Abstract
Solar irradiance forecasting is essential in renewable energy grids amongst others for back-up programming, operational planning, and short-term power purchases. This study focuses on forecasting hourly solar irradiance using data obtained from the Southern African Universities Radiometric Network at the University of Pretoria radiometric station. The study compares the predictive performance of long short-term memory (LSTM) networks, support vector regression and feed forward neural networks (FFNN) models for forecasting short-term solar irradiance. While all the models outperform principal component regression model, a benchmark model in this study, the FFNN yields the lowest mean absolute error and root mean square error on the testing set. Empirical results show that the FFNN model produces the most accurate forecasts based on mean absolute error and root mean square error. Forecast combination of machine learning models' forecasts is done using convex combination and quantile regression averaging (QRA). The predictive performance we found is statistically significant on the Diebold Mariano and Giacomini-White tests. Based on all the forecast accuracy measures used in this study including the statistical tests, QRA is found to be the best forecast combination method. QRA was also the best forecasting model compared with the stand-alone machine learning models. The median method for combining interval limits gives the best results on prediction interval widths analysis. This is the first application of LSTM on South African and African solar irradiance data to the best of our knowledge. This study has shown that providing adequate and detailed evaluation metrics, including statistical tests in forecasting gives more insight into the developed forecasting models.
Highlights
The growth of industrialisation globally is resulting in the depletion of fossil fuels which are used for electricity generation [1]
This paper compares the predictive performance of long shortterm memory (LSTM) networks, support vector regression (SVR), and feed forward neural networks (FFNN) model for forecasting short-term solar irradiance
3) Forecasts based on linear quantile regression averaging are unbiased, while those from feed forward neural networks and long short-term memory are all biased with more under-predictions compared to over-predictions
Summary
The growth of industrialisation globally is resulting in the depletion of fossil fuels which are used for electricity generation [1]. Gensler et al [19] in their study, discussed different ANN and deep neural network (DNN) architectures in the field of solar power forecasting Different models such as physical forecasting model, MLPNN, LSTM, deep belief networks (DBN) and Auto-Encoders were employed and compared. The results of the study showed that deep learning algorithms have better solar power forecasting performance compared to ANN and physical models on the data from 21 solar power plants in Germany. In this study, [23] used three methods, seasonal autoregressive fractionally integrated moving average (SARFIMA), harmonically Coupled SARIMA (HCSAFRIMA) and Regression model with SARFIMA error terms (SARFIMAX) to address the long-range dependence inherent in the solar irradiance data from three radiometric stations in South Africa. Results from this study showed that long memory is anti-persistent in all models except one that showed persistence
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
