Machine learning-assisted performance prediction for dual PV-TEG systems with a triangular-elliptical cooling channel

Abstract

Optimized thermal management of PV (photovoltaic) units is key to improving their energy conversion efficiency and maintaining stable operation. This study presents a novel lightweight cooling system featuring a triangular cavity with an inner elliptical section, integrated with two thermoelectric generator (TEG) units, to improve the performance of dual PV modules. Using finite element method, the channel's cooling performance is investigated by varying the flow Reynolds number (Re from 100 to 500), the height of the inner elliptical section (h from 0.01H to 0.3H), the radius of the elliptical curvature (r from 0.01H to 0.2H), and the width of the elliptical gap (l from 0.2H to 0.5H). Machine learning-based regression is used to predict cooling performance of the triangular channel with inner elliptic part. The top-performing model is then combined with the PV-TEG unit to forecast the performance of the dual PV-TEG system. The most pronounced enhancement in cooling is observed when varying the flow Reynolds number, followed by changes in the height and radius of the elliptical curvature. The width of the inner elliptical section has the minimal effect on cooling performance. The optimal cooling performance is obtained with geometric parameters (h, r, l) = (0.3H, 0.2H, 0.5H), whereas the absence of the inner elliptical section leads to the poorest cooling. With the best cooling configuration, PV-cell temperatures decrease by 7.4 degrees C for PV-r and 3.1 degrees C for PV-l compared to the worst-case. The temperature difference between PV-r and PV-l is 4.8 degrees C under the worst cooling conditions, but reduces to 0.5 degrees C in the optimal scenario. Feature importance analysis showed that the Reynolds number has the dominant effect on heat transfer, explaining 57%-78% of the observed variance, whereas geometric parameters have a relatively minor impact. Cross-validation further confirmed that support vector regression (SVR) provides the most consistent generalization performance, highlighting its reliability for practical applications. The optimal configuration reduces the channel volume by 35.6% compared to a design without the inner elliptical section.