Machine Learning Supported Real-Time Ventilation Control in Sewer Networks

Sulfide-induced sewer corrosion is a critical issue faced by water utilities with a worldwide annual cost of many billions of US dollars. With a warm and humid climate and seawater toilet flushing, Hong Kong is most susceptible to sewer corrosion. Ventilation is among the key technologies widely used to mitigate sewer corrosion. It draws fresh air into, and foul air out of sewers, thus reduces the humidity and hydrogen sulfide levels in sewer air thus mitigating sewer corrosion. However, constant airflows are typically used, resulting in inadequate airflows in some periods and excessive airflows in others, leading to ineffective corrosion mitigation and/or waste of energy. Real-time control of sewer ventilation, based on rapidly changing in-sewer and atmospheric conditions, is required to address these limitations. The recent rapid development in low-cost Internet-of-Things sensors have made network-wide in-sewer monitoring possible, supporting a data-driven approach to real-time sewer ventilation control. The overarching objective of this project is to develop the framework and advanced data analytics to achieve real-time sewer ventilation control towards effective corrosion mitigation with minimised energy consumption.Adopting a novel model predictive control architecture, we will develop (1) predictive data analytics, by combining the traditional statistical models with the contemporary Recurrent Neural Networks, to predict key disturbance variables in the control horizon; (2) descriptive data analytics, using the Graph Neural Networks (GNN) making use of the structural similarity between a GNN and a sewer network, to predict the effect of control actions and predicted disturbance variables on the controlled variables and hence on the control objectives; (3) prescriptive data analytics, based on a combination of evolutionary algorithms and reinforcement learning, to generate and optimise the control policies. We will develop mechanistic-model-based digital twins of sewer networks to generate data, complementary to those collected from collaborating industry partners and from a pilot sewer, to support the above-described analytics. We will demonstrate the methodology, the models and the control algorithms using the pilot sewer and the digital twins.This is, to the best of our knowledge, the first study internationally on real time control of sewer ventilation, and involves the development of innovative descriptive, predictive, and prescriptive data analytics. The project outcomes will enable water utilities to achieve better corrosion control outcomes with reduced economic and environmental costs.With strong backgrounds in both sewer management and control engineering, the PI and Co-I are well positioned to lead the project to success.