Application of explainable machine learning for estimating direct and diffuse components of solar irradiance

Area of Science:

• Renewable Energy
• Atmospheric Science
• Machine Learning

Background:

• Accurate measurement of diffuse horizontal irradiance (DHI) and direct normal irradiance (DNI) is vital for solar energy applications.
• Current instruments often prioritize global horizontal irradiance (GHI) due to the high cost of DNI/DHI measurement devices.
• Existing solar decomposition models attempt to derive DNI and DHI from GHI, with varying success.

Purpose of the Study:

• To develop and evaluate a novel machine learning-based approach for separating direct and diffuse solar irradiance from GHI.
• To assess the performance of gradient boost machine learning models in solar irradiance decomposition.
• To identify key meteorological parameters influencing the accuracy of DNI and DHI estimation.

Main Methods:

• Utilized a novel separation approach employing machine learning algorithms with 1-minute temporal resolution data.
• Trained and compared three gradient boost machine learning models using data from 10 global stations with diverse climates.
• Applied Shapley Additive Explanations (SHAP) to interpret model parameters and their interactions.

Main Results:

• The CatBoost machine learning model significantly outperformed existing solar decomposition models across all tested stations.
• CatBoost achieved the lowest root mean squared error (RMSE) of 8.73% in direct normal irradiance (DNI) calculations.
• Humidity was identified as a crucial parameter for accurate estimation of both direct normal irradiance (DNI) and diffuse horizontal irradiance (DHI).

Conclusions:

• Machine learning, particularly the CatBoost model, offers a highly effective method for solar irradiance decomposition.
• The developed approach provides a cost-effective alternative for obtaining crucial DNI and DHI data for solar energy applications.
• Understanding the role of meteorological factors like humidity is essential for advancing solar energy resource assessment.