Hybrid LSTM–Physics Electro-Thermal Modeling of Samsung INR21700-53G

Lithium-ion batteries constitute the cornerstone technology for contemporary energy storage systems, particularly in automotive applications, due to their superior energy density, extended operational lifetime, and exceptional efficiency characteristics. However, thermal management represents a critical engineering challenge, as heat generation during electrochemical cycling significantly impacts performance, safety margins, and long-term reliability. Elevated operational temperatures necessitate sophisticated thermal management architectures, while the complex coupling between electrochemical and thermal phenomena presents substantial computational modeling challenges. This research addresses the critical need for accurate thermal prediction models in lithium-ion batteries by developing and validating two distinct modeling approaches for voltage and power estimation applied to a Samsung 53G 21700 Li-ion cell. The investigation represents a pioneering application of machine learning techniques to battery thermal modeling, introducing Long Short-Term Memory (LSTM) neural networks as an innovative alternative to conventional electrochemical modeling approaches. Two methodologies were implemented and systematically compared: first, a comprehensive lumped-parameter model incorporating coupled electrochemical and thermal physics within a unified COMSOL computational domain, utilizing experimentally-derived material properties; second, an innovative hybrid approach employing LSTM neural networks for electrical power estimation in MATLAB, with subsequent thermal simulation performed in COMSOL utilizing the computed power profiles as thermal boundary conditions. The development of the Machine Learning (LSTM) approach originated from analysis conducted at multiple discharge rates at different temperatures (25°C and 40°C) using classical 1 RC branch ECM approach, which revealed substantial voltage estimation errors due to electrical hysteresis effects at low State of Charge conditions, propagating into thermal prediction inaccuracies. The LSTM-based approach demonstrated superior thermal prediction accuracy by learning battery behavior patterns directly from experimental data rather than relying on predetermined cell parameters, effectively capturing complex hysteresis dynamics. Quantitative results at 25°C demonstrate the LSTM-based approach achieved RMS temperature error of 0.302°C during validation test through drive cycle testing confirming practical applicability compared to 0.475°C for the lumped parameter approach. The hybrid LSTM based model appears to simulate temperature better than the lumped model, but the lumped model requires a third of the computational time and its results are more consistant, always overestimating the temperature while the LSTM based tends to overestimate temeprature in discharge and underestimate in charge. This research contributes significantly to the advancement of battery thermal modeling methodologies through the novel integration of machine learning techniques with traditional finite element thermal analysis. The methodology exhibits versatility and adaptability for various battery configurations and operating conditions, with applications extending to diverse lithium-ion battery systems and broader thermal management challenges in energy storage applications.