Comprehensive Comparison of Machine Learning and Kalman Filter Battery State of Charge Estimators

This paper provides a comprehensive review of five different methods for estimating state of charge (SoC): feedforward neural network (FNN), nonlinear autoregressive with exogenous inputs (NARX) neural network, long short-term memory (LSTM) neural network, extended Kalman filter (EKF), and unscented Kalman filter (UKF). The analysis was conducted using a comprehensive experimental dataset for 21700 cylindrical lithium ion cells from an electric vehicle. The dataset covers a wide range of operational conditions, including three vehicle payloads (80 kg to 1000 kg), six temperatures (−20°C to 40°C), and different drive cycles. Ten different measurement offset error values are added to the data, capturing sensor error that may be present in the vehicle, and incorrect initial SoC estimates are also considered. The computational efficiency and memory requirements of each algorithm when deployed on an NXPő S32K146 Evaluation Board and NXPő GreenBox 3 Real-Time development platform were also explored, which is critical for real-time applications in vehicle control units. The results show that the NARX model consistently outperformed the other models, with an RMSE of 1.70%, while other models ranged between 2% and 3%. All algorithms demonstrated robustness to sensor errors, with current, voltage, and temperature offsets increasing estimation errors by no more than 0.5%. The NARX and LSTM models showed over 90% faster convergence times compared to other methods. Computational efficiency was similar across all models, with execution times ranging from 1 to 2 ms for estimating SoC of 10 cells at 1 Hz on the GreenBox platform. Overall, the findings suggest that the NARX model offers the best trade-off considering memory usage, execution time, and accuracy among the analyzed models, being a reliable choice for SoC estimation in electric vehicles.