State of Charge Estimation Across Dynamic and Sub-Zero Thermal Conditions for Electric Vehicles

Accurate State of Charge (SOC) estimation is critical for ensuring the performance, safety, and longevity of lithium-ion battery systems, particularly under varying thermal and dynamic operating conditions such as in electric vehicles. Despite recent advances, existing deep learning approaches often rely on data collected in controlled, constant-temperature settings, limiting their real-world applicability. This study introduces a deep learning framework that leverages a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture, enriched with domain-specific features and tailored for realistic EV environments. In this model, in addition to voltage, current, and temperature, it incorporates dV/dt, integrated current, and their exponential moving averages to capture both transient and long-term electrochemical behavior. Comprehensive benchmarking against LSTM, GRU, and LSTM-DNN architectures is conducted using fixed-temperature training data and evaluated under varying ambient conditions ranging from −20°C to +25°C. Results show the proposed model consistently outperforms its counterparts, achieving lower RMSE and MAE across all scenarios, well bounded under 1.09% and 0.85%. These findings establish a robust, feature-enhanced, and temperature-resilient solution for real-time SOC estimation in modern battery management systems. Experimental results confirm that the proposed hybrid model consistently outperforms its counterparts.