Forecasting Hourly Solar Radiation Using Artificial Intelligence Techniques

Abstract

Recently, photovoltaic (PV) cell technology has witnessed significant developments and improvements. However, solar power is still unattractive to some consumers due to its unpredictability. Therefore, accurate prediction of solar irradiance continues to be critical for stable solar power. In this article, a proposed methodology to improve PV systems' reliability and stability, using state-of-the-art machine learning techniques, is introduced. Time-series long short-term memory (LSTM) network, convolutional LSTM (ConvLSTM), convolutional neural network (CNN), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) regression models were used to predict the city of Johannesburg's solar irradiance. These models were separately trained using the historical meteorological data of Johannesburg. Using the normalized root mean square error (nRMSE) as a performance indicator, the ConvLSTM model outperformed all the other models. When the models were individually trained with 80% of one-year input data, the results obtained showed that ConvLSTM recorded an nRMSE value of 4.05%. Results obtained using LSTM, XGBoost, and SVM models came next with nRMSE values of 4.72%, 6.63%, and 7.8%, respectively. The RF model gave an nRMSE value of 19.8%, whereas the CNN time-series model recorded an nRMSE value of 12.61%. The best experimental result obtained was an nRMSE value of 1.51% when the ConvLSTM model was trained at 500 epochs using 80% of ten years of actual historical solar irradiance as the dataset. These results suggest that the adoption of the ConvLSTM model for solar irradiance forecasts by solar farm operators in South Africa could improve electricity generation stability for PV connection to the power grid.