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

This study developed a full-scale transfer-learning workflow integrated with an anaerobic-digestion-specific feature engineering pipeline for methane prediction in newly commissioned anaerobic digesters with limited monitoring data. The performance of anaerobic digestion is significantly influenced by heterogeneous substrates, fluctuating operating conditions, and time-dependent microbial dynamics, resulting in nonlinear and site-specific methane production behaviors. To address this, we developed a comprehensive biogas machine learning (ML) pipeline using long-term monitoring data from a data-rich mesophilic digester, employing an ensemble of ML models, including Random Forest, XGBoost, and LightGBM. The optimized model was then applied to a newly commissioned digester with limited and incomplete datasets, without additional hyperparameter optimization. Remarkably, Hyperparameter-transfer-based transfer learning achieved prediction accuracy comparable to that of a fully re-optimized target-domain model, with an R2 of 0.88, while reducing model development time from 59.1 s to approximately 5 s. The transferred ensemble model achieved an RMSE of 171.70 m3/day and an MAE of 138.58 m3/day, compared with 151.48 m3/day and 121.43 m3/day, respectively, for the fully tuned target-domain model. Overall, this framework offers a practical and computationally efficient method for early-stage methane prediction, facilitating the rapid deployment of reliable forecasting tools in newly commissioned anaerobic digesters.