Towards foundational (meta)models of water distribution networks with Graph Neural Networks
Speaker: Riccardo Taormina
Abstract: Water distribution networks (WDNs) and urban drainage systems (UDSs) are critical components of urban infrastructure, ensuring the availability of quality drinking water and the efficient removal of wastewater, respectively. These systems can be effectively modelled as graphs, where nodes represent junctions or manholes and edges symbolize connecting pipes or channels. This graph-based representation lays the groundwork for the utilization of Graph Neural Networks (GNNs).
After introducing water networks and their graph-based representation, we will delve into using GNNs for surrogate or meta modelling, with the aim of building foundational AI models for water network analysis. Resorting to simulations, rather than real data, is necessary due to several challenges associated with direct data acquisition from water networks. These challenges include the low density of sensors within the networks, high uncertainty, and the safety-critical nature of water infrastructure that severely hinders data sharing.
We will then showcase our recent advancements on using GNN approaches for surrogate modelling of WDNs. The discussion will focus on the use of edge-based GNNs, which outperform node-based alternatives as they align more closely with the underlying physical processes governing water distribution. Lastly, we explore the concept of transferability within our GNN models. Transferability is crucial for the development of foundational models as it ensures that the models can be applied across different networks without the need for extensive retraining. We find that edge-based GNNs show promising signs of transferability, suggesting that these models can serve as robust tools for predicting and managing water network behaviours across various contexts.
Faster and Transferable Urban Drainage Simulations with Graph Neural Networks
Speaker: Alexander Garzón
Abstract: Simulating urban drainage systems for tasks like optimization and real-time control can be slow and impractical with traditional models. This talk presents a novel approach using Graph Neural Networks (GNNs) to create faster, more efficient, and transferable metamodels. These metamodels leverage the system’s inherent structure to reduce training data and time, enabling rapid simulations even for large and complex drainage systems. Additionally, the GNN architecture allows for efficient transfer learning, meaning the model can be adapted to new unseen systems with minimal additional training data. This opens exciting possibilities for improved urban drainage management across diverse cityscapes and infrastructure conditions.
Relating complex network theory metrics with discolouration activity in water distribution systems
Speaker: Greg Kyritsakas
Abstract: Complex network theory (CNT) is an emerging research area that provides methods for representing interacting systems as networks. With CNT, a system is represented by a graph network where information is transferred to different entities represented as nodes, through several connections represented as edges. Water distribution networks (WDN), that consist of pipes (edges) connecting different nodes are suitable for CNT applications. Past work in this area, used CNT centrality metrics for understanding WDN complexity and for optimising their design, finding that edge betweenness is the main CNT metric for analysing them. These results open the possibility to explore centrality metrics as a tool that relates the complexity of the WDNs with degradation of the water quality. In this work, we examine the relationship between CNT metrics and discolouration parameters, such as iron, manganese, and turbidity, measured from samples taken in multiple WDNs of a UK utility. Once the CNT metrics are calculated, self-organising maps (SOMs) are applied for the identification of similarities and correlations between CNT metrics and water quality parameters. Based on these results, the discolouration risk of the utility’s pipes could be estimated. This metric could be then used for supporting decisions for WDNs’ proactive management and guarantee high quality water to consumers.