Identification of a spatial distribution threshold for the development of a solar radiation model using deep neural networks

Abstract

We propose an approach to develop a solar radiation model with spatial portability based on deep neural networks (DNNs). Weather station networks in South Korea between 33.5–37.9° N latitude were used to collect data for development and internal testing of the DNNs, respectively. Multiple sets of weather station data were selected for cross-validation of the DNNs by standard distance deviation (SDD) among training sites. The DNNs tended to have greater spatial portability when a threshold of spatial dispersion among training sites, e.g. 190 km of SDD, was met. The final formulation of the deep solar radiation (DSR) model was obtained from training sites associated with the threshold of SDD. The DSR model had RMSE values <4 MJ m ^−2 d ^−1 at external test sites in Japan that were within ±6° of the latitude boundary of the training sites. The relative difference between the outputs of crop yield simulations using observed versus estimated solar radiation inputs from the DSR model was about 4% at the test sites within the given boundary. These results indicate that the identification of the spatial dispersion threshold among training sites would aid the development of DNN models with reasonable spatial portability for estimation of solar radiation.

Keywords

• solar radiation
• spatial portability
• artificial intelligence
• empirical model
• crop model