Flexible HI-Based Approach to SOH Estimation in Energy Storage Systems

Lithium-ion batteries (LIBs) play a pivotal role in energy storage systems (ESSs) and electric vehicles (EVs); however, they inevitably undergo capacity fading and performance degradation during long-term operation. Accordingly, accurate state-of-health (SOH) estimation is essential for reliable battery management. In this study, refined health indicators (HIs) are defined from charge– discharge voltage–time characteristics, and a flexible SOH prediction framework is proposed. Aging data obtained from INR21700-33J cells over 1,400 cycles are analyzed to extract multiple HIs, including mean voltage falloff (MVF), voltage interval of equal discharging/charging time difference (VIEDTD/VIECTD), and time interval of discharging/charging voltage difference (TIEDVD/TIECVD). These indicators are further subdivided using voltage resolutions of 0.1 V and 0.01 V, enabling robust HI extraction under data imbalance. To compensate for missing HI data caused by external operational factors in practical ESS environments, a denoising autoencoder (DAE)-based interpolation method is employed to preserve temporal degradation trends. Subsequently, recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) models are applied for SOH prediction. Among them, the GRU model achieves the best performance, with an MAE of 2.24, RMSE of 2.74, and R² of 0.97, demonstrating improved accuracy and robustness under incomplete operational data conditions.