Reliable and efficient solar radiation estimation with the insights of XAI.

The unpredictability of solar energy has led to reliability and integration problems that require costly and technically complex solutions in the electrical grid. Solar resource availability and energy generation are highly influenced by local climate variables, like atmospheric temperature, humidity, wind and pressure. If there is generational uncertainty, it is challenging to calculate economic criteria such as energy costs and returns, which impacts the feasibility study of a solar power plant. Also, it is very difficult and costly to maintain pyranometers at the locations. To forecast solar irradiance and the power generation at any location, machine learning (ML) techniques can be used. The present work deals with determining the influence of different parameters in predicting the solar radiation by ML models. A comparison study of six regression models: Ada Boosing Regressor (ABR), Gradient Boosting Regressor (GBR), Random Forest Regressor (RFR), Decision Tree Regressor (DTR), Linear Regression (LR), and Extreme Gradient Boosting Regressor (XGBR), shows that RFR gave the highest regression score of 0.9028. This also recorded the Mean Absolute Error (MAE) of 0.6198 and Mean Squared Error (MSE) of 1.348. The next best regression score produced by the Gradient Boosting Regressor(GBR) with value of 0.891. This is 1.18% lower than the RFR. For the RFR regression analysis, an Explainable AI (XAI) model used to interpret the results using Local Interpretable Model-agnostic Explanations (LIME) for local surrogacy and Shapely for global surrogacy. Both the LIME and Shapely interpretations shows that the parameter temperature has the highest correlation with the radiation. The paper would benefit from a more explicit statement of what is new compared to prior studies.