Data-Driven Modeling and Experimental Investigation of Anaerobic Digestion of Wastewater Sludge

Abstract

This thesis investigates two key strategies for improving the performance of sludge anaerobic digestion (AD): hydrothermal pre-treatment (HTP) and anaerobic co-digestion (An-CoD) with food waste (FW). The research integrates machine learning (ML) and hybrid modeling with experimental validation to optimize process parameters, predict methane yield, and support data-driven decision-making in sludge treatment and energy recovery systems. The first research phase focuses on HTP to enhance sludge solubilization and subsequent biogas production. A wide range of ML algorithms, including artificial neural networks (ANN), decision-tree models (Random Forest, XGBoost), support vector machines, and K-nearest neighbors, were trained on nearly two decades of published data. ANNs delivered the best performance for solubilization prediction, while decision-tree models excelled in methane yield forecasting. Heating time was found to be as influential as or more than holding time, challenging previous assumptions that overlooked heating rate. The heating method itself (e.g., microwave, radio frequency, steam) was shown to have minimal impact on AD outcomes. Experimental validation using a Box-Behnken Design (BBD) confirmed model predictions, with XGBoost achieving best fitting for solubilization and methane yield. Moderate pre-treatment conditions were found optimal, whereas severe conditions reduced methane yield. The second phase addresses sludge–FW co-digestion. ML models were developed using published batch data to predict methane yield and determine optimal mixing ratios. A regularized ANN performed well and was validated with eleven lab-scale tests, identifying ~20% sludge and 80% FW as optimal. For continuous flow, four reactors were operated under various conditions derived from the City of Kingston’s biosolids plan. Experimental and literature time-series data were used to develop a physics-informed long short-term memory (PI-LSTM) model. Incorporating mechanistic equations into the loss function, the model outperformed standard LSTM. This research is the first to integrate ML with a large-scale HTP/AD dataset and to apply physics-informed deep learning for continuous co-digestion. It offers validated, scalable models and actionable insights for optimizing sludge treatment, advancing AD system modeling toward real-time, sustainable biosolids management.