Forecasting hydrogen production through electrolysis powered by concentrated solar power plant using artificial neural network

Abstract

In today's world, artificial intelligence has become a vital utility technology with the potential to benefit various industries and research endeavors significantly. One such application lies in harnessing solar thermal energy for green hydrogen purposes. Hence, this study aims to examine the feed-forward back-propagation network (FFBPN) in the context of a Dish/Stirling powered electrolysis system for hydrogen production, utilizing time series data from over twenty locations in Morocco. The FFBPN model was developed to evaluate the impact of various input parameters, including direct normal irradiation (DNI) at different geographical locations, on hydrogen production from the Dish/Stirling powered electrolysis system. This model was trained using different training algorithms, namely Levenberg-Marquardt (LM), Fletcher-Powell Conjugate Gradient (CGF), One Step Secant (OSS), and Scaled Conjugate Gradient Back-propagation (SCG), to identify the most effective approach for predicting green hydrogen production. By using a variety of training algorithms and evaluating the models using specified metrics, the study aimed to determine the most suitable and effective training approach for the given data. The different results indicate that, among all the locations examined in eastern Morocco, Figuig and Bouarfa cities have been identified as the most appropriate locations for implementing the proposed system, yielding the highest annual net electric energy output of 83.52 GWh/yr and 81.92 GWh/yr, respectively. Furthermore, the system allowed the production of over 1462 tons/yr of green hydrogen, supported by a total installed capacity of 50 MWe. Furthermore, the statistical analysis reveals that the Levenberg-Marquardt (LM) algorithm, using 33 neurons, outperformed others, exhibiting the lowest errors and the highest R² value during both training and testing. Specifically, during training, the metrics of RMSE, MRE, COV, and R² recorded values of 0.582, 0.395, 0.510, and 0.99999, respectively, whereas during testing they were 0.633, 0.474, 0.560, and 0.99999, respectively. The FFBPN application stands as a pioneering and effective model to predict green hydrogen production from the Dish/Stirling system.