Hiba Qasim Mohammed
Day-ahead forecasting of global horizontal irradiance (GHI) is required for photovoltaic planning, operational scheduling, and grid integration, especially in countries where strong solar resources coexist with marked climatic heterogeneity. In Iraq, the southern, central, and northern regions differ substantially in temperature, humidity, rainfall, wind, and seasonality, and these contrasts are expected to affect how forecasting models behave and transfer between regions. In this study, an explainable forecasting framework was developed for Basra, Najaf, and Erbil using daily NASA POWER solar and meteorological data for 2020–2025. Next-day GHI was defined as the target variable, and a leakage-controlled predictor set was constructed from solar, meteorological, temporal, lagged, and rolling features. Six machine-learning models were evaluated against a persistence benchmark: linear regression, random forest, support vector regression, XG Boost, Light GBM, and a feed-forward artificial neural network. The lowest test error in 2025 was obtained by linear regression (MAE = 0.486 kWh/m²/day, RMSE = 0.723 kWh/m²/day, MAPE = 14.974%, R² = 0.860), although its advantage over XG Boost was not statistically significant. Similar behavior was confirmed by walk-forward testing, which indicated that forecast quality was governed more by feature design than by model complexity. The lowest errors were obtained in Najaf and Basra, whereas the highest errors were observed in Erbil, where stronger seasonality and higher precipitation were present. The explainability analysis further indicated that same-day GHI, lagged and rolling irradiance terms, seasonal indicators, surface pressure, temperature, and humidity contributed most strongly to the forecasts. It was therefore demonstrated that, once temporally consistent features were constructed without leakage, a simple transparent model could perform competitively with more complex ensemble approaches across the main climatic regions of Iraq.