Discharging performance prediction of experimentally tested sorption heat storage materials with machine learning method

Abstract

In this study, the usability of the machine learning method in predicting the discharge performance of experimentally tested sorption heat storage materials was investigated. Experimental data was obtained from a lab scale fixed-bed thermochemical heat storage unit. 9 candidate composites were tested under different inlet conditions. Based on the experimental data, moisture sorption rates, heat output, exergy output and energy storage densities were determined. For the 6 cycles testing, highest average heat and exergy output were obtained with vermiculite/LiCl composite with the values of 0.83 kW and 0.013 kW, respectively. On the other hand, P-CaCl2 was found as the most durable material in terms of energy storage density (296 → 209 kWh/m3). A multilayer perceptron artificial neural network was established to evaluate measured data and its prediction performance was extensively studied. In the model 54 experimental data sets were utilized, consisting of 6 cycles testing of 9 different composite sorbents. Levenberg-Marquardt algorithm was benefited as the training one in the artificial neural network model established and the Tan-Sig and Purelin functions were selected as the transfer one in the multilayer neural network with 7 neurons in the hidden layer. According to the mathematical definition of the discussed statistical metrics, experimental data were used to compare them to the predicted output in order to verify the reliability of the proposed ANN model; and the analysis of the model was performed by examining the coefficient of determination, mean squared error, and deviation values, which were assumed as performance parameters, in detail. The deviation rate between the prediction values acquired from the artificial neural network and the practical data was determined as less than ±5 %. The acquired findings showed that artificial neural networks, which is one of the common machine learning algorithms, is a preferable method that can be employed to estimate the discharge performance of sorption heat storage materials.