results demonstrated a maximum absolute estimation error within 0.0254 and estimation time reaches the millisecond level. All evaluation metrics outperform baseline models, demonstrating its high-speed estimation performance, high accuracy, and usability, which makes promising for future applications in SOE estimation for real-world electric vehicle battery management system.