Parallel Recurrent Artificial Neural Networks for Electric Vehicle Battery State of Health Estimation

Li-ion battery State of Health (SOH) estimation is a crucial function of the Electric Vehicle (EV) Battery Management System. This is because of the unpredictable performance of Li-ion battery cells once the nominal capacity drops below 70% (due to exposure to numerous cycles). Artificial Neural Networks (ANNs) have gained popularity for SOH estimation in recent years due to their high flexibility and low complexity. The possibility of using parallel recurrent architectures in ANNs for SoH estimation is investigated in this paper. Gated Recurrent Unit (GRU-RNN) architecture was used for the parallel recurrent layers, due to its simplicity and good SoH prediction capability as seen in recent literature. The charging profile of B0005, B0006, B0007 and B0018 batteries from the NASA Ames Prognostics Center of Excellence (PCoE) dataset were used for training and testing the ANNs. The time intervals between certain points in the charging voltage profile (3.8 to 3.9V, 3.9 to 4.0V and 4.0 to 4.1V) and the time interval between 0.1 to 0.05A of the charging current profile were used as input features. The obtained results show that the proposed model has low testing dataset Mean Squared Error (MSE) (0.0299%) and good generalization when compared to the conventional GRU-RNN (0.352% MSE), parallel Bidirectional GRU-RNN (0.0360% MSE), parallel Long Short Term Memory configuration (0.0549% MSE), Bidirectional GRU-RNN (0.035% MSE) and GRU-RNN with attention (0.0448% MSE). Overall, the proposed model can accurately predict the SoH of the Li-ion batteries upon successful implementation on an EV, resulting in better consumer safety.