Estimation of Global Horizontal Irradiance Using Machine Learning

Abstract

Estimation of Global Horizontal Irradiance Using Machine Learning newline newlineWith the rapid advancement of solar power, its integration into the smart grid faces challenges due to intermittency, making accurate Global Horizontal Irradiance (GHI) prediction crucial. This study proposes a feature selection method using variance inflation factor and mutual information (VIF-MI) to identify key meteorological variables. A novel stack ensemble with an extra tree (SE-ET) regressor is introduced, incorporating Decision Tree, Random Forest, LGBM, CatBoost, and XGBoost as base learners, with Ridge and Extra Tree regressors as meta learners. The SE-ET model significantly reduces prediction errors, improving MAE by 16-60%, RMSE by 7-64%, and MAPE by 20-75%. newlineThe second objective explores tree-based ensemble models for GHI estimation. Evaluating 18 ML regressors and 10 feature selection methods, the Extra Tree Regressor achieves the lowest MAE (0.9938) and RMSE (2.5327). Feature selection improves predictions, reducing MAE and RMSE by 13.15% and 5.24% using Mutual Information. Influential features such as solar zenith angle, diffuse horizontal irradiance, and direct normal irradiance are identified using SHAP values for better interpretability. newlineThe final objective develops time series forecasting models using past GHI data from 2017-2019 for four districts. Auto-SARIMA, Facebook Prophet, and Neural Prophet models are compared, with Auto-SARIMA achieving the lowest RMSE (10.70-21.62 W/m²) and MAE (9.222-19.31 W/m²), proving its superior forecasting ability. This study confirms that ensemble techniques and optimized models enhance solar energy planning by minimizing prediction errors and improving reliability. newline newline