A prediction model for a concentrating solar thermoelectric generator using artificial neural networks and extreme learning machines

Abstract

Abstract The current numerical simulation tools used to optimize the performance of concentrating solar thermoelectric generators are extremely time consuming, and consequently require expensive computational energies. Furthermore, they are incapable of considering the effects of diverse real-life operating conditions on the performance of the system. Additionally, they sometimes neglect temperature dependency in the thermoelectric semiconductors and base their studies on just unicouple thermoelectric cells to avoid the further complexity of the numerical computation. These factors limit the flexibility of optimization studies that can be conducted on solar thermoelectrics; hence, limiting the insights that can be drawn to design high performing solar thermoelectric generators. This work is the first of its kind to introduce artificial neural networks and extreme learning machines as a substitute to these numerical methods to accelerate and ease the design process of solar thermoelectric generators. The data generation process is conducted using a 3-dimensional numerical model developed in ANSYS numerical solver and the optimized parameters include the high-temperature material content, semiconductor height and area, concentrated solar irradiance, cooling film coefficient, wind speed, and ambient temperature – on the system performance. A full-scale customized thermoelectric module comprising 127 thermocouples is designed and integrated in an optical concentrator for solar power generation while considering temperature dependency in all thermoelectric materials. Results depict that the geometry and operating condition optimization improved the system power and efficiency by 42.02% and 82.23%, respectively. Furthermore, the artificial neural network had the highest regression of 95.82% with the least mean squared error of 2.71 \(\times\) 10− 5 in learning the numerical-generated data set while performing 389 and 203 times faster than the numerical method in forecasting the system power and efficiency, respectively. Finally, methods of manufacturing the optimized thermoelectric module using 3-dimensional printing are discussed.