Machine Learning and Multilayer Perceptron Enhanced CFD Approach for Improving Design on Latent Heat Storage Tank
Yuanji Li,Xinyu Huang,Xiaoyong Huang,Xinyu Gao,Rukun Hu,Xiaohu Yang,Ya-Ling He
DOI: https://doi.org/10.1016/j.apenergy.2023.121458
IF: 11.2
2023-01-01
Applied Energy
Abstract:Solid-liquid phase change heat storage is an important method to solve the mismatch of the generation and usage of waste heat, which is conducive to decarbonization and energy conservation. However, there always exists inhomogeneity for the melting and temperature of the phase change material (PCM) in solid-liquid phase change heat storage. Changing the shape of the heat storage tank will change the distance of heat transfer to the phase change material, thus changing the heat storage performance. Therefore, this paper designed ten shapes of trapezoidal tank with different proportions of upper and lower radius. And the heat storage performance of ten cases were investigated through the designed numerical model, whose accuracy is convinced through comparison with experimental results. Many indexes are proposed to evaluate the properties of the latent heat thermal energy storage (LHTES) from the view of whole region, uniformity among subregions, and the storage heat performance of integrity. Except for drawing mass fraction, temperature, and streamline contours, the temperature proportion and the average velocity are proposed to quantitatively evaluate the temperature distribution and convection intensity. The results showed that the cases with larger upper radius are shorter than cases with larger lower radius. In addition, the shortest melting time is 17210 s, which is about 39.2% shorter by 11080 s than that of basic case. The evaluation from the view of uniformity also indicated that properly increasing the PCM amount is beneficial for improving the melting uniformity. The most uniform temperature and mass fraction are case 5 and case 4, respectively, with 12.92% and 38.28% improvement. Finally, one multilayer perceptron (MLP) network model is built to predict the melting fraction and total heat storage. After training, one model with a deviation lower than 10% between simulation and prediction is set up.