Machine Learning Performance Prediction of a Solar Photovoltaic-Thermoelectric System with Various Crystalline Silicon Cell Types

Abstract

Since the solar photovoltaic-thermoelectric (PV-TE) is an upcoming technology, the current literature on PV-TEs have failed to thoroughly investigate the effects of different solar cell types on the PV-TE’s performance. Such parametric study becomes necessary since the properties of the solar cell, characterized by the cell temperature coefficient and reference efficiency, determine the overall performance of the PV-TE. Further, the design information obtained from numerical solvers is minimal due to the high time and computational energy required to optimize the photovoltaic-thermoelectric (PV-TE) performance. This problem severely affects the ease with which useful design information can be drawn from current methods to design high-performance PVs. Additionally, to reduce the complexity of the existing numerical model, the previous models have introduced some principal assumptions that have significantly affected the accuracy of the numerical results. Finally, the numerical model has neglected several real-life operating conditions since they increase the model complexity. For the first time, deep neural networks are proposed to predict the photovoltaic-thermoelectric performance designed with 3 different crystalline solar cells as a perfect replacement for the inefficient numerical methods used to analyze the hybrid system. After that, the data generated by the numerical solver is fed to an optimum 3-layer deep neural network with 20 neurons per layer to forecast the hybrid system’s performance efficiently. The training of the neural network is governed by Bayesian regularization and the performance of the model is evaluated using the mean squared error, coefficient of determination, and training time. Results show that the deep neural network accurately learned the results that took the conventional numerical solver 1,600 mins to generate in just 12 min and 27 s, making the proposed network 128.51 times faster than the traditional numerical method. Furthermore, the accuracy of the network is demonstrated by the low mean squared error of $3.34\times {10}^{-8}$ and high training and testing regression of 100% and 99.98%, respectively.