Retraining prior state performances of anaerobic digestion improves prediction accuracy of methane yield in various machine learning models

• Published in: Applied energy Vol. 298; p. 117250
• Main Authors: Park, Jun-Gyu, Jun, Hang-Bae, Heo, Tae-Young
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.09.2021
• Subjects: 1-step ahead; Anaerobic digestion; Machine learning; Methane yield; Retraining; pH; Methane yield; Retraining; 1-step ahead; Anaerobic digestion; Machine learning
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2021.117250

[Display omitted] •Various ML models using 1-step ahead with retraining method were applied.•Retraining prior state performance helps ML models imitate microbial memories.•Transition period could be properly predicted by retraining process.•pH only input data shows possibility of practical application. The prediction of anaerobic digestion (AD) performance using numerical models, which are based on mathematics and kinetics, is being challenged by poor mechanistic understanding and the non-linear relationships between performance and operational parameters. This study demonstrated that various machine learning (ML) models using the 1-step ahead with the retraining method, which utilized AD performance data from prior states, can improve the prediction accuracy of ML models. For the four types of ML models studied, the 1-step ahead with the retraining method could improve the root mean square errors by 32–49% compared to the conventional multi-step ahead method, which was particularly noteworthy during the transition period when AD reactors were faced with loading shocks and showed inhibited methane yields. Moreover, the 1-step ahead with the retraining method showed the potential of achieving accurate predictions using a single input parameter, pH, which was considerably less labor-intensive to monitor than the other parameters often required in AD models (e.g., VSS). As such, the 1-step ahead with retraining method is suitable for efficient real-time prediction of AD performance in real-world operations.