Electrical efficiency of the photovoltaic/thermal collectors cooled by nanofluids: Machine learning simulation and optimization by evolutionary algorithm

Photovoltaic/thermal (PV/T) are high-tech devices to transform solar radiation into electrical and thermal energies. Nano-coolants are recently considered to enhance the efficiency of PV/T systems. There is no accurate model to predict/optimize the PV/T systems' electrical efficiency cooled by nano-coolants. Therefore, this research employs machine-learning approaches to simulate PV/T system electrical performance cooled by water-based nanofluids. The best topology of artificial neural networks, least-squares support vector regression, and adaptive neuro-fuzzy inference systems (ANFIS) are found by trial-and-error and statistical analyses. The ANFIS is found as the best method for simulation of the electrical performance of the considered solar system. This approach predicted 200 experimental datasets with the absolute average relative deviation (AARD) of 13.6%, mean squared error (MSE) of 2.548, and R-2 = 0.9534. Furthermore, the ANFIS model predicts a new external database containing 63 samples with the AARD=15.21%. The optimization stage confirms that 30 lit/hr of water-silica nano-coolant (3wt%, 12.5 nm) at radiation intensity of 788.285 W/m(2) is the condition that maximizes electrical efficiency. In this optimum condition, the enhancement in the PV/T electrical efficiency is 27.7%. Finally, the fabricated ANFIS model has been utilized for generating several pure simulation predictions that have never been published before. (C) 2021 The Authors. Published by Elsevier Ltd.

Automated summary

This study utilized machine-learning approaches to simulate the electrical performance of PV/T systems cooled by water-based nanofluids, finding the ANFIS as the most effective method. The optimized condition with 30 lit/hr of water-silica nano-coolant at a radiation intensity of 788.285 W/m(2) maximized electrical efficiency by 27.7%. The ANFIS model successfully predicted a large amount of experimental data and an external database with a low average relative deviation and high R-squared value.