Prediction of latent heat storage transient thermal performance for integrated solar combined cycle using machine learning techniques

Abstract

The objective of this work is to develop a prediction methodology for latent heat storage thermal performance to reduce computational efforts. Thermal storage can effectively compensate for the limitations of solar energy by extending operation time and leveling fluctuations. Although latent heat storage is advantageous compared to sensible heat storage for its large energy density and stable output temperature, its thermal performance has strong nonlinearity. With the conventional numerical simulations, it takes large computational time to obtain the one year or longer thermal performance of the latent heat storage. In this study, the exit steam enthalpy of latent heat storage for an integrated solar combined cycle (ISCC) is predicted using machine learning techniques. As latent heat storage is used to store solar thermal energy, the inlet steam properties such as enthalpy, pressure, and flow rate are continuously altered. Heat charged-discharged history in the regression model inputs significantly improves the prediction accuracy. The regression models are trained with transient heat transfer analysis results based on meteorological data of Bawean Island, Indonesia. Seven regression algorithms (Gaussian process regression (GPR), kernel approximation, ensemble of trees, support vector machines (SVM), artificial neural networks (ANN), linear regression, and regression trees) are compared and GPR is selected for its root mean square error (RMSE) of 16.9 kJ/kg or less. The RMSE values are equivalent to Bawean results even when the GPR is applied to two other locations: Kupang and Cikarang. Meanwhile, an ensemble classifier, which is selected from nine classification algorithms (Ensemble classifiers, ANN, SVM, decision trees, nearest neighbor classifiers, naive bayes classifiers, kernel approximation classifiers, discriminant analysis, and logistic regression classifiers), is used to determine when to terminate the latent heat storage operation. Owing to its short calculation time, the developed methodology is advantageous at the early design stages of latent heat storage, especially when multiple locations need to be compared.