Application of Machine Learning (ML) for Enhancing the Transient Performance of Thermal Energy Storage (TES) Platforms Using Radial Basis Function (RBF)
Academic Article
Overview
Research
Identity
Additional Document Info
Other
View All
Overview
abstract
Thermal energy storage (TES) can be utilized as supplemental platforms for improving operational reliability and systemic efficiency in variety of industries, such as for reducing water usage in power production (food-energy-water/ FEW nexus), chemical and agro-process industries and for improving sustainability (e.g., desalination), etc. Phase change materials (PCMs) can be used in TES due to their high latent heat storage capacity during phase transformation. Inorganic PCMs typically have the highest latent heat capacity and are attractive for their ability to store the larger quantities of thermal energy in small form factors while conferring respectable power ratings (however, they suffer from compromised reliability issues, that often arise from the need for subcooling). Subcooling (also known as supercooling) is a phenomenon where the temperature needs to be reduced substantially below the melting point to initiate solidification. A technique for obviating subcooling issues is to allow a small portion of the PCM to remain un-melted. This allows the PCM to initiate nucleation from the un-melted portion of PCM (this is termed as the cold finger technique). Thus, reliability is enhanced at the expense of substantial reduction in storage capacity. A fundamental challenge for using this technique is the inability to reliably predict and control the amount of melt fraction in the total volume of the PCM (such that a target amount of the PCM remains solidified or un-melted at the end of each melt-cycle during repeated melting and solidification of the total mass of PCM). However, using Machine Learning (ML) techniques, this deficiency can be addressed by reliably predicting and thus controlling the amount of melt fraction in the total volume of the PCM with a higher accuracy than conventional techniques (such as using multi-physics-based models or numerical solvers). Conventional techniques for predicting transient characteristics in real time control schemes typically leverage multi-physics-based models that are often effective only for a narrow range of operating conditions with concomitant disadvantages: they are highly sensitive to small variations in the measurement uncertainties and are therefore susceptible to large levels of error in the real time predictions (and are unreliable for implementation in diverse range of operating conditions). In this pioneering study, nearest neighbor search processes (such as radial basis functions) were utilized along with machine learning (ML) algorithm using a training data set to predict the PCM melt fraction and to demonstrate the feasibility (and efficacy) of this approach. This technique is simple to implement and is device independent as well as robust (i.e., it can be deployed successfully even under conditions where the sensors malfunction, such as thermocouples that are off-calibration). This technique was demonstrated successfully for predicting the melt fraction of a PCM with high accuracy and robustness. With this method, the melt fraction of a PCM can be accurately determined, which allows the maximum thermal capacity of a PCM to be utilized while mitigating reliability issues (such as subcooling) and enhancing the thermodynamic efficiencies of the TES platforms. Melting experiments were performed using a digital camera (for video recording) and a graduated cylinder containing PCM for monitoring the transient values of the melt fraction based on the height of the liquid phase of the PCM in the cylinder. An array of 3 thermocouples was mounted at specific heights within the body of the PCM to monitor the temperature transients at these specific location during the propagation of the melt front within the PCM. In the final stages of the melting process, the predictions from the ML algorithm was found to be more accurate (90~95% accuracy) than that of the conventional techniques based on physics-based solvers (~60% accuracy). The accuracy of the ML algorithm was low at smaller melt fractions (~30%) and improved substantially at higher melt fractions (~95%). Furthermore, the accomplishments of this study display the feasibility of a RBF ML method which can be implemented for the accurate prediction and control of a real world stochastic system which can exhibit nonlinear and chaotic dynamics which change over time.
