Machine learning models to quantify the influence of PM10 aerosol concentration on global solar radiation prediction in South Africa

Abstract

Solar prediction models are essential in developing countries such as South Africa as most meteorological stations are unable to consistently measure this quantity. Quantification of solar radiation is of great significance for the adoption and application of renewable energy systems. Machine learning models are infamous for their high prediction capacity and low input requirements. This study's foremost intention was to investigate the efficacy of using the most widespread machine learning techniques for solar radiation estimation across South Africa, by introducing PM10 air pollutant concentration to generalized, readily available meteorological datasets. While assessing the performance of various input parameter configurations, the techniques which were evaluated include; Artificial Neural Network (ANN), Support Vector Regression (SVR), General Regression Neural Network (GRNN), and Random Forest (RF), of which ANNs proved the most appropriate for predicting global solar radiation as indicated by the high correlation coefficients (R2 = 0.99852) and low prediction errors (RMSE = 0.22627). The results described in this work indicate that machine learning models performed excellently for hybrid models as opposed to empirical models established in South Africa. ANN models which include PM10 concentration data profoundly improved the performance of relative humidity models and reduced overall error measures for models of various input parameter configurations. In addition, a poor correlation between PM10 concentration and air temperature was observed. This work suggests that the use of generalized, hybrid ANN models to predict solar radiation across South Africa are more functional than empirical modeling and this is indicated by the high prediction accuracy and low computational effort. The suggested model suffices as an accurate prediction model which will allow for a holistic understanding of the solar capacity available across this country while encouraging the implementation and investigation of sustainable, renewable energy technologies.

Topics & Concepts