The paper builds a day-ahead solar-irradiance forecast of a given site in Jordan (32.5566° N, 35.8442° E) using the publicly reproducible NASA POWER day-to-day surface meteorology and solar-flux data. The work was implemented by learning calendar/seasonal encodings, physically inspired interactions, rolling statistics, and irradiance lags, and compares three learners that are trained using the same, time-ordered protocol feed-forward back-propagation (FFBP), a convolutional feed-forward network (CFBP; Conv1D + Dense), and ε-SVR with an RBF kernel. SVR has the highest precision on the held-out test block, with RMSE = 0.3799 kWh m–2 day–1, MAE = 0.2499, MAPE = 7.02%, nRMSE = 0.071, and R2 = 0.9651; FFBP and CFBP have greater errors and smaller values of R 2. These findings can be explained by diagnostic plots: SVR tracks time series peaks and troughs most closely and has the smallest parity scatter and error distribution whereas CFBP configuration suppresses high-frequency variability. A manual historical check using one month of context to predict 2014-02-01 reproduces the same ranking and shows SVR’s near-zero bias, indicating robust transfer to an unseen period. Overall, under the same engineered features and split, SVR provides the most reliable day-ahead irradiance signal for this site, with ANNs offering acceptable but less precise alternatives.
Shatnawi et al. (Mon,) studied this question.