Design of a hybrid mechanistic/Gaussian process model to predict full-scale wastewater treatment plant effluent

Abstract

This paper presents the design of a hybrid model of a wastewater treatment plant (WWTP), which is meant to improve the quality of effluent prediction. By combining mechanistic, i.e. activated sludge model, and data-driven model it is expected to retain physical transparency and achieve good prediction accuracy. For the data-driven model, a state-of-the-art machine learning approach based on Gaussian process (GP) model was applied. GP models systematically address model uncertainty when lacking identification data and are applicable also for small data-sets, which both are encountered in WWTP modelling. Serial and parallel hybrid structures were designed to address the challenges of missing input data, insufficient mechanistic model accuracy and demanding model parameter estimation. Results of full-scale effluent predictions show that, by applying hybrid models, the accuracy of the model is improved. Good results were obtained also for default values of activated sludge model parameters, which significantly simplifies the model design process.