A prediction model for the performance of solar photovoltaic-thermoelectric systems utilizing various semiconductors via optimal surrogate machine learning methods

This research focuses on finding the best surrogate performance prediction model for a solar photovoltaic-thermoelectric (PV-TE) module with different semiconductor materials. The need for a surrogate machine learning model arises due to the inefficiency of the current numerical simulations used to assess the performance of the device. The study introduces several surrogate machine learning models, such as recurrent, time delay, and regular artificial neural networks (ANNs), that are trained using expensive finite element generated data when the operating parameters of the system are altered. These parameters include the optical concentration ratio, cooling coefficient, wind speed, air temperature, glass emissivity, semiconductor dimensions, external load resistance, and thermoelectric current. Despite the time-intensive and costly data generation method, 714 datapoints were produced and utilized to train the surrogate machine learning models for improved performance prediction and optimization. The results indicate that the optimal machine learning model for solar PV-TE performance modelling was the ANN architecture with two hidden layers and five neurons per layer. Furthermore, lithium nitride oxide PV-TE showed a 65% improvement over the popular bismuth telluride PV-TE when tested under 25 Suns. The surrogate ANN also outperformed the conventional numerical simulations by 10,000 times using the same computing resources. Finally, the study suggests the potential of recurrent and time delay neural networks for modelling time series PV-TE data in the future.