Deep Learning to Enhance Transient Thermal Performance and Real-Time Control of an Energy Storage (TES) Platform

Phase Change Materials (PCMs) are used for thermal energy storage (TES) due to their high latent heat capacity. While inorganic PCMs have the ability to store large amounts of thermal energy (as they possess higher values of latent heat capacity) they need significant supercooling to initiate solidification - which compromises their operational reliability (compared to organic PCMs). 'Cold Finger Technique'/ CFT (wherein a portion of the PCM is maintained in the solid-state as the un-melted PCM serves to initiate nucleation for subsequent solidification of the rest of the PCM) enables higher operational reliability at the expense of a marginal loss of storage capacity (from the portion of the PCM that remains un-melted). In order to apply CFT, one possible approach is to target a desired melt fraction in the early stages of operation for avoiding catastrophic failure (i.e., avoiding 100% melting). The objective is to reach as close to 100% melting as possible - without completely melting the whole mass of PCM - using a reliable and predictable method. Hence, at any instant, a reliable way is desired to predict the time remaining to reach a particular melt fraction of the PCM (during the melting cycle). In this study we determine the efficacy of employing Artificial Intelligence (AI) based techniques, especially Deep Learning, for predicting the time remaining to reach a target melt fraction at any instant during the melting cycle. The time history of the PCM melt fraction and temperature transients at multiple locations within the TES platform can thus be correlated by employing an Artificial Neural Network (ANN) to achieve this goal. The aim of this study is to develop and deploy ANN based prediction tools that can enable real-time predictions (e.g., when the melt cycle should be stopped and freezing cycle should be initiated) apriori based on instantaneous values of temperature measured during operation of a TES platform. The data for training and evaluating the efficacy of the ANN model predictions is obtained from PCM melting experiments performed in this study using PureTemp29. Initially a solid mass of PCM is melted (using a nichrome wire heater mounted at the bottom) in a graduated cylinder with three thermocouples recording the temperatures at distinct locations along the height of the cylinder. An automated digital data acquisition (DAQ) system is used to record the transient temperature profiles from these thermocouples. The height of the liquid meniscus (melted PCM) is recorded using a digital image acquisition apparatus (and also monitored periodically from the measuring cylinder containing the PCM) until the PCM is completely melted. The results show that the predictions of the ANN model are most accurate in the later stages of the melting cycle. The error decreases as the melt fraction approaches the target melt. For a target melt of 90%, the error in predictions in the last 1800 seconds of the melting cycle is in the range 150-350 seconds.