Abstract
The fact that South Africa still relies on coal-fired power generation leads to greenhouse gases emissions and the intermittent shortages of power that result in rotating power outages. With the growth of solar photovoltaic systems as an alternative source of energy, the intermittency of these systems poses some uncertainty in the generation of power. Precise short-term prediction of the solar PV performance is thus a key element in enhancing energy management, operational efficiency and grid reliability.
This study set out to create and comparatively evaluate machine learning models in short-term solar PV power prediction. An integrated research design was followed incorporating a systematic literature review to determine socio-economic benefits and performance-affecting factors and quantitative modelling with the help of secondary weather data and solar generation. Pre-processing of the data and feature engineering, as well as exploratory analysis, were done before developing the model. Three regression methods; Time Series, Random Forest and Support Vector Regression were trained and measured in terms of standard error and goodness-of-fit.
According to the literature review, precise forecasting will improve energy management, minimize the cost of operation, and ensure greater reliability of the system. Empirical evidence indicate that the Random Forest model was more successful than the Time Series and Support Vector Regression models in terms of predictive performance and strength whereas the Time Series model was moderately successful and the Support Vector Regression model the least successful.
These findings show that ensemble learning methods are appropriate in short-term solar power prediction, and support the significance of proper model selection in renewable energy. A better forecasting accuracy can help in a more credible integration of solar energy in the dynamic energy mix of South Africa.