Machine learning for predicting and optimizing the performance of a commercial-scale anaerobic digester with diverse feedstocks and operating conditions

Abstract

The improper disposal of food waste and livestock manure poses significant environmental risks, including nutrient pollution, water contamination, and greenhouse gas emissions. Anaerobic digestion (AD) provides a sustainable pathway for converting organic waste into biogas while reducing environmental impacts. However, optimizing AD performance at a commercial scale remains challenging due to feedstock variability, operational complexity, and time-dependent dynamics. In this study, we analyzed six years of data from a commercial-scale AD system processing 18 types of food and manure waste to develop machine learning (ML) models for predictive analysis and process optimization. Three key outputs, total gas production, methane percentage, and H₂S content, were predicted using Random Forest (RF), Artificial Neural Networks (ANN), and XGBoost. RF consistently yielded the highest performance with accuracy of 0.91 (gas production), 0.93 (methane), and 0.91 (H₂S). Feature importance analysis revealed that time-series factors (e.g., rolling averages of previous days), pH, temperature, and hydraulic retention time (HRT) significantly influenced model accuracy. Notably, feedstocks such as dairy manure and pineapple waste exhibited strong correlations with both gas yield and H₂S fluctuations. Optimization using Particle Swarm Optimization and Simulated Annealing demonstrated the potential to improve biogas production by up to 12 % and reduce H₂S levels by as much as 65 % through adjusted operating conditions. These findings highlight the value of ML in not only forecasting AD performance with high accuracy but also in identifying operational strategies to enhance system efficiency and stability. This work provides actionable insights for the data-driven management of commercial-scale AD systems.