Exploring efficacy of machine learning (artificial neural networks) for enhancing reliability of thermal energy storage platforms utilizing phase change materials

Abstract

Thermal Energy Storage (TES) platforms mitigate the paucity between peaks in consumption and supply, i.e., they absorb thermal energy during periods of excess supply and release thermal energy during periods of deficit. Phase change materials (PCMs) have attracted significant attention over recent years due to their efficacy in improving performance and reliability of TES platforms. High latent heat values accruing from the use of PCMs enable improved storage densities which in turn yield devices with compact form factors for TES applications. Typically, inorganic PCMs afford higher latent heat values than organic PCMs, yet often at the detriment of compromised reliability. A crucial issue with inorganic PCMs is the higher degree of supercooling (also known as “subcooling”) required to initiate nucleation (which degrades their reliability, net energy storage capacity, and power rating of the TES platform). “Cold Finger Technique (CFT)” can be used to mitigate these issues where a small portion of the total mass of PCM in the TES platform is left in solid state (in order to facilitate the spontaneous nucleation). Therefore, reliability issues are ameliorated by using CFT but at a marginal cost to the net storage capacity while power rating of the TES remains almost unaffected. In this study, machine learning (ML) techniques were leveraged to exploit the capability of CFT more effectively. Temperature transients from PCM melting experiments are used to investigate the capability of this deep learning technique (i.e., using multi-layer perceptron model or “MLP”) in order to predict the required time to reach the designated melt-fraction of the PCM. The results show that the Artificial Neural Network (ANN) model designed and implemented in this study is capable of predicting the time required to reach pre-designated value of melt-fraction with outstanding accuracy (e.g., 90% melt-fraction or 95% melt-fraction, that is specified by the user). The mean error of the predictions is calculated and is expected to be less than 10 min, especially for an interval of 30 min before the TES platform reaches the desired value of the melt fraction (i.e., 90% melt-fraction) for a total time spanning 2∼3 h. However, this approach is more susceptible to the fidelity of the data-set utilized for training the ANN (MLP) algorithm.