Leveraging Machine Learning (ML) for Enhancing Sustainability, Reliability, Robustness and Resilience in Thermal Energy Storage (TES) Applications Using Phase Change Materials (PCM)

Abstract

A Machine Learning (ML) algorithm was utilized to enhance the efficacy, sustainability, reliability, and resilience of a simple Thermal Energy Storage (TES) platform that utilized Phase Change Materials (PCM) and leveraged the Cold Finger Technique (CFT). Experiments were conducted at constant values of power input while video monitoring and recording of the melting of PCM in a measuring cylinder. Temperature transients were recorded by three thermocouples immersed in the PCM (termed as “PCM thermocouples”). Three additional thermocouples were mounted on the outer surface of the cylinder (termed as “surface thermocouples”).

Each set of transient temperature data (for a chosen constant power input condition) was used for training the ANN algorithm. The training data set consisted of the transient temperature recorded by the three thermocouples for a particular constant power input condition (either the PCM thermocouples or the surface thermocouples). Experimental validation of the numerical predictions was performed for a chosen value of power input. By comparing with the experimental measurements, the predictions from the ANN model (trained on a different value of power input) were used to calculate the error in the predicted values. The ANN model was used to predict the time-remaining to reach a particular value of melt-fraction. To further reinforce the value of this model, linear prediction models were implemented (ranging from 0-40%, 0-50% and 0-60% melt-fraction). The predictions from each of these linear models were then projected all the way to 100% melt-fraction.

Overall, the predictions from the ANN model were observed to be more reliable (yet less accurate in a few situations) compared to that of the linear prediction models. Since overprediction in the predicted values can lead to catastrophic failure of the CFT strategies in TES, underprediction afforded by the ANN models are therefore more reliable. This demonstrated the efficacy of the chosen ANN model for implementing CFT reliably and precisely for TES platforms that utilize PCMs. This study also demonstrates that surface thermocouple measurements can be used for training the ANN models for easier deployment and more reliable predictions (provided that the training data is obtained at a higher power input condition).