Exploring available input variables for machine learning models to predict biogas production in industrial-scale biogas plants treating food waste

Abstract

Anaerobic digestion (AD) of food waste (FW) has been widely used in China and shows great potential in both recovering renewable energy and reducing carbon emissions. Unfortunately, many digesters suffer from unexpected perturbations and low biogas production due to the specific characteristics of food waste. Modeling of AD is crucial to better understand the process and improve efficiency. In this study, five machine learning (ML) algorithms were used to build models to predict biogas production in an industrial-scale biogas plant treating FW. Three categories of routine monitoring indicators (feed amount, feedstock properties, and digester properties), individually or collectively, were used as the input variables. The results showed that the random forest (RF) model achieved the best performance with an average R2 of 0.74 when all the indicators were contained in the dataset. Feature importance analysis revealed that the significance descended in the order of feed amount (45.9%), digester properties (38.6%), and feedstock properties (15.4%). The performance of the predictive models degraded with lag time except for the RF model, which showed the potential to predict biogas production of the next day (R2 = 0.73). The study verified the feasibility of ML models in predicting biogas yield using routine monitoring data from industrial-scale biogas plants treating FW and suggested that the monitoring indicators and frequency be redesigned to build smarter ML predictive models and improve system efficiency.