Performance prediction of ZVI-based anaerobic digestion reactor using machine learning algorithms

The use of zero-valent iron (ZVI) to enhance anaerobic digestion (AD) systems is widely advocated as it improves methane production and system stability. Accurate modeling of ZVI-based AD reactor is conducive to predicting methane production potential, optimizing operational strategy, and gathering reference information for industrial design in place of time-consuming and laborious tests. In this study, three machine learning (ML) algorithms, namely random forest (RF), extreme gradient boosting (XGBoost), and deep learning (DL), were evaluated for their feasibility of predicting the performance of ZVI-based AD reactors based on the operating parameters collected in 9 published articles. XGBoost demonstrated the highest accuracy in predicting total methane production, with a root mean squared error (RMSE) of 21.09, compared to 26.03 and 27.35 of RF and DL, respectively. The accuracy represented by mean absolute percentage error also showed the same trend, with 14.26%, 15.14% and 17.82% for XGBoost, RF and DL, respectively. Through the feature importance generated by XGBoost, the parameters of total solid of feedstock (TSf), sCOD, ZVI dosage and particle size were identified as the dominant parameters that affect the methane production, with feature importance weights of 0.339, 0.238, 0.158, and 0.116, respectively. The digestion time was further introduced into the above-established model to predict the cumulative methane production. With the expansion of training dataset, DL outperformed XGBoost and RF to show the lowest RMSEs of 11.83 and 5.82 in the control and ZVI-added reactors, respectively. This study demonstrates the potential of using ML algorithms to model ZVI-based AD reactors. (C) 2020 Elsevier Ltd. All rights reserved.

Automated summary

The study evaluates three machine learning algorithms for predicting the performance of ZVI-based AD reactors and finds XGBoost to have the highest accuracy. Parameters like total solid of feedstock, sCOD, ZVI dosage, and particle size were identified as dominant factors affecting methane production. Deep learning outperformed XGBoost and random forest with the lowest root mean squared errors for predicting cumulative methane production.