Assessment and optimization of co-digestion of grass silage and animal slurry

Pasture-based beef systems can provide cattle slurry and grass silage for anaerobic digestion (AD). However, there are still considerable gaps in the full-scale anaerobic co-digestion of these feedstocks, particularly the capacity to predict AD performance under a range of grass silage and cattle slurry characteristics, and how AD can be integrated into livestock farming. The objectives of this PhD research included: (1) application of the Anaerobic Digestion Model No. 1 (ADM1) to simulate the full-scale co-digestion of grass silage and cattle slurry, and determination of the most influential parameters; (2) assessment of the performance of simplified ADM1 and two machine learning algorithms in simulating biogas and methane production from a full-scale AD plant co-digesting different agricultural feedstocks; and (3) assessment of the integration of AD into a livestock farming system in terms of feedstock provision, greenhouse gas emissions (GHG), digestate management and economic viability. Results indicated that a simplified version of the ADM1 was able to accurately depict the biogas and methane production as well as general trends of pH and ammonium nitrogen from a full-scale AD plant co-digesting grass silage and cattle slurry. Results from substrate composition variability indicated that the variations in crude carbohydrates, proteins and lipids concentrations did not significantly affect biogas and methane yields across the data sets analyzed. In contrast, carbohydrate degradability emerged as the most significant parameter in explaining the variability in biogas and methane production. For a full-scale AD plant co-digesting various agricultural feedstocks, a simplified ADM1 model (ADM1-R3) and two machine learning (ML) algorithms, random forest (RF) and long short-term memory (LSTM), demonstrated high accuracy in simulating biogas and methane production. Unlike ADM1, which required detailed feedstock characterization, the ML algorithms achieved comparable performance using only fresh feedstock quantities and volatile solids (VS) as inputs. LSTM exhibited the highest computational demand, with simulation times 141 times longer than ADM1-R3 and 11 times higher than RF. Among the tested input variables, maize silage (fresh quantity and volatile solids) was identified as the most influential feature in the ML models.

Funder

Teagasc Walsh Scholars Programme

Publisher

University of Galway

Rights

Attribution-NonCommercial-NoDerivatives 4.0 International

cbn

Collections

University of Galway Theses (PhD Theses)