Application of machine learning techniques to model a full-scale wastewater treatment plant with biological nutrient removal

A full-scale biological nutrient removal wastewater treatment process was simulated using artificial intelligence. In wastewater treatment plants, adaptive machine learning models can reduce process disruptions and generate savings through optimized operation. Machine learning is also useful when simulating processes that are particularly complex and where the physio-chemical interactions are not well understood, such as biological nutrients removal. Current models in literature only focus on the prediction of a small number of effluent parameters using a direct input-output approach. This paper presents a machine learning ensemble model that combines artificial neural networks, adaptive neuro-fuzzy inference systems, and support vector regression to predict 15 process parameters that include biomass properties, operation parameters, and effluent characteristics. A historical dataset between 2010 and 2020 was used to develop and validate the model. The model features a six-stage modular model structure where each parameter was predicted using a separate model and based on the preceding predicted parameters. The average correlation coefficient, normalized root mean square error, and symmetric mean absolute error of 69%, 0.06%, and 7.5%, respectively. The ensemble approach improved the average prediction accuracy over individual base models by 5%. The model developed in this study was more versatile than other machine learning models in the literature and relatively reduced the ambiguity of black-box data-driven models.

Automated summary

In this study, a full-scale biological nutrient removal wastewater treatment process was simulated using artificial intelligence. The researchers developed an ensemble model that combined artificial neural networks, adaptive neuro-fuzzy inference systems, and support vector regression to predict 15 process parameters. The model improved prediction accuracy and reduced ambiguity compared to other machine learning models.