Machine Learning Approach for Short- and Long-Term Global Solar Irradiance Prediction

Solar radiation data forecasting algorithms are important, especially in developing countries, as vast solar power plants cannot measure reliable and constant solar irradiance. The challenges of solar irradiance prediction may be resolved by machine learning using weather datasets. This study emphasises the daily and monthly global solar radiation data predictions of three locations, Pretoria, Bloemfontein, and Vuwani, at different provinces in South Africa with various solar radiation distributions. The study evaluated five different machine learning models. Forecasting models were established to evaluate global solar radiation, focusing on input data. The selected forecast models are centered on their ability to perform with time series data. These models use five years of data from meteorological parameters, such as global horizontal irradiance (GHI), relative humidity, wind speed and ambient temperature between 1 January 2018 and 31 December 2022. The datasets from these meteorological parameters are utilised for training and testing the employed algorithms, which are examined using five statistical metrics. Moreover, the inconsistency of the solar irradiance time series was equally assessed using the clearness index. The results from this study demonstrate that the R2 value recording 0.866 datasets in Bloemfontein of random forest algorithm presents the highest performance during the training processes for all models studied, while the random tree in Vuwani showed the lowest performance of R2 of 0.210 with other algorithms in testing processes. Additionally, the maximum solar radiation was found in December for both Pretoria and Bloemfontein, recorded as 5.347 and 5.844 kWh/m2/day, respectively, while it was 4.692 kWh/m2/day at Vuwani in January. Similarly, the average clearness index of 0.605, 0.657 and 0.533 are obtained at Pretoria, Bloemfontein, and Vuwani, respectively. Among the three sites under study, the solar radiation and clearness index are higher in Bloemfontein. Therefore, the proposed algorithms could be used conveniently for short- and long-term solar power plants in South Africa.