Machine Learning Methods for Wastewater Treatment Plants

Modern wastewater treatment plants base their biological processes on advanced control systems which ensure compliance with discharge limits and minimize energy consumption by responding to information from on-line probes. The correct probe readings are particularly crucial for intermittent aeration controllers, which rely on real-time measurements of ammonia and oxygen in biological tanks. These data are an important resource for developing artificial intelligence algorithms that can identify process or sensor anomalies. However, using anomaly detection and classification algorithms in real-time wastewater treatment is challenging due to the multiclass and imbalanced nature of the problem, the difficulty in obtaining labeled data from real plants, and the complex and interdependent mechanisms that govern biological processes. This thesis introduces a solution that uses machine learning to detect anomalies within wastewater treatment plants, focusing on activated sludge compartments and systems that utilize intermittent aeration based on ammonia and oxygen measurements. The study analyzes the main anomalies that may arise in such systems (including both sensor inaccuracies and process-related issues), explores the features that can enable their detection using only common available measurements, and develops a multiclass classification model and suitable post-processing and automation strategies. Among the tested models, the best-performing were tree-based algorithms, particularly gradient boosting methods such as LightGBM. This model was implemented in real plants as a Decision Support System that can alert plant operators, and subsequently integrated into a new aeration controller that automatically reacts to events without the need of operator intervention, improving operational efficiency and reaching a recall of 82% and a precision of 75%. To address the scarcity of labeled data, an active learning methodology was employed, specifically uncertainty sampling, to iteratively select the most informative samples for annotation. This approach enabled efficient model adaptation to new plants with minimal labeling effort. Tests on operational plants showed significant improvements in anomaly detection, reducing labeling time and achieving optimal performance with only 6% of labeled data.