Feature Engineering and Supervised Machine Learning to Forecast Biogas Production during Municipal Anaerobic...

Area of Science:

• Environmental Engineering
• Biotechnology
• Data Science

Background:

• Municipalities co-digest offsite wastes to increase biogas production and meet sustainability goals.
• Operational challenges arise from the heterogeneity of waste streams in anaerobic digestion.
• Predicting and modeling complex microbial processes in anaerobic digestion is difficult.

Purpose of the Study:

• To predict biogas production using historical operational data from a water resource recovery facility (WRRF).
• To assess the feasibility of forecasting biogas flow for a 24-hour horizon without additional data collection.
• To compare machine learning models for biogas production prediction.

Main Methods:

• Utilized two datasets: daily lab/operational data (n=1089) and minute-by-minute SCADA data (n=491,761).
• Developed and compared multilayer perceptron (MLP), tree-based, and multiple linear regression models.
• Focused on feature engineering from raw SCADA outputs for model improvement.

Main Results:

• A multilayer perceptron (MLP) neural network model significantly outperformed other models.
• The MLP model using SCADA data achieved an R² of 0.78 and a 13.4% mean absolute percentage error.
• Adding daily lab analyses did not substantially improve biogas flow prediction accuracy.

Conclusions:

• Minute-scale SCADA data can effectively forecast biogas production in municipal co-digestion.
• Feature engineering from SCADA data is crucial for accurate prediction.
• This predictive capability is a foundational step towards developing a digital twin for anaerobic digestion processes.