Performance Comparison of Recurrent Neural Network Variants for State of Charge Prediction in Lithium-Ion Batteries

This paper presents a comparative study of three widely used recurrent neural network (RNN) variants-Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Gated Recurrent Unit (GRU)-for predicting the State of Charge (SoC) of lithium-ion batteries in electric vehicles (EVs). A key contribution of this work is the evaluation of these models under varying environmental and operational conditions, specifically using battery data (INR18650-20R) collected at three different ambient temperatures: $0^{\circ} \mathrm{C}, 25^{\circ} \mathrm{C}$, and $45^{\circ} \mathrm{C}$. The models are trained using the Dynamic Stress Test (DST) drive cycle and tested on two unseen drive cycles-Federal Urban Driving Schedule (FUDS) and the Supplemental Federal Test Procedure (US06)-without retraining, demonstrating their generalization capability. The performance of each model is assessed using standard evaluation metrics, including Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the Coefficient of Determination ($\mathrm{R}^{\mathbf{2}}$). Results indicate that all three RNN architectures provide accurate SoC predictions, with notable differences in performance under different conditions. This study provides valuable insights into the robustness and effectiveness of RNN-based approaches for real-time SoC estimation in EV battery management systems.