A Three-Stage Hierarchical Meta-Random Forest Stacking for One-Hour-Ahead Global Horizontal Irradiance Forecasting

Abstract

Accurate short-term solar irradiance forecasting is crucial for PV generation scheduling and energy management, however, hierarchical ensemble models can achieve optimistic performance if the temporal information flow is not maintained during meta-learning. This study presents a three-stage Meta-Random Forest (Meta-RF) stacking framework for one-hour-ahead global horizontal irradiance (GHI) forecasting. In the first stage, Lasso Regression, Ridge Regression, Support Vector Regression, and Decision Tree Regression are used; in the intermediate stage, Gradient Boosting Regression, XGBoost Regression, and AdaBoost Regression are employed; and in the final stage, Random Forest is used for meta-regression. A temporally ordered out-of-fold learning procedure is used to produce meta-features without direct propagation of in-sample fitted predictions across successive learning levels. The five years of hourly solar–meteorological data are divided chronologically into training (2016–2018), validation (2019), and independent testing (2020), where the hyperparameters are configured using the validation data. The complete Meta-RF framework outperforms the evaluated individual learners and intermediate stacking configurations in terms of MAE and RMSE, with an MAE of 1.13 W/m², RMSE of 3.41 W/m² and R² of 0.9998. Stage-wise analysis also shows the progressive improvement in forecasting with the hierarchical learning architecture, and measured–predicted characteristics and confidence-interval analysis provide evidence of consistency of the fitted relationship. The proposed framework is a temporally structured hierarchical learning approach for accurate one-hour-ahead GHI forecasting and short-term solar-energy applications.