Efficient Machine Learning-Based Prediction of Solar Irradiance Using Multi-Site Data
Hassan N. Noura, Zaid Allal, Ola Salman, Khaled Chahine
Photovoltaic panels have become a promising solution for generating renewable energy and reducing our reliance on fossil fuels by capturing solar energy and converting it into electricity. The effectiveness of this conversion depends on several factors, such as the quality of the solar panels and the amount of solar radiation received in a specific region. This makes accurate solar irradiance forecasting essential for planning and managing efficient solar power systems. This study examines the application of machine learning (ML) models for accurately predicting global horizontal irradiance (GHI) using a three-year dataset from six distinct photovoltaic stations: NELHA, ULL, HSU, RaZON+, UNLV, and NWTC. The primary aim is to identify optimal shared features for GHI prediction across multiple sites using a 30 min time shift based on autocorrelation analysis. Key features identified for accurate GHI prediction include direct normal irradiance (DNI), diffuse horizontal irradiance (DHI), and solar panel temperatures. The predictions were performed using tree-based algorithms and ensemble learners, achieving R2 values exceeding 95% at most stations, with NWTC reaching 99%. Gradient Boosting Regression (GBR) performed best at NELHA, NWTC, and RaZON, while Multi-Layer Perceptron (MLP) excelled at ULL and UNLV. CatBoost was optimal for HSU. The impact of time-shifting values on performance was also examined, revealing that larger shifts led to performance deterioration, though MLP performed well under these conditions. The study further proposes a stacking ensemble approach to enhance model generalizability, integrating the strengths of various models for more robust GHI prediction.