A Compact Unified Methodology via a Recurrent Neural Network for Accurate Modeling of Lithium-Ion Battery Voltage and State-of-Charge

This work investigates an approach to combining accurate lithium-ion battery (LIB) dynamic modeling and effective state-of-charge (SOC) prediction at various operating conditions using a structured recurrent neural network (RNN). The RNN model is trained with drive cycle data so that model parameters do not have to be determined with characterization tests, as is typically necessary for an equivalent circuit model (ECM). The RNN is also able to capture the Butler-Volmer (BV) relationship for the charge-transfer voltage drop current dependency and the lithium-ion diffusion process, two characteristics which are challenging to capture with an ECM. This work proposes a compact unified methodology of two RNNs (current-based & power-based) incorporating Gated Recurrent Unit (GRU) and Deep Feature Selection (DFS) structures. Both RNNs accurately model LIB dynamic responses including battery nonlinear behavior at different temperatures, while the power-based RNN also exhibits effective SOC prediction capability. The power based RNN is also shown to accurately predict battery state of charge versus time for a drive cycle, which is useful for vehicle range prediction. Both RNN models can also be used as a LIB simulator in model-based design and especially for hardware-in-loop (HIL) applications to test battery management systems and other electronic components.