Machine-learning-based predictive control of anaerobic digestion considering meteorological conditions

Abstract

Efficient control of the anaerobic digestion process in municipal wastewater treatment plants poses an engineering challenge due to the high inertia of the system and the impact of uncontrolled external disturbances. This paper pre - sented a novel approach to optimising feedstock composition (the ratio of primary to excess sludge) to maximise biogas production. Unlike classical control methods, the proposed solution accounts for the impact of exogenous factors ambient temperature and precipitation which determine the physicochemical parameters of raw sludge. To build an effective model on a limited dataset, a spatial disaggregation technique was applied, merging data from four parallel digestion chambers. A hybrid decision-making system was developed, integrating a Random Forest model with a genetic algorithm (GA). The resulting predictive model achieved accuracy ( R 2 = 0.77, MAPE = 7.33) under LODO validation, indicating that incorporating time delays in weather data allows for precise prediction of efficiency drops. The application of the evolutionary algorithm in a one-day numerical case study enabled the determination of optimal operational setpoints, indicating the potential for increased energy yield under difficult environmental conditions. Rather than a real-time autonomous control system, the presented solution constitutes a predictive optimisation framework and a functional module for a decision support system (DSS), enabling a transi - tion from reactive to prescriptive strategies.