Abstract:
Renewable energy forecasts are critical to renewable energy grids and backup
plans, operational plans and short-term power purchases. This dissertation
focused on forecasting solar irradiance at one radiometric station in South
Africa using high-frequency data obtained from the Vuwani radiometric station
(USAid Venda). The aim of this dissertation was to compare the predictive
performance of the Genetic Algorithm (GA), recurrent neural networks
(RNN) and k-nearest neighbour (KNN) models in forecasting short-term solar
irradiance where KNN is used as a benchmark model. From the results
it is discovered that the RNN is the best forecasting model in terms of the
relative mean absolute error (rMAE). The forecasts of the machine learning
algorithms combined using convex combination technique and quantile
regression averaging (QRA) found that QRA is the best model. Predictive
interval widths analysis with 95% level of confidence was performed and the
results showed that QRA over RNN is the best model for forecasting solar
irradiance when looking at the PICP and PANAW. The Diebold-Mariano
test discovered that the tests fall between the -1.96 and 1.96 range, which
tells us that it accepts the null hypothesis. The Murphy diagram presented
and showed the 95% pointwise confidence intervals. The study will have an
impact on the South African power utility decision-makers to align electricity
demand and its supply in an efficient way that promotes potential economic
growth and environmental sustainability.