Accurate prophecy of photovltaic-segmented thermoelectric generator's performance using a neural network that feeds on finite element-generated data. (December 2022)

• Record Type: Journal Article
• Authors: Maduabuchi, Chika; Alanazi, Mohana; Alzahmi, Ahmed
• Is Part Of: Sustainable energy, grids and networks. Volume 32(2022)
• Publication Date: 2022-12
• Subjects: Solar energy -- Segmented photovoltaic–thermoelectric -- Geometry optimization -- Transient analysis -- Finite element method -- Bayesian regularized artificial neural networks; Renewable energy sources -- Periodicals; Smart power grids -- Periodicals; Electric power systems -- Periodicals; 333.794
• DOI: 10.1016/j.segan.2022.100905
• Languages: English
• ISSNs: 2352-4677

Abstract

Abstract: To further enhance the photovoltaic–thermoelectric system efficiency, this paper proposes a new hybrid system design comprising a segmented thermoelectric generator and aluminum heat sink directly lapped to the back plate of a photovoltaic cell operating under Nigerian transient and fluctuating weather conditions. The performance evaluation and optimization of the hybrid system design is conducted using a numerical model setup in ANSYS software and the optimized parameters include the thermoelectric leg height and cross-sectional area, skutterudite content, ceramic height, fin and fin base heights, convective film coefficient, and solar concentration ratio while the performance indices are the system power generation rate and the system efficiency. Finally, a Bayesian regularized artificial neural network with 10 neurons in the hidden layer is proposed to overcome the lengthy computational time and energy needed to conduct the numerical-inspired system optimization. Results are that the proposed system design improved the efficiency of the conventional photovoltaic cell integrated with regular unsegmented thermoelectric generators by 96% at a solar concentration of 5. Additionally, the numerical-inspired optimization was able to improve the conventional system power output and efficiency by 45.7% and 45.9%, respectively, compared to the unoptimized system when operated under peak sunshine conditions. Finally, the Bayesian regularized neural network with a low mean squared error of 9.7 × 10 −14 after 471 iterations and perfect regression correlations for training, testing, and all processes provided a perfect fit of the numerical-generated data, being 201 times faster than the conventional numerical optimization scheme. Graphical abstract: Highlights: PV-STEG is proposed to improve the efficiency of traditional PV–TEGs. ANNs are used to overcome the limitations of FEMs in analyzing PV-STEG. PV-STEG is 97% more efficient than traditional PV–TEG. Bayesian regularized ANN is 201 times faster than FEM in optimizing PV-STEG. Nigeria transient weather conditions severely affects PV-STEG performance.