Sewer digital twins stop retraining per pipe: the FLOOD operator predicts across unseen pipes and networks, about 5,500-fold faster on one pipe
Synopsis
Sewer hydraulic digital twins have been limited by the cost of solving the Saint-Venant equations, while existing machine-learning surrogates need retraining for each new pipe or flow regime. FLOOD learns the solution operator of the Saint-Venant equations with a multi-branch neural operator, taking pipe geometry, hydraulic parameters and water levels at both manholes as inputs and returning depth and flow rate. On pipes excluded from training it reaches coefficients of determination of 1.000 for depth and 0.999 for flow rate, speeds up single-pipe online prediction by about 5,500-fold, and also serves roughness calibration and leakage localization.
Interpretation
A sewer hydraulic digital twin can now be trained once and applied to unseen pipes, networks and flow conditions without retraining for each pipe. Earlier machine-learning surrogates were typically case- and task-specific, requiring extensive retraining for new systems or whenever system characteristics changed, for example through ageing or renewal. The operator was trained only on synthetic data driven by simple sinusoidal inflows but tested on about 0.4 billion points from real infrastructure, covering dry- and wet-weather flows, pump-driven high-frequency fluctuations, backwater effects and severe surcharge that were not in training.
The operator maps pipe geometry, hydraulic parameters and water-level time series at both ends of a pipe directly to depth and flow rate at a queried location and time. The multi-branch neural operator splits the solution operator into a finite set of branches by diameter, 14 in this study following the Hong Kong design standard, each tied to one nominal diameter. Each branch uses a Fourier feature embedding layer followed by a single hidden layer, averaging about 280k trainable parameters; at inference the branch is selected deterministically from the pipe diameter.
The same operator also supports inverse and diagnostic tasks, including pipe roughness calibration and leakage localization with volume estimation. Conventional roughness calibration relies on non-differentiable Saint-Venant solvers and must approximate gradients through repeated full-domain simulations. On Pipes #3 and #4, excluded from training, the operator gave Manning roughness values of 0.025 and 0.015 against references of 0.025 and 0.016, with calibration speedups of 228,284-fold and 1,182,962-fold; on a 397-pipe Greater Bay Area network, five pipes with imposed infiltration took the top five leakage evidence scores, with leaked-volume relative errors below 2.5%.
Perspective
The demonstrated scope is one-dimensional Saint-Venant hydraulics in circular pipes, evaluated on three monitored pipes, one synthetic surcharge pipe, and two networks of about 485 pipes in Kowloon City, Hong Kong and about 397 pipes in the Greater Bay Area; leakage localization is a conceptual demonstration assuming water levels are measured at all manholes.
Under sparse monitoring, performance falls with sensor density: at 60-95% monitored manholes the method achieved more than a 500-fold speedup while maintaining a flow-rate coefficient of determination of 1.0 and a normalized mean absolute error below 0.7%; at around 40% it still achieved a 174-fold speedup; at 30% or lower the fidelity suits only applications accepting moderate accuracy. Training data come from simple sinusoidal inflows, and extension to non-circular conduits, control structures and sewer-overland interactions remains to be done.
