Eur. Phys. J. Plus (2022) 137: 474
https://doi.org/10.1140/epjp/s13360-022-02666-y

Regular Article

Unleashing deep neural network full potential for solar radiation forecasting in a new geographic location with historical data scarcity: a transfer learning approach

¹ Mechanical Design and Production Engineering Department, Faculty of Engineering, Cairo University, 12613, Giza, Egypt
² Mechanical Power Engineering Department, Faculty of Engineering, Cairo University, 12613, Giza, Egypt

^b bassem.akoush@eng.cu.edu.eg
^d mhd.zidan17@cu.edu.eg

Received: 17 January 2022
Accepted: 30 March 2022
Published online: 17 April 2022

Abstract

More grid-connected solar thermal and photovoltaic power plants are coming online every year, which necessitates precise irradiance forecasters for plant management. A major share of these plants is installed at remote sites with no long-term meteorological records to develop precise site-specific global horizontal irradiance (GHI) forecasters, and the process of prior collection of those records is both expensive and time-consuming. This study proposes a unique transfer learning approach for training one-hour-ahead forecasters of hourly average GHI at new locations with limited data records by refining trained recurrent neural network-based models at other locations with abundant data. The methodology is demonstrated by considering Cairo (Egypt) as the source site and the trained models are tuned using limited datasets from five other locations in Tunisia, Morocco, and Jordan. The approach was found valid for all targeted locations, with mean absolute errors (MAEs) and root mean square errors (RMSEs) lower than 46.4 and 73.4 W/m², respectively, and coefficients of determination (R²) higher than 93.7. The proposed models also outperformed two baseline persistence models by reducing MAEs and RMSEs by at least 34 and 44%, respectively, while holding more uniformly distributed and less clearness index-dependent residuals. The size of the source site’s dataset and the type of input predictors were, respectively, found as the least and most influential parameters on the models’ performances. This approach makes GHI forecasting accessible to practitioners and enhances the control of power plants since their start date.