Data-driven biogas production forecasting: a time series case study on a full-scale plant

The transition towards renewable resources constitutes a focal point of the modern energy question. Hydroelectric, solar and wind power are the most popular approaches, but here another alternative will be examined: anaerobic digestion (AD) on full-scale plants. Microbe-mediated digestion of organic waste serves a dual purpose, acting simultaneously as a biowaste treatment method and as a biofuel generation process, yielding biogas. Given the considerable complexity of this process, this thesis addresses the implementation of machine learning systems (ML) for forecasting the productivity of an existing full-scale biogas plant. Five candidate models (Ridge, ElasticNet, K-Nearest Neighbors, Random Forest and Histogram-based Gradient Boosting) were benchmarked against naive baselines, across two forecasting strategies (Recursive and Direct) and two temporal resolutions (1-hour and 6-hour), for a fixed 24-hour horizon. Models were tuned via time-series cross-validation and Random Search, then evaluated on two distinct test sets using NMAE, RelMAE, and correlation metrics. Findings reveal that 6-hour resolutions consistently outperformed 1-hour ones, and the Direct strategy generalized more robustly than the Recursive one, which suffered from error propagation. ElasticNet (6h-Direct) achieved the best balance of predictive stability and computational efficiency, though only barely matching naive baselines on point-wise error. Finally, SHAP interpretability analysis validated that candidate forecasters rely on physically coherent operational drivers, supporting their practical utility for plant monitoring and decision support.