Data-driven modeling of anaerobic digestion processes with a view to process control: a first approach

Abstract

This thesis presents a data-driven modeling approach for predicting the dynamics of anaerobic digestion processes with the aim of informing process control strategies. Three neural network architectures—NARX and two ways of using LSTM— are first evaluated using a simple system—emulated by a mass balance of a second-order reaction—as a baseline, including a sensitivity analysis of parameters such as the number of excitations, sampling rate, step duration, and network complexity. The insights obtained from this analysis are then applied to guide the training of networks for predicting the behavior of anaerobic digestion, as emulated by the Hill and AM2 models. Although reliable multi-step forecasts can be achieved with abundant training data, this study examines how prediction accuracy varies as the available data and model complexity change. The results indicate that it is possible to train data-driven models that mimic the nonlinear dynamics of anaerobic digestion, considering the trade-offs between data availability and model complexity. These findings provide a foundation for subsequent investigations focused on integrating such models into control frameworks, particularly Model Predictive Control (MPC).