PhD Thesis

Contributions to the real-time monitoring and control of water systems

Work default illustration


  • Started: 08/03/2022
  • Thesis project read: 25/07/2022


This thesis presents several contributions to the state of the art of real-time monitoring and control of water networks.

These contributions are mainly focused on data-driven methodologies, which are characterized by their independence from hydraulic network models. Nonetheless, advancements has been also made in model-based and mixed model-based/data-driven strategies. The former relies on hydraulic simulations to gather knowledge of the network dynamics, whereas the latter reduce the dependence in the model through the integration of learning algorithms.

Regarding the field of monitoring, this thesis focuses on water distribution networks, responsible of delivering clean water to consumers. Two primary challenges are addressed: the localization of water leaks and the strategic placement of sensors.

About leak localization, an initial two-stage data-driven approach is proposed. First, a quadratic programming problem estimates the complete hydraulic network state, given by the hydraulic head (pressure + elevation) at the nodes. This process, denoted as Graph-based State Interpolation (GSI), only uses data from pressure sensors and the structure of the network. Then, a comparison between leak and leak-free estimated states is performed to indicate the leak location. Several improvements are conceived with the aim of strengthening the weak spots of the estimation and localization schemes. First, learning schemes are applied or combined with GSI to improve the node-level localization accuracy, leading to introduce GSI-DL, which uses Dictionary Learning to learn from the estimated residuals, LL-GSI, which adds an adaptive learning layer to GSI, improving the localization during the online application of the method, and DeepFGSI, which relaxes GSI to convert its solution to generate an interpolation matrix (FGSI), which is then used to add network information to the layers of a Deep Learning scheme. Then, the interpolation process is improved by considering the physics behind the dynamics of a water distribution network, leading to the development of the Analytical Weights GSI (AW-GSI) strategy. Finally, the availability of additional sources of hydraulic information in the network, such as flow sensors or demand meters, is not considered by base GSI to improve the estimation. Thus, several sensor fusion strategies are conceived, with the most prominent leveraging the Unscented Kalman Filter algorithm, leading to the UKF-GSI approach.

From the sensor placement perspective, a purely data-driven scheme employing genetic algorithms (GA) is developed to minimize a structural metric related to node-sensor distances. Further refinements include incorporating a hydraulic-based metric, derived from the state reconstruction errors obtained by comparing generated hydraulic state data and the estimated state through the set of sensors encoded by each individual of the population in the GA. The hydraulic data generation is performed through a UKF-based approach.