Comparative Machine Learning Bias Correction of Satellite-Derived Solar Irradiance at Tropical Stations in Central Java, Indonesia
Abstract
Satellite-derived Global Horizontal Irradiance products carry systematic bias in tropical regions, so site-level correction is a prerequisite for their use where ground networks are sparse. This study compares bias-correction methods for NASA POWER daily irradiance in Central Java, Indonesia, and asks how much reported accuracy reflects algebraic dependence between predictors and the ground reference. Ten PUSDATARU stations were analysed over 2020–2024 (17,072 station-days; 8,459 from five exactly co-located stations). Random Forest, XGBoost, LightGBM, Elastic Net, and Quantile Mapping were trained chronologically (2020–2022/2023/2024) against persistence and climatology baselines, with preprocessing fitted on training data only, paired significance tests, leave-one-station-out validation, and SHAP attribution. Random Forest reduced RMSE from 3.398 to 2.674 MJ m⁻² day⁻¹ (21.3%), lowered mean bias from 0.988 to 0.221, and raised R² from 0.266 to 0.546; across ten stations XGBoost reduced error from 5.921 to 2.831. Differences among tree ensembles were not statistically robust, and the weighted ensemble gave no gain at the highest inference cost. Adding ground-observed sunshine as a predictor raised the coefficient of determination to 0.9994, identified as recovery of the Ångström–Prescott expression defining the reference (70.9% attribution). Near-perfect accuracies against sunshine-derived references therefore measure leakage, not skill; correction also failed to transfer spatially, motivating sub-degree geostationary products as the next step.
Related
- Offline Approximate Value Iteration: A Reinforcement Learning Approach for Satellite Constellation Deployment
- Learning-Based Spectrum Cartography in Low Earth Orbit Satellite Networks: A Survey
- A Sensor-Adaptive Incremental Learning Framework for Artifact Detection in Satellite Precipitation Data
- Adaptive Beam Hopping and Power Control for Dual-Layer Over-the-Air Online Federated Learning in LEO Satellite Networks
- A Prediction--Correction Analysis of Two-Way Block Splitting in Distributed Learning
BibTeX
@article{riyanto2026comparative,
title = {{Comparative Machine Learning Bias Correction of Satellite-Derived Solar Irradiance at Tropical Stations in Central Java, Indonesia}},
author = {Eko Riyanto and Iwan Setiawan and Uky Yudatama},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1778},
url = {https://doi.org/10.63158/journalisi.v8i5.1778},
note = {CC BY},
}
RIS
TY - JOUR
TI - Comparative Machine Learning Bias Correction of Satellite-Derived Solar Irradiance at Tropical Stations in Central Java, Indonesia
AU - Eko Riyanto
AU - Iwan Setiawan
AU - Uky Yudatama
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1778
UR - https://doi.org/10.63158/journalisi.v8i5.1778
AB - Satellite-derived Global Horizontal Irradiance products carry systematic bias in tropical regions, so site-level correction is a prerequisite for their use where ground networks are sparse. This study compares bias-correction methods for NASA POWER daily irradiance in Central Java, Indonesia, and asks how much reported accuracy reflects algebraic dependence between predictors and the ground reference. Ten PUSDATARU stations were analysed over 2020–2024 (17,072 station-days; 8,459 from five exactly co-located stations). Random Forest, XGBoost, LightGBM, Elastic Net, and Quantile Mapping were trained chronologically (2020–2022/2023/2024) against persistence and climatology baselines, with preprocessing fitted on training data only, paired significance tests, leave-one-station-out validation, and SHAP attribution. Random Forest reduced RMSE from 3.398 to 2.674 MJ m⁻² day⁻¹ (21.3%), lowered mean bias from 0.988 to 0.221, and raised R² from 0.266 to 0.546; across ten stations XGBoost reduced error from 5.921 to 2.831. Differences among tree ensembles were not statistically robust, and the weighted ensemble gave no gain at the highest inference cost. Adding ground-observed sunshine as a predictor raised the coefficient of determination to 0.9994, identified as recovery of the Ångström–Prescott expression defining the reference (70.9% attribution). Near-perfect accuracies against sunshine-derived references therefore measure leakage, not skill; correction also failed to transfer spatially, motivating sub-degree geostationary products as the next step.
SN - 2656-5935
SN - 2656-4882
ER -