EPANET Hydraulic Model Overview

• EPANET Hydraulic Model is a computational tool that simulates steady-state and extended-period flows in pressurized water distribution systems using coupled nonlinear equations and conservation laws.
• It employs iterative methods like Newton–Raphson to solve nodal mass balance and head-loss equations, ensuring unique and physically correct solutions.
• The model supports calibration and operational optimization, with recent developments in physics-informed machine learning reducing simulation times by up to 100×.

The EPANET hydraulic model is a foundational computational framework for analyzing steady-state and extended-period flows in pressurized water distribution systems (WDS). EPANET's mathematical core rests on the solution of coupled nonlinear algebraic equations representing physical conservation laws in large-scale urban and industrial pipe networks. The model serves both as a reference for traditional hydraulic analysis and as a benchmark for the development of physics-informed machine learning emulators.

1. Mathematical Formulation and Governing Equations

A WDS is modeled as a connected directed graph $G = (V, E)$, in which $V$ denotes junctions (including reservoirs, tanks, consumer nodes) and $E$ denotes links (pipes, valves, pumps). EPANET’s steady-state hydraulic solver seeks a vector of nodal pressure heads $H = (H_i)_{i \in V}$ and link flows $Q = (Q_{ij})_{(i, j) \in E}$ satisfying the following system (Strotherm et al., 11 Jun 2025):

• Nodal Mass Balance (Continuity):

$$\sum_{j \in \mathcal N(i)} Q_{ji} = d_i\,,\quad \forall i \in V$$

where $d_i$ is the specified demand (positive for withdrawal, negative for supply) at node $i$.

• Head-Loss (Energy) Law for Each Pipe:

$$H_i - H_j = r_{ij} f(Q_{ij})\,,\quad \forall (i,j) \in E$$

with $r_{ij}$ the resistance coefficient (function of length $V$0, diameter $V$1, roughness $V$2) and $V$3 the empirical nonlinear head-loss function. EPANET supports both Darcy–Weisbach ($V$4) and Hazen–Williams ($V$5) formalisms:

$V$6

(Hazen–Williams form, in standard units).

The unknowns $V$7 are constrained by prescribed heads on reservoirs/tanks, specified demands $V$8, and network topology. The system thus contains $V$9 mass-balance and $E$0 head-loss equations linking $E$1 unknowns.

2. Existence, Uniqueness, and Well-Posedness

Rigorous guarantees of physically correct solutions are essential for model reliability. It has been proven that (Strotherm et al., 11 Jun 2025):

• Existence: Given a connected network, prescribed reservoir heads, and strictly increasing, continuous head-loss functions, there exists at least one set of nodal heads and pipe flows simultaneously satisfying all mass-balance and energy equations.
• Uniqueness: If in addition $E$2 is strictly monotonic (true for standard pipe laws), the solution is unique.

Proofs invoke monotonicity and convexity properties, showing that the residual mapping is coercive and continuous, which enables application of monotone operator theory and topological degree arguments. Only configurations with full boundary data (all reservoir heads and consumer demands prescribed) result in a nonsingular Jacobian for the Newton–Raphson method underlying EPANET’s solver. Improper specification (e.g., unconnected components or ill-posed sets of unknowns) can yield singular systems and algorithmic failure.

3. Computational Solution and Numerical Schemes

EPANET employs a Newton–Raphson-style iterative method to solve the nonlinear system. The implementation leverages the symmetric positive-definite character of the Jacobian (when using Darcy–Weisbach) or quasi-symmetric properties (Hazen–Williams), enabling fast (quadratic) convergence near the solution (Strotherm et al., 11 Jun 2025). Valid choices of unknowns are:

• Unknown heads at all consumer nodes plus flows on all edges.
• Unknown heads (eliminating flows via demand equations).
• Flows on a spanning-tree subset of pipes plus unknown node heads.

Proper initialization, e.g., via spanning-tree-based flow allocation, ensures a full-rank Jacobian on the first iteration. When boundary conditions are chosen to match the incidence matrix structure guaranteed by the uniqueness theorem, solver convergence is robust.

4. Integration in Practical Model Calibration and Operations

EPANET is widely deployed for both planning and operational optimization. In system calibration, measured flow and pressure data (often from SCADA and field monitoring) are used to tune pipe roughness coefficients by minimizing the deviation between measured and simulated heads across the network. For example, calibrated roughness values for various pipe material groups—such as εₘₛ𝑐ₗ = 10.6 mm (MSCL), ε_𝑔𝑟𝑝 = 2.91 mm (GRP), εᴅɪᴄʟ = 0.44 mm, and εₘₚⱽᴄ = 0.01 mm (mPVC)—were applied in a real network calibration (Wang et al., 2022). The calibration minimizes

$E$3

where $E$4 and $E$5 are observed and modeled pressures.

Pump setpoint optimization can be realized by iteratively updating pump controls to ensure customer service constraints (e.g., minimum pressure at all outlets) while minimizing energy consumption. For instance, dynamic adjustment of pump setpoints based on simulated outlet pressures and field-measured demand profiles resulted in 4.7% energy and greenhouse gas savings for a high-pressure irrigation system (Wang et al., 2022). Control rules were implemented via the EPANET [RULES] block or external scripting.

5. Surrogate Modeling: Machine Learning Emulation

Recent research has focused on physics-informed emulation of EPANET hydraulics using graph neural networks (GNNs), enabling rapid scenario analysis for large-scale networks (Ashraf et al., 2024). In such approaches, the network is represented as a graph where node and edge features encode demands, flows, and resistance coefficients. The emulator enforces mass continuity and head-loss constraints via a combination of local graph convolutions and global physics correction steps:

• Mass continuity at node $E$6:

$E$7

• Head-loss law per edge:

$E$8

The GNN recursively updates candidate flows and heads, applying max-aggregation and message passing, with a final correction step that reconstructs physically consistent heads and flows in closed form. Surrogate emulators match EPANET output to within 0.00–0.36% for demands and 0.002–0.015% for heads (MRAE), while achieving up to 100× acceleration on large networks (Ashraf et al., 2024).

6. Model-Building Guidelines and Best Practices

Based on theoretical and empirical analysis, the following best practices are advised when developing EPANET-based hydraulic models (Strotherm et al., 11 Jun 2025, Wang et al., 2022):

• Fix all reservoir and tank heads; do not treat them as free unknowns in the solve.
• Prescribe all nodal demands; solve for consumer heads and pipe flows.
• Ensure connectivity; avoid isolated network loops with zero demand.
• Use the Darcy–Weisbach law for strict symmetry and guaranteed positive-definite Jacobians.
• Initialize flows along a spanning tree to guarantee full-rank Newton iterations.
• Calibrate pipe roughness using time-averaged squared error between observed and simulated pressures.
• In operational deployment, implement feedback control/rule-based strategies for pumps through EPANET [RULES] or external automation.

7. Limitations and Extensions

EPANET’s standard formulation is restricted to steady-state hydraulics under the assumption of fixed (pressure-independent) demands, constant-head reservoirs, and standard pipe laws. Extensions—such as pressure-dependent demand models, explicit representation of pumps, valves, or leaks—require modifications to the mathematical structure or solution algorithm. Machine learning emulators currently replicate only core EPANET physics for pipes and reservoirs; extension to full asset classes or time-dependent operation remains an active research topic (Ashraf et al., 2024). Outlier errors in surrogate flows are rare (<5%) and have minimal impact on predicted heads or demands (Ashraf et al., 2024).

---

References:

• (Strotherm et al., 11 Jun 2025): "Existence and Uniqueness of Physically Correct Hydraulic States in Water Distribution Systems"
• (Ashraf et al., 2024): "Physics-Informed Graph Neural Networks for Water Distribution Systems"
• (Wang et al., 2022): "Improved Pump Setpoint Selection Using a Calibrated Hydraulic Model of a High-Pressure Irrigation System"