Generalized real-time state of health estimation for lithium-ion batteries using simulation-augmented multi-objective dual-stream fusion of multi-Bi-LSTM-attention

Abstract

To maintain the safe and reliable operation of lithium-ion batteries and manage their timely replacement, accurate state of health (SOH) estimation is critically important. This paper presents a novel deep-learning framework based on multi-loss optimized dual stream fusion of attention integrated multi-Bi-LSTM networks (multi-ABi-LSTM), for generalized real-time SOH estimation of lithium-ion batteries. Battery sensor data is first preprocessed utilizing novel energy discrepancy aware variable cycle length synchronization and grid encoding schemes to achieve generalizability considering battery sets with different discharge profiles and then passed through two parallel networks: overlapped data splitting (ODS)-based attention integrated multi-Bi-LSTM network (ODS-multi-ABi-LSTM) and past cycles’ SOHs (PCSs)-based attention integrated multi-Bi-LSTM (PCS-multi-ABi-LSTM) network. The complementary features extracted from these two networks are effectively combined by a proposed fusion network to achieve high SOH estimation accuracy. Furthermore, a lithium-ion battery simulation model is employed for data augmentation during training, enhancing the generalizability of the proposed data-driven model. The suggested technique outperforms previous methods by a remarkable margin achieving 0.716% MAPE, 0.005 MAE, 0.653% RMSE, and 0.992 $R^2$ on a combined dataset consisting of four different battery sets with varying specifications and discharge profiles, indicating its generalization capability. Appliances using lithium-ion batteries can adopt the proposed SOH prediction framework to predict battery health conditions in real-time, ensuring operational safety and reliability.