water

Article
An Artificial Compressibility Method for 1D
Simulation of Open-Channel
and Pressurized-Pipe Flow
Ben R. Hodges
National Center for Infrastructure Modeling and Management, Center for Water and the Environment,
The University of Texas at Austin, Austin, TX 78758, USA; [email protected]




Received: 15 May 2020; Accepted: 12 June 2020; Published: 17 June 2020


Abstract: Piping systems (e.g., storm sewers) that transition between free-surface flow and
surcharged flow are challenging to model in one-dimensional (1D) networks as the continuity
equation changes from hyperbolic to elliptic as the water surface reaches the pipe ceiling. Previous
network models are known to have poor mass conservation or unpredictable convergence behavior at
such transitions. To address this problem, a new algorithm is developed for simulating unsteady 1D
flow in closed conduits with both free-surface and surcharged flow. The shallow-water (hydrostatic)
approximation is used as the governing equations. The artificial compressibility (AC) method
is implemented as a dual-time-stepping discretization for a finite-volume solver with timescale
interpolation used for face reconstruction. A new formulation for the AC celerity parameter is
proposed such that the AC celerity matches the equivalent gravity wave speed for the local hydraulic
head—which has some similarities to the classic Preissmann Slot used to approximate pressurized
flow in conduits. The new approach allows the AC celerity to be set locally by the flow (i.e.,
non-uniform in space) and removes it as a free parameter of the AC solution method. The derivation
of the AC method provides for only a minor change in the form of the solution equations when
a computational element switches from free-surface to surcharged. The new solver is tested for both
unsteady free-surface (supercritical, subcritical) and surcharged flow transitions in a circular pipe
and is implemented in an open-source Python code available under the name “PipeAC.” The results
are compared to laboratory experiments that include rapid flow changes due to opening/closing of
gates. Results show that the new algorithm is satisfactory for 1D representation of unsteady transition
behavior with two caveats: (i) sufficient grid resolution must be applied, and (ii) the shallow-water
equation approximations (hydrostatic, single-fluid) limit the accuracy of the solution with regards to
the celerity of the turbulent unsteady bore that propagates upstream. This research might benefit
any piping network model that must smoothly handle unsteady transitions from free surface to
surcharged flow.

Keywords: Saint-Venant equations; surcharged pipe; Preissmann slot; stormwater model; SWMM

1. Introduction
There is an ongoing need for city-wide models of stormwater networks to facilitate real-time
modeling and operational control [1–3], design for urban growth [4–6], low-impact development [7–9],
flood mitigation [10–12], runoff water quality [13], and address impacts of the changing climate [14,15].
Such models must be computationally efficient to allow a large number of simulations for
calibration [16], design optimization [17,18], Monte-Carlo uncertainty analysis [19], and simulation
of stochastic rainstorms instead of a single design storm [20,21]; however, the models must also be
detailed and comprehensive is their treatment of the piping system for an entire city [22]. Two major

Water 2020, 12, 1727; doi:10.3390/w12061727 www.mdpi.com/journal/water

Water 2020, 12, 1727 2 of 33

challenges for system-wide modeling of municipal stormwater networks are: (i) numerical stability
near a pressure celerity shock occurring at the transition between open-channel and surcharged
flow [23,24], and (ii) efficient parallel computation for large systems [25,26]. This paper focuses on
the modeling the transitions, with best practices for parallel computation considered in the algorithm
design. The efficacy of a new method for simulating one-dimensional (1D) unsteady transitions
between supercritical, subcritical, and surcharged flows is presented in detail herein. Analyses of
parallel performance reserved for future work.
The new algorithm uses the concept of “Artificial Compressibility” (AC) whose roots are in
Chorin’s classic approaches for treating the combined hyperbolic/elliptic nature of the incompressible
Navier-Stokes equations [27,28]. The AC method has been demonstrated as effective for parallel
solution of two-dimensional (2D) and three-dimensional (3D) incompressible flow (e.g., [29,30]).
The present work appears to be the first providing an AC method for solving 1D closed-conduit water
flow. Beyond this new application, other novel contributions include (i) use of the recent “timescale
face reconstruction” method [31] applied to surcharged pipe flow, (ii) a formulation of the artificial
compressibility parameter that varies in space to ensure pressure celerity is at gravity-wave speeds,
(iii) an implicit time-linearized friction term built into the Runge-Kutta 4th-order (RK4) time-march
that ensures flow reversals have friction opposing the end-of-step flow direction, and (iv) the ability to
handle unsteady transitions between open-channel and surcharged sections with minimal changes in
the numerical solution algorithm.
The algorithm developed herein is called “PipeAC,” and is based on a 1D finite-volume
formulation of the shallow-water and incompressible flow equations with added artificial
compressibility terms. A 3-level backwards implicit stencil is used for real-time discretization [32].
Runge-Kutta 4th-order explicit time-stepping is used for time-marching in pseudo-time. A timescale
spatial discretization is applied for interpolations from finite-volume elements to faces [31]. The present
code is tested for circular pipes with open channel (subcritical, supercritical) and surcharged flow, but
the underlying algorithm is not strictly limited to pipe flow. PipeAC is coded in Python (v3.7) and is
available as public domain software [33]. Investigation of the practical implementation of the method
in a large network requires further development of an efficient compiled code, which is an ongoing
project of the NCIMM research team in developing a new computational engine for the widely-used
Storm Water Management Model (SWMM) [31,34].
This paper provides a background on modeling in Section 2, which motivates using artificial
compressibility for pipe network simulation. In Section 3, the governing equations for finite-volume
artificial compressibility are derived. The numerical approach of PipeAC is described in Section 4.
Conditions for unsteady flow test cases developed from the experiments of Trajkovic Ivetic et al. [35] are
outlined in Section 5. Results and discussion are provided in Sections 6 and 7, respectively. Conclusions
in Section 8 summarize the present findings.

2. Background

2.1. Challenges of Surcharged Pipe in a Stormwater Drainage Network

Stormwater systems are generally designed to carry the hydrological runoff from a “design storm”
as a gravity-driven flow with a free surface. Design conditions requiring a full pipe, i.e., pressurized
or “surcharged,” are usually reserved for a few locations where pipes are either deeply buried or
the pipe size for free-surface flow is impractical. However, the spatial and temporal variability of
many real-world storms will exceed the design storm and cause surcharged conditions at a variety
of locations across a system. Surcharged pressures can lead to water levels that back up through the
curb inlets and cause flooding on the street, so modeling the occurrence and system behaviors for the
transitions between free-surface and surcharged flow is critical to many of the stormwater modeling
applications discussed in Section 1, above.
Water 2020, 12, 1727 3 of 33

Where storm sewers flow with an unpressurized free surface the hydrostatic Saint-Venant
equations provide the relationships between mass and momentum transport. In contrast, closed
conduit flow (i.e., surcharged pipe) can be considered either compressible flow for transient
analysis [36,37] or as a series of quasi-steady incompressible flow solutions as in the gradient
algorithms used in the widely-used EPANET code for water distribution systems [38,39]. A challenge
for large-scale network modeling is that rapid events can be significant to operations [40–42],
but are difficult to represent in system-wide models [43–45]. Modeling pressure transients in
a surcharged pipe moving at acoustic velocities requires numerical methods that are stable at large
Courant-Friedrichs-Lewy (CFL) numbers or use very small time steps [46,47]. Such transient solvers
have yet to be demonstrated capable of solving large stormwater networks, so efficiently handling
the incompressibility approximation—as in the widely-used SWMM code [48]—remains an ongoing
challenge. The problem with incompressibility is that it treats the pressure celerity as infinite rather
than as an acoustic speed. To maintain mass conservation throughout a system the pressure wave
indicating that mass moves into a surcharged pipe section must be instantaneously transmitted through
all the connected surcharged sections. In effect, the use of the incompressibility constraint neglects
nonlinear advection for 1D equations in the surcharged pipe; i.e., u∂u/∂x = 0 for velocity (u) in
a uniform diameter pipe with incompressible flow.

2.2. Hyperbolic (Free Surface) vs. Elliptic (Surcharged) Continuity Equations

Stormwater network models must be able to handle both surcharged and free-surface flows.
Invoking the incompressibility constraint at the boundary between a section of free-surface flow and
a section of surcharged pipe flow has the mathematical effect of changing the continuity equation from
hyperbolic to elliptic, which can be readily illustrated with two common 3D forms of mass conservation:

∂η ∂η ∂η ∂u ∂v ∂w
hyperbolic: +u +v =w , elliptic: + + = 0,
∂t ∂x ∂y ∂x ∂y ∂z

where u,v,w are Cartesian velocities in x,y,z directions and η is a free surface elevation. The hyperbolic
form is derived from continuity and the kinematic boundary condition (integrated over depth) [49].
In contrast, the elliptic form (without a free surface) arises from requiring a divergence-free velocity
field for an incompressible flow [50]. The hyperbolic equation is prognostic as it can be integrated
forward in time from some known past state to predict a future state. In contrast, the elliptic equation is
diagnostic in that it is a constraint that must be instantaneously satisfied everywhere at every instant in
time; i.e., it cannot be simply integrated forward in time but is inherently both a local and (integrated)
global constraint on the fluid behavior. In contrast to continuity, the conservation of momentum is
inherently prognostic for both free surface and surcharged flows; thus, coupling the hyperbolic free
surface continuity with momentum provides an equation set that is marched forward in time for both
velocity and volume, whereas coupling the elliptic incompressible continuity with momentum requires
solution of a time-marching equation for velocity with a local/global volume constraint.
The distinction between hyperbolic and elliptic forms is important as numerical solution methods
for hyperbolic problems generally perform poorly on elliptic problems and vice-versa. The hyperbolic
equation can be thought of as providing “local” solutions, where effects at point x propagate
some limited distance ∆x in time step ∆t based on celerity C , whereas the elliptic equation are
“non-local” requiring any change at x must instantaneously affect the entire domain. There are two
canonical approaches to this problem for stormwater systems: (i) introduction of an approximate
hyperbolic equation for the incompressible pipe flow, and (ii) switching the governing equations to the
incompressible form at the boundaries between free-surface and surcharged flow. The “Preissmann
Slot” [51,52] is an example of the former approach and the head perturbation equation for surcharged
flow in SWMM5 is an example of the latter (see Equations (3)–(25) in Rossman [48]).
Water 2020, 12, 1727 4 of 33

2.3. The Preissmann Slot: A Hyperbolic Approximation of Incompressible Surcharged Pipe

The Preissmann Slot approximates a surcharged closed conduit as a pipe with narrow, longitudinal
slot at the crown that has walls extending infinitely high on either side. This imaginary pipe-slot
combination can never be surcharged. Instead, there will always be a free surface somewhere in
the slot whose gradients provide the driving force in the surcharged pipe. An advantage of this
approach is that the Saint-Venant equations apply in both free-surface and surcharged sections [53,54],
which theoretically allows a smooth solution across the transition. However, a pressure celerity shock
at the transition may still exist if (i) the hydrostatic pressure rises rapidly in the surcharged section,
or (ii) the transition from open to surcharged includes a hydraulic jump. Furthermore, the slot alters
local mass conservation by allowing a small amount of water to be stored/transported in the slot
volume above the pipe; however, global mass conservation can be sustained if the underlying algorithm
is mass conservative. The key algorithmic point is that pressure celerity in the Saint-Venant solution
will be at a wave speed that is slower than the acoustic velocity. As noted in Cunge and Wegner [51],
the wave speed for the Preissmann Slot is a function of gravity, the pipe cross-sectional area, and the
slot width.
Numerical approaches that directly use the incompressibility constraint might seem less ad hoc
than the Preissmann Slot, but that viewpoint is arguable. Incompressibility replaces the very fast
(but finite) real-world pressure celerity with an infinite celerity—an instantaneous transmission of
information that is clearly non-physical. The Preissmann Slot treats the pressure celerity by slowing it
down below the acoustic speed—while clearly not the correct speed it is, at least, a plausible physical
speed. In effect, the incompressibility approximation imagines the water as a solid that can be pushed
through a pipe without deformation whereas the Preissmann Slot imagines the fluid as somewhat
more compressible than water. Thus, incompressibility and the Preissmann Slot are merely alternative
approximations of the true pressure celerity. Neither is correct, and whether a slower celerity or
an infinite celerity is preferable depends on the problem being investigated, the numerical algorithm,
and the types of errors that are acceptable. Interestingly, the Artificial Compressibility (AC) method
provides some insights into the relationship between the Preissmann Slot and incompressible flow
methods, as discussed below.

2.4. Artificial Compressibility for a Hyperbolic Continuity Solution That Is Exactly Incompressible
Chorin [27] proposed the Artificial Compressibility (AC) method as a way to change the form of
the steady incompressible equations from elliptic to hyperbolic for a more tractable numerical solution
while still solving the incompressible equations. Chorin’s innovation was to see that artificial unsteady
terms could be added to the steady-state incompressible continuity and momentum equations to
create equations that could be time-marched in pseudo-time, τ, i.e., adding terms of the form ∂Ψ/∂τ,
where Ψ is merely a place-holder for solution variables in the present exposition. The steady-state
incompressible solution is recovered by marching the AC momentum and continuity equations forward
in pseudo-time until ∂Ψ/∂τ = 0. Thus, the AC approach creates a prognostic equation for continuity
in pseudo-time that will satisfy the diagnostic constraint of incompressibility when converged to the
pseudo-time steady state.
The steady-state incompressible AC method of Chorin [27] was later extended to both
compressible and unsteady flows [55–57]. For unsteady flows the AC method is sometimes known as
a “dual time-stepping” approach [58] as it includes a time-march in real-time and a separate time-march
in pseudo-time. The “outer-loop” solution over real-time step ∆t for step n to n + 1 requires some
number of “inner-loop” iterations in ∆τ, where these inner-loop iterations are a consistent set of
hyperbolic equations that are not subtime steps of ∆t. That is, ∑ ∆τ > ∆t is allowed (although not
required). For a fully-converged solution, the AC method is theoretically equivalent (to machine
precision and numerical truncation error) to any other numerical solution of the incompressible flow
equations—which is not true for the Preissmann Slot.
Water 2020, 12, 1727 5 of 33

Both the Preissmann Slot and AC methods use solutions of hyperbolic equations for time-marching,
the difference is that the AC method can be converged to ensure local mass-conservation (through
inner iterations) whereas the Preissmann Slot will be locally non-conservative with a single step due
to the change in storage in the slot. However, if inner-loop iterations are cut off when ∂Ψ/∂τ > 0,
mass will not be exactly conserved in the AC method. Interestingly, an argument can be made that the
Preissmann Slot is effectively the first ∆τ step of an AC method where the AC parameter is tuned to
the same celerity implied by the Preissmann Slot width. Alternatively, one might propose an inner
iteration of the Preissmann Slot to obtain a method similar to AC. This speculative idea revisited in the
discussion in Section 7.3 as it perhaps is easier to see after the new ideas for the AC celerity parameter
are presented in Section 3.3.

2.5. The Relationship between the Artificial Compressibility Method and Elliptic Solvers
The AC method is an alternative to requiring an elliptic solver for the coupled diagnostic
continuity and prognostic momentum equations of unsteady incompressible flow. The approaches
are mathematically distinct but share some similarities [59,60]. The elliptic problem implies a matrix
inversion over the entire domain, which requires some inner-loop iterations in a time-marching
solution. Similar to the AC pseudo-time stepping, each inner-loop iteration in an elliptic solver is
an approximate solution with some error that is characterized by failure to exactly conserve mass
(i.e., a residual). Both AC and elliptic methods can theoretically converge to conservation at machine
accuracy, but typically inner-loop iterations are stopped when the solution reaches some acceptable
measure of the global residual. The key difference between the methods is that an elliptic solver uses
the characteristics of the matrix to determine the solution perturbation at each iteration whereas the
AC approach propagates information through the domain as artificial pressure waves. It can be argued
that the artificial pressure waves are more tractable for the mathematically stiff systems of stormwater
networks where (i) surcharging is variable in both space and time, and (ii) identifying well-posed
boundary conditions at each open-channel/surcharge interface can be a problem.
The AC approach is appealing because it is (i) simple to code, (ii) can use widely-tested hyperbolic
numerical schemes, (iii) does not require a matrix inversion, and (iv) is stable as long as the artificial
compressibility parameter is chosen carefully and the pseudo-time-stepping satisfies a CFL condition.
However, the AC method can require small steps in pseudo-time, which can be computationally
expensive—especially in an unsteady scheme with dual time-stepping. In general, over some interested
real-time duration t1 to t2 an AC method will require more total outer and inner-loop iterations than
an elliptic solver (hence more floating point operations) to reach a desired convergence norm. In the
1980s, elliptic solvers for the incompressible flow equations received a boost from Patankar’s SIMPLE
algorithm [61], which has roots in Chorin’s projection method [28] as a splitting form of incompressible
solution. Adoption of efficient elliptic solvers was supported by the LINPACK and later LAPACK
software libraries for matrix solutions. While AC methods never entirely disappeared from the
literature, their advancement for incompressible flow was overwhelmed by advances in elliptic
solvers through the 1990s and early 2000s. Nevertheless, the overall simplicity and hyperbolicity of
AC schemes saw their continued use well into the 1990s in adaptive multigrid [62] and unsteady
flows [63–65].

2.6. Recent Development of the Artificial Compressibility Method

It appears that Sotiropolous in the late 1990s was the first to note that AC methods hold particular
promise for vector and parallel computing of incompressible flows [30]; an idea that was further
developed in a series of papers over a decade [66–68]. For parallel computing, the advantage of
AC is that it replaces the global solution of an elliptic matrix inversion with local hyperbolic wave
propagation. This “local” limitation means that communication between processors is readily predicted
and controlled—A critical issue today because the death of Moore’s law for computer clock speed [69]
implies that high-performance computing on large problems is limited by the communication
Water 2020, 12, 1727 6 of 33

bandwidth between processors rather than the number of floating point operations [70]. Indeed,
The rapid increase in parallel computing has also seen a resurgence of interest in AC methods since 2004,
e.g., [71–73]. Research has included multifluid flows [74,75], compressible flow [76], incompressible
flow [77], shear-thickening and shear-thinning fluids [78], and a link-wise discretization similar to
Lattice-Boltzmann methods [79]. Foundations for characteristic-based methods were developed as
an an early part of this work [80–82], and have been studied for higher accuracy discretizations [83,84].
A relatively new addition has been an entropically-damped formulation [85], which substitutes
a thermodynamic approximation for the simple artificial compressibility of Chorin [27].
Indeed, from these rather modest foundations the research using AC methods has exploded over
the last five years. More than 200 papers were retrieved by database searches for 2015–2020 using
the terms artificial compressibility, pseudo compressibility, and dual time stepping. These studies
include analysis of the relationships between AC and fractional step methods [86,87], studies of
turbulent flow [88–90], development of discontinuous Galerkin methods [91–93], parallel solution of
large-scale problems [94–96], multi-fluid simulations [97], further advances to entropically-damped
AC methods [98–100], implementation high-order numerical schemes [101–103], use of characteristic
methods [104–106], application for non-hydrostatic effects [107,108], magneto-hydrodynamic
simulations [109–111], solution with lattice Boltzmann methods [112], and solution by smoothed
particle hydrodynamics [113,114]. Note the above are examples and should not be considered an
exhaustive list of recent work. For those seeking deeper theoretical understanding, the mathematical
foundations of AC are derived in detail by Lemarie-Rieusset [115].
In summary, the AC method is an established numerical approach with a long history in 2D and
3D incompressible flow simulation. Arguably, the AC method holds potential for being efficient for
solution of large 1D pipe networks solved on massively-parallel computers. However, only a single
work, that of Bassi et al. [116], uses the AC method in a 1D model, and none of the applications
in the literature consider the use of AC for transitions between free-surface and pressurized flow
(as studied herein).
The remainder of this paper demonstrates that the AC method can be effectively used to model
the unsteady transitions between supercritical, subcritical, and surcharged flows that commonly occur
in stormwater systems.

3. Governing Equations

3.1. Finite-Volume Continuity and Momentum

To avoid the complexities of dealing with the short time and space scales of non-hydrostatic
pressure [117], models for free surface flows commonly use the “shallow water” equations that include
the Boussinesq, hydrostatic, and incompressibility approximations, e.g., [118]. Following the approach
of Hodges [119], the incompressible, hydrostatic, 1D equations for mass and momentum conservation
can be written for a finite volume with a free surface as
∂Ae
Le = Qu − Qd + qe Le , (1)
∂t
Z
∂Qe
Le = β u Qu Uu − β d Qd Ud + gAu ηu − gAd ηd + g ( Ad − Au ) η ( x )λ( x ) dx − gVe S f (e) , (2)
∂t L

where Q is a flow rate, U is a corresponding average velocity, A is a cross-sectional flow area, η is

the water surface elevation, g is gravity, L is an element length, q is a leakage inflow per unit length
along a volume, and subscripts e, u, d represent the finite-volume element, the nominal upstream face,
and the nominal downstream face, respectively. The coefficient β is the standard nonlinear velocity
integration coefficient (e.g., [119]). A uniform approximation of β = 1 is typically used for 1D models,
and is adopted herein so that β can be dropped from subsequent equations. The term S f in Equation (2)
is the friction slope that represents the normalized energy dissipation rate over the finite-volume
element. The η ( x )λ( x ) term is a quadrature function introduced by Hodges [119] that represents the
Water 2020, 12, 1727 7 of 33

nonlinear effects of spatial variability of both the free surface and the conduit or channel shape on the
integrated piezometric pressure over a finite-volume element. By definition, λ is a weighting function
that must satisfy
Z
λ( x ) dx = 1.
L

The choice of different ηλ discretizations leads to different approximations of the effects of

spatially-varying piezometric pressure over a control volume. Derivation of the above is presented in
detail by Hodges [119].
The above are a finite-volume form of the Saint-Venant equations for free-surface flow, but they
can be readily converted to equations for pressurized pipe flow if we neglect non-hydrostatic pressure
and consider a piezometric pressure defined as P ≡ ρgH. Thus, by simply substituting η = H
we arrive at an equivalent set of equations that apply to a finite volume of pressurized pipe with
piezometric head H.

3.2. Artificial Compressibility for Continuity

The AC method introduced by Chorin [27] can be represented as a pseudo-time derivative of
density added to the governing equations. For continuity this becomes:

1 ∂
Z
∂Ae
ρ̃ dV + Le = Qu − Qd + qe Le (3)
ρ0 ∂τ V ∂t

where ρ0 is the reference or “incompressible” density of the fluid, ρ̃ is an artificial density, and τ is
pseudo-time. A more recent alternative approach is the “entropically-damped artificial compressibility”
(EDAC) formulation that adds a scaled Laplacian of pressure to continuity [85,98,99]. For simplicity,
the present work will use the classic AC approaches of Chorin [27] and Rogers et al. [32] that are
consistent with Equation (3).
The AC method requires that as τ → ∞ there must be a steady-state (in pseudo-time) such that
∂/∂τ → 0 and incompressible continuity, Equation (1), is recovered. As described in Breuer and
Hänel [120], we introduce an artificial equation of state p̃ = γ2 ρ̃, where p̃ is a pressure and γ is the
artificial speed of sound (units of LT−1 ), i.e., the propagation speed of a pressure disturbance across
pseudo-time τ. Note that the literature generally uses β instead of γ in the artificial equation of state,
but we reserve β for the momentum coefficient in Equation (2), which is common nomenclature for 1D
open-channel flows. It follows that the pseudo-time derivative in Equation (3) can be written as

1 ∂
Z
p̃ dV.
ρ0 γ2 ∂τ V

As ∂/∂τ → 0 we require p̃ → p + c, where p is the true pressure and c is a constant, which implies
spatial gradients are preserved, i.e., ∂ p̃/∂x = ∂p/∂x. Note that c = 0 will generally be obtained when
the boundary conditions for p̃ are Dirichlet. Consistent with the open-channel Saint-Venant equations,
the combination of hydrostatic pressure and potential energy can be represented as piezometric
pressure (i.e., ρgη), so an artificial piezometric pressure can be defined as P̃ ≡ p̃ + ρ0 gz where z is
the vertical elevation above the baseline z = 0. It follows that p̃ = P̃ − ρ0 gz. In general, it will be
convenient to work in terms of an artificial piezometric head (i.e., the equivalent static water elevation,
or η, in a free surface flow), i.e., H̃ ≡ P̃ (ρ0 g)−1 . With the above, Equation (3) for continuity becomes

gLe ∂
∂Ae
2
H̃e Ae + Le = Qu − Qd + qe Le (4)
γ ∂τ ∂t

Although the above was originally derived for an open-channel flow, it becomes an equation for
a closed-conduit flow by requiring ∂Ae /∂t = 0.
Water 2020, 12, 1727 8 of 33

3.3. Choosing an Artificial Speed of Sound

The choice of γ, the artificial speed-of-sound (or pressure celerity), in Equation (4) plays
a role in the robustness of an AC method. Typically γ is taken as a uniform, constant parameter
(e.g., [32,60,66,121]). Indeed, a global γ = 1 is often a default in artificial compressibility studies
(e.g., [30,56,122,123]), although higher values are also found (e.g., [62,78,124]). A wide range of
studies fail to document the values used for γ; likely because it formally controls convergence rates
rather than accuracy (e.g., [73,77,81,83,120])—although it can also affect the simulation result for
upwind schemes [124]. Methods have been developed to select an optimum global γ based on
flow characteristics and/or model time step (e.g., [63,64,125–127]), or using mesh spacing [72], or as
a preconditioning parameter for a solution matrix [128]. For the present purposes, a key observation
is that from a mathematical standpoint γ is a coefficient of the ∂/∂τ term that is an ad hoc addition
to continuity; thus, γ is not related to any gradients in space and can be considered an arbitrary
and independent value for any control volume. Although not expressly stated, this idea is implicit
in [74] where γ is a local dynamic function of the flow for an “artificial fluid” with “pseudotime
stretching.” For the present work, it is useful to follow [74] in treating γ as a local function of space
and pseudo-time, which we will represent as γe for a finite-volume element.
Let us introduce a length scale ` that is a local function of the flow geometry and pressure
head (as described further below) and hence denoted as `e for a finite-volume element. Multiplying
continuity, Equation (4), by `e /Le , provides

Qu − Qd
 
g`e ∂
∂Ae
H̃e Ae = `e − + + qe . (5)
γe2 ∂τ ∂t Le

Although `e is arguably arbitrary, defining it as a modified hydraulic depth (varying with both
pseudo-time and location) has some particular advantages. We will let `e represent the hydraulic
depth associated with an imagined free surface flow for η = H̃ where the closure of the conduit is
neglected. To elaborate, let z(c)e be the crown height of a closed-conduit element (i.e., the elevation of
the top inside that is representative of the entire element). Define B(m)e as the maximum breadth of the
conduit, which occurs at some elevation z(m) near the conduit midpoint. Let Be be the actual breadth
(topwidth) at the free surface when H̃e < z(c)e . Furthermore, let A(m)e represent the cross-sectional
area of the conduit below the location of maximum breadth, z(m) . It follows that a modified hydraulic
depth can be defined as

Ae Be−1
(
H̃e < z(m)e
`e ≡ h  i (6)
H̃e − z(m)e B(m) + A(m)e B(−m1)e H̃e ≥ z(m)e

The above assigns `e for a free surface flow below the maximum conduit width using the
traditional definition of hydraulic depth. However, for a free surface flow above the maximum
conduit width and for any closed conduit flow, `e is the hydraulic depth as imagined for a flow with
free surface ηe = H̃e in an open channel with vertical sidewalls beginning at the widest elevation of
the conduit. The purpose of this approach is that we can define a local non-dimensional number Fe for
each element as
γe
Fe ≡ p (7)
g`e

where Fe is a form of Froude number relating the AC pressure celerity (γe ) to the gravity wave speed
at hydraulic depth `e . We can now write the AC continuity equation as

Qu − Qd
 
∂ ∂Ae
H̃e Ae = `e Fe2


− + + qe . (8)
∂τ ∂t Le
Water 2020, 12, 1727 9 of 33

In dealing with liquid flow in a pipe, it is reasonable to take Fe = 1, which is equivalent

p
to γe = g`e . That is, the AC pressure celerity is selected to locally slow the speed of acoustic
pressure waves in pseudo-time to the speed of an equivalent gravity wave on hydraulic depth `e .
The advantage of this approach is it automates the selection of γe and ensures a local value that
is consistent with the value H̃e , which thus allows smooth transition from open-channel flow to
surcharged flow. This approach also makes it trivial to choose a pseudo time-step size that satisfies the
local CFL condition for the artificial pressure wave.

3.4. Artificial Compressibility for Momentum

Derivation of the AC equation for momentum is simpler as it requires adding only a pseudo-time
derivative of the integrated momentum, Qe Le , to Equation (2) and replacing the piezometric head
(η) with the AC head ( H̃). The integrated piezometric pressure term involving quadrature of η and
λ can be represented by the T(2,0) approach of Hodges [119], which approximates the piezometric
pressure distribution across a finite-volume element by a parabola with a uniform λ( x ) = L− 1
e .
This approximation modifies the coefficient of the gA H̃ terms in Equation (2) and provides a source
term of
g


Ke ≡ Ad H̃u + 4 H̃e − Au H̃d + 4 H̃e . (9)
6

The friction term ( gVS f ) from Equation (2) is represented by a model for the friction slope of
S f = FU |U | or gVS f = gFLQ |U | where F is a dimensional friction coefficient that depends on the
type of turbulence model applied. Basic friction models are
( ñ2
: Chezy-Manning model
R4/3
F≡ h
fP
(10)
2gDh : Darcy-Weisbach model

where ñ is the Gauckler-Strickler-Manning open-channel friction coefficient, Rh is the hydraulic radius,

f P is the Darcy friction factor for pipe flow, and Dh is the hydraulic diameter. For the present version
of PipeAC, only the Chezy-Manning model is used.
With the above substitutions, Equation (2) can be written as
( )
∂Qe ∂Qe 1 5 5
=− + Qu Uu − Qd Ud + gAu H̃u − gAd H̃d + Ke − gFe Le Qe |Ue | (11)
∂τ ∂t Le 6 6

where β = 1 is implied. The 5/6 coefficient for the gAu Hu and gAd Hd terms—as compared
R
to unity implied by Equation (2)—arise from the T(2,0) approximation of the ηλdx term of
Equation (2), which is consistent with the Ke source term of Equation (9). The detailed derivation is
presented by Hodges [119].

3.5. Solution Variables

For the present work, it is convenient to take Qe and either Ae or He as the principal solution
variables, (as discussed in Section 4.2, below) with the provision that the element velocity is recovered
by Ue = Qe A− 1
e and the element volume by Ve = Ae Le . Furthermore, as ∂/∂τ → 0 we have H̃ → H + c,
where H is the true piezometric head and c is a constant. If physical Dirichlet boundary conditions
on H̃ are provided we will have c = 0 and H̃ = H. Thus, in the discrete equations it is convenient to
replace H̃ with H, keeping in mind that H only represents the true piezometric head at convergence to
a pseudo steady-state. For simplicity, PipeAC is confined to fixed geometry where Le and geometry
relationships of Ae (z) are static.
Water 2020, 12, 1727 10 of 33

4. Numerical Method

4.1. Overview
The new PipeAC numerical method is designed to solve the dual time-stepping, finite-volume,
1D problem posed by Equations (8) and (11) for both free-surface and surcharged flow in a closed
conduit. In the following, the semi-discrete form of these prognostic equations is presented in
Section 4.2, the discrete form for auxiliary terms in Section 4.3, interpolation of element values to
faces in Section 4.4, handling of hydraulic jumps in Section 4.5 and a Runge-Kutta 4th order (RK4)
time-marching scheme is presented in Section 4.6. Details of the linearized friction term in the RK4 are
presented in Appendix A).

4.2. Semi-Discrete Governing Equations

The real-time derivatives in continuity, Equation (8), and momentum, Equation (11), for an AC
method can be approximated with a semi-discrete stencil of the general form

∂φ 1  
= a1 φn+1 + a2 φn + a3 φn−1 + ... (12)
∂t ∆t

where superscripts indicate discrete time levels. Using a j ∈ {3/2, −2, 1/2, 0, 0, ...} provides a 3-level
backwards stencil as applied by Rogers et al. [32]. Similarly, a 2-level backwards stencil is defined as
a j ∈ {1/2, −1/2, 0, 0, ...}. Using a general backwards stencil of NB terms the pseudo-time semi-discrete
finite-volume formulation of Equation (11) can be written as

∂Qe a 1 n +2− j
∂τ
= − 1 Qne +1 −
∆t ∆t ∑ a j Qe − gFQe |Ue | + Se (13)
j=2,NB

where the source term is

 
1 5
Se ≡ Qu Uu − Qd Ud + g [ Au Hu − Ad Hd ] + Ke . (14)
Le 6

Note that the friction term gFQe |Ue | is separated from the Se source term in this nomenclature
as friction is handled implicitly within the pseudo-time marching (as discussed in Section 4.6 below
and derived in Appendix A) whereas Se is strictly the explicit source term. PipeAC is coded with both
2-level and 3-level implicit stencils, but only the 3-point stencil was used in the present work.
The semi-discrete form of continuity, Equation (8), can be manipulated into different forms
for open-channel versus surcharged conduits. Open-channel continuity presents a challenge as
He = f ( Ae ), which implies a nonlinear term in the ∂/∂τ on the LHS of Equation (8) that will depend
on conduit geometry. For simplicity, in open-channel flows we seek an artificial compressibility
equation of the form ∂Ae /∂τ = S. The reason for using Ae rather than He as the solution variable for
open-channel flow is that small changes in He can cause relatively large changes in Ae (depending
on conduit shape). In contrast, small changes in Ae will always result in small changes in He for any
practical conduit cross-section. Using the chain rule and product rule for differentiation, it can be
shown that
∂H A ∂A ∂H
=G where G≡A + H.
∂τ ∂τ ∂A
The semi-discrete form of Equation (8) for an open channel using a general backwards stencil
is then
( )
∂Ae `e Fe2 a 1 n +1 1 n +2− j Qu − Qd
∆t j=∑
= − Ae − ak Ae + + qe (15)
∂τ Ge ∆t 2,N
Le
B
Water 2020, 12, 1727 11 of 33

where G is considered a geometric coefficient that can be treated explicitly in the time-marching scheme.
For a surcharged conduit, continuity is simpler as Ae is constant and hence is factored out of
the ∂/∂τ term in Equation (8). Furthermore, the real-time derivative of Ae is also identically zero in
a closed conduit. The resulting surcharged conduit artificial compressibility equation is

`e Fe2 Qu − Qd
 
∂He
= + qe . (16)
∂τ Ae Le

Equations (13), (15) and (16) constitute the dual-time marching system of PipeAC that is applicable
in a system with mixed open-channel and surcharged conduits. From any discrete time step n in real
time we take pseudo-time steps m = {1, 2, 3...} until we reach a pseudo-time steady-state (∂/∂τ = 0),
at which point we will have converged to the discrete real-time march equations for the unknown
n + 1 step:
( )
1 n +2− j
momentum: Qne +1 =
a1 ∑ ak Qe − ∆tgFQe |U | + ∆tSe , (17)
j=2,NB
( )
1 n +2− j ∆t
open channel: Ane +1 =
a1 ∑ ak Ae +
Le
( Qu − Qd ) + ∆tqe , (18)
j=2,NB

surcharged conduit: 0 = Qnu+1 − Qnd +1 + qne +1 Le . (19)

Equation (17) is the converged, real-time, unsteady momentum solution that applies to every
element (both open and surcharged). In contrast only one of Equations (18) or (19) is the converged,
real-time, unsteady continuity solution on a computational element. In PipeAC, an element is evaluated
as being either “open-channel” or “surcharged” for selecting the governing continuity equation at the
beginning of each pseudo-time step. Thus, an element cannot change status during substeps of the RK4
advance (Section 4.6) but can change status during at the end of a ∆τ step of the pseudo-time march.

4.3. Discrete Equations, Other Terms

To complete the set of semi-discrete, pseudo-time marching Equations (13), (15) and (16) requires
geometric definitions and interpolations of finite-volume element values (subscript e) to face values
(subscripts d, u). The interpolation approach is discussed in Section 4.4, below. A summary of the
geometrical relationships used on elements to close the equation set are:

H = f ( A) : open channel geometry,

B = f ( A) : open channel top width,
Lw = f ( A) : open channel wetted perimeter,
Lc = f (shape) : surcharged conduit circumference,
Rh = AL− w
1 : hydraulic radius, (20)
Dh = ALc 1− : hydraulic diameter,
∂H/∂A = f ( A) : open channel head-geometry relationship,
` = see Equation (6) : modified hydraulic depth,
U = QA−1 : average velocity.

Geometry definitions for the finite-volume elements in PipeAC are considered characteristic
values; e.g., the value of Ae for a finite-volume element is considered an average for the volume rather
than the value precisely at the element center. Thus, the volume of water in an element is recovered
with Le Ae without reference to the slope of the pipe element or the slope of the free surface.
Water 2020, 12, 1727 12 of 33

4.4. Timescale Interpolation from Finite-Volume Elements to Faces

Face values for a finite volume, e.g., Qu , Ad , are determined by interpolation from the
finite-volume element values. Here we follow the “timescale” approach [31], where the celerities
of information from the cell center to the upstream and downstream faces are defined for each
element and used for face interpolation. To briefly explain, it is convenient to temporarily abandon
the e, u, d subscripting that provides an element-centered view of the discretization and adopt the
conventional i subscript as the index for a element center, with i − 1 as an adjacent upstream element
and i + 1 as a downstream element. For the i element with the i − 1/2 face in the upstream direction,
the information celerity (C ) from the cell center to the upstream face is
p
Ci−1/2[+] ≡ Ui − g `i (21)

where the [+] indicates the information approaching the i − 1/2 face is from the [+] downstream
p
side and g` is the effective wave propagation speed at modified hydraulic depth `. Similarly,
the information speed from the i face to the i + 1/2 face is given by
p
Ci+1/2[−] ≡ Ui + g `i . (22)

where the [−] indicates information approaching the i + 1/2 face from the [−] upstream side.
The provisional timescale of information transit from the i cell to a face is given by, e.g.,

Li
Ti+1/2[−] = (23)
2Ci+1/2[−]

which is subject to limiters Tmax and Tmin such that where T < Tmin the computed T is replaced by
Tmin . Similarly, where T > Tmax we apply T = Tmax . These limiters prevent negative, zero, or infinite
timescales. With the above, the timescale interpolation of any variable (φ) from element values φi and
φi+1 to the face value φi+1/2 is given by linear weighting in timescales:
! !
Ti+1/2[+] Ti+1/2[−]
φi+1/2 = φi + φi+1 . (24)
Ti+1/2[+] + Ti+1/2[−] Ti+1/2[+] + Ti+1/2[−]

Timescale interpolation is used for φ ∈ { Q, A, T } whereas H interpolation to the cell face

uses conventional linear spatial interpolation based on the adjacent element lengths, e.g., Li−1 and
Li for the Li−1/2 face. Face velocities are consistent with the interpolated flowrate and areas as
Ui+1/2 = Qi+1/2 Ai−+11/2 . The timescale approach is advantageous as it smoothly switches between
upwind weighted interpolation in supercritical flow to more centrally-weighted interpolation in
subcritical flow. The spatial interpolation for H to the faces (rather than timescale interpolation) follows
recommendations in Hodges and Liu [31], where it was noted that use of timescale interpolation for
water surface elevation led to numerical instabilities that could not be readily controlled. The linear
spatial interpolation for H ameliorates the problem, but still has a tendency to propagate energy to
the subgrid scale causing V-shaped grid scale oscillations—which are annoying but not necessarily
destabilizing. This problem is ameliorated in free-surface flows with a flowrate oscillation damping
filter (Q-filter), which does not affect mass conservation but merely acts to dissipate momentum that is
being passed down into grid-scale oscillations [31]. The Q-filter was used in the present work, but it
was discovered that a similar damping filter was also needed for the surcharged pressure head in
pipes. An equivalent H-filter can be derived by simply replacing Q with H in the flowrate oscillation
damping filter presented in Hodges and Liu [31]—see their Equations (18)–(22). Herein, the H-filter
only applies to surcharged cells and does not alter mass conservation.
Water 2020, 12, 1727 13 of 33

4.5. Hydraulic Jumps

The PipeAC numerical method follows Hodges and Liu [31] in handling hydraulic jump
conditions in an open channel. Briefly, a jump occurs when an upstream element i is supercritical
and the downstream element i + 1 is subcritical. When this occurs the face i + 1/2 is characterized
by two values of A and two values of H. The H on the supercritical side is the sum of the upstream
flow depth and the conduit invert at the i + 1/2 face. The subcritical side extrapolates the downstream
element H to the jump face. In steeply sloping geometry it is possible to have a jump occur where
the upstream Hi+1/2 would be higher than the downstream Hi+1/2 ; in such cases the jump is simply
neglected and the standard interpolation approach is used. A limiter is invoked on the downstream
jump elevation when the effective jump height would cause the downstream total energy head to
exceed the upstream total energy head. This condition can occur as an occasional numerical artifact of
simple extrapolation from the downstream H to the subcritical jump face. In this approach a hydraulic
jump will always occur on a face, and each finite-volume element is either fully subcritical or fully
supercritical. More sophisticated jump approaches are discussed and implemented in the SvePy
open-channel code [129], but have not been investigated with PipeAC.

4.6. The RK4 Pseudo-Time March

The semi-discrete equations are marched forward in pseudo-time with a 3-level backwards
implicit stencil in real time, similar to Rogers et al. [32]. A unique addition is the use of an implicit
treatment of the linearized friction term in momentum, as described in Appendix A. The implicit
friction treatment ensures that dissipation is always opposite the direction of the flow in any substep
of the RK4, which ensures stability in reversing flows. The general RK4 approach for a step from m to
m + 1 in pseudo-time interval ∆τ can be presented as

∆τ  (1) 
φ m +1 = φ m + k + 2k(2) + 2k(3) + k(4) (25)
6
where

k(1) = CΛ [S (φm ) + Γφm ] , (26)

∆τ (1)
   
(2) m m
k = CΛ S φ + k + Γφ , (27)
2
∆τ (2)
   
(3) m m
k = CΛ S φ + k + Γφ , (28)
2
h   i
k(4) = CΛ S φm + ∆τk(3) + Γφm . (29)

The RK4 terms have different definitions for the time-evolution equations for flow (Q),
open-channel area ( A) and surcharged-conduit head ( H ). For clarity in the following the e subscripts
that indicate values at the element center are omitted, and thus are implied for any term without a u or
d subscript. The Γ coefficients for the linear implicit terms are
a1 a1
ΓQ = − − gF |U | ; ΓA = − ; ΓH = 0
∆t ∆t
where a1 is the weighting of the n + 1 term in the backwards real-time stencil of Equation (12). The Λ
weighting coefficients are

1
Λ (r ) ≡ for r ∈ { A, H, Q} .
1 − ∆τΓ(r)
Water 2020, 12, 1727 14 of 33

The C represents the AC coefficients, which are given as

`F 2 `F 2
CQ = 1 ; CA = ; CH =
G A
where F = 1 is used for the present work, as discussed after Equation (8). Finally, the source terms are
 
1 1 5
SQ = −
∆t ∑ aj Q n +2− j
+
L
Qu Uu − Qd Ud + g [ Au Hu − Ad Hd ] + Ke ,
6
(30)
j=2,NB
1 Qu − Qd
SA = −
∆t ∑ a j A n +2− j +
L
+ qe , (31)
j=2,NB
Qu − Qd
SH = + qe . (32)
L
The a j are the backwards real-time stencil weights that are coefficient for the known real-time n
and n − 1 values of Q and A, respectively. In the present work, lateral leakage fluxes (qe ) are set to
zero throughout.

5. Test Case

5.1. Geometry of Test Setup

To test the new algorithm, we compare PipeAC performance to laboratory experiments reported by
Trajkovic et al. [35] for unsteady flows transitioning between open-channel flow and surcharged-pipe
flow. These data were also use by Leon et al. [130] for model validation. The data from
Trajkovic et al. [35] figures for Type A and Type D experiments have been digitized for the present
comparisons (available for download as provided in Hodges [33]). Their experimental apparatus is
a ∼10 m pipe with 0.1 m inside diameter. An undershot gate at the upstream pipe end is used to control
flow from a constant-head tank. The downstream end uses either a supercritical overflow to another
tank (Case A), or a pressurized reservoir to surcharge the downstream end of the pipe (Case D). Pressure
taps along the length of the pipe are used to measure flow head. A summary of the pipe geometry and
pressure measurement locations through the test section between gates is provided in Table 1.

Table 1. Distance (m) from upstream gate (G1) to pressure measurement taps P2–P8 and downstream
gate (G2) in Trajkovic et al. [35] experiments.

G1 P2 P3 P4 P5 P6 P7 P8 G2
0 2.5 3.7 4.7 5.7 6.7 7.7 8.2 8.3

Unfortunately, geometry details in Trajkovic et al. [35] were insufficient to exactly replicate their
downstream boundary conditions, so the present work extended the pipe by 1.0 m and added a 5.0 m
buffer domain consisting of a 0.5 m diameter pipe whose crown was coincident with the crown of
the downstream end of the experimental pipe. The buffer domain has zero slope and with the larger
diameter it allows subcritical flow conditions at the numerical model outlet, which ensures a fixed
head boundary condition can be readily applied. The Gauckler-Strickler-Manning friction coefficient
in the test section was 0.008, following that reported by Trajkovic et al. [35]. The buffer domain used
a friction coefficient of 0.016 to help damp wave reflections off the downstream boundary. By using
the buffer domain, the supercritical flow cases had downstream H boundary conditions that did not
affect flow through the downstream gate G2.

5.2. Flow Conditions for Tests

The Trajkovic et al. [35] Type A and D experiments were originally conducted by establishing
a baseline flowrate in the system with gate G1 (upstream) at a fixed opening for the duration of
Water 2020, 12, 1727 15 of 33

the experiments and gate G2 (downstream) set at an initial height to establish a steady flow over
an interval of time. Once the stable flow is obtained, G2 was closed “instantaneously,” held closed for
some duration, and then reopened to a height that varied by experiment. A truly instantaneous gate
closing/opening is impossible in the real world and can cause numerical problems in a simulation,
so an arbitrary gate closing speed of 0.1 m/s was used in the simulations. This allows the G2 gate to
fully open or close in 1.0 s. Quantitative conditions for the test cases are provided in Table 2. For further
details of the experimental apparatus and testing conditions please see Trajkovic et al. [35].

Table 2. Type A and Type D experiments reported in [35], as tested with PipeAC. Model downstream
H is measured above the lowest point in the test pipe (not including buffer domain). * See discussion
in text as how values for Q and bed slope were obtained for Case A4.

Upstream Downstream G1 G2 Initial Initial Interval G2 Closed G2 Reopen

Bed
Type Q H Height Height of Steady Flow Interval Height
Slope
(m3 s−1 ) (m) (m) (m) (s) (s) (m)
A1 0.027 0.0013 0.03 0.014 0.100 120 30.0 0.008
A2 0.027 0.0013 0.03 0.014 0.100 120 30.0 0.015
A3 0.027 0.0013 0.03 0.014 0.100 120 30.0 0.028
A4 0.014 * 0.0013 * 0.03 0.014 0.100 120 30.0 0.100
D1 0.014 0.0015 0.107 0.015 0.033 240 16.0 0.033

The numerical simulations use fixed Q upstream boundary conditions to match the measured
values from the experiments of Trajkovic et al. [35]. The downstream H boundary conditions are
enforced at the end of the buffer domain.
Note that for Case A4 the experimental Q and S0 were not provided in Trajkovic et al. [35],
but from their reported data it is clear that one of these was not the same as those used in Cases
A1–A3. The peak pressure at gauge P7 in Trajkovic et al. [35]—shown in their Figure 3 —is different
from the peak pressure at the same gauge in their Figure 4. Such a discrepancy is only possible if
the experiments used different flow rates and/or slopes. As shown in Section 6 below, the present
simulations provide reasonable estimates of peak pressures and the rate of pressure increase at P7
for Cases A1–A3, so several simulations at varying flow rates and slopes were conducted to examine
the likely flow rate and slope used in Case A4. The simulation that matched the experimental results
most closely used the same Q as Cases A1–A3 with the bed slope of Case D1, as presented in Section 6,
below. It seems likely that these are the correct values for Case A4 and were simply omitted in
Trajkovic et al. [35].

5.3. Consistent Data Comparisons

The data analyses in Trajkovic et al. [35] are reported in terms of the pressure head measured from
the sloping bottom of the pipe at individual sensors. For the numerical model, we use He − zb(e) − δe
where He is the simulated piezometric head on the computational element containing the reported sensor
location, zb(e) is the inside bottom elevation of pipe at that location, and δe is a sensor head offset shown
in Table 3. The δe accounts for the fact that the computational cell centers were not exactly at the sensor
location and thus there is an offset of the baseline along the sloping pipe. Rather than attempting to
estimate the geometry effect of the displacement, the small values of δe were obtained by minimizing the
difference between the simulation and the experiments at the end of the steady-flow ramp-up period.

Table 3. Sensor head adjustment, δe , between the model and experimental sensors for different bed
slopes. Values for P3—P7 are in meters of head.

Cases Slope P3 P4 P5 P6 P7
A1, A2, A3 0.027 – – 0.0050 – 0.0060
A4, D1 0.014 0.0100 0.0095 0.0095 0.0095 0.0090
Water 2020, 12, 1727 16 of 33

5.4. Calibration of Drag Model for Undershot Gate

The test cases use two undershot gates to control the flow conditions. From the data in
Trajkovic et al. [35] it is impossible to reconstruct an accurate estimation of the frictional losses
through these gates. Lacking any other data, these are modeled using a standard discharge equation
following Crane Co. [131] and Brater et al. [132]:
!−1/2
A2
2 g
2g ∆H
p
Q = Cv A g where Cv ≡ Cd 1 − Cd
A21

Here A g is the open cross sectional area under the gate, A1 is the upstream cross-sectional
area approaching the gate, ∆H is the change in piezometric head across the gate, Cv is the effective
discharge coefficient that includes effect of approach velocity and Cd is the gate discharge coefficient
when approach velocity is negligible. For simplicity, a closed gate is modeled as a gate with leakage
through a minimal area of 10−5 m2 . The only calibration in the model was in adjustment of CD . Tested
values of CD were {0.66, 0.76, 0.86}. The differences between the results were relatively small, but 0.76
was deemed superior and used in the present work. There did not seem to be any value in further
refining the estimate.

5.5. Grid Refinement Test Setup

The baseline discretization used in the simulations was 40 cells over the 8.3 m between gates G1
and G2. Grid refinement behavior was tested with Case A3 using 5, 10, 20, and 80 cells in the test
section. The discretization of all sections for different simulations are provide in Table 4. Simulations
with more than 80 cells in the test section were not conducted as a cell length less than the pipe diameter
(0.1 m) inherently violates the hydrostatic approximation of the Saint-Venant equations that requires
horizontal length scales to be greater than vertical lengths scales. Results from the grid refinement
study are found in Section 6.5, below.

Table 4. Discretization of simulations. N is the number of cells in the section, ∆x is the length of each cell.

Upstream Test Section Downstream

Cases Buffer
of G1 G1 to G2 of G2
N:∆x (m) N:∆x (m) N:∆x (m) N:∆x (m)
A3a 2:0.750 5:1.66000 2:0.7000 5:1.000
A3b 2:0.750 10:0.83000 2:0.7000 5:1.000
A3c 3:0.500 20:0.41500 3:0.4667 10:0.500
A1, A2, A3, A4, D1 6:0.250 40:0.20750 6:0.2333 20:0.250
A3e 12:0.125 80:0.10375 12:0.1167 40:0.125

6. Results

6.1. Temporal Evolution of Piezometric Head

The test cases described in Section 5 have been simulated in the PipeAC finite-volume code.
The general progression of the piezometric head in the simulations follows the illustrations in
Figures 1 and 2 for Case A3 and D1, respectively. An initial steady supercritical flow through upstream
gate G1 is achieved (t < 0) with gate G2 being fully open (Cases A1–A4) or partially open (Case D1, see
Table 2). With gate G2 fully open the initial steady flow is supercritical throughout the entire test section
of the pipe (i.e., between the gates), as shown in Figure 1a in the t = 0 s line. For gate G2 partially open
(Case D1) the gate partially restricts the flow forming a transitional steady-state profile, as shown for
the t = 0 line in Figure 2a. At t = 0 gate G2 is rapidly closed. A bore subsequently propagates upstream
until the gate is re-opened, as shown by the piezometric head profiles in Figure 1a for 0 < t ≤ 30 s
and in Figure 2a for 0 < t ≤ 16 s. In Case A3 the gate G2 re-opens to 0.028 m at time t = 30 s, and for
Case D1 the gate re-opening is to 0.033 m at time t = 16 s. There is a first transitional phase from
Water 2020, 12, 1727 17 of 33

30 < t < 31 s in Case A3 (Figure 1b) in which the upstream bore holds position and there is a rapid
decrease in the surcharge pressure moving up the channel. In Case D1, Figure 2b, this transition phase
is longer—from 16 < t < 22 s—and shows the bore continuing to move slightly upstream with the
surcharge pressure rapidly depleted and the water surface elevation dropping relatively evenly along
the channel between the bore and gate G2. In the second transition phase, for Case A3, Figure 1c, the
water surface elevation drops below surcharge and a secondary bore that crosses into surcharge forms
and propagates up the channel. Note that the original upstream bore appears to hold its position and
slowly decline in magnitude. The effect is similar in for Case D1, Figure 2c, except that the secondary
bore is not surcharged. By t = 36 s in Case A3 and t = 32 s in D1 (Figures 1d and 2d, respectively),
both systems transition into a simple drawdown to the steady-state conditions with their respective
gate restrictions at G2. Unfortunately, the experimental results reported in Trajkovic et al. [35] do not
have sufficient detail to re-create the evolution of the piezometric profiles illustrated in Figures 1 and 2.

Distance (m) from G2 (negative is upstream)

6 5 4 3 2 1 0 1 2
(a) t=0
0.3 t=8
t=16
0.2 t=23
t=30
0.1 gate
P5 closes at t=0
0.0 P7

(b) t=30
Piezometric head (m)

0.3 t=30.4
t=30.7
0.2 t=31
0.1 gate opening
30<t<31
0.0
(c) t=31
0.3 t=32
t=34
0.2 t=36
0.1 gate open
P5
0.0
(d) t=36
0.3 pipe cro
wn t=40
t=50
0.2 pipe t=65
invert
t=80
0.1 t=120
0.0
6 5 4 3 2 1 0 1 2
Distance (m) from G2 (negative is upstream)
Figure 1. Profiles of piezometric head evolution for simulation of Case A3: (a) is progression from
steady flow just before t = 0 s (with gate fully open) showing upstream propagation of upstream bore
after gate closes at t = 0 s, (b) is initial slumping of surcharged flow as the gate is opened to 0.028 m
at t = 30 s, (c) secondary transition phase with secondary bore, and (d) is drawdown to the steady
condition with gate partially open. P5 and P7 are locations of pressure sensors where experimental
data were reported.
Water 2020, 12, 1727 18 of 33

Distance (m) from G2 (negative is upstream)

6 5 4 3 2 1 0 1
0.20
(a)
0.15
t=0
0.10 t=4
t=8 P3 gate
0.05 t=12 P4 closes
t=16 P5 at t=0
P6 P7
0.00
0.20
(b)
Piezometric head (m)

0.15
0.10 t=16
t=18
0.05 t=20 gate opening
t=22 16<t<17
0.00
0.20
(c)
0.15
0.10 t=22
t=25
0.05 t=28 gate open
t=32
0.00
0.20
(d)
0.15 pipe crow
n
t=32
0.10 t=35 pipe inv
t=40 ert
0.05 t=50 gate open
t=80
0.00
6 5 4 3 2 1 0 1
Distance (m) from G2 (negative is upstream)
Figure 2. Profiles of piezometric head evolution for simulation of Case D1: (a) is progression from
steady flow just before t = 0 s (with gate open to 0.033 m) showing upstream propagation of hydraulic
jump after gate closes at t = 0, (b) is initial slumping of subcritical flow shortly after gate re-opening
at t = 16 s, (c) secondary transition phase with secondary bore, and (d) is drawdown to the steady
condition. P3–P7 are locations of pressure sensors where experimental data were reported.

6.2. Comparison of Simulations and Experiments for Cases A1–A3

We can make direct comparisons between simulations and experiments at sensors where data
were reported in Trajkovic et al. [35]. For the A1–A3 experiments, a comparison of the experimentally-
measured and simulated head at gauge P7, which is 0.6 m upstream of gate G2 (see Table 3), is shown
in Figure 3. The simulations track well with the filling of the pipe, including the magnitude of the
sudden pressure drop and subsequent recovery shortly after the gate is opened.
Results further upstream in the pipe, at pressure gauge P5, are shown in Figure 4. It can be seen
that the bore at the leading edge of the pressure signal in the simulations moves upstream slightly
faster than that measured in the experiment. Note that the A1–A3 numerical simulations are identical
prior to t = 30 as they all use identical initial conditions (supercritical steady flow with gate G1
at 0.014 m) and have gate G2 closed at exactly the same time. The differences between simulations
after t = 30 are due to different re-opening heights for gate G2 (see Table 2). For Cases A1 and A2,
the pressure-drop magnitude and duration at the gate opening are well-represented in the simulation.
In contrast Case A3 has a good representation only for the pressure-drop magnitude, but shows
an exaggerated duration for the pressure reduction after the abrupt opening of the gate. The increased
Water 2020, 12, 1727 19 of 33

duration is related to the propagation of the secondary bore through P5 for 31 < t < 34 as shown in
Figure 1c. There we see a rapid change from surcharge (lines t = 31 s to t = 32 s), followed by the
secondary bore propagating upstream (lines t = 32 s to t = 34 s). The implication of the discrepancy
between the observations and the model is that the secondary bore in the experiments moves more
rapidly upstream than the model can represent—hence the longer duration of the opening pulse at P5
in the model.
0.25
A1 sim
0.20 A1 exp
Pressure head (m) at P7

A2 sim
A2 exp
0.15 A3 sim
A3 exp
0.10 pipe interior crown

0.05 gate
opened
0.00
0 10 20 30 40 50
Time (s) from G2 closing
Figure 3. Pressure head at P7 for simulations and experiments of Cases A1–A3.

0.25
A1 sim
0.20 A1 exp
Pressure head (m) at P5

A2 sim
A2 exp
0.15 A3 sim
A3 exp
0.10 pipe interior crown

0.05 gate
opened
0.00
0 10 20 30 40 50
Time (s) from G2 closing
Figure 4. Pressure head at P5 for simulations and experiments of Cases A1–A3.

6.3. Comparison of Simulations and Experiments for Case A4

For Case A4, the bed slope is about half that of Cases A1–A3 (see Table 2). In Figure 5, the bore
at P6 and P7 (near to G2) are well-represented, but as the bore propagates upstream sequentially
through P5, P4, and P3 there is an increasing phase error—the simulation bore propagates upstream
faster than the experiment. It can also be observed that the simulation bore has a sharper front in the
simulations than the experiments—a discrepancy that is small at P7 but increases moving upstream
from P6 to P3.
0.14
0.12 P7 sim
P7 exp
Pressure head (m) at P7

0.10 pipe interior P6 sim

crown
P6 exp
0.08 P5 sim
0.06 P5 exp
P4 sim
0.04 gate P4 exp
opened P3 sim
0.02 P3 exp
0.00
0 10 20 30 40 50 60
Time (s) from G2 closing
Figure 5. Pressure head at at sensors along the pipe length for Case A4.
Water 2020, 12, 1727 20 of 33

6.4. Comparison of Simulations and Experiments for Case D1

Case D1 is quite different from Cases A1–A4 as the pipe downstream of the gate and in the
buffer zone are surcharged in D1, versus the supercritical drawdown into the buffer domain in
A1–A4. Unfortunately, Trajkovic et al. [35] did not give a measurement of the surcharged head in the
downstream end of the pipe. A number of test cases were conducted and compared to the experimental
data. As shown in Figure 6, the results at the reported sensors in Trajkovic et al. [35] is quite reasonable
for a downstream head that is only slightly greater than the pipe crown elevation at the downstream
end of the pipe (see Downstream H in Table 2). For this experiment the downstream gate G2 is initially
partially open, which provides a standing hydraulic jump as the initial condition (see t = 0 line in
Figure 2a). The gate is then rapidly closed for only 16 seconds before being re-opened to its initial
height. It can be seen that the model provides reasonable representation of the bore propagating
upstream prior to the gate opening. The drawdown after the gate opened shows the model emptying
the upstream test section faster than the observed system—which might be attributable to the drag
coefficient used for the gate. The gate CD calibration study (Section 5.4) was conducted using data for
Cases A1–A3, which have a steeper pipe slope and a greater pressure drop across the gate than Case D1.

Time (s) from G2 closing

0 10 20 30 40 50 60 70
(a) P7 sim
0.10 P7 exp
0.05
0.00
Pressure head (m)

(b) P6 sim
0.10 P6 exp
0.05
0.00
(c) P5 sim
0.10 pipe interior crown P5 exp
0.05 gate
opened
0.00
0 10 20 30 40 50 60 70
Time (s) from G2 closing
Figure 6. Pressure head at sensors along pipe for Case D1.

6.5. Grid Refinement Results

A grid refinement study (Section 5.5) was conducted to compare with the experimental results
at P7 and P5 for Case A3, similar to Figures 3 and 4. Results of the grid refinement study are shown in
Figures 7 and 8. It can be seen that results from N = 20 to N = 80 are reasonably similar—although the
bore front for N = 20 is less steep. The N = 10 simulation is significantly degraded in quality, but still
shows the same basic flow features. The coarse N = 5 simulation is unable to represent the sharp bore
front at P7 but instead has a pressure spike when the pipe transitions to surcharged at P7, and another
pressure spike when gate is opened. Note that these coarse-grid simulations never show the correct
surcharged conditions at P5.
Water 2020, 12, 1727 21 of 33

0.30
N=5 sim
0.25 N=10 sim

Pressure head (m) at P7

N=20 sim
0.20 N=40 sim
N=80 sim
0.15 A3 exp
0.10 pipe interior crown
0.05 gate
opened
0.00
0 10 20 30 40 50
Time (s) from G2 closing
Figure 7. Pressure head at P7 sensor 0.6 m upstream of G2 for grid refinement simulations of Case A3.
See Table 4 for details.

0.30
N=5 sim
0.25 N=10 sim
Pressure head (m) at P7

N=20 sim
0.20 N=40 sim
N=80 sim
0.15 A3 exp
0.10 pipe interior crown
0.05 gate
opened
0.00
0 10 20 30 40 50
Time (s) from G2 closing
Figure 8. Pressure head at P5 sensor 2.6 m upstream of G2 for grid refinement simulations of Case A3.
See Table 4 for details.

7. Discussion

7.1. Interpretation of Results

Overall, the results presented in Figures 3–6 show reasonable agreement between the PipeAC
model and the experiment. Perfect agreement is not expected as Trajkovic et al. [35] reported trapped air
pockets and multiphase bubbly regions in these dynamic flows. It is likely that the present simulations
are at the limit of what can be captured for pipe flow undergoing rapid unsteady changes across
supercritical, subcritical, and surcharged regimes when the simulation model assumes a single liquid
that is governed by the hydrostatic approximation.
From the grid refinement study in Figures 7 and 8, it appears that it takes at least N = 20 grid
cells to get a reasonable representation of the unsteady dynamics of associated with the gate rapidly
closing and opening. From Table 4, N = 20 corresponds to an element length slightly more than 4× the
pipe diameter of 0.1 m. At coarser resolutions, PipeAC remains stable but the dynamics are poorly
represented. It is not surprising that the N = 5 simulation does quite poorly. As noted in Section 4.5
and discussed in detail in Hodges and Liu [31], the hydraulic jump discretization used herein will
force a supercritical-to-subcritical condition to occur on the face of a finite-volume. This simplifies
the numerical method as an entire computational cell is either subcritical or supercritical—there
are no “transitional” elements. With sufficiently fine grid resolution, forcing the jump to occur on
a cell face has marginal effect on the results. However, at coarse grid resolution this approach can
be troublesome. In the present simulations, with N = 5 there are only 4 locations where jumps are
allowed: x = {−6.64, −4.98, −3.32, −1.67} as measured from the downstream gate G2 and used in
Figure 1. Thus, it is impossible for location P7 (at x = −0.6 m from G2) to see a hydraulic jump in the
N = 5 simulation.
Water 2020, 12, 1727 22 of 33

The behavior at the bore front as it moves upstream is particularly interesting. In Figures 3, 5
and 6a it can be seen that at P7, only 0.6 m upstream from gate G2, the steep bore seen in the experiments
is well-represented by the simulation. As the bore moves upstream—i.e., at P5 in Figure 4 and in
the evolution across sensors shown in Figure 5—the experimental data shows the sharp corner of
the bore progressively erodes, and the observed bore propagation speed is slower than that in the
simulation. To understand these phenomena, it is useful to consider the bore propagation shown
in the grid refinement tests of Figures 7 and 8 for sensors P7 and P5, respectively. At P7, we see
the sharp experimental front is well-represented by tests N = {40, 80}, but the N = {5, 10, 20} show
a shallower slope at the front. However, moving upstream to P5 we see the experimental front has
a shallower slope—which is well-represented by the N = 20 simulations but not by the finer N = {40, 80}
simulations. Indeed, at P5 there it is clear that the coarser N = 20 simulation is “better” than the finer
resolution simulations. This “better” agreement at coarser resolution is consistent with prior analysis
of a hydrostatic model in representing non-hydrostatic waves [133]. Briefly—the real waves in the
experiment are inherently non-hydrostatic and include dispersion, whereas the hydrostatic simulation
neglects dispersion but includes nonlinear steepening. At coarse grid resolution, numerical dispersion
and numerical dissipation can act similar to physical dispersion in modifying the wave, hence the
excellent N = 20 results from about 14 < t < 30 in Figure 8 are serendipitous. Numerical error at
a particular grid scale can approximate the missing non-hydrostatic physics—giving the “right answer
for the wrong reason.” At finer resolution, numerical dissipation is decreased and the wave front
becomes steeper, as seen in Figure 8. Furthermore, Figures 4 and 5 also show that the hydrostatic model
has a slightly faster wave speed as the bore propagates upstream, which is expected in a low-order
hydrostatic model.
The behavior of Case A3 at sensor P5, as shown in Figures 4 and 8, illustrates the limitations of
the PipeAC algorithm for rapid changes with the flow near the surcharge threshold. The experimental
data for Case A3 at sensor P5 show the opening of the gate causes a rapid pressure drop at P5 such
that the head drops below surcharged. We do not have information from the experiments as to
whether or not the pressure drop actually allowed air to enter the head space at P5, or if the pipe
remained full with a slightly negative pressure. The PipeAC model effectively assumes that air can
move into the head space whenever the piezometric pressure drops below pipe crown elevation,
which is an inherent limitation of the present single-fluid model. Future work should consider
implementation of approaches that correctly handle sub-atmospheric pressure [134–136]. However,
from the point of view of large-scale network models, the abrupt transient and pressure drops are
handled reasonably—causing neither catastrophic instabilities nor degrading the simulation quality
after the abrupt transient. Note that the model handles the abrupt transient better at sensor P7,
as shown in Figure 3, where the pressure drop is only slightly below surcharge in Case A3.
A final feature that can be noted in the simulation results are oscillations after the return to
surcharge at sensor P5 for Case A3. Incipient oscillations for the N = 40 resolution are seen in the
Case A3 line in Figure 4 for 33 < t < 37 after the abrupt return to surcharge. The oscillations
are also visible in the N = 80 grid refinement test, as shown Figure 8. The N = 20 test only briefly
crosses to surcharged after the gate opens at P5 and shows no oscillations. For the N = {40, 80} tests,
the oscillations persist for only a few seconds and are likely an artifact of the impulse that occurs when
the solution crosses rapidly from free-surface flow to surcharged. That these oscillations occur only
when the flow abruptly crosses to weakly surcharged is an indication that these are likely induced
by the change of the continuity solution from A to H, i.e., as in Equations (15) and (16), respectively.
It seems that some form of smoothing for the transition between these equation forms could be useful
in reducing these oscillations.

7.2. Implications for Large-Scale Network Models

The algorithm developed for numerical simulation in PipeAC is a step towards large-scale
1D modeling of stormwater networks. Consistent with this goal, the method uses “no-neighbour”
Water 2020, 12, 1727 23 of 33

spatial interpolation, which ensures that a computational element only needs to know values on
its adjacent faces and the faces only need to know the adjacent elements; i.e., the discretization of
an element does not include a neighbour element and the discretization of a face does not include
a neighbour face [31]. The no-neighbour approach inherently limits a code to 1st and 2nd-order spatial
discretizations, but reduces the communication burden associated with decomposing a system for
parallel computing [70]. Arguably, the PipeAC method could be improved in its ability to handle
abrupt transitions by inclusion of higher-order discretization methods, but that would increase the
communication burden in the parallel network model that is under development.
At this point, it is not clear that the PipeAC algorithm will be suitable for the entire solution of
a city-scale stormwater network. The NCIMM research team’s working hypothesis (which remains
to be demonstrated) is that it will be more effective to use a standard Saint-Venant solver—such
as SvePy [31,129]—for all cells H < zc − δH , where zc is the elevation of the inside crown of the
pipe and δH is a small length scale. That is, the standard hyperbolic solver for the Saint-Venant
equations should be used up until the flow is about to transition to surcharge, and then the dual
time-stepping AC method is invoked. However, this idea remains speculative and there remain
a number of issues associated with artificial shocks and oscillations that might be produced when
switching the solution algorithm.

7.3. Possible Approaches Melding Artificial Compressibility and the Preissmann Slot
Another area that should be investigated is the use of the AC method with limited pseudo-time
stepping. In the present work, the inner loop (RK4 pseudo-time stepping) is iterated until L1 , L2 and L∞
norms for ∆Q and ∆H are all below cutoff values of 5 × 10−8 , 10−8 , 10−7 , respectively, or the number


of iterations reaches 100 (details are available in Hodges [33]). For the simulations herein the relative
error in volume conservation is O(10−6 ) or better. However, the importance of this level of mass
conservation is questionable. Note that a single time step of a Preissmann Slot method (Section 2.3)
is similar to the first iteration of an AC method (Section 2.4) in that it solves a hyperbolic continuity
equation using an artificial wave speed. Indeed, as shown in Section 3.3, the novel representation of the
AC pressure celerity γ using the hydraulic depth (`e ) and the Froude number Fe = 1 ensures the AC
pressure celerity is exactly a gravity wave speed. Hence, the first AC iteration in the PipeAC method
is effectively a Preissmann Slot approximation where the slot width is set to obtain the gravity wave
speed for the local piezometric head, H. Further investigation is needed to understand the relationship
between the number of AC iterations (or the norms used for convergence), the selection of Fe , and how
permutations in these affect the results in a large-scale network simulation. Limiting the number of
AC pseudo time-steps has the potential for making the solver much faster—albeit by losing some mass
conservation. This idea has previously been investigated for coastal ocean flows [71], but has not been
considered in 1D network modeling.

8. Conclusions
This paper presents a new approach to an old numerical method—the artificial compressibility
(AC) approach of Chorin. It is demonstrated that a dual-timestepping AC approach can be used with
the 1D Saint-Venant equations and 1D incompressible flow closed-conduit equations to solve unsteady
flows that transition between supercritical, subcritical, and surcharged regimes. The formulation
of the AC governing equations allow the transition from open-channel flow to surcharged-pipe
flow to be made with only small changes in the numerical algorithm—i.e, switching from solution of
cross-sectional area to piezometric head as the solution variable for continuity. Satisfactory comparisons
are made to previously published experimental measurements of unsteady transitional flows in
a circular pipe.
The new finite-volume model, PipeAC, uses a low-order numerical scheme (timescale
interpolation) to translate cell-centered values to faces—which makes the algorithm a building
block towards efficient large-scale computation of city-wide stormwater networks. Two numerical
Water 2020, 12, 1727 24 of 33

contributions that have wider implications are: (i) demonstration that the artificial compressibility
parameter—which has previously been considered a somewhat arbitrary value associated with solution
convergence rates—can be formulated in terms of an equivalent gravity-wave speed and hence the first
iteration is mathematically equivalent (for surcharged pipe) to a Preissmann Slot time step; (ii) a dual
time-stepping method using a Runge-Kutta time advance in pseudo time can be implemented with
an implicit treatment of the friction term to ensure the correct direction of friction (opposing velocity)
in each substep of the RK advance.
The known limitations of the present work are (i) the underlying equations are hydrostatic,
single-fluid, and 1D, which inherently cannot exactly represent multi-fluid interactions of air and
water at the front of an unsteady bore or its correct celerity, and (ii) sufficient grid resolution must
be applied to reasonably capture the shock front of a bore. The overall computational efficiency of
the new method remains an unknown issue to be further investigated. It is speculated that artificial
compressibility will be most useful as a technique that is selectively used in closed-conduit flow for
surcharged pipe and where pipe sections are nearing surcharge.
The benefit of this research is that it provides a simple, consistent, and mass conservative approach
to handling the switch between hyperbolic and elliptic equations that occurs when the free-surface
flow in a pipe becomes constrained by the pipe ceiling. This issue has previously been a challenge to
stormwater modeling that causes either loss of mass conservation or slow convergence of the solution.

Data: The PipeAC code and all test case data are available through Hodges [33] at https://doi.org/10.
18738/T8/5GJ6ST.
Funding: This article was developed under Cooperative Agreement No. 83595001 awarded by the
U.S. Environmental Protection Agency to The University of Texas at Austin. It has not been formally reviewed by
EPA. The views expressed in this document are solely those of the author and do not necessarily reflect those of
the Agency. EPA does not endorse any products or commercial services mentioned in this publication.
Acknowledgments: The author gratefully acknowledges the assistance of Ehsan Madadi Kandjani in refactoring
the SvePy code as the foundations for the PipeAC code and finding the Trajkovic et al. [35] test case to use for
testing and validation.
Conflicts of Interest: The author declares no conflict of interest. The funders had no role in the design of the
study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to
publish the results.

Abbreviations
The following abbreviations and mathematical notation are used in this manuscript:

Acronyms:
AC artificial compressibility
A1, A2, A3, A4 test cases
D1 test case
G1, G2 moveable gates in experiment
LHS left-hand-side
NCIMM National Center for Infrastructure Modeling and Management
P2,...P7 pressure sensors
RK4 Runge-Kutta 4th order
SvePy Saint-Venant Equations Python (code)
1D, 2D, 3D one-, two-, and three-dimensional
Greek letters:
β nonlinear velocity integration coefficient
δH small length scale (m)
∆t real time step (s)
∆τ pseudo-time step (s̃)
η water surface elevation (m)
Water 2020, 12, 1727 25 of 33

γ artificial compressibility speed of sound

Γ weighting of implicit terms in RK4
λ piezometric pressure quadrature function
Λ implicit coefficient in RK4
ρ density
τ pseudo-time
φ placeholder for a range of variables
φ̃ value of φ in AC iteration
Latin letters:
a constants of backwards time stencil
A area, cross-sectional (m2 )
B breadth of conduit (m)
c integration constant
C coefficient of AC in RK system
C information celerity to a cell face (m s− 1 )
d downstream face (as subscript)
Dh hydraulic diameter (m)
e element (as subscript)
F friction function
F Froude number
g gravity (9.81 m s−2 )
G area-head function (m)
H piezometric head (m)
`e modified hydraulic depth (m)
`i distance from cell center to face (m)
k substep source term in RK4
K source term from piezometric pressure quadrature
Le length, streamwise (m)
Lw wetted perimeter (m)
n time step
N number of computational elements in test section
q lateral inflow per unit length (m2 s−1 )
Q flowrate (m3 s−1 )
p pressure (kg m−1 s−2 )
P piezometric pressure (kg m−1 s−2 )
Rh hydraulic radius (m)
S explicit source terms
Sf friction slope
t real time
T timescale of information transport (s)
u upstream face (as subscript)
U velocity, cross-sectional averaged (m s−1 )
V volume (m3 )
x streamwise coordinate direction
z vertical coordinate direction
zc elevation of pipe crown

Appendix A. Implicit Linearized Friction Term in an Explicit RK4

The friction source term in the momentum equation can be troubling when the flow slows and
reverses direction. In the time step of a reversal, a simple explicit treatment can result in friction
direction that is opposite of the final flow direction—an effect that can cause friction to artificially
increase the reversal velocity [137]. To counteract this effect we can use an implicit discretization of
a linearized friction term within an otherwise explicit RK4 advance. This ensures that the friction at
each RK4 substep is always counteracting the predicted velocity direction at the end of that substep.
Water 2020, 12, 1727 26 of 33

Consider a differential equation of the form

∂φ
= S. (A1)
∂τ
A general RK4 time advance from the m pseudo-time step to the m + 1 pseudo-time can be
written as

k (1) = S ( φ m ) , (A2)
∆τ (1)
 
k (2) = S φ m + k , (A3)
2
∆τ (2)
 
k (3) = S φ m + k , (A4)
2
 
k(4) = S φm + ∆τk(3) , (A5)
∆τ n (1) o
φ m +1 = φ m + k + 2k(2) + 2k(3) + k(4) , (A6)
6
where the superscripts in parentheses indicate the the successive approximation steps for k. The above
implies four estimates of φm+1 as

φ(1) = φm + ∆τk(1) , (A7)

φ(2) = φm + ∆τk(2) , (A8)
(3) m (3)
φ = φ + ∆τk , (A9)
φ(4) = φm + ∆τk(4) . (A10)

In the standard RK4 system only the φ(3) estimate is actually computed—in the source for
Equation (A5)—but the other φ values are implied.
Now let us consider an equation of the form

∂φ
= S + Γφ (A11)
∂τ
such that the source includes linear (or linearized) function of φ. In particular, we are interested
in a linearized drag term (a function of Q) which we would like to handle implicitly to ensure the
direction of drag is consistent with the final direction of the flow. This approach is similar to the
“constrained friction” method [129], but is simpler in its implementation. With a linear term, the φ(1)
approximation of Equation (A7) might be written in an implicit form
h i
φ(1) = φm + ∆τ S (φm ) + Γφ(1) . (A12)

After some algebra, the equation for φ(1) can be written as

∆τ
φ (1) = φ m + [S (φm ) + Γφm ] , (A13)
1 − ∆τΓ

which implies an RK4 step of

φ(1) = φm + ∆τk(1) (A14)

where
1
k (1) = [S (φm ) + Γφm ] . (A15)
1 − ∆τΓ
Water 2020, 12, 1727 27 of 33

Applying the same approach for φ(2) , φ(3) and φ(4) an RK4 system with the following definitions
of k results in
1
k (1) = [S (φm ) + Γφm ] (A16)
1 − ∆τΓ
∆τ (1)
   
1
k (2) = S φm + k + Γφm (A17)
1 − ∆τΓ 2
∆τ (2)
   
1
k (3) = S φm + k + Γφm (A18)
1 − ∆τΓ 2
1 h   i
k (4) = S φm + ∆τk(3) + Γφm (A19)
1 − ∆τΓ

which can be represented compactly as Equations (26)–(29). Note that the above is derived for the
pseudo-timestep of τ as used in the present work, but is readily extended to an explicit RK4 march in
real time (t).

References
1. Wong, B.P.; Kerkez, B. Real-time control of urban headwater catchments through linear feedback:
Performance, analysis, and site selection. Water Resour. Res. 2018, 54, 7309–7330, doi:10.1029/2018WR022657.
2. Kim, H.I.; Keum, H.J.; Han, K.Y. Real-time urban inundation prediction combining hydraulic and
probabilistic methods. Water 2019, 11, 293, doi:10.3390/w11020293.
3. Rjeily, Y.A.; Abbas, O.; Sadek, M.; Shahrour, I.; Chehade, F.H. Model predictive control for optimising the
operation of urban drainage systems. J. Hydrol. 2018, 566, 558–565, doi:10.1016/j.jhydrol.2018.09.044.
4. Kong, F.; Ban, Y.; Yin, H.; James, P.; Dronova, I. Modeling stormwater management at the city district level
in response to changes in land use and low impact development. Environ. Model. Softw. 2017, 95, 132–142,
doi:10.1016/j.envsoft.2017.06.021.
5. Jia, N.; Sitzenfrei, R.; Rauch, W.; Liang, S.; Liu, Y. Effects of urban forms on separate drainage systems:
A virtual city perspective. Water 2019, 11, 758, doi:10.3390/w11040758.
6. Deng, S.; Zhang, X.; Shao, Z.; Yan, W.; Chai, H.; He, Q. An integrated urban stormwater model system
supporting the whole life cycle of sponge city construction programs in China. J. Water Clim. Chang.
2019, 10, 298–312, doi:10.2166/wcc.2018.197.
7. Chui, T.F.M.; Liu, X.; Zhan, W. Assessing cost-effectiveness of specific LID practice designs in response to
large storm events. J. Hydrol. 2016, 533, 353–364, doi:10.1016/j.jhydrol.2015.12.011.
8. Morsy, M.M.; Goodall, J.L.; Shatnawi, F.M.; Meadows, M.E. Distributed Stormwater Controls for Flood
Mitigation within Urbanized Watersheds: Case Study of Rocky Branch Watershed in Columbia, South
Carolina. J. Hydrol. Eng. 2016, 21, 05016025, doi:10.1061/(ASCE)HE.1943-5584.0001430.
9. Goncalves, M.L.R.; Zischg, J.; Rau, S.; Sitzmann, M.; Rauch, W.; Kleidorfer, M. Modeling the Effects of
Introducing Low Impact Development in a Tropical City: A Case Study from Joinville, Brazil. Sustainability
2018, 10, 728, doi:10.3390/su10030728.
10. Bisht, D.S.; Chatterjee, C.; Kalakoti, S.; Upadhyay, P.; Sahoo, M.; Panda, A. Modeling urban floods
and drainage using SWMM and MIKE URBAN: A case study. Nat. Hazards 2016, 84, 749–776,
doi:10.1007/s11069-016-2455-1.
11. Zhu, J.; Dai, Q.; Deng, Y.; Zhang, A.; Zhang, Y.; Zhang, S. Indirect Damage of Urban Flooding: Investigation
of Flood-Induced Traffic Congestion Using Dynamic Modeling. Water 2018, 10, 622, doi:10.3390/w10050622.
12. Duan, H.F.; Gao, X. Flooding control and hydro-energy assessment for urban stormwater drainage systems
under climate change: Framework development and case study. Water Resour. Manag. 2019, 33, 3523–3545,
doi:10.1007/s11269-019-02314-8.
13. Maharjan, B.; Pachel, K.; Loigu, E. Modelling stormwater runoff, quality, and pollutant loads in a large
urban catchment. Proc. Est. Acad. Sci. 2017, 66, 225–242, doi:10.3176/proc.2017.3.02.
14. Alamdari, N.; Sample, D.J.; Steinberg, P.; Ross, A.C.; Easton, Z.M. Assessing the Effects of Climate Change
on Water Quantity and Quality in an Urban Watershed Using a Calibrated Stormwater Model. Water
2017, 9, 464, doi:10.3390/w9070464.
Water 2020, 12, 1727 28 of 33

15. Andimuthu, R.; Kandasamy, P.; Mudgal, B.V.; Jeganathan, A.; Balu, A.; Sankar, G. Performance of urban
storm drainage network under changing climate scenarios: Flood mitigation in Indian coastal city. Sci. Rep.
2019, 9, 7783, doi:10.1038/s41598-019-43859-3.
16. Tscheikner-Gratl, F.; Zeisl, P.; Kinzel, C.; Leimgruber, J.; Ertl, T.; Rauch, W.; Kleidorfer, M. Lost in calibration:
why people still do not calibrate their models, and why they still should—A case study from urban drainage
modelling. Water Sci. Technol. 2016, 74, 2337–2348, doi:10.2166/wst.2016.395.
17. Ngamalieu-Nengoue, U.A.; Javier Martinez-Solano, F.; Iglesias-Rey, P.L.; Mora-Melia, D. Multi-objective
optimization for urban drainage or sewer networks rehabilitation through pipes substitution and storage
tanks installation. Water 2019, 11, 935, doi:10.3390/w11050935.
18. Wu, J.; Kauhanen, P.G.; Hunt, J.A.; Senn, D.B.; Hale, T.; McKee, L.J. Optimal selection and placement of
green infrastructure in urban watersheds for PCB control. J. Sustain. Water Built Environ. 2019, 5, 04018019,
doi:10.1061/JSWBAY.0000876.
19. Knighton, J.; White, E.; Lennon, E.; Rajan, R. Development of probability distributions for urban hydrologic
model parameters and a Monte Carlo analysis of model sensitivity. Hydrol. Process. 2014, 28, 5131–5139,
doi:10.1002/hyp.10009.
20. Knighton, J.O.; Walter, M.T. Critical rainfall statistics for predicting watershed flood responses: rethinking
the design storm concept. Hydrol. Process. 2016, 30, 3788–3803, doi:{10.1002/hyp.10888}.
21. Yu, J.; Qin, X.; Chiew, Y.M.; Min, R.; Shen, X. Stochastic optimization model for supporting
urban drainage design under complexity. J. Water Resour. Plan. Manag. 2017, 143, 05017008,
doi:10.1061/(ASCE)WR.1943-5452.0000806.
22. Shamsi, U.S.; Gamble, B.; Koran, J. Cincinnati’s SWMM Model: A Journey Through Time. J. Water
Manag. Model. 2016, 25, UNSP C398, doi:10.14796/JWMM.C398.
23. Vasconcelos, J.G.; Eldayih, Y.; Zhao, Y.; Jamily, J.A. Evaluating Storm Water Management Model accuracy in
conditions of mixed flows. J. Water Manag. Model. 2018, 26, C451, doi:10.14796/JWMM.CA51.
24. An, H.; Lee, S.; Noh, S.J.; Kim, Y.; Noh, J. Hybrid Numerical Scheme of Preissmann Slot Model for Transient
Mixed Flows. Water 2018, 10, 899, doi:10.3390/w10070899.
25. Burger, G.; Sitzenfrei, R.; Kleidorfer, M.; Rauch, W. Parallel flow routing in SWMM 5. Environ. Model. Softw.
2014, 53, 27–34. doi:10.1016/j.envsoft.2013.11.002.
26. Sadler, J.M.; Goodall, J.L.; Behl, M.; Morsy, M.M.; Culver, T.B.; Bowes, B.D. Leveraging open source
software and parallel computing for model predictive control of urban drainage systems using EPA-SWMM5.
Environ. Model. Softw. 2019, 120, 104484.
27. Chorin, A.J. A numerical method for solving incompressible viscous flow problems. J. Comput. Phys.
1967, 2, 12–26.
28. Chorin, A.J. Numerical solution of the Navier-Stokes equations. Math. Comput. 1968, 22, 745–762,
doi:S0025-5718-1968-0242392-2.
29. Malan, A.G.; Lewis, R.W.; Nithiarasu, P. An improved unsteady, unstructured, artificial compressibility,
finite volume scheme for viscous incompressible flows: Part II. Application. Int. J. Numer. Methods Eng.
2002, 54, 715–729, doi:10.1002/nme.443.
30. Sotiropoulos, F.; Constantinescu, G. Pressure-based residual smoothing operators for multistage
pseudocompressibility algorithms. J. Comput. Phys. 1997, 133, 129–145, doi:10.1006/jcph.1997.5662.
31. Hodges, B.R.; Liu, F. Timescale interpolation and no-neighbour discretization for a 1D finite-volume
Saint-Venant solver. J. Hydraul. Res. 2019, doi:10.1080/00221686.2019.1671510.
32. Rogers, S.; Kwak, D.; Kiris, C. Steady and unsteady solutions of the incompressible Navier-Stokes equations.
AIAA J. 1991, 29, 603–610. doi:10.2514/3.10627.
33. Hodges, B.R. PipeAC code and supporting data. In Texas Data Repository Dataverse; 2020.
doi:10.18738/T8/5GJ6ST.
34. Hodges, B.R.; Liu, F.; Rowney, A.C. A new Saint-Venant solver for SWMM. In New Trends in Urban
Drainage Modeling, Proceedings of the 11th International Conference on Urban Drainage Modelling, Palermo, Italy,
23–26 September 2018; Mannina, G., Ed.; Springer Nature: Basel, Switzerland, 2018; pp. 582–586.
35. Trajkovic, B.; Ivetic, M.; Calomino, F.; D’Ippolito, A. Investigation of transition from free surface to
pressurized flow in a circular pipe. Water Sci. Technol. 1999, 39, 105–112.
36. Wylie, E.B.; Streeter, V.L. Fluid Transients; FEB Press: Ann Arbor, MI, USA, 1983.
Water 2020, 12, 1727 29 of 33

37. Larock, B.E.; Jeppson, R.W.; Watters, G.Z. Hydraulics of Pipeline Systems; CRC Press: Boca Raton, FL, USA,
2000.
38. Todini, E.; Pilati, S. Network modelling and solutions. In Computer Applications in Water Supply: Systems
Analysis and Simulation; Coulbeck, B.; Orr, C.H., Eds.; Research Studies Press, Ltd.: Philadelphia, PA, USA,
1988; Volume 1.
39. Rossman, L.A. EPANET—An advanced water quality modeling package for distribution systems.
In Proceedings of the American Water Works Association 1993 Annual Conference—Proceedings: Water
Quality, New York, NY, USA, 19–23 June 1993; pp. 411–418.
40. Vasconcelos, J.G.; Wright, S.J. Experimental investigation of surges in a stormwater storage tunnel. J. Hydraul.
Eng. ASCE 2005, 13, 853–861, doi:10.1061/(ASCE)0733-9429(2005)131:10(853).
41. Zhou, L.; Liu, D.; Karney, B. Investigation of hydraulic transients of two entrapped air pockets in a water
pipeline. J. Hydraul. Eng. ASCE 2013, 139, 949–959. doi:10.1061/(ASCE)HY.1943-7900.0000750.
42. Vasconcelos, J.G.; Wright, S.J. Anticipating transient problems during the rapid filling of
deep stormwater storage tunnel systems. J. Hydraul. Eng. ASCE 2017, 143, 06016025.
doi:10.1061/(ASCE)HY.1943-7900.0001250.
43. Fernandez-Pato, J.; Garcia-Navarro, P. A pipe network simulation model with dynamic transition between
free surface and pressurized flow. Procedia Eng. 2014, 70, 641–650. doi:10.1016/j.proeng.2014.02.070.
44. Leon, A.S.; Oberg, N.; Schmidt, A.R.; Garcia, M.H. Illinois Transient Model: simulating the flow dynamics in
combined storm sewer systems. J. Water Manag. Model. 2011, 23–39, doi:10.14796/JWMM.R241-02.
45. Pachaly, R.L.; Vasconcelos, J.G.; Allasia, D.G.; Tassi, R.; Bocchi, J.P.P. Comparing SWMM 5.1 Calculation
Alternatives to Represent Unsteady Stormwater Sewer Flows. J. Hydraul. Eng. ASCE 2020, 146, 04020046.
46. Garcia-Navarro, P.; Alcrudo, F.; Priestley, A. An implicit method for water flow modelling in channels and
pipes. J. Hydraul. Res. 1994, 32, 721–742, doi:10.1080/00221689409498711.
47. Casulli, V.; Stelling, G.S. A semi-implicit numerical model for urban drainage systems. Int. J. Numer.
Methods Fluids 2013, 73, 600–614, doi:10.1002/fld.3817.
48. Rossman, L.A. Storm Water Management Model Reference Manual, Volume II—Hydraulics; Technical Report
EPA/600/R-17/111; US EPA Office of Research and Development, Water Systems Division: Cincinnati, OH,
USA, 2017.
49. Casulli, V.; Cattani, E. Stability, Accuracy and Efficiency of a Semi-Implicit Method for Three-Dimensional
Shallow Water Flow. Comput. Math. Appl. 1994, 27, 99–112.
50. Batchelor, G.K. An Introduction to Fluid Dynamics; Cambridge University Press: Cambridge, UK, 2000.
51. Cunge, J.A.; Wegner, M. Intégration numérique des équations d’écoulement de Barré de Saint-Venant par un
schéma implicite de différences finies. La Houille Blanche 1964, 33–39, doi:10.1051/lhb/1964002. (In French)
52. Malekpour, A.; Karney, B.W. Spurious numerical oscillations in the Preissmann Slot method: Origin and
suppression. J. Hydraul. Eng. 2016, 142. doi:10.1061/(ASCE)HY.1943-7900.0001106.
53. Fuamba, M. Contribution on transient flow modelling in storm sewers. J. Hydraul. Eng. 2002, 40, 685–693,
doi:10.1080/00221680209499915.
54. Vasconcelos, J.G.; Wright, S.J.; Roe, P.L. Improved simulation of flow regime transition in
sewers: Two-component pressure approach. J. Hydraul. Eng. ASCE 2006, 132, 553–562,
doi:10.1061/(ASCE)0733-9429(2006)132:6(553).
55. Merkle, C.L.; Athavale, M. Time-accurate unsteady incompressible flow algorithms based on artificial
compressibility. AIAA J. 1987, 87, 397–407.
56. Rogers, S.E.; Kwak, D.; Kiris, C. Numerical solution of the incompressible Navier-Stokes equations for
steady-state and time-dependent problems. In Proceedings of the Tenth Australasian Fluid Mechanics
Conference, Melbourne, Australia, 11–15 December 1989; pp. 3.17–3.20.
57. Kwak, D. Computation of Viscous Incompressible Flows; NASA Technical Memorandum 101090; NASA Ames
Research Center: Moffett Field, CA, USA, 1989.
58. Weiss, J.M.; Amith, W.A. Preconditioning applied to variable and constant density flows. AIAA J. 1995,
33, 2050–2057, doi:10.2514/3.12946.
59. Langtangen, H.P.; Mardal, K.A.; Winther, R. Numerical methods for incompressible viscous flow.
Adv. Water Resour. 2002, 25, 1125–1146. doi:10.1016/S0309-1708(02)00052-0.
Water 2020, 12, 1727 30 of 33

60. Tamamidis, P.; Zhang, G.Q.; Assanis, D.N. Comparison of pressure-based and artificial compressibility
methods for solving 3D steady incompressible viscous flows. J. Comput. Phys. 1996, 124, 1–13.
doi:10.1006/jcph.1996.0041.
61. Patankar, S.V. Numerical Heat Transfer and Fluid Flow; McGraw-Hill: New York, NY, USA, 1980.
62. Lin, S.Y.; Wu, T.M. An adaptive multigrid finite-volume scheme for incompressible Navier-Stokes equations.
Int. J. Numer. Methods Fluids 1993, 17, 687–710, doi:10.1002/fld.1650170804.
63. Mateescu, D.; Païdoussis, M.P.; Bélanger, F. A time-integration method using Artificial Compressibility for
unsteady viscous flows. J. Sound Vib. 1994, 177, 197–205.
64. McHugh, P.R.; Ramshaw, J.D. Damped Artificial Compressibility iteration scheme for implicit calculations
of unsteady incompressible flow. Int. J. Numer. Methods Fluids 1995, 21, 141–153.
65. Shen, J. On a new pseudocompressibility method for the incompressible Navier-Stokes equation.
Appl. Numer. Math. 1996, 21, 71–90, doi:10.1016/0168-9274(95)00132-8.
66. Lin, F.B.; Sotiropoulos, F. Strongly-coupled multigrid method for 3-D incompressible flows using near-wall
turbulence closures. J. Fluids Eng. 1997, 119, 314–324. doi:10.1115/1.2819136.
67. Ge, L.; Sotiropoulos, F. 3D unsteady RANS modeling of complex hydraulic engineering flows. I: Numerical
model. J. Hydraul. Eng. ASCE 2005, 131, 800–808, doi:10.1061/(ASCE)0733-9429(2005)131:9(800).
68. Tang, H.S.; Sotiropoulos, F. Fractional step artificial compressibility schemes for the unsteady incompressible
Navier-Stokes equations. Comput. Fluids 2007, 36, 974–986. doi:10.1016/j.compfluid.2006.01.019.
69. Flamm, K. Has Moore’s Law Been Repealed? An Economist’s Perspective. Comput. Sci. Eng. 2017, 19, 29–40.
70. Morales-Hernández, M.; Sharif, M.B.; Gangrade, S.; Dullo, T.T.; Kao, S.C.; Kalyanapu, A.; Ghafoor, S.K.;
Evans, K.J.; Madadi-Kandjani, E.; Hodges, B.R. High performance computing in water resources
hydrodynamics. J. Hydroinform. 2020, in press.
71. Madsen, P.; Schaffer, H. A discussion of artificial compressibility. Coast. Eng. 2006, 53, 93–98.
doi:10.1016/j.coastaleng.2005.09.020.
72. Ohwada, T.; Asinari, P. Artificial compressibility method revisited: Asymptotic numerical method for
incompressible Navier-Stokes equations. J. Comput. Phys. 2010, 229, 1698–1723. doi:10.1016/j.jcp.2009.11.003.
73. Vermeire, B.C.; Loppi, N.A.; Vincent, P.E. Optimal Runge–Kutta schemes for pseudo time-stepping with
high-order unstructured methods. J. Comput. Phys. 2019, 383, 55–71, doi:10.1016/j.jcp.2019.01.003.
74. Nourgaliev, R.R.; Dinh, T.N.; Theofanous, T.G. A pseudocompressibility method for the numerical
simulation of incompressible multifluid flows. Int. J. Multiph. Flow 2004, 30, 901–937.
doi:10.1016/j.imultiphaseflow.2004.03.010.
75. Golay, F.; Helluy, P. Numerical schemes for low Mach wave breaking. Int. J. Comput. Fluid Dyn. 2007,
21, 69–86. doi:10.1080/10618560701343382.
76. Pattinson, J.; Malan, A.G.; Meyer, J.P. A cut-cell non-conforming Cartesian mesh method for compressible
and incompressible flow. Int. J. Numer. Methods Eng. 2007, 72, 1332–1354. doi:10.1002/nme.2048.
77. Razavi, S.E.; Zamzamian, K.; Farzadi, A. Genuinely multidimensional characteristic-based scheme for
incompressible flows. Int. J. Numer. Methods Fluids 2008, 57, 929–949. doi:10.1002/fld.1662.
78. Keslerova, R.; Kozel, K. Numerical study of steady and unsteady flow for power-law type generalized
Newtonian fluids. Computing 2013, 95, S409–S424. doi:10.1007/s00607-013-0301-9.
79. Asinari, P.; Ohwada, T.; Chiavazzo, E.; Di Rienzo, A.F. Link-wise artificial compressibility method.
J. Comput. Phys. 2012, 231, 5109–5143. doi:10.1016/j.jcp.2012.04.027.
80. Nithiarasu, P.; Liu, C.B. An artificial compressibility based characteristic based split (CBS) scheme for steady
and unsteady turbulent incompressible flows. Comput. Methods Appl. Mech. Eng. 2006, 195, 2961–2982.
doi:10.1016/j.cma.2004.09.017.
81. Shapiro, E.; Drikakis, D. Non-conservative and conservative formulations of characteristics-based
numerical reconstructions for incompressible flows. Int. J. Numer. Methods Eng. 2006, 66, 1466–1482.
doi:10.1002/nme.1606.
82. Su, X.; Zhao, Y.; Huang, X. On the characteristics-based ACM for incompressible flow. J. Comput. Phys.
2007, 227, 1–11, doi:10.1016/j.jcp.2007.08.009.
83. Orang, A.A.; Razavi, S.E.; Pourmirzaagha, H. Computational study of incompressible turbulent flows with
method of characteristics. J. Comput. Appl. Math. 2014, 259, doi:10.1016/j.cam.2013.05.019.
84. Abdollahi, I.; Zamzamian, K.; Fathollahi, R. High-accuracy upwind method using improved characteristics
speeds for incompressible flows. Int. J. Numer. Methods Fluids 2016, 80, 476–489. doi:10.1002/fld.4088.
Water 2020, 12, 1727 31 of 33

85. Clausen, J.R. Entropically damped form of artificial compressibility for explicit simulation of incompressible
flow. Phys. Rev. E 2013, 87, 013309, 1–12. doi:10.1103/PhysRevE.87.013309.
86. Bevan, R.L.T.; Boileau, E.; van Loon, R.; Lewis, R.W.; Nithiarasu, P. A comparative study of fractional step
method in its quasi-implicit, semi-implicit and fully-explicit forms for incompressible flows. Int. J. Numer.
Methods Heat Fluid Flow 2016, 26, 595–623. doi:10.1108/HFF-06-2015-0233.
87. Bevan, R.L.T.; Nithiarasu, P. A dual time stepping approach to eliminate first order error in fractional
step methods for incompressible flows. Int. J. Numer. Methods Heat Fluid Flow 2016, 26, 556–570.
doi:10.1108/HFF-03-2015-0090.
88. Yang, L.; Badia, S.; Codina, R. A pseudo-compressible variational multiscale solver for turbulent
incompressible flows. Comput. Mech. 2016, 58, 1051–1069. doi:10.1007/s00466-016-1332-9.
89. Loppi, N.A.; Witherden, F.D.; Jameson, A.; Vincent, P.E. A high-order cross-platform incompressible
Navier-Stokes solver via artificial compressibility with application to a turbulent jet. Comput. Phys. Commun.
2018, 233, 193–205. doi:10.1016/j.cpc.2018.06.016.
90. Dupuy, D.; Toutant, A.; Bataille, F. Analysis of artificial pressure equations in numerical simulations of
a turbulent channel flow. J. Comput. Phys. 2020, 411, 109407. doi:10.1016/j.jcp.2020.109407.
91. Zhang, F.; Cheng, J.; Liu, T. A direct discontinuous Galerkin method for the incompressible Navier-Stokes
equations on arbitrary grids. J. Comput. Phys. 2019, 380, 269–294. doi:10.1016/j.jcp.2018.11.033.
92. Manzanero, J.; Rubio, G.; Kopriva, D.A.; Ferrer, E.; Valero, E. An entropy-stable discontinuous
Galerkin approximation for the incompressible Navier-Stokes equations with variable density and artificial
compressibility. J. Comput. Phys. 2020, 408, 109241. doi:10.1016/j.jcp.2020.109241.
93. Hajihassanpour, M.; Hejranfar, K. An implicit dual-time stepping high-order nodal discontinuous Galerkin
method for solving incompressible flows on triangle elements. Math. Comput. Simul. 2020, 168, 173–214.
doi:10.1016/j.matcom.2019.08.011.
94. Obrecht, C.; Asinari, P.; Kuznik, F.; Roux, J.J. Thermal link-wise artificial compressibility method: GPU
implementation and validation of a double-population model. Comput. Math. Appl. 2016, 72, 375–385.
doi:10.1016/j.camwa.2015.05.022.
95. Yin, L.; Yang, C.; Ma, S.Z.; Zhang, K.K. Parallel and fully implicit simulations of the thermal convection in
the Earth’s outer core. Comput. Fluids 2019, 193, 104278. doi:10.1016/j.compfluid.2019.104278.
96. Nishiguchi, K.; Bale, R.; Okazawa, S.; Tsubokura, M. Full Eulerian deformable solid-fluid interaction
scheme based on building-cube method for large-scale parallel computing. Int. J. Numer. Methods Eng.
2019, 117, 221–248. doi:10.1002/nme.5954.
97. Niu, Y.Y.; Weng, C.H. A development of a sharp interface AUSMD scheme for an incompressible
preconditioning multi-fluid model. Comput. Fluids 2019, 192, 104269, doi:10.1016/j.compfluid.2019.104269.
98. Delorme, Y.T.; Puri, K.; Nordstrom, J.; Linders, V.; Dong, S.; Frankel, S.H. A simple and efficient
incompressible Navier-Stokes solver for unsteady complex geometry flows on truncated domains.
Comput. Fluids 2017, 150, 84–94. doi:10.1016/j.compfluid.2017.03.030.
99. Kajzer, A.; Pozorski, J. Application of the Entropically Damped Artificial Compressibility model to
direct numerical simulation of turbulent channel flow. Comput. Math. Appl. 2018, 76, 99–1013.
doi:10.1016/j.camwa.2018.05.036.
100. Ramachandran, P.; Puri, K. Entropically damped artificial compressibility for SPH. Comput. Fluids 2019,
179, 579–594. doi:10.1016/j.compfluid.2018.11.023.
101. Loy, K.C.; Bourgault, Y. On efficient high-order semi-implicit time-stepping schemes for
unsteady incompressible Navier-Stokes equations. Comput. Fluids 2017, 148, 166–184.
doi:10.1016/j.compfluid.2017.02.017.
102. Shah, A.; Fayyaz, H.; Rizwan, M. Fourth-order central compact scheme for the numerical
solution of incompressible Navier-Stokes equations. Int. J. Comput. Math. 2017, 94, 2492–2507.
doi:10.1080/00207160.2017.1284315.
103. Guermond, J.L.; Minev, P. High-order adaptive time stepping for the incompressible Navier-Stokes equations.
SIAM J. Sci. Comput. 2019, 41, A770–A788, doi:10.1137/18M1209301.
104. Teschner, T.R.; Konozsy, L.; Jenkins, K.W. Predicting non-linear flow phenomena through different
characteristics-based schemes. Aerospace 2018, 5, 22, doi:10.3390/aerospace5010022.
105. Adibi, T.; Adibi, O.; Razavi, S.E. A characteristic-based solution of forced and free convection in closed
domains with emphasis on various fluids. Int. J. Eng. 2019, 32, 1689–1695. doi:10.5829/ije.2019.32.11b.20.
Water 2020, 12, 1727 32 of 33

106. Parseh, K.; Hejranfar, K. Unsteady preconditioned characteristic boundary conditions for direct numerical
simulation of incompressible flows. AIAA J. 2020, 58, 1476–1489. doi:10.2514/1.J058279.
107. Zalesny, V.B.; Gusev, A.V.; Fomin, V.V. Numerical model of nonhydrostatic ocean dynamics based
on methods of artificial compressibility and multicomponent splitting. Oceanology 2016, 56, 876–887.
doi:10.1134/S0001437016050167.
108. Marsaleix, P.; Michaud, H.; Estournel, C. 3D phase-resolved wave modelling with a non-hydrostatic ocean
circulation model. Ocean Model. 2019, 136, 28–50. doi:10.1016/j.ocemod.2019.02.002.
109. Zhang, J.K.; Li, B.W.; Dong, H.; Luo, X.H.; Lin, H. Analysis of magnetohydrodynamics (MHD) natural
convection in 2D cavity and 3D cavity with thermal radiation effects. Int. J. Heat Mass Transf. 2017,
112, 216–223. doi:10.1016/j.ijheatmasstransfer.2017.04.105.
110. Wang, W.; Li, B.W.; Varghese, P.L.; Leng, X.Y.; Tian, X.Y. Numerical analysis of three-dimensional MHD
natural convection flow in a short horizontal cylindrical annulus. Int. Commun. Heat Mass Transf.
2018, 98, 273–285. doi:10.1016/j.icheatmasstransfer.2018.09.009.
111. Hamdi, M.; Elalimi, S.; Ben Nasrallah, S. Double-population link-wise artificial compressibility method
for simulating two-dimensional magnetoconvection. Math. Problems Eng. 2019, 2019, 7834731.
doi:10.1155/2019/7834731.
112. Dubois, F.; Lallemand, P.; Obrecht, C.; Tekitek, M.M. Lattice Boltzmann model approximated with finite
difference expressions. Comput. Fluids 2017, 155, 3–8, doi:10.1016/j.compfluid.2016.04.013.
113. Rouzbahani, F.; Hejranfar, K. A truly incompressible smoothed particle hydrodynamics based on artificial
compressibility method. Comput. Phys. Commun. 2017, 210, 10–28, doi:10.1016/j.cpc.2016.09.008.
114. Lee, C.H.; Gil, A.J.; Ghavamian, A.; Bonet, J. A Total Lagrangian upwind Smooth Particle Hydrodynamics
algorithm for large strain explicit solid dynamics. Comput. Methods Appl. Mech. Eng. 2019, 344, 209–250.
doi:10.1016/j.cma.2018.09.033.
115. Lemarie-Rieusset, P.G., Artificial compressibility. In The Navier-Stokes Problem in the 21st Century; CRC Press:
Boca Raton, FL, USA, 2016; Chapter 19, pp. 633–663.
116. Bassi, F.; Massa, F.; Botti, L.; Colombo, A. Artificial compressibility Godunov fluxes for variable density
incompressible flows. Comput. Fluids 2018, 169, 186–200, doi:10.1016/j.compfluid.2017.09.010.
117. Wadzuk, B.; Hodges, B. Hydrostatic versus nonhydrostatic Euler-Equation modeling of nonlinear internal
waves. J. Eng. Mech. 2009, 135, 1069–1080, doi:10.1061/(ASCE)0733-9399(2009)135:10(1069).
118. Liu, F.; Hodges, B.R. Applying microprocessor analysis methods to river network modelling.
Environ. Model. Softw. 2014, 52, 234–252, doi:10.1016/j.envsoft.2013.09.013.
119. Hodges, B.R. Conservative finite-volume forms of the Saint-Venant equations for hydrology and urban
drainage. Hydrol. Earth Syst. Sci. 2019, 23, 1281–1304, doi:10.5194/hess-23-1281-2019.
120. Breuer, M.; Hänel, D. A dual time-stepping method For 3-D, viscous, incompressible vortex flows.
Comput. Fluids 1993, 22, 467–484, doi:10.1016/0045-7930(93)90020-A.
121. Kim, W.W.; Menon, S. An unsteady incompressible Navier-Stokes solver for large eddy
simulation of turbulent flows. Int. J. Numer. Methods Fluids 1999, 31, 983–1017,
doi:10.1002/(SICI)1097-0363(19991130)31:6<983::AID-FLD908>3.0.CO;2-Q.
122. Gaitonde, A.L. A dual-time method for two-dimensional unsteady incompressible
flow calculations. Int. J. Numer. Methods Fluids 1998, 41, 1153–1166,
doi:10.1002/(SICI)1097-0207(19980330)41:6<1153::AID-NME334>3.0.CO;2-9.
123. Zhao, Y.; Tai, C.H. Higher-order characteristics-based method for incompressible flow computation on
unstructured grids. AIAA J. 2001, 39, 1280–1287, doi:10.2514/2.1470.
124. Pan, D.; Chakravarthy, S. Unified formulation for incompressible flows. In Proceedings of the 27th Aerospace
Sciences Meeting, Cambridge, MA, USA, 25–29 June 1989; Number AIAA-89-0122, pp. 1–11.
125. Malan, A.G.; Lewis, R.W.; Nithiarasu, P. An improved unsteady, unstructured, artificial compressibility,
finite volume scheme for viscous incompressible flows: Part I. Theory and implementation. Int. J. Numer.
Methods Eng. 2002, 54, 695–714. doi:10.1002/nme.447.
126. Nithiarasu, P. An efficient artificial compressibility (AC) scheme based on the characteristic based split (CBS)
method for incompressible flows. Int. J. Numer. Methods Eng. 2003, 56, 1815–1845, doi:10.1002/nme.712.
127. Ramshaw, J.D.; Mousseau, V.A. Damped Artificial Compressibility method for Steady-state low-speed flow
calculations. Comput. Fluids 1991, 20, 177–186, doi:10.1016/0045-7930(91)90019-E.
Water 2020, 12, 1727 33 of 33

128. Cabuk, H.; Sung, C.H.; Modi, V. Explicit Runge-Kutta method for 3-dimensional internal incompressible
flows. AIAA J. 1992, 30, 2024–2031, doi:10.2514/3.11175.
129. Hodges, B.R.; Liu, F. Algorithms in the SvePy Solver for the Unsteady Saint-Venant Equations; Technical Report,
Texas Data Repository Dataverse; Center for Infrastructure Modeling and Management, University of Texas
at Austin: Austin, TX, USA, 2019. doi:10.18738/T8/ETSJDJ.
130. Leon, A.S.; Ghidaoui, M.S.; Schmidt, A.R.; Garcia, M.H. Application of Godunov-type schemes to transient
mixed flows. J. Hydraul. Res. 2009, 47, 147–156, doi:10.3826/jhr.2009.3157.
131. Crane Co. Flow of Fluids through Valves, Fittings, and Pipe; Technical Paper No. 410; Crane Co.:
Chicago, IL, USA, 1976.
132. Brater, E.F.; King, H.W.; Lindell, J.E.; Wei, C.Y. Handbook of Hydraulics for the Solution of Hydraulic Engineering
Problems, 7th ed.; McGraw-Hill: Boston, MA, USA, 1996.
133. Hodges, B.R.; Laval, B.; Wadzuk, B.M. Numerical error assessment and a temporal horizon for internal
waves in a hydrostatic model. Ocean Model. 2006, 13, 44–64, doi:10.1016/j.ocemod.2005.09.005.
134. Vasconcelos, J.G.; Wright, S.J. Comparison between the two-component pressure approach and current
transient flow solvers. J. Hydraul. Res. 2007, 45, 178–187, doi:10.1080/00221686.2007.9521758.
135. Sanders, B.F.; Bradford, S.F. Network Implementation of the Two-Component Pressure Approach for
Transient Flow in Storm Sewers. J. Hydraul. Eng. ASCE 2011, 137, 158–172.
136. Kerger, F.; Archambeau, P.; Erpicum, S.; Dewals, B.J.; Pirotton, M. A fast universal solver for 1D continuous
and discontinuous steady flows in rivers and pipes. Int. J. Numer. Methods Fluids 2011, 66, 38–48,
doi:10.1002/fld.2243.
137. Li, Z.; Hodges, B.R. Model instability and channel connectivity for 2D coastal marsh simulations.
Environ. Fluid Mech. 2019, 19, 1309–1338, doi:10.1007/s10652-018-9623-7.

© 2020 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).