Hyperparameter-transfer learning framework for methane prediction in data-limited full-scale anaerobic digesters

Area of Science:

• Environmental Engineering
• Biotechnology
• Machine Learning Applications

Background:

• Anaerobic digestion performance is complex, affected by substrate variability, operating conditions, and microbial dynamics.
• Predicting methane production in new anaerobic digesters is challenging due to limited initial monitoring data.
• Nonlinear and site-specific methane production behaviors require advanced modeling techniques.

Purpose of the Study:

• To develop a transfer-learning workflow for accurate methane prediction in newly commissioned anaerobic digesters.
• To integrate an anaerobic digestion-specific feature engineering pipeline into the machine learning model.
• To reduce the computational cost and time required for developing predictive models for new digesters.

Main Methods:

• Developed a machine learning pipeline using long-term data from a mesophilic digester.
• Employed an ensemble of machine learning models: Random Forest, XGBoost, and LightGBM.
• Applied hyperparameter-transfer-based transfer learning to a new digester with limited data.

Main Results:

• Transfer learning achieved prediction accuracy (R² of 0.88) comparable to a fully re-optimized model.
• Model development time was reduced from 59.1 seconds to approximately 5 seconds.
• The transferred ensemble model showed RMSE of 171.70 m³/day and MAE of 138.58 m³/day.

Conclusions:

• The developed framework provides a computationally efficient method for early-stage methane prediction.
• Transfer learning facilitates rapid deployment of reliable forecasting tools for new anaerobic digesters.
• This approach addresses the challenge of limited data in newly commissioned anaerobic digestion facilities.