AIgua | Water consumption analysis and leak prediction

The Catalan region has faced increasingly severe droughts in recent years, with water reservoir levels reaching critical lows in 2024 and prompting a State of Emergency. While strict consumption restrictions were imposed on residents and businesses, a far more significant problem remained unaddressed: the region's inefficient water distribution infrastructure was losing over 50% of water before it even reached households.

This massive infrastructure wastage represented an urgent opportunity for conservation, especially during a water crisis. The challenge was to develop a system that could:

Traditional leak detection methods were labor-intensive, requiring physical inspection of kilometers of underground pipes—an impractical approach for systematic monitoring. A data-driven solution was needed to transform the newly available sensor data into precise leak locations.

I partnered with Aigües de Vic, the local water utility, to develop an innovative machine learning approach to leak detection. The core idea was to leverage Graph Neural Networks (GNNs) to model the city's water distribution system as a graph, with pipes represented as edges and junctions (along with households) as nodes.

The system development followed these key steps:

For the leak simulations, I used a specific technique where each potentially leaking pipe was modeled by dividing it into two segments and introducing a water-consuming junction between them. This allowed for precise control over leak locations and magnitudes in the training data.

The GNN architecture was particularly well-suited for this problem because it could capture the complex interdependencies in the water network. Water pressure and flow changes propagate through the network in complex ways that depend on the network topology—exactly the kind of relationships that graph neural networks excel at learning.

The implementation journey progressed through several stages, with significant work on developing an end-to-end data processing pipeline:

This sophisticated pipeline operated largely automatically, creating a continuous flow of processed data from real-world sensors to simulation environments and finally to the GNN-based analysis system. The ability to maintain this data flow without constant manual intervention was crucial for the project's feasibility.

The technical implementation utilized Python's ecosystem of data science and hydraulic modeling tools:

The GNN model itself was trained on the processed data to predict leak probabilities across the pipe network. It encoded the topological relationships between junctions, pipes, and sensors, learning to recognize the hydraulic signatures associated with leaks in different locations.

While the project showed promising results in controlled simulations, the transition to real-world deployment revealed significant challenges that ultimately prevented full implementation:

These challenges highlighted a fundamental issue in many data science projects: the gap between clean, theoretical models and messy real-world data. The hydraulic simulations that formed our training data represented an idealized version of the network that diverged significantly from reality.

The constantly evolving nature of the water network posed a particularly difficult challenge. As new construction expanded the city's infrastructure, the network model would quickly become outdated, requiring continuous updates to maintain accuracy.

Although the AIgua project couldn't be fully implemented in its intended form, it yielded several valuable outcomes:

While the GNN-based leak detection system demonstrated accuracy on simulated networks, the reality gap proved too significant for immediate deployment. This outcome illustrates the concept of a "successful failure"—a project that doesn't achieve its original goal but still delivers substantial value through the insights gained and problems identified.

The project highlighted the need for better data quality control and infrastructure modeling practices before advanced analytics can be successfully applied—insights that are themselves valuable for future water conservation efforts.

The AIgua project offered several important lessons for applying data science to infrastructure problems:

Perhaps the most important lesson was the recognition that foundational data infrastructure must be solid before advanced ML techniques can be effectively applied. No matter how sophisticated the algorithm, its performance is bounded by the quality of its inputs.

Based on the insights gained from this project, several promising directions for future work emerge:

The water conservation challenge in Catalonia remains urgent, and the fundamental approach of using Graph Neural Networks for leak detection remains promising. With improved data quality and more accurate network models, this approach could still deliver significant water savings and help build more resilient water infrastructure for drought-prone regions.

The techniques developed in this project could also be extended to other infrastructure monitoring applications, such as power grid management or transportation networks, where similar graph-structured data and physical processes exist.