The Pruned Network Approach: A feature optimizing based deep learning framework to estimate hydrogen production potential from geothermal energy

The world's increasing energy demand and the need for sustainable energy alternatives have accelerated efforts to find efficient hydrogen production methods from renewable sources. Geothermal Power Plants (GPPs) offer significant potential for green hydrogen production due to their stable and high thermal output. However, the complexity and nonlinearity of geothermal data make precisely predicting hydrogen production challenging. This study addresses this gap by presenting an innovative framework, the Pruned Network Approach (PNA), that combines feature selection with advanced deep learning models to improve predictive accuracy and computational efficiency. First, the Least Absolute Shrinkage and Selection Operator (LASSO) method is used to reduce redundant input attributes from geothermal datasets. Then, unlike previous studies in the literature, optimized Supervised Deep Neural Networks (S-DNNs), Extreme Gradient Boosting (XGBoost) and Deep Belief Networks (DBNs) models are used in the present study. By this way, a novel predictive framework (Pruned Network Approach) for hydrogen production estimation using geothermal energy sources was developed. The key novelty of the present study is optimizing computational efficiency and predictive accuracy through an ensemble of S-DNNs, XGBoost and DBNs. The primary aim of the study is to provide a technical report and analytical methodology to readers, manufacturers, and scientists who specialize in underground (geothermal) energy. As a result of the analyses, the developed models significantly improved over existing methods (with a R 2 value of 0.9875 and RMSE value of 4.01162). This framework aimed to provide a scalable and accurate tool for renewable hydrogen production forecasting, considering the growing global energy demand. • A novel Pruned Network Approach was developed. • LASSO was used to reduce the number of variables and determine important parameters. • S-DNN, XGBoost and DBN models were employed for estimating hydrogen production data. • Supervised and unsupervised approaches were used together with the LASSO-DBN-DNN model. • Computational efficiency and predictive accuracy were optimized by models of S-DNN, XGBoost and DBN.