Performance assessment of advanced data-intelligence models for solar radiation forecasting: a case study in a high solar potential region

Solar Radiation (SR) is of great importance in the fields of water engineering, energy engineering, and climate sustainability. By knowing the potential of solar energy, this energy can be converted into useful energy forms such as heat and electricity by determining the technical and economic suitability of technologies in a region. In this study, three different machine learning (ML) models, Long Short-Term Memory (LSTM) deep learning architecture, multivariate adaptive regression (MARS), and M5 model tree (M5Tree), were used for accurate prediction of SR with a few months ahead inputs. Additionally, the Kruskal-Wallis (KW) test was used to check the robustness of the model results in the study. The estimated SR data examined in the study are two important components of solar energy and these are daily Downward Shortwave Radiation (DSR) and Downward Longwave Radiation (DLR) data. The data was obtained from Satellite-based NASA/POWER for five Iraqi stations: Baghdad, Basrah, Duhok, Kirkuk, and Ramadi. The data covers the years 2013-2020. The input data were obtained by examining the autocorrelation between the lag-time data and using the month information, which is the periodicity component. As a result, among the models evaluated, the LSTM model yielded the most accurate forecasts for both Shortwave Downward Irradiance and Longwave Downward Irradiance, as indicated by lower RMSE and MAE values compared to MARS and M5Tree. Furthermore, the study demonstrated that incorporating periodicity significantly enhances the accuracy of SR estimates.