Data-driven modeling for fast and accurate transient thermal predictions in shell-and-tube latent thermal energy storage devices

Traditional numerical simulation methods have long struggled with high computational costs when predicting the melting behavior of phase change materials in latent thermal energy storage (LTES) devices. This study presents a reduced-order prediction model that combines proper orthogonal decomposition (POD) with Bayesian neural networks (BNN) for efficient forecasting of melting behavior in a shell-and-tube LTES device. A three-dimensional unsteady numerical model was established and validated using experimental data, and spatiotemporal datasets were generated across various operational conditions. The POD approach was employed to reduce the dimensionality of flow field by extracting dominant modes, while operating parameters such as time, inlet temperature, and inlet flow rate were mapped to reduced-order coefficients through the BNN, enabling fast field reconstruction under different operating conditions. Results showed that the proposed data-driven model achieves high prediction accuracy. The mean absolute errors for temperature and liquid fraction were 0.29 K and 1.15 × 10⁻³ under a specific operating condition, respectively. As for varying operating conditions, temperature predictions exhibited mean absolute errors of 4.3–12.3 K at monitoring points, while liquid fraction predictions had mean absolute errors of 0.044–0.094, with melting time deviations of approximately 4.0 %. Furthermore, the model provides a computational speedup of approximately 10⁴ times compared to traditional CFD simulations. Overall, the data-driven approach proposed in this work exhibits potential for providing accurate and computationally feasible transient thermal predictions in LTES devices.