Artificial neural network for predicting annual output energy of building-integrated photovoltaics based on the 2-terminal perovskite/silicon tandem cells under realistic conditions

A high-efficiency perovskite/silicon tandem solar cell has been a promising candidate for building-integrated photovoltaics. This work presents a deep learning approach to predicting the annual output energy harvested by the 2-terminal perovskite/silicon tandem solar cells, thereby optimizing the tandem structure design. The data set for training and validating an artificial neural network (ANN) is the Atlas-simulated results of the tandem cell with the various perovskite layer's bandgap and thickness under the real-world conditions composed of the solar spectrum, incident spectral angle, and solar module temperature in a particular month of a year for a specific direction. Consequently, we reveal the significant influence of solar spectral shape on the ANN performance. The proposed ANN model has a mean square error of 1.26 and a correlation coefficient of 0.99979. Based on the spectral and environmental database in Gifu (Japan) in 2015, we predict that the optimal perovskite layer's bandgap and thickness are 1.72 eV and 680 nm for the east, south, and west facades, 1.73 eV and 700 nm for the rooftop, respectively. Consequently, the highest annual output energy obtained is 282.54, 105.07, 174.71, and 90.79 kWh/m2 for the rooftop, the east, south, and west facades, respectively.