Implementation of Machine Learning Models for Predicting Hydrogen Production from Renewable Energy

Abstract

The world continues to be largely reliant on fossil fuels, such as coal, oil and natural gas, which are major contributors of greenhouse gases, air pollution and climate change caused by the release of gases like CO2, NO o and SO2. Even though renewable energy sources such as solar and wind provide a cleaner substitute, their nature as intermittent and weather-dependent sources become a big problem in terms of large scale and long-term energy storage making traditional battery systems economically impractical. Consequently, the utilization of the surplus renewable energy through the process of water electrolysis to create green hydrogen has become a viable and alternative way of storing energy in the long term. This work suggests a two-stage machine learning-based predictive model of hydrogen production when using renewable energy, based on real-world working data of a 40.5 MW grid-connected photovoltaic (PV) power facility. The initial step involves predicting photovoltaic power output from meteorological variables using multiple regression and deep learning models, such as Support Vector Regression (SVR), Random Forest, Decision Tree, and Gated Recurrent Units (GRU). The second stage is to incorporate the predicted solar energy into a hybrid electrochemical model of hydrogen production, and the same machine learning models serve as data-driven correction models to more effectively predict hydrogen yield. It is a two-stage method that integrates both physical modelling and machine learning to model nonlinear system behaviour and real-world losses of operation. The findings show that SVR was always better in both phases than the other models with an R 2 value of 0.965 in photovoltaic power prediction and 0.968 in hydrogen production prediction. The given framework minimized the mistake in the production of hydrogen annually to about 3 percent, which is much better than theoretical models and deep learning alternatives. Moreover, a Sobol based global sensitivity analysis revealed that Global Horizontal Irradiance (GHI) made the greatest contribution to the uncertainty in the hydrogen production process, then AC power output and temperature at the module. The outcomes of these studies support the premise that the two stages proposed framework is a viable, precise, and large-scale solution to real-time forecasting, optimization of the system, and successful grid integration of green hydrogen system.