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

Authors: KANG, D - Seoul University; KIM, K S - Seoul University; KIM, J - Rural Development Administration - Korea; LEE, C K - Rural Development Administration - Korea; MARUYAMA, A - National Agriculture And Food Research Organization (NARO), Agricultrual Research Center; BERESFORD, R M - New Zealand Institute Of Plant & Food Research; Fleisher, David

Submitted to: Environmental Research Letters
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/6/2023
Publication Date: 9/28/2023
Citation: Kang, D., Kim, K., Kim, J., Lee, C., Maruyama, A., Beresford, R., Fleisher, D.H. 2023. Identification of a spatial distribution threshold for the development of a solar radiation model using deep neural networks. Environmental Research Letters. 18(10). https://doi.org/10.1088/1748-9326/acf6d4.
DOI: https://doi.org/10.1088/1748-9326/acf6d4

Interpretive Summary: Crop yield at a particular field is highly dependent on weather characteristics, including solar radiation. Having access to high quality solar radiation data is very important because it helps farmers and scientists estimate how productive that land could be for growing different plants. However, solar radiation data is not usually available at most farm locations. A new method for estimating the quantity of solar radiation available at a particular agricultural location was developed to address this need. The method uses mathematical models and historical data records that were previously measured at different locations. This so-called Deep Solar Radiation (DSR) model was shown to have small errors in estimating solar radiation quantities. These errors tended to be small even when this DSR model was used to make predictions at locations far away from where the historical records were available. This means the model can successfully be used to provide growers, crop consultants, and farmers with accurate estimates of solar radiation amounts that can be used to evaluate field productivity. This model can also be used for those growers interested in evaluating the potential use of solar energy collection systems.

Technical Abstract: Empirical models have been used to estimate global solar radiation, which is a key variable to assess the productivity of photovoltaic generation and farmland for crop production. Spatial portability of simple empirical models is usually uncertain. Alternatively, a deep neural network (DNN) can be used to develop a solar radiation model with greater spatial portability than simple empirical approaches. The objectives of this study were to examine the hypothesis that the spatial portability of the neural network model would be affected by the spatial distribution of training sites and to develop the Deep Solar Radiation (DSR) model. The weather station network operated by the Korea Meteorological Administration (KMA) was used to collect weather data for training and validation of the DNNs. A gridded weather data product was also used to evaluate the DSR model in North-East Asia. Multiple sets of weather stations were selected for cross-validation using the standard distance deviation (SDD) to assess the impact of the spatial distribution among training sites on spatial portability of the DNNs. The DSR model was chosen to have the smallest error statistics among the DNNs obtained from the cross-validation. The spatial variation in the error statistics for the DNNs tended to decrease with increasing value of SDD for the training sites. The estimates of annual power generation (APG) for the DSR model had relatively small errors over large regions including China, Japan, Korea, and Mongolia, even though the neural network was trained using weather stations only in Korea. For example, the normalized value of the root mean square error for APG was < 0.05 and <0.1 in 31 and 63 percent in the region of interest, respectively. Our results demonstrated that assessment of the spatial distribution of weather stations can guide development of a deep neural network model for estimation of solar radiation with acceptable spatial portability.