Machine learning modeling of reversible thermochemical reactions applicable in energy storage systems

Abstract

BackgroundHeat/cooling generation through reversible thermochemical reactions provides high potential for application in thermochemical energy storage systems. However, analyzing these systems through traditional physics-based models and experimental studies imposes high financial and computational costs. MethodIn the present study, by using five powerful machine learning algorithms, including k-nearest neighbors (KNN), least absolute shrinkage and selection operator (LASSO), extreme gradient boosting (XGBoost), support vector machines (SVM) and Bayesian ridge accurate models are developed to predict the dynamic behavior of a reversible thermochemical reactions based system. This research uses five machine-learning algorithms to reach the best formula and model for the NH4NO3 and the KOH. Using the Bayesian ridge and the KNN, respectively, create the formula and make the model for the NH4NO3 by KOH and water. FindingsUsing the polynomial regression, degree 4 for the NH4NO3 temperature by the KOH and water temperatures can be reached as the best R-Squared, 0.938, and MSE is 3.898504. This dataset is time series, so using the XGBoost (eXtreme Gradient Boosting) can be computed the future of NH4NO3 temperature, which is done in this study. The KNN (K = 5) is suitable for making the model by the polynomial regression degree 2, the R-Squared is 0.999, and MSE is 0.009042. The Bayesian ridge has the best R-Squared for creating the formula for the KOH temperature by the NH4NO3 and water temperatures; the R-Squared is 0.949, and MSE is 16.900704. The R-Squared of the KNN (K = 5) degree 3 is 0.984, and MSE is 5.355852, which has the best value among all methods for making a model. The KOH has a time series dataset, so using the XGBoost can be computed the future of the KOH temperature.