Comparative data-driven modeling of thermal energy storage using artificial neural networks and multiple linear regression

Efficient thermal energy storage (TES) is essential for enhancing the reliability and sustainability of solar thermal systems, particularly under fluctuating solar radiation conditions. This study investigates the predictive performance of different data-driven approaches—Multiple Linear Regression (MLR), Multilayer Perceptron (MLP), and Radial Basis Function (RBF) neural networks—for estimating stored thermal energy in a solar thermal tank. Experimental data were obtained from a controlled solar simulator setup that incorporated halogen-lamp irradiation, a flat-plate collector, and nanoparticle-enhanced water as the storage medium. Eight independent variables—including collector inlet and outlet temperatures, tank and ambient temperatures, flow rate, solar radiation, nanoparticle concentration, and specific heat capacity—were used as model inputs. Results show that the MLP model significantly outperformed both MLR and RBF, achieving the highest correlation coefficient (R = 0.647), and the lowest RMSE (346.35) and MBE (152.49), demonstrating superior accuracy and generalization. By contrast, MLR exhibited limited predictive power due to its linear assumptions, while RBF suffered from high testing error and poor generalization. These findings underscore the suitability of neural network models, particularly MLP, for capturing the nonlinear dynamics of TES systems, providing a robust framework for system optimization and improved energy management strategies.