Explainable short-term solar irradiance forecasting across diverse climates for consumer-centric energy management

The rapid growth of solar energy integration in modern power systems has increased the demand for reliable short-term forecasting to support efficient operation and energy management. Accurate one-hour-ahead prediction of global horizontal irradiance (GHI) is particularly important for applications such as load scheduling, energy storage management, and grid coordination. This study develops an interpretable machine-learning framework for one-hour-ahead GHI forecasting under different climatic conditions within India. Hourly solar irradiance and meteorological data from the NASA Prediction of Worldwide Energy Resources (NASA POWER) database were considered for four geographically and climatically distinct locations: Chennai, Guwahati, Jaipur, and Shimla, covering the period from 2022 to 2025. The forecasting target was explicitly defined as the GHI value at the subsequent hour ( t + 1). A systematic feature-selection procedure combined Pearson correlation filtering with mutual-information ranking to reduce redundancy and retain the most informative predictors. The candidate predictor space was reduced from 41 numerical features to 26 after correlation filtering and subsequently to 15 predictors using mutual-information ranking. A time-aware modelling strategy was adopted, with observations from 2022–2024 used for training and the year 2025 reserved exclusively for independent testing. Six forecasting models, namely Ridge Regression, Support Vector Regression, Random Forest, XGBoost, LightGBM, and Long Short-Term Memory (LSTM), were evaluated using RMSE, MAE, R 2 , and sMAPE. The results show that the ensemble-based models provided the strongest overall performance on the unseen 2025 test period. XGBoost achieved the lowest test RMSE of 67.49 W/m 2 and the highest R 2 of 0.9400, while Random Forest achieved the lowest test MAE of 21.80 W/m 2 . LightGBM provided comparable performance, with a test RMSE of 67.75 W/m 2 and an R 2 of 0.9396. A feature-group ablation analysis further demonstrated the contribution of the selected predictor groups. Removal of the temporal–cyclical group produced the largest increase in RMSE, from 66.460 to 70.905 W/m 2 , corresponding to an increase of 4.445 W/m 2 , followed by the temperature and GHI lag–difference groups. The regional and seasonal analyses further indicate that the forecasting framework captures the characteristic diurnal behaviour of GHI across the four study locations and different seasonal conditions. Overall, the combination of physically informed feature engineering, systematic feature selection, comparative model evaluation, and feature-group ablation provides an interpretable framework for one-hour-ahead GHI forecasting and offers a data-driven basis for renewable-energy planning and energy-management applications.