A Neural Network Approach for Health State Estimation of Lithium-Ion Batteries Incorporating Physics Knowledge

The assessment of the State of Health (SOH) of lithium-ion batteries is paramount to ensuring the safety and reliability of battery management systems. Numerous researchers have employed Equivalent Circuit Models (ECM) and data-driven methodologies to estimate SOH. Each methodology has its merits and drawbacks, yet their integration poses substantial challenges. This paper proposes a novel approach for SOH estimation that synthesizes ECM with data-driven techniques. Initially, parameters for a second-order ECM are identified utilizing the voltage rebound characteristics of lithium-ion batteries. Subsequently, a predictive model is established employing a Long Short-Term Memory (LSTM) neural network. Finally, features extracted from the ECM and the dataset are utilized as inputs for the LSTM neural network to predict SOH. The efficacy of the proposed technique is corroborated by datasets from NASA and CALCE. Results indicate that the novel method's maximum Root Mean Square Error (RMSE) is confined to 0.79%, and the Mean Absolute Error (MAE) is limited to 0.47%. Compared to other methods, this approach exhibits faster convergence, higher precision, and enhanced generalizability.