Predictive Performance Comparison of Artificial Neural Network and Ensemble Models for Methane Gas Prediction in a Real-Scale Biogas Facility

Biogas, a renewable energy, results from complex biochemical processes, primarily comprising methane, carbon dioxide, oxygen, and trace gases. Inef-ficient management in biogas facilities leads to fermentation slowdowns and reactor failures. The utilization of Artificial Intelligence strategies enables a flexible and precise control process, achieving an optimal balance between anaerobic performance and biogas production.
Utilizing two datasets, the study examines input variables (CO2, O2, pH, H2S, temperature, fluid/cattle manure, poultry manure, clean water) meas-ured daily on a particular day and output methane gas amounts measured af-ter 30 days, while the second dataset includes more input variables measured over a period of 30 days. In this study, the predictive accuracy of methane gas alone, as well as in combination with oxygen and carbon dioxide is ex-amined. A variety of statistical and machine learning models were used, in-cluding Artificial Neural Networks, Support Vector Regression, Polynomial Regression, Random Forest, Gradient Boosting, Ridge, LASSO, and Elastic Net Regression. ANN produces better results than statistical and machine learning models, exhibiting RMSE of 0.1, MAE of 0.08, and MSE of 0.01. Applied ANN model excels at predicting methane, carbon dioxide, and oxy-gen together compared to methane gas alone.