A hybrid approach of anaerobic digestion model no. 1 and machine learning to model and optimize continuous anaerobic digestion processes

Abstract

Anaerobic digestion is a promising approach to dispose of biodegradable waste and wastewater, generating biogas as an alternative energy resource. This work proposed a so-called M-CADM1 for continuous anaerobic digestion simulation, which combined the machine learning and anaerobic digestion model No.1 (ADM1). The detailed reaction path and intermediate products in different stages of anaerobic digestion are specified in ADM1. The kinetic parameters were modified by machine learning. The characteristics (elemental composition) of feedstocks were used to predict kinetic parameters. A total of 75 biomass samples were used to establish for machine learning models. Five element contents (C, H, O, N, S), feedstock feed rate, and anaerobic digestion temperature were used as the input. The kinetic parameters were set as output. The sensitivities of 17 kinetic parameters were evaluated. 7 kinetic parameters with the highest sensitivities were selected as ADM1 model inputs by sensitivity analysis. The R² and RMSE were used as the index to evaluated the accuracy of machine learning model. The best R² and RMSE reached 0.84 and 0.196. The TIC was used as the index to evaluated the accuracy of M-CADM1. By comparing the simulated value with the experimental value, the accuracy of the overall M-CADM1 expressed by TIC of kitchen waste was 0.036. The organic acid content and pH in the reactor were considered as indicators to study the accuracy and stability of the M-CADM1. Trends in organic acids, free ammonia or hydrogen inhibition, and pH were consistent with experimental continuous anaerobic digestion results.