Temperature prediction in EV battery cells under varying ambient temperatures and charge-discharge conditions using finite element method and neural networks

• A hybrid FEM–Neural Network approach accurately predicts EV battery cell temperature in real time. • FEM simulates thermal behavior under varying charge-discharge cycles and ambient temperatures. • The best model (2 hidden layers, 512 neurons) achieves high accuracy ( R ² up to 0.98, RMSE 0.58). • The method significantly reduces computational time, enabling efficient thermal management in EVs. Accurate temperature prediction in electric vehicle (EV) battery cells is crucial for optimizing performance, ensuring safety, and extending lifespan. Addressing the limitations of existing static-boundary models, this study presents a novel hybrid framework that combines the Finite Element Method (FEM) and Neural Networks (NNs) to model the thermal behavior of lithium-ion battery cells under simultaneous fluctuations in charge-discharge cycles and ambient temperatures. The temperature range of 293.15–313.15 K (20–40 °C) is selected, as it represents typical EV operating conditions, enhances model accuracy, and supports safe and efficient thermal management. To ensure model robustness, the FEM physics-based model was first calibrated against experimental data. Subsequently, the data from FEM simulations, covering a 2000 s duration and 1 s interval over the range of 293.15–313.15 K (20–40 °C) for a total of 42,000 samples, was randomly partitioned into training (60%), validation (20%), and testing (20%) subsets to ensure generalization to unseen data. Key inputs include ambient temperature, current, state of charge (SoC), and voltage. Among the tested models, Model 1 (two hidden layers, 512 neurons) outperformed the others, achieving the lowest MAE (0.45, 0.38, 0.42), the lowest RMSE (0.58), and the highest R² values (0.98, 0.97, 0.96). While simpler models trained faster, they lacked accuracy; deeper models offered no significant improvement. Model 1 optimized the accuracy-efficiency trade-off, requiring just 2.5 min for training and 0.002 s for inference. This high speed and precision confirm its viability as a practical solution for real-time EV thermal management.