Enhancing solar irradiance prediction for sustainable energy solutions employing a hybrid machine learning model; improving hydrogen production through Photoelectrochemical device

Utilizing solar energy can meet global energy demands and mitigate the environmental impact of fossil fuels, particularly in combating global warming. Given the various factors influencing solar irradiation, accurate prediction of direct normal irradiance (DNI) is crucial. This study aims to create innovative hybrid forecasting models using a unique combination of techniques, including complete ensemble empirical decomposition with additive noise (CEEMDAN), sample entropy (SE) clustering, grey wolf optimizer (GWO), and support vector regression (SVR) algorithm. Data from Jiangsu province in China for four seasons are analyzed, and the SVR hybrid model outperforms the multi-layer perceptron (MLP) and long short-term memory (LSTM) models, earning its selection for development. The CEEMDAN-SE-GWO-SVR model exhibits the lowest error and highest efficiency, with average values for the four seasons' coefficient of determination (R2), root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE) at 0.97, 43.25, 3.35, and 25.80, respectively. Additionally, a simulation using a photoelectrochemical (PEC) device validates the reliability of the machine-learning method in generating hydrogen from solar irradiation employing one-day ahead prediction. During peak production hours, from 7:00 AM to 9:00 PM, the maximum hydrogen production rate of 57.5 μg/s is observed. These results support the proposal of the model for accurate direct normal irradiance prediction.