Neural-accelerated numerical model for packed bed latent heat storage system

• A neural accelerated numerical model for latent heat storage systems is proposed. • It fuses fast, less accurate model with slow, accurate model via a neural network. • It is 764 times faster than the slow and accurate model with error of below 0.235%. • Offers good generalization, interpretability, and strong extrapolation capability. • It is suitable for model-based optimization and control applications. Developing accurate and computationally efficient dynamic models for packed-bed latent-heat storages (PBLHS) is crucial for reliably predicting their performance across different operating scenarios and enabling their use in planning and real-time control. In this study, a novel neural accelerated numerical model for PBLHS is proposed by coupling a neural network (NN) into a coarsely discretized equations of the Continuous-solid Phase (CP) model. The embedded NN predicts the surface temperature of the phase change material (PCM) given the fluid temperature and enthalpy of the PCM as inputs, which the CP model fails to capture. This allows the neural accelerated model to replicate the accuracy of a high-fidelity and computationally expensive model namely Concentric Dispersion (CD) model. An innovative data generation process to generate training data for NN involving both CD and CP model is proposed. Two versions of neural accelerated model are proposed, one with conventional NN and another using NN with a custom activation function. Both versions demonstrate an excellent accuracy, achieving MSE as low as 0.117°C, R 2 values closer to 0.995 and error percentage below 0.394 % compared to the highly accurate CD model. As for computational efficiency, the proposed models achieved 342 times and 764 times acceleration respectively. The gain in more acceleration for the later version of the proposed model is achieved through the use of a compact architecture that benefits from the custom activation function, while also enhancing model explainability. These results highlight the model’s suitability for scenarios demanding both high accuracy and computational efficiency.