Enhanced accuracy in solar irradiance forecasting through machine learning stack-based ensemble approach
Muhammad Sabir Naveed, Hafiz M.N. Iqbal, Muhammad Fainan Hanif, Jianli Xiao, Xianhua Liu, Jianchun Mi
Abstract
Accurate solar irradiance (SI) prediction is vital for optimizing solar photovoltaic systems. This study addresses shortcomings in existing forecasting methods by exploring advanced machine-learning techniques using meteorological satellite data. We develop three novel models for SI forecasting: Stack-based Ensemble Fusion with Meta-Neural Network (SEFMNN), Extreme Gradient Boosting-Squared Error (XGB-SE), and Extreme Learning Machine (ELM). These models predict All-sky and Clear-sky shortwave solar irradiance across three Chinese provinces (Guangdong, Shandong, and Zhejiang) and one Saudi Arabian province (Najran). The SEFMNN model combines Artificial Neural Network (ANN), Random Forest (RF), and Support Vector Machine (SVM) to improve prediction accuracy. The XGB-SE model employs a specialized loss function to manage extreme values in historical data. These models are designed to mitigate overfitting and data inconsistency while balancing computational efficiency and predictive accuracy. Comparative analysis reveals that SEFMNN and XGB-SE outperform the ELM model, with SEFMNN achieving an R2 of 0.9979, MAE of 0.0231, and MSE of 0.0020 in Najran. This demonstrates that SEFMNN significantly enhances solar irradiance forecasting, aiding efficient solar photovoltaic system planning and operation.