A transferable multi-state estimation framework for lithium-ion batteries based on sparse electrochemical parameters

Reliable estimation of internal battery states is critical for ensuring safety and performance in electric vehicles and energy storage applications. However, conventional methods often struggle with issues of generalization and scalability. This study proposes a transferable multi-state estimation framework of state of health (SOH), state of charge (SOC) and state of energy (SOE), integrating sparse electrochemical parameters identified only at a few representative aging points with deep learning and transfer learning. These sparse parameters are reconstructed into continuous pseudo parameter sequences. These sequences are used to drive a hybrid model that combines a convolutional neural network (CNN) and a bidirectional long short-term memory network (BiLSTM) for SOH estimation, achieving a minimum mean absolute error (MAE) of 0.2883 %. A two-stage simulation correction strategy significantly reduces simulation errors by over one order of magnitude. Based on the corrected data, SOC and SOE are estimated using support vector regression (SVR) models, achieving mean absolute errors below 1.5 %. A hybrid transfer-based strategy enables rapid model adaptation across battery types and operating conditions with limited labeled data. The proposed framework achieves a strong balance between physical consistency and data-driven adaptability, providing a robust and scalable solution for intelligent battery state estimation.