A Comparison of Navier Stokes and Network Models To Predict Chemical Transport In Municipal Water Distribution Systems

We investigate the accuracy of chemical transport in network models for small geometric configurations. Network model have successfully simulated the general operations of large water distribution systems. However, some of the simplifying assumptions associated with the implementation may cause inaccuracies if chemicals need to be carefully characterized at a high level of detail. In particular, we are interested in precise transport behavior so that inversion and control problems can be applied to water distribution networks. As an initial phase, Navier Stokes combined with a convection-diffusion formulation was used to characterize the mixing behavior at a pipe intersection in two dimensions. Our numerical models predict only on the order of 12-14 % of the chemical to be mixed with the other inlet pipe. Laboratory results show similar behavior and suggest that even if our numerical model is able to resolve turbulence, it may not improve the mixing behavior. This conclusion may not be appropriate however for other sets of operating conditions, and therefore we have started to develop a 3D implementation. Preliminary results for duct geometry are presented. Introduction In this paper, we investigate the accuracy of chemical transport within water distribution system network models by applying high-fidelity computations to model individual network components. Network models have been used for years to simulate chemical transport within water distribution systems in order to better manage chlorine distribution and water quality. Tools such as EPANET have been developed to efficiently predict hydraulic flow and chemical transport behavior within a system. Due to the extreme size and detail of real-world datasets, simplifying assumptions must be applied to fluid flow and chemical transport in order to reduce computational demands. However, these simplifications often reduce the accuracy of a network model since the impact of small-scale phenomenon is neglected. We believe that the resolution of such small-scale phenomenon is necessary in order to more accurately characterize and remediate a water distribution system during a contamination event. This work is motivated by the need to utilize novel mathematical algorithms within numerical simulation to support water distribution system management during a contamination event. Our goal is to improve the mitigation performance by leveraging highfidelity modeling to more accurately characterize the transport of contaminants during an 1Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed-Martin Company, for the United States Department of Energyunder Contract DE-AC04-94AL85000 event. In response to this need, Laird et al. [9] have developed inversion algorithms to reconstruct a contamination event assuming that measurements of contaminant concentrations are available at sparsely located sensors, and Berry et al. [1] have developed combinatorial methods to provide optimal sensor placement strategies. Although these algorithms have been numerically verified, their utility is limited by the accuracy of the numerical models used to predict or characterize chemical transport. High-fidelity computational fluid dynamics (CFD) tools have advanced significantly, enabling the characterization of complex phenomena ranging from chemical reactions in chemical vapor depositions [14] to highly turbulent flows using Direct Numerical Simulation (DNS) and Large Eddy Simulation (LES) [10,11,12] . Combined with large computational resources, these CFD tools can more accurately characterize fluid flow and chemical mixing in complex geometries. To date, very little work has been performed to characterize the complex, fine-scale chemical transport behavior within the turbulent flows of a water distribution system. For practical purposes, modeling such detail is unnecessary, and perhaps more importantly, too expensive for entire networks. However, the state-of-the-art network model is only as good as the accuracy of its underlying assumptions for the smallest, fundamental physics within these networks. By carefully evaluating the fluid flow behavior at this fundamental level, we hope to extract certain corrections to help improve the accuracy of chemical transport in network models. During the initial phase of this investigation, we investigate the cross-joint intersection of four pipes (two inflowing, two outflowing). The standard network simulator assumes an even distribution of chemical within the joint based on an instantaneous mixing assumption. Although intuitively we can predict less than perfect mixing under prescribed inlet conditions, it is difficult to quantify the behavior with sufficient accuracy to develop and apply first-order corrections within a network model. We attempt to characterize this mixing behavior assuming no chemical reaction in the system, although chemical reactions can be handled in our current formulation and may be included in the future. The remainder of the paper presents the mathematical formulation for the high fidelity computational fluid dynamics. Numerical and experimental results are presented. Mathematical Formulation The final goal of this investigation is to develop a first-order correction for mixing phenomena in network models by more accurately characterizing the chemical transport behavior within important pipe geometries through the use of 3D turbulent Navier-Stokes simulation. To accomplish this goal, we first test smaller components (i.e. pipe joints) in 2D under the assumption of laminar flow or using a Reynolds Averaged NavierStokes (RANS) model without resolving the small-scale, unsteady fluctuations of a 3D fully turbulent flow. Eventually, we will build up to a highfidelity characterization consisting of fully-developed 3D turbulence. The underlying physics for these studies is based on the Navier-Stokes equations and convection-diffusion equations. Navier Stokes is based on conservation of mass and momentum. First the conservation of momentum can be formulated as follows: where is the density, u is the velocity vector, t is time, T = -P I + is the stress tensor for a Newtonian fluid, P is the hydrodynamic pressure, is the viscous stress tensor, and g is the gravitational force. Conservation of mass is as follows: Chemical transport is formulated as a separate convection-diffusion equation: where C represents concentration, Deff is the effective diffusion coefficient, and u is the velocity field. Numerical Implementation The incompressible Navier Stokes equations are solved using MPSalsa [15,16]. Our implementation of incompressible Navier Stokes is designed to handle large datasets on massively parallel computers in addition to an efficient implementation for complex dynamics in three dimensions. The code uses a Petrov Galerkin finite element formulation for unstructured meshes and is capable of solving steady state and transient problems using a fully implicit time integration. The nonlinear systems are solved using inexact Newton methods which in turn depend on Krylov based linear solvers. This implementation (MPSalsa) is capable of resolving turbulence using a variety of Reynolds averaged Navier Stokes (RANS) and Large Eddy simulation (LES) based methods and can handle chemical reactions [8]. Using DNS or LES to more accurately evaluate mixing phenomenon within a water distribution network is computationally demanding, even when modeling is focused on small components of the network. To resolve turbulent flow, a high-resolution 3D mesh is necessary to capture the small-scale effects of turbulence, which results in extensive computational processing and memory use. To alleviate these computational demands and establish insight during the initial stages of this investigation, we limited flow and transport to 2D recognizing that turbulence would not be captured properly. For the purposes of the 2D investigation, we employed an 1152 element mesh on 8 to 16 computational processors to simulate the mixing phenomenon within a 2 inch cross joint (see Figure 1). The first step in modeling cross-joint mixing was to develop 2D flow inside a hypothetical pipe joint within which chemical tracer could be added to assess the amount of mixing at the Figure 1 Schematic of pipe cross-joint type joint. This flow field was developed by incorporating oscillatory (sinusoidal) boundary conditions at the inlets shown in Figure 1, thereby emulating turbulent-like flow. Inlet bulk velocities were prescribed at 0.78 meters per second, resulting in an average pipe Reynolds number of ~44,000 based on the characteristics of water at 25 C (i.e. viscosity of 8.9 x 10 kg/m-s, density of 997.0 kg/m) and a 2 inch (.0508 meter) pipe diameter. Figure 2 illustrates the velocity field 10 seconds into the 20 second simulation, at which point 32.5 pipe volumes of water have passed through the joint. Figure 2 – Velocity field within pipe cross-joint type at 10 seconds simulation time Although the Reynolds number is well in excess of 2000 suggesting turbulent flow, the 2D model is not capable of fully resolving turbulence, as shown in Figure 2. MPSalsa accounts for subgrid scale turbulent kinetic energy enabling it to compute a turbulent viscosity term(see Figure 3), and therefore is able to predict the general fluid flow characteristics for this geometry. By definition, this turbulent viscosity represents the turbulent transfer of momentum by eddies. In the case of chemical transport, the term can also be thought of as a coefficient of enhanced diffusion due to mechanical mixing. MPSalsa utilizes the turbulent viscosity to calculate an effective diffusivity Deff Figure 3 – Turbulent viscosity within pipe cross-joint type at 10 seconds of simulation where Dmolecular is the coefficient of molecular diffusion and Sc represents the dimensionless Schmidt number. Therefore, in regions of high turbulent viscosity, one can expect mechanical mixing to dominate the diffusion process. Coupling the velocity field u and effective diffusivity above, MPSalsa uses the convection-diffusion equation (3) to simulate chemical transport.