Optimization-Based Comparative Study of Machine Learning Methods for the Prediction of Hydrogen Production From Biogas

Production of hydrogen from biogas is indeed a promising approach to address the various sustainability challenges such as reducing greenhouse gas (GHG) emissions and replacing fossil fuels with renewable energy resources. Despite the growing significance of renewable energy, optimization of hydrogen production from biogas with machine learning (ML) techniques remain underexplored. This study aims to fill the research gap by analyzing and evaluating various ML methods including linear regression (LR), K-nearest neighbors (KNNs), random forest (RF), and convolutional neural networks (CNNs) in modeling hydrogen production from biogas. The models were trained on the data obtained through sensitivity analysis of process conditions with Aspen Plus (v14, AspenTech) simulation. Hyperparameter tuning was performed to enhance and optimize the prediction capabilities of the models. Performance analysis of the models indicates R-squared (R²) values of 0.98 (RF), 0.88 (LR), 0.72 (CNN), and 0.89 (KNN) and mean squared error (MSE) values of 0.0695 (RF), 0.5138 (LR), 0.90 (CNN), and 0.5241 (KNN), respectively. To further optimize hydrogen production, the RF model was chosen due to its high R² and low MSE value, indicating superior predictive performance. This model was then used as a surrogate for fitness function evaluations in two optimization frameworks based on the genetic algorithm (GA) and Nelder–Mead (NM) methods. Optimization of input parameters using surrogate-based methodology resulted in an increase in hydrogen production by 25%. This approach provides a platform for plant-level implementation, realizing the concept of Industry 4.0 in biogas processing for hydrogen production.