Knowledge-enhanced data-driven modeling of wastewater treatment processes for energy consumption prediction
Louis Allen, Joan Cordiner
Abstract
Rising energy usage in wastewater treatment processes (WWTPs) poses pressing economic and environmental challenges. Machine learning approaches to model these complex systems have been limited by highly non-linear processes and high dataset noise. To address this, we introduce a novel Knowledge-Enhanced Graph Disentanglement framework for Energy Consumption Prediction (KEGD-EC) that leverages causal inference and graph neural networks. This work combines specific knowledge of causal relationships with a disentangled graph convolutional network architecture to facilitate accurate predictions. In a study on a WWTP in Melbourne, we demonstrate a 59.7% reduction in root mean squared error in energy consumption prediction using KEGD-EC compared to the next best model. We show that causal models built using domain knowledge outperform data-driven causal discovery models for complex systems. These results signify a step forward in applying machine learning to complex manufacturing processes, with the integration of causal knowledge into deep learning architectures posing a promising area of research for predictive analytics in manufacturing. • Rising energy use in wastewater treatment demands efficient solutions. • Our novel framework, KEGD-EC, combines causal knowledge with graph machine learning. • This enables us to reduce energy consumption forecasting error by 60.88% from previous work. • KEGD-EC remains reliable even with noisy data, making it viable for real-world use. • The impact of KEGD-EC includes lowering energy bills and promoting sustainable operation of WWTPs.