Accurate measurements of diffuse irradiance are essential to design a solar photovoltaic system. However, in-situ radiation measurements in Colombia, South America, can be limited by the costs of the implementation of meteorological stations equipped with a pyranometer mounted on a sun tracker with a shading device, which is required to measure diffuse irradiance. Furthermore, the databases found in Colombia contain missing data, which raises the need for implementing models that are trained with very few features. In this paper, we introduce a methodology based on simple angle calculations and a regression model to predict half-hourly diffuse horizontal solar irradiance from only the measure of global horizontal irradiance and a geographic coordinate as inputs. Using measurements taken from the national solar radiation database for 6 different sites in Colombia and state-of-the-art machine learning models for regression, we validated the accuracy prediction of the proposed methodology. The results showed a prediction error ranging from 5.86 to 9.36 [W/m 2], and a coefficient of determination ranging from 0.9974 to 0.9983. The data-set used along with the feature engineering process and the deep neural network model created can be found in a Github repository referenced in the paper.
Keywords: Clearness index; Diffuse horizontal irradiance; Gaussian processes; Machine learning; Random forest; Sun position angles.
© 2021 Published by Elsevier Ltd.