Adaptive battery SOC estimation: Temporal attention-enhanced BiLSTM and quantized model comparison for scalable edge solutions

Lithium-ion batteries are the dominant energy storage technology in modern energy systems, making the accurate state-of-charge (SOC) estimation essential for effective battery management. Reliable SOC estimation helps determine remaining battery capacity, improves operational safety, and supports real-time energy control. However, SOC estimation is challenging due to the nonlinear electrochemical behavior of lithium-ion batteries and performance degradation caused by aging and temperature variations. This study proposes a Bidirectional Long Short-Term Memory network with a Temporal Attention Mechanism (BiLSTM-TAM) to enhance SOC estimation accuracy across a wide temperature range from −20 °C to 40 °C. The model is evaluated under several driving conditions, including the Urban Dynamometer Driving Schedule (UDDS), the US06 aggressive driving cycle, a mixed driving cycle that combines multiple standard cycles, and the Hybrid Pulse Power Characterization (HPPC) test commonly used for battery modeling. Model performance is assessed using root-mean-square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R 2 ). Results demonstrate that the proposed BiLSTM-TAM model improves SOC estimation accuracy and robustness. To enable practical deployment, post-training quantization (PTQ) is applied to reduce model size and computational requirements while maintaining predictive performance, making it suitable for edge-based electric vehicle applications.