Thermodynamic Modelling with Machine Learning for Industrial-Scale Biogas Production with CHP Recovery

Shirin Naderi, Amirhossein Ahmadi — Mon, 05 Oct 2026 00:00:00 +0200

Most process models of anaerobic-digestion (AD) biogas plants either resolve the biochemistry in detail while treating the combined heat and power (CHP) unit with fixed heating values, or resolve the thermodynamics while neglecting how operating conditions affect biological conversion. This study presents an open-source, digital-twin-oriented simulation framework that couples a multi-factor steady-state kinetic model of a co-digestion reactor (cattle manure, food waste and maize silage) with an exergy model of the biogas and CHP combustion, evaluated directly with the PYroMat property library. After calibration against 19 steady-state operating periods from two independent laboratory CSTR studies, the kinetic model reproduced the measured specific methane yield (SMY) with a cross-validated R² of 0.89, whereas an uncalibrated ADM1 implementation under-predicted the same data. The framework was applied to a 365-day industrial run of a 0.6 MWel plant, a 4000-point machine-learning (ML) dataset and a 10,000-run Monte Carlo study. Gradient Boosting surrogates reproduced the mechanistic model with R² = 0.990 for SMY and 0.995 for CHP exergy efficiency, values that reflect surrogate fidelity rather than accuracy against plant data. Sobol' sensitivity analysis showed that digester temperature and organic loading rate govern SMY, whereas feedstock blend and organic loading rate govern exergy efficiency, a ranking confirmed by ML permutation importance. At the nominal point, the plant reaches 82.5% energy efficiency but only 40.7% exergy efficiency, mainly because recovered heat carries little exergy. All code, data and figures are released as a reproducible artefact.

Journal of Transactions in Systems Engineering

Thermodynamic Modelling with Machine Learning for Industrial-Scale Biogas Production with CHP Recovery