⚡ electrical engineering

CIDE: Calibration identiability-governed with structural diagnosis of water distribution networks

This study applies a calibration identifiability-governed framework (CIDE) to a water distribution network, demonstrating that while the model achieves high predictive accuracy for pressures and flows, the strong correlation between demand and leakage renders them structurally unidentifiable, necessitating external telemetry analysis to estimate physical leakage.

Original authors: Alexandre da Silva Gonçalves, Guilherme da Cruz dos Reis, Kamila Almeida dos Santos, Tatiane Souza Rodrigues Pereira, Klebber Teodomiro Martins Formiga

Published 2026-10-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alexandre da Silva Gonçalves, Guilherme da Cruz dos Reis, Kamila Almeida dos Santos, Tatiane Souza Rodrigues Pereira, Klebber Teodomiro Martins Formiga

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Water flows through miles of pipes to reach our taps, but the network delivering it is often a mystery. Engineers build digital maps of these systems to predict how water moves, where pressure drops, and how much is lost to leaks. These models are essential for fixing bursts, planning upgrades, and ensuring every household gets enough water. However, a digital map is only as good as the data fed into it. The pipes inside the ground may be older or rougher than recorded, the amount of water people use changes constantly, and hidden leaks can drain the system. When engineers try to tune their models to match real-world measurements, they face a tricky problem: different combinations of these hidden factors can produce the exact same result on the surface. A drop in pressure could mean a pipe is rough, or it could mean people are using more water, or it could mean a leak is sucking water away. Without a way to tell these causes apart, a model might fit the data perfectly while being completely wrong about what is actually happening underground.

A team of researchers from the Federal University of Goiás in Brazil tackled this puzzle in a real neighborhood called Guanabara, in the city of Goiânia. This area is fed by a single large turbine that controls the water pressure, acting like a giant valve that adjusts the flow throughout the day. The researchers wanted to know if they could use the data from this specific setup to figure out exactly how much water was being used versus how much was leaking, and how rough the pipes had become over time. They developed a new method called CIDE, which acts like a strict rulebook for calibration. Instead of letting the computer guess and adjust everything at once, this method forces the model to treat the known pressure settings as fixed facts and uses the water flow measurements as a single, honest check on the system. They split their data into two groups: two days to teach the model and six different days to test it, ensuring the model wasn't just memorizing the answers but actually learning how the system worked.

The results showed that the model could predict water pressure and flow with impressive accuracy. When tested on the days it had never seen before, the model's predictions for water flow were off by only about 1.04 liters per second, a small margin of error for such a complex system. The pressure readings were also very close to reality, with an average difference of less than one meter of water column. This proved that the model was useful for predicting how the system would behave under similar conditions. However, when the researchers tried to interpret the specific numbers the model gave for pipe roughness and leakage, the story changed. The model calculated a roughness factor that made sense, suggesting the pipes were indeed a bit rougher than the original records indicated. But when it came to leakage, the model insisted the amount was effectively zero.

This zero result was not a sign that the pipes were perfect. Instead, it revealed a fundamental limitation in what the data could tell them. The mathematical analysis showed that the model could not distinguish between water lost to leaks and water used by residents at night. The two factors were so intertwined in the data that increasing the estimated leak and decreasing the estimated night-time usage produced the exact same pressure readings. The model simply chose the path of least resistance, setting the leak to zero because the flexible demand numbers could absorb the error. To get a real sense of the leakage, the researchers had to step outside the model and look at the flow data directly. By analyzing the minimum flow during the night, they estimated that physical leakage likely sits between 6% and 32% of the total water entering the system. The wide range of this estimate depends entirely on how much water people are assumed to use at night, a detail the sensors could not measure.

The study concludes that while the model is excellent for predicting pressure and flow, it cannot definitively say how much water is leaking versus how much is being consumed. The researchers argue that the flow measurement at the inlet is the most critical piece of information, acting as the only true balance sheet for the system. If engineers try to force the model to find a specific leak amount using only pressure sensors, they risk getting a false answer. The real value of this work lies in knowing what the model can and cannot do. It can tell operators if the system is working as expected and if the pressure is sufficient for the neighborhood, but it cannot, on its own, pinpoint the exact volume of lost water without additional, independent measurements. The approach used here offers a clear, honest way to calibrate water networks, ensuring that decisions about repairs and resource management are based on what the data can truly support, rather than on a false sense of precision.

Technical Summary: CIDE – Calibration Identifiability-Governed with Structural Diagnosis of Water Distribution Networks

Problem Statement
Hydraulic calibration of water distribution networks (WDNs) traditionally relies on minimizing the difference between observed and simulated pressures and flows. However, this approach often fails to distinguish between parameters that produce similar system responses, such as nodal demands, pipe roughness, and leakage. When demand, leakage, and roughness cause collinear effects on measured pressures, or when flow data is used simultaneously to balance the system and evaluate the fit, the resulting model may achieve numerical closeness without physical adequacy. Furthermore, standard practices often reuse information (e.g., using inflow to reconstruct demand and then as an observation) or treat operational set-points as free variables, leading to over-optimistic error metrics and misinterpretation of causal mechanisms.

Methodology: The CIDE Framework
The authors propose the CIDE (Calibration Identifiability-governed with structural diagnosis) framework, applied to the Guanabara sector in Goiânia, Brazil. This sector is controlled by a pressure-regulating turbine, represented in EPANET as a pressure-reducing valve (PRV). The methodology follows a six-step workflow designed to separate information roles and govern interpretation through identifiability analysis:

Information Separation: Operational data is strictly categorized before inversion. The downstream pressure profile (set-point) is imposed as a fixed boundary condition via the PRV, not treated as a variable to be estimated. Inflow is included only once as an observation with associated error, preventing its dual use for demand closure and objective function minimization.
Data Partitioning: Two days (Saturday and Wednesday) are reserved for parameter estimation, while six subsequent days are used exclusively for out-of-sample validation without re-estimation.
Parsimonious Modeling: The hydraulic model incorporates Pressure-Dependent Demand (PDD), distributed leakage (proportional to pipe length and pressure), and a four-block demand shape. Parameters are shared across estimation days.
Robust Inversion: A nonlinear least-squares optimizer with a reflective trust region and a robust loss function (soft-L1) is used. Multiple starting points (reference state and Latin hypercube samples) ensure convergence to a global minimum.
Hydraulic Guards: Simulations must satisfy strict criteria: finite results, active control in 100% of hours, downstream pressure error < 0.25 m, and mass balance error < 0.02 L/s.
Identifiability Diagnosis: Post-convergence, a dimensionless Jacobian is computed, weighted by observational uncertainties and parameter ranges. Singular Value Decomposition (SVD) determines the effective rank and local parameter correlations.

Key Contributions
The paper's primary contribution is the protocol rather than a new optimization algorithm. Key innovations include:

Strict Role Assignment: Classifying the downstream pressure as a boundary condition and the inflow as a single observation prevents the calibration from compensating for structural errors via control variables or double-counting data.
Identifiability-Governed Interpretation: The framework explicitly checks whether the data can distinguish parameters. It accepts that a model can be predictively useful (low error) while being structurally uninterpretable (parameters are collinear).
Physical vs. Numerical Admissibility: The study separates numerical convergence from physical plausibility, using the SVD of the dimensionless Jacobian to identify when parameters are "weakly separable" (correlations > 0.99) even if the rank is full.
Independent Leakage Estimation: Instead of relying on the inverse problem to separate leakage from demand, the authors use a minimum-night-flow decomposition of the telemetry (independent of the inversion) to provide a physical estimate of leakage bounds.

Results
Applied to the Guanabara sector, the CIDE framework yielded the following:

Performance: The calibrated model achieved a Root-Mean-Square Error (RMSE) of 0.798 m for pressure and 1.040 L/s for flow during calibration. Validation RMSE for flow decreased by 31.9% compared to the unadjusted model, and the Nash-Sutcliffe efficiency was 0.830.
Parameter Estimates: The Hazen-Williams multiplier ( $m_C$ ) was estimated at 0.808, falling within an admissible range. However, the leakage term converged to a value below $10^{-3}$ L/s (effectively zero).
Identifiability: While the local rank was full (7/7), the correlation matrix revealed extreme collinearity (|r| ≈ 0.99) between the leakage term and the demand shape/scale parameters. Consequently, the partition between consumption and leakage was deemed unidentifiable. The model could predict aggregate flows and pressures, but could not causally assign the inflow between demand and loss.
Physical Leakage Bounds: Independent analysis of the minimum night flow indicated that physical leakage likely lies between 6% and 32% of the inflow, depending on the assumed fraction of legitimate night demand. This range is absorbed by the flexible demand terms in the inverse model, explaining why the calibration required no positive leakage term.
Service Assessment: The model identified that 18.7% of node-hours fell below the 15 m required pressure threshold, primarily during night modulation when the set-point is 7 m.

Significance and Claims
The paper argues that a low aggregate error does not authorize the interpretation of roughness, demand, and leakage as distinct physical properties if the data lacks sufficient independent excitation. The CIDE framework demonstrates that:

Predictive Utility $\neq$ Structural Interpretability: A model can successfully reproduce observed cycles and validate well on out-of-sample data while failing to uniquely identify the underlying physical mechanisms (e.g., leakage vs. demand).
Legitimate Outcomes: It is a valid and necessary outcome of a calibration process to issue a "structural diagnosis" where parameters are rejected or deemed uninterpretable, rather than forcing a parameter set that fits the data but lacks physical meaning.
Data Usage: The flow measured at the inlet is the critical information for arbitrating loss estimates, while pressure data validates the model structure. Using flow as a boundary condition in isolated sectors removes its ability to constrain the demand-leakage separation.

The authors conclude that for sectors with inlet flow measurement, the flow should enter the calibration once as an observation with error to fix the demand level, while pressures should be used to validate the model and locate structural errors. The study does not claim to have solved the leakage estimation problem for this specific network but rather provides a rigorous protocol to determine when such estimation is possible and when it is not.

More like this