Full-scale dynamic anaerobic digestion process simulation with machine and deep learning algorithms at intra-day resolution

Machine learning algorithms have been proven to be effective in predicting characteristic process variables of the anaerobic digestion process. However, industrial application has rarely been investigated, and the most effective algorithms for typical operating conditions have not been defined. Thus, 13 machine learning, deep learning and statistical algorithms were applied to three full-scale datasets at intra-day resolution. A systematic procedure was applied for reliable data preparation and hyperparameter optimization. Methane yield was predicted one step, 12 h and 24 h in advance. Results indicate that random forest and long short-term memory neural networks are the most robust algorithms, while further linear models can be advantageous in specific situations. Previous step methane yield and fed volatile solids are, in general, the most relevant parameters, while further laboratory measurements can be advantageous at high feed quantities. Data preparation is crucial to allow less complex models (such as linear models) to perform well. This study defines appropriate machine learning algorithms and essential measurements for characteristic process conditions at different data resolutions, when predicting dynamic intra-day methane production of industrial-scale anaerobic digestion processes, as a reliable basis for model-based process monitoring and control. • Random Forest and LSTM Neural Networks are recommended for AD process prediction. • Linear models show high performance difference between validation and test datasets. • Only few measurements are highly influential for prediction of methane production. • Data preparation parameters are highly influential in model performances.