Abstract
Accurately predicting biogas production for real-time applications remains a challenge in anaerobic digestion (AD) due to the process's complexity and dynamic nature. While mechanistic models are essential for understanding and modelling AD processes, however they are highly non-linear and depend on detailed feedstock characterisation and parameter calibration. In contrast, data-driven models do not rely on predefined equations and rather use process data to capture the system's underlying dynamics. This study compares mechanistic and data-driven models for biogas prediction using lab-scale data. A state estimation framework with a rolling window was used for the mechanistic model, based on biomass and substrate concentrations with Haldane kinetics, achieved an accuracy of (R² = 0.91). A Long Short-Term Memory (LSTM) model with Bayesian Optimisation for hyperparameter optimisation, trained on the same data showed superior performance (R² = 0.93–0.98) and captured temporal dependencies inherent to the AD process. The LSTM model was further applied to industrial data, maintaining high accuracy (R² = 0.95–0.97) and demonstrating its scalability. Its strong predictive capabilities, combined with practicality for real-time applications, make it a promising tool for optimising operations in large-scale AD plants.