Deep learning framework designed for high-performance lithium-ion batteries state monitoring

Accurate state of charge (SOC) estimation is crucial for ensuring the safety of batteries, especially in real-time battery management system (BMS) applications. Deep learning methods have become increasingly popular, driving significant advancements in battery research across various fields. However, their accuracy is limited due to the nonlinear adverse driving conditions batteries experience during operation and an over-reliance on raw battery information. In this work, a deep-stacked denoising autoencoder is established for a long short-term memory model that incorporates a transfer learning mechanism to estimate and study the SOC from an electrochemical perspective. More importantly, this proposed model is designed to extract and optimize the electrochemical features from the training data on a secondary scale, improving noise reduction and the precision of initial weights. This adaptation allows for accurate SOC estimation of batteries while minimizing interference and divergence. For large-scale applicability, the proposed model is tested with high-performance lithium-ion batteries featuring different morphologies under a range of complex loads and driving conditions. The experimental results highlight the distinct behaviors of the tested batteries. Moreover, the performance of the proposed model demonstrates its effectiveness and outperforms existing models, achieving a mean absolute error of 0.04721% and a coefficient of determination of 98.99%, facilitating more precise state monitoring of batteries through secondary feature extraction.