A simplified electrochemical lithium-ion batteries model based on physics-informed LSTM_Res network

Abstract

Due to the high computational complexity of traditional electrochemical models, applying them to real-time battery management systems (BMS) is challenging. Most simplified electrochemical models lack sufficient accuracy due to assumptions and errors in the simplified solutions of partial differential equations. To strike a balance between complexity and accuracy, a Long Short-Term Memory Network with Residual Electrochemical Model (LrEM) for lithium-ion batteries is proposed. In this model, solid-phase lithium-ion transfer is approximated using a Long Short-Term Memory Network with Residuals (LSTM_Res), while lithium-ion transfer in the electrolyte is modeled using a standard Long Short-Term Memory (LSTM) network. The networks within the LrEM are trained by the underlying physical mechanisms in lithium-ion batteries. The results demonstrate that the LrEM can precisely predict the concentration of lithium ions in both solid and electrolyte phases, as well as accurately forecast the voltage under various current rates. Notably, the LrEM can complete 8000 s of charge and discharge dynamics in just 1.499 s, making it feasible for implementation in future real-time BMS. This advancement significantly enhances the mass transfer understanding and prediction in lithium-ion batteries, paving the way for more efficient and accurate battery management.