Algorithm-oriented benchmarking of deep learning and hybrid architectures for robust SOC estimation in electric vehicle batteries

Accurate state-of-charge (SOC) estimation is a key requirement for the safe and efficient management of lithium-ion batteries in electric vehicles, especially under varying thermal and dynamic operating conditions. This study presents a comprehensive, algorithm-oriented assessment of several deep learning and hybrid SOC estimation architectures—including feedforward neural networks (FNN), gated recurrent networks (GRU), long short-term memory networks (LSTM), temporal convolutional networks (TCN), and their hybrid combinations—using a multi-temperature dataset collected at 10 °C, 25 °C, and 40 °C under diverse dynamic load profiles and standardized drive cycles such as UDDS, HWFET, US06, and LA92. All architectures were trained and evaluated under a unified preprocessing and training configuration to ensure methodological consistency and a fair basis for comparison. The evaluation highlights how different recurrent, convolutional, and hybrid architectures respond to thermal variations and dynamic load transitions, revealing model-specific strengths and limitations under realistic operating conditions. Among the evaluated models, the hybrid FNN + GRU architecture demonstrated the most reliable overall performance, achieving an RMSE of 1.11 % and reducing peak estimation errors to 3.6 % under nominal temperature conditions. SOC-zone analysis further showed characteristic error amplification at low and high SOC levels, emphasizing the importance of architectures capable of capturing nonlinear boundary dynamics. Computational benchmarking indicated that hybrid structures—particularly FNN + GRU—also provide an advantageous balance between estimation accuracy and inference speed, supporting their suitability for embedded Battery Management Systems (BMSs) with real-time constraints. Overall, this study contributes a unified evaluation framework that simultaneously addresses thermal robustness, dynamic load variability, SOC-dependent behavior, and computational efficiency, offering practical guidance for selecting reliable and deployable SOC estimation models for next-generation electric vehicle BMSs.