Experimental evaluation of SOC estimation performance under realistic thermal conditions, load dynamics, and dataset complexity in electric vehicle batteries

Accurate State-of-Charge (SOC) estimation is essential for ensuring the safety, efficiency, and longevity of lithium-ion batteries in electric vehicles (EVs). However, achieving reliable SOC prediction under real-world conditions remains challenging due to nonlinear electrochemical behavior, temperature dependence, and dynamic load variability. This study presents a comprehensive evaluation of SOC estimation performance under diverse operating conditions, focusing on the combined effects of temperature, current profile complexity, and SOC range. To acquire the experimental dataset, a test setup was designed, and experiments were conducted at 10°C, 25°C, and 40°C for charge–discharge and Hybrid Pulse Power Characterization (HPPC) tests. These datasets encompass a wide range of operational scenarios, ranging from low-rate and steady-state conditions to highly dynamic load pattern representative of actual EV operations. Two data-driven approaches, Feedforward Neural Network (FNN), and Random Forest (RF), were trained and evaluated across all datasets to investigate their estimation robustness and generalization capability. The results demonstrate that both models achieve their highest accuracy under nominal thermal conditions (25°C) and within the mid-SOC range (20–80%), whereas their performance deteriorates significantly under extreme temperatures and irregular current profiles. The FNN model consistently outperformed the RF, yielding lower maximum errors and smoother error distributions. These findings underscore the importance of dataset structure, thermal conditions, and SOC range in determining estimation accuracy, and highlight the necessity of realistic, scenario-driven testing for robust algorithm validation. By providing one of the most comprehensive experimental assessments of SOC estimation under realistic EV operating conditions, this study establishes a solid foundation for developing adaptive algorithms and more reliable battery management systems.