Urban stormwater and combined sewer networks are among the most topologically complex infrastructure systems managed by municipal engineers — tens of thousands of pipes, manholes, inlets, pumps, and overflows interconnected in a graph structure where hydraulic conditions at any node propagate through the network according to pipe gradient, diameter, roughness, and connectivity. Traditional approaches to overflow prediction have relied on threshold-based rules (e.g., trigger an alert when upstream gauge exceeds 80% pipe capacity) that fail to capture the spatial propagation dynamics that determine whether a local anomaly becomes a network-level event.
Graph Neural Networks (GNNs) offer a fundamentally different approach. By representing the drainage network as a directed graph — nodes are manholes and inlets, edges are pipe segments with hydraulic attributes (diameter, length, gradient, Manning's n) — a GNN can learn the spatial correlation structure of hydraulic conditions across the network topology. A spatiotemporal GNN adds temporal convolution across the time dimension, enabling the model to learn how hydraulic pulses propagate through the network during storm events.
This case study describes deployment of a spatiotemporal GNN (ST-GNN using ChebConv spatial convolution and TCN temporal layers) across a 2,400-node combined sewer network in a mid-sized UK urban catchment (population 180,000). Training data comprised five years of SCADA telemetry from 142 depth sensors and 18 flow meters, combined with NEXRAD-equivalent rainfall radar data at 500m resolution and 5-minute timestep. The model was trained to predict combined sewer overflow (CSO) events at the 34 permitted discharge points in the network at 6-, 24-, and 48-hour lead times.
Model performance at 48-hour lead time achieved a precision of 0.94, recall of 0.88, and F1 score of 0.91 on the held-out test set — substantially outperforming the operational threshold-based system (precision 0.61, recall 0.79) currently used by the utility. At 6-hour lead time (operationally most critical for public notification), precision reached 0.97 with recall of 0.93. False positive rate — a key operational metric given the cost of unnecessary field crew dispatch — was reduced from 22% to 4%.
The ST-GNN model is now integrated into the Smart Water Platform's Stormwater Intelligence Module, running in real-time inference mode on 15-minute SCADA telemetry updates. Predicted CSO events are georeferenced against the GIS network layer and trigger automated regulatory notification workflows to the Environment Agency's EMMA reporting system, satisfying the real-time reporting obligations of the EPR permit without manual operator intervention.