The operation of wastewater networks is challenged by several factors, such as the rapid population growth and the increased load on the infrastructure due to changing weather conditions. A way to handle the increased load on the sewer infrastructure (without substantial capital investment) is to better utilize the available capacity via real-time modelling and control.

Assimilating data into the modelling and using real-time sensor and forecast feeds provide adaptation to changing conditions and the environment. Predictive control in wastewater networks is widely used, however, often underpinned by the fact that well-maintained models are required to set them into operation. This thesis is concerned with the data-driven modelling and predictive control of combined wastewater and stormwater networks.

Our contribution is to bridge the gap between hydraulic and hydrologic modelling, typically used individually as two disciplines but both necessary for predictive waste-water control. To tackle the challenges associated with the infiltrating flows, we propose to learn the dynamic effect of the wet- and dry-weather through the variations of easy commissionable level sensors distributed through the network. We adopt a Gaussian process-based control framework, where the Gaussian process model is obtained via training on the residuals generated between the real-time level measurements and the pre-identified physically-based model of the sewer hydraulics.

The practical feasibility of the closed-loop control has been tested in high-fidelity simulation environments and also on an experimental laboratory setup, while the prediction capabilities have been validated through the course of a pilot project with a real-world water utility. Compared to rule-based control methods, we achieved a cumulative overflow decrease ranging from 10 to 28 percent under extreme load over a one-month study period.

After the defence there will be a reception in the foyer at Fredrik Bajers Vej 7A – all are welcome!