Incorporating advanced machine learning algorithms into solar power forecasting in off-grid hybrid renewable systems
Shiyi Tian, Xuechun Liu
Abstract
Accurate forecasting of Direct Normal Irradiance (DNI) is essential for the efficient planning and performance of off-grid renewable energy microgrids. This study introduces a novel hybrid model that integrates the Mountain Gazelle Optimizer with Transformer architecture (MGO-Transformer) to significantly improve the accuracy of DNI prediction. Applied to Kuqa, Xinjiang, a region with abundant solar resources, the model outperforms conventional approaches, achieving a coefficient of determination of 0.998, root mean square error of 19.94, and mean absolute percentage error of 0.12. Using the enhanced forecasting output, an optimized off-grid microgrid is designed to meet the area’s electricity and hydrogen demands. The proposed system incorporates photovoltaic (PV) panels, wind turbines, an electrolyzer, hydrogen storage, lithium-ion batteries, and a converter. Simulation results indicate an annual energy production of 1,671,030 kWh, with PV contributing 97.7% of the total output. The electrolyzer produces 27,329 kilograms of hydrogen per year, meeting the full hydrogen demand of the system. Economic analysis reveals strong performance metrics, including a Levelized Cost of Energy of 1.93 $/kWh and a Levelized Cost of Hydrogen of 5.26 $/kg. This work underscores the value of machine learning-based forecasting in optimizing microgrid configurations for sustainable and economically viable energy and hydrogen production in remote regions.