Internal electrochemical variables-attention long short-term memory based on meta-learning and augmentation for battery state of health estimation

Accurate State of Health (SoH) estimation is critical for battery management in industrial applications. This paper proposes the Internal Electrochemical Variable-Attention Long Short-Term Memory (IEV-ALSTM), a physics-informed machine learning (PIML) method for SoH estimation. A Single Particle Model (SPM) incorporating Solid Electrolyte Interphase (SEI) growth, lithium plating, Loss of Active Material (LAM), and thermal effects is developed using PyBaMM and validated against a public Nickel-Manganese-Cobalt (NMC) battery aging dataset. The model serves two roles: generating simulation data for conditions not represented in experimental training sets, and extracting internal electrochemical variables that encode aging-relevant degradation information. A three-step Meta-Learning feature selection process identifies the two most informative internal electrochemical variables from a large candidate pool. These are combined with statistical features extracted from voltage, current, and temperature signals as inputs to the Attention-LSTM (ALSTM) model. The proposed IEV-ALSTM achieves MSE improvements of 90% and 79%, over Liu 2025, and a purely data-driven ALSTM, respectively.