Volume 48

Evaluation of Hyperparameter Optimization in Machine Learning Models for CH4 and H2 Production Prediction based on Supercritical Water Gasification Chiagoziem C. Ukwuoma1, Dongsheng Cai2*, Mmesoma P. Chukwuemeka3, Blessing O. Ayeni4, Ibrahim Kunle Adefarati5, Huan Yang6 Olusola Bamisile7 Qi Huang8

Abstract

Renewable energy continues to rise, with supercritical water gasification (SCWG) emerging as a potential biofuel production technique. Conventional approaches for estimating CH4 and H2 generation in complex systems frequently fail due to the complicated and dynamic nature of these processes. Machine learning (ML) has emerged as a disruptive technology in various industries, including energy, where it is used to optimize operations and improve prediction accuracy. Conventional techniques lack the versatility and scalability of ML models, resulting in less accurate and efficient prediction capabilities. This gap emphasizes the importance of incorporating machine learning into the energy domain, notably for optimization and prediction in SCWG processes. Furthermore, for any machine learning model, determining the appropriate hyperparameter setting has a direct and significant influence on its performance. In this study, we investigate the influence of three distinct types of hyperparameter optimization techniques on CH4 and H2 production prediction based on supercritical water gasification. Grid Search Optimization, Random Search Optimization, and Bayesian Optimization were analyzed utilizing six machine learning models: Ridge Lasso, Elastic Net, Decision Tree, Random Forest, and XGBoost, as well as two ensemble models: linear-based and tree-based models. The dataset from supercritical water gasification of Yimin lignite was used based on three evaluation metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R2). The results were evaluated based on 5-fold cross-validation and results show that the Random Forest stood out with an R2 of 0.995, RMSE of 0.109, and MAE of 0.085 for the CH4 production prediction while recording an R2 of 1.00, RMSE of 0.112 and MAE of 0.088 for H2 production prediction. The analysis carried out in this study shows that the choice of optimization technique does not significantly impact the performance of the deployed models, which indicates that the hyperparameter space is relatively well-behaved for this CH4 and H2 Production Prediction based on Supercritical Water Gasification, and thus even simpler optimization strategies like Random Search can perform nearly as well as more sophisticated ones like Bayesian Optimization. The result implies that if computation time or resources are a factor, Random Search would be a more efficient approach without a significant trade-off in model performance

Keywords Hyperparameter Optimization, Supercritical Water Gasification, Hydrogen Production, Methane Production, Machine Learning