A Bayesian Optimized Deep Learning Approach for Accurate State of Charge Estimation of Lithium-Ion Batteries Used for Electric Vehicle Applications

Abstract

The rapid growth of electric vehicles (EVs) has sparked significant interest in battery technology, particularly in monitoring the State of Charge (SOC). Accurate SOC estimation is critical for safe and efficient battery operation. While various estimation methods exist, more research is needed to adapt to diverse lithium-ion battery chemistries and operating conditions. Deep learning (DL) has shown considerable promise in improving SOC estimation accuracy; however, selecting optimal hyperparameters remains a persistent challenge. This paper presents an automated hyperparameter tuning method using Bayesian optimization applied to three deep learning architectures: Long ShortTerm Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM (BiLSTM). The models incorporate battery data including current, voltage, temperature, capacity, and average voltage and current values to enhance estimation accuracy. Tested across varying temperature condit4ions using the Panasonic 10°C dataset, the proposed approach demonstrates that BiLSTM models with 70 hidden neurons achieve SOC predictions with Root Mean Square Error (RMSE) below 0.02, significantly improving reliability in battery management systems. A CNN2D extension achieves the best performance with RMSE of 0.0094, outperforming all other evaluated models.