Abstract
Solar energy is a clean, renewable energy source (RES) that is readily available and universally accessible. Despite the environmental consequences of fossil fuels, reliance on them for power generation has yet to decline significantly. Although recent advances in photovoltaic (PV) cell technology have significantly reduced the cost of solar panels, solar power remains unappealing to some consumers because of the fluctuating solar radiation. As a result, accurate prediction and analysis of Global Horizontal Irradiance (GHI) patterns for stable solar power generation remain critical for improved PV power output. This study presents nine state-of-the-art Artificial Intelligence (AI)-based models for predicting and analyzing the patterns of GHI in multiple locations in South Africa. The performance of various methods used for GHI prediction is forecasting horizon. The data used for the experiments for training the proposed Globa GHI prediction models are historical meteorological data for different cities and locations in South Africa, such as Johannesburg, Durban, Limpopo, and Cape Town. The first contribution of this work is using and comparing five deep learning algorithms and four machine learning algorithms with different dataset sizes to predict GHI over different horizons of 5-minute, 10-minute, 15-minute, 30-minute, and 60-minute. The results (outputs) were analyzed to determine the most accurate method. The comparison was made based on different performance measures using Coefficient of determination (R²), Root Mean Square Error (RMSE), and Normalized Root Mean Square Error (NRMSE) as performance evaluation metrics, Deep learning (DL) models such as the Convolutional Long Short-Term Memory (ConvLSTM) network, Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM) model often outperformed traditional machine learning models in this study's experiments. However, ensemble learning models such as Random Forest (RF) and Extreme Gradient Boosting (XGBoost) sometimes outperformed the deep learning models. When all models’ performance was compared, the DL models typically outperformed the traditional ML models. The best result obtained was an RMSE value of 6.99 at the 5-minute prediction horizon with the LSTM model trained with 80% of the ten years of the dataset collected from Limpopo.
Another contribution of this work is determining the best location for installing a solar PV plant. The prediction errors from the trained models with 80% of ten years of Limpopo data yielded the lowest errors in most prediction intervals. At those time horizons, the model's prediction accuracy was higher when trained with Limpopo data. Cape Town came next, then Durban, and finally Johannesburg. Based on these findings, it is suggested that if solar plants in South Africa apply
V
the LSTM deep learning network to their real-time simulations, they can take corrective actions in advance to counteract the effects of solar energy's stochastic nature. This approach may imply establishing a more stable and intelligent connection to the power grid.