"Hybrid battery thermal management system assessment using shape-stabilized phase change material"

The growing deployment of lithium-ion batteries in electric vehicles, stationary energy storage, and portable electronics has intensified the need for accurate thermal monitoring and predictive management strategies. This report investigates an integrated approach that couples advanced thermal modeling with data-driven forecasting to enhance battery safety, performance, and lifespan. A comprehensive review of conventional thermal management methods, including phase change materials (PCMs), liquid cooling, and heat pipe systems, establishes the limitations of purely passive or purely active solutions. Special emphasis is placed on Shape-Stabilized PCMs (SS-PCMs), which offer leakage-free operation and stable thermal conductivity, yet require tailored integration to manage transient temperature spikes. This work explores a hybrid battery thermal management system (BTMS) that combines SS-PCMs with active liquid cooling and heating. Building upon prior work, which developed a detailed CFD 3D model of the hybrid BTMS to optimize SS-PCM properties, this project introduces a simplified 2D transient numerical model. The new model reduces computational cost dramatically from 31 hours to 7.15 hours, while maintaining high fidelity, with a maximum temperature deviation of 0.5 °C and an average error of 0.2 °C, corresponding to relative errors of 0.6% and 0.1%, respectively. Based on this efficient model, a full-scale battery pack consisting of 39 cylindrical LIB cells is simulated, with explicit integration of active system control, ambient thermal boundary conditions, and cell-to-cell temperature variation. To support data-driven approaches, a fully automated workflow is developed, generating a comprehensive dataset covering a wide range of C-rate profiles and boundary conditions. To enhance predictive thermal control, this thesis evaluates three recurrent neural network (RNN) architectures: Simple RNN, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). A complete preprocessing pipeline, feature engineering strategy, and hyperparameter tuning framework are implemented to optimize predictive accuracy while avoiding overfitting. Among the tested models, GRU and LSTM achieve the highest accuracy, with GRU outperforming LSTM by producing fewer error peaks and reducing computation time. Ultimately, the GRU-based reduced-order model (ROM) cuts thermal prediction time from 5.60 hours to only 3 minutes, including both training and prediction, thereby providing a lightweight, data-driven surrogate capable of approximating thermal responses based on new ambient conditions. The proposed methodology establishes the foundation for a digital twin of hybrid BTMS with SS-PCM, demonstrating that physics-based modeling and data-driven forecasting can be tightly integrated into a scalable framework. Beyond predictive accuracy, the ROM opens pathways toward self-optimizing thermal management systems that learn and adapt in real time. The thesis outlines opportunities for future work, including the integration of physics-informed neural networks loop based on Bernardi equation, predictive modeling of PCM state for optimization properties, reinforcement learning-based active control strategies, experimental validation, and further generalization of the model to different current cycles and ambient conditions.