Eltahir et al. (2026) Horizon-Dependent Solar Irradiance Forecasting with Boosted Trees, and Seasonal Baselines Based on Measurements in Sudan

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Sensors
• Year: 2026
• Date: 2026-09-11
• Authors: E.I. Eltahir, Devrim Akgün, Sohaib Ashri, Ahmad M. Khachan, Elfatih A. A. Elsheikh, Ceyda Aksoy Tırmıkçı
• DOI: 10.3390/s26185778

Research Groups

Not specified

Short Summary

The study proposes a leakage-safe, horizon-specific forecasting framework for solar irradiance, demonstrating that boosted trees are most effective for short-term predictions while seasonal combinations prevail at longer lead times.

Objective

• To develop a reproducible and leakage-safe benchmarking framework for solar irradiance forecasting across multiple prediction horizons to avoid overstating the performance of complex models.

Study Configuration

• Spatial Scale: Not specified
• Temporal Scale: Prediction horizons of 10 min, 1 h, 3 h, 6 h, 12 h, and 24 h

Methodology and Data

• Models used: XGBoost, CatBoost, validation-weighted convex forecast combinations, and seasonal reference models.
• Data sources: Recent observations and meteorological variables (specific dataset not named in text).
• Validation Approach: Expanding-window validation and a predefined canonical test support to prevent data leakage.

Main Results

• 10-minute horizon: XGBoost achieved the lowest RMSE (10.78 W/m²), with statistically significant improvement over reference models.
• 1-hour horizon: CatBoost achieved the lowest RMSE (16.08 W/m²).
• 3, 12, and 24-hour horizons: Seasonally guided forecast combinations provided the lowest RMSE.
• 6-hour horizon: XGBoost and the two-day seasonal-mean combination ranked first by RMSE, though the gain over the seasonal mean resulted in a higher MAE.
• Statistical Significance: Improvements over seasonal references at horizons longer than 10 minutes were not statistically significant after adjustment.

Contributions

• Provides a reproducible, leakage-safe framework for horizon-dependent solar irradiance forecasting.
• Highlights the risk of overestimating complex model performance due to weak baselines or unreliable validation.
• Establishes a clear performance boundary where seasonal patterns outweigh complex machine learning models as the prediction horizon increases.

Funding

Not specified

Citation

@article{Eltahir2026HorizonDependent,
  author = {Eltahir, E.I. and Akgün, Devrim and Ashri, Sohaib and Khachan, Ahmad M. and Elsheikh, Elfatih A. A. and Tırmıkçı, Ceyda Aksoy},
  title = {Horizon-Dependent Solar Irradiance Forecasting with Boosted Trees, and Seasonal Baselines Based on Measurements in Sudan},
  journal = {Sensors},
  year = {2026},
  doi = {10.3390/s26185778},
  url = {https://doi.org/10.3390/s26185778}
}