A Multi-Head Attention ResCNN–BiGRU Model for Robust SOC Estimation in EVs Lithium-Ion Batteries Using Real-World Driving Data

Precise assessment of the state of charge (SOC) is essential for ensuring the safety, efficiency, and dependability of electric vehicles (EVs). However, achieving high accuracy remains challenging due to the complex nonlinear characteristics of lithium-ion batteries (LIBs), which are strongly influenced by ambient temperature fluctuations, dynamic load variations, and aging effects. To address these challenges, this study proposes a sophisticated hybrid deep learning (DL) framework that combines convolutional neural networks (CNNs) with residual connections (ResCNN), bidirectional gated recurrent units (Bi-GRU), and a multi-head attention (MHA) mechanism. In the proposed ResCNN-Bi-GRU-MHA model, the ResCNN layer implements residual connections to strengthen CNN-based spatial feature extraction, allowing for more stable training and richer feature representations, while the Bi-GRU layer captures temporal dependencies; subsequently, both features are dynamically integrated through theMHAto improve prediction accuracy. The framework is validated using real-world field data gathered from BMW i3 EV driving trips outfitted with a 60 Ah LIB pack. A comparative analysis is performed to determine the influence of the MHA by comparing the hybrid ResCNN-Bi-GRU model with and without the MHA. The established results demonstrate that the ResCNN-Bi-GRU-MHA model achieves superior predictive accuracy, with a root mean squared error (RMSE) of 6.0476% and a mean absolute error (MAE) of 4.5550%. Furthermore, the proposed framework outperforms conventional approaches, including feedforward neural networks (FFNN), long short-term memory (LSTM), gated recurrent units (GRU), and extreme learning machines (ELM). These findings point out the ResCNN-Bi-GRU-MHA model’s robustness for real-time SOC estimation in battery management systems (BMS).