Modelling anaerobic digestion of sewage sludge: mechanistic models vs machine learning

Abstract

Anaerobic digestion processes are one of the technologies most used by wastewater treatment plants (WWTPs) to stabilize and decrease the organic content of sludge. This process decreases the costs of disposal while increasing the energetic efficiency of WWTPs. In order to optimize this process, three model approaches were implemented. First, we calibrated and validated the anaerobic digestion model no.1 (ADM1) using data from an anaerobic lab digester treating sewage sludge (Phases I, II, III), and further receiving glycerol pulses (Phases IV, V). Then, to optimize the calibration and parameter estimation, an iterative procedure was applied by minimizing the root mean square error (RMSE). The second approach consisted of applying a machine learning (ML) model to the biogas and methane produced. The results showed that the ADM1 model adjusted well to the experimental results, especially to biogas, methane and pH. The optimization routine was useful to identify the most sensitive parameters, improving model calibration. Overall, the ML approach was more reliable to predict anaerobic reactors performance but did not respond so well to process perturbations (glycerol pulses).