Facilitating the predictions of batch and continuous anaerobic digestion processes performance with statistical tools

Anaerobic digestion (AD) is a promising technology for waste management and renewable energy production. Determining the biomethane potential (BMP) of a material is crucial when considering it as feedstock for the digester. The practical BMP of a material and its degradation kinetics can be derived from the batch BMP test, typically taking at least 30 days, which can be onerous to the industrial operator. Many studies have attempted to predict BMP test results by building regressions between various feedstock physiochemical characteristics and BMP test result. However, these methods primarily predict the ultimate biogas yield of the BMP test and are unable to capture the reaction kinetics. Part I of this study proposed a method to predict the BMP test result of a material, not only the ultimate biogas yield but also the degradation kinetics, which was achieved by discovering a model that describes the biogas production well and then inferring parameters of this model from feedstock’s physiochemical characteristics. The machine learning (ML) model, Decision Tree, was adopted to predict the parameters of the time series model, first-order Autoregressive Model, from the characteristics including total solids, volatile solids, total volatile fatty acid, total ammonia nitrogen, chemical oxygen demand, alkalinity, elemental composition, pH, heavy metals, etc. The DT trained and tested by leave-one-out cross validation with 25 BMP test data had a mean absolute percentage error (MAPE) of 45.63% on the BMP test results, showing unsatisfactory prediction accuracy and unreliable feature importance analysis results. To obtain sufficient data for ML model training and avoid the consistency issues with BMP test data collected from diverse sources, a data augmentation method employing response surface design was proposed. With sufficient data for model training, eXtreme Gradient Boosting (XGBoost) models with three important features determined via feature importance analysis could predict biogas production model parameters with R²values above 0.99 on the test set. Despite the strong regression capabilities of ML models compared to simple statistical model, their explainability remains a challenge. The current popular methods for ML model interpretation focus on feature importance analysis. In this study, Meijei G-functions were used to interpret the predictions of the XGBoost model mathematically to enhance the accessibility and transparency of the black- box ML model to domain experts as a tool for material realistic BMP prediction. The general predictability of the mathematical metamodel was tested using 13 BMP test data sourced from the literature, all within the applicability range of the trained ML model, resulting in a mean absolute error of 38.074 mL CH ₄/g volatile solid added and MAPE of 15.424%. Besides the BMP of a material, operational parameters of the digester are critical to the performance of a continuous AD process. The application of Anaerobic Digestion Model No.1 (ADM1) for continuous AD simulation, which assists in decision-making in the AD industry by predicting digester performance under various operational schemes, is often hindered by its calibration difficulties, especially with limited data. Part II of this study presents a Bayesian inference-based framework to reliably calibrate ADM1 using only initial-stage digester data. A sequence of sensitivity analysis (SA) was applied to identify the most influential kinetic parameters and initial values to be calibrated. SA results revealed that steady-state biogas production was collaboratively influenced by the disintegration rate, hydrolysis rates, and initial concentrations of acetate degraders, cations, and anions. In contrast, Total Ammonia Nitrogen and pH of digestate were predominantly influenced by initial values of cations and anions. These findings challenge the common practice in ADM1 studies of only calibrating kinetic and stoichiometric parameters. Then, using biogas production and digestate data from less than two hydraulic retention times and informative prior distributions determined from domain knowledge, seven the most influential uncertain inputs were calibrated. The calibrated model predicted the steady state performance satisfactorily. The 95% credible intervals of the calibrated model encompassed 66.047% of the 10- day moving average trendline of the daily biogas flow data and all of the steady- state digestate pH and total chemical oxygen demand data.