You are on page 1of 24

Computer methods

in applied
mechanics and
engineering
ELSEVIER Comput. Methods Appl. Mech. Engrg. 167 (1998) 167-190

Three-d imensional analysis of flow in the spiral casing of a


reaction turbine using a differently weighted Petrov Galerkin
method
P.K. Maji, G. Biswas*

Received 6 March 1998

Abstract

A weighted residual based finite element method has been used to predict the flow structure and the head loss in the spiral casing of a
reaction turbine. An explicit Eulerian velocity correction scheme has been employed to solve the complete Reynolds-averaged Navier-
Stokes equations. A Differently Weighted Petrov Galerkin (DWPG) method has been used for spatial discretization. The simulation has been
performed for high Reynolds numbers. The head loss in the spiral casing has been calculated for Re = 5 X IO’ and Re = IOh. The velocity
field and the pressure distribution inside the spiral casing corroborate the trends available in literature. The velocity field reveals the
existence of a free-vortex character of the global flow field. A strong secondary flow is evolved on the cross-stream plane due to the
centrifugal forces which act at right angles to the main flow. 0 1998 Elsevier Science S.A. All rights reserved.

1. Introduction

In order to reduce the design time, an efficient CFD tool such as three-dimensional viscous flow analysis, may
be used for the initial optimization of the geometry of hydraulic components. The numerical simulation of
viscous flow is capable of accurately predicting the overall turbine efficiency, and can be used to solve a variety
of unsteady flow problems such as draft tube surging, pressure oscillations due to the wake effect of the wicket
gate on the runner, and the flow structure in the spiral casing.
The role of the spiral casing (Fig. 1) in a Francis or Kaplan turbine is to distribute the water, as evenly as
possible, to the stay vanes and wicket gates and then to the turbine runner. In a well-designed spiral casing the
pressure head of fluid should be made available to the runner with minimum loss, hence the analysis of the flow
through a spiral casing is important for the design of efficient hydraulic turbines. Very few articles are available
in the open literature on the computational analysis of this problem, possibly because the subject is of direct
commercial interest. Some pioneering investigations have been carried out by Ulrich [ 11. Vu and Shyy [2,3],
Shyy and Vu [4] and Mrsa [5]. In the present work, a detailed numerical simulation of turbulent flow through a
spiral casing has been attempted.
The hydraulic loss from the spiral casing inlet to the distributor outlet sometimes amounts to twice that in a
runner and a reduction of this loss is necessary for improving the overall efficiency of a hydraulic turbine.
Various experimental studies suggest that the loss is dependent on generation of secondary flow. In a spiral
casing, secondary IIow is generated due to flow curvature and viscous forces. This has been observed
experimentally by Tabakoff et al. [6] and Kurokawa and Nagahara [7].

* Corresponding author.

0045-7825/98/$19.00 0 1998 Elsevier Science S.A. All rights reserved


PII: SOO45-7825(98)00000-O
168

Fig. I. Spiral casing of a hydraulic turbine.

Ulrich [l] has modeled the viscous flow in a spiral casing for the first time and described the flow structure.
He has used Galerkin’s finite element technique for his analysis. However, his studies were restricted to the low
Reynolds number regime. Thereafter, Shyy and Vu [Xl, Shyy and Braaten [9], Vu et al. [lo], Vu and Shyy [ll],
Vu and Shyy [2], Vu and Shyy [3], Shyy and Vu [ 121, and Shyy and Vu [4] have conducted several important
investigations for predicting the flow characteristics and energy losses in different parts of the hydraulic turbine.
In their study, full Reynolds averaged Navier-Stokes equations, closed with k--E turbulence model [13], were
solved for high Reynolds numbers. The numerical procedure adopted for all the investigations is essentially the
same as that used by Shyy and Vu [S] and Shyy and Braaten [9], where a pressure correction type semi-implicit
finite volume/finite difference formulation for arbitrary curvilinear coordinates has been adopted.
In order to set up a computational model for the spiral casing, the casing and the distributor should be treated
in a coupled manner since both closely affect each other. It has been found from the work of Shyy and Vu ]4]
that if the presence of the distributor is not suitably accounted for, the flow in the spiral casing can become
persistently oscillatory due to the lack of dissipation of the kinetic energy in the exit region.
The aim of the present work is to develop a numerical tool which is capable of predicting the flow structure
and pressure drop characteristics in a given spiral casing. The spiral casing, considered here, is designed on the
basis of constant velocity at each cross-section. The variation of cross-sectional area A(8) along the peripheral
direction (0) is shown in Fig. 2. A Differently Weighted Petrov Galerkin (DWPG) based finite element
formulation has been employed to solve the three-dimensional Reynolds averaged Navier-Stokes equations in
the geometry of interest.
P.K. Muji, G. Biswas I Cornput. Methods Appl. Mech. Engrg. 167 (1998) 167-190 169

2. Governing equations and the solution scheme

The flow is considered to be viscous, incompressible and turbulent. The computational domain is discretized
into small curvilinear hexahedral elements (Fig. 3). The velocity components and the pressure are collocated at
each node of the element. The Reynolds averaged Navier-Stokes equations for incompressible flows, with an
extra term incorporating the porous medium treatment based on Darcy’s law, have been used here as governing
equations. These equations together with the mass conservation equations can be written in the Cartesian tensor
form as
au,
ax,
Du,_ ““++-- , [ (l+ (2)
Dr - - i)x,

The subscripts i and j can take the values 1, 2, or 3 in three coordinate directions as x, , x2 and xj, respectively.
The x,, x, and x, are the X, y and z of Cartesian coordinates and the velocity components u,, u2 and uj are the
U, u and W, respectively. The equations have been written in the dimensionless form which facilitates
generalization to a large range of problems. Here

Re = W,,Df v

where D is the characteristic length, v is the kinematic viscosity and W,, is the inlet velocity. The variables x,, II,,
r and p are non-dimensionalized from dimensional variables using D, W,, D/W,, and pW,?,, respectively. The
non-dimensional turbulent eddy viscosity v,,, is given by Prandtl’s mixing length model as

(3)

This is also known as a zero equation model, because the eddy-viscosity is directly prescribed, without

Fig. 3. Grid for the spiral casing


Stay vane

Guide vane

Flow
J /
Stoy vane

(a) Plan view (b) Elevation view


Fig. 1. Plan and &cation of a spiral casing.

involving the solution of any equation. The zero equation models obviously have several shortcomings which
are indicated in the turbulent flow literature. Since the geometry itself is extremely complex in this problem. at
this stage the turbulent flow has been dealt with a simple model. Albeit, this simplitication, all the subtle features
of engineering importance are highlighted through this simulation. In any case, in Eq. (3) I,, is given by h6,
where 6 is the non-dimensional radius of the duct at any section and A is a constant (chosen as 0.085).
The term K is the non-dimensionalized Darcy’s coefficient whose value indicates the resistance due to the
distributor region. As such, the distributor has twenty pairs of airfoils which control the mass flux distribution of
fluid (Fig. 4). It would be extremely difficult, and probably unnecessary to include each individual airfoil effect
into the problem. In order to predict the flow behavior inside the spiral casing, the distributor zone has been
modeled as a porous medium in which the effect of individual airfoils has been assumed to be smeared out.

3. Eulerian velocity correction approach

This method is essentially based on the projection scheme of Chorin [ 141 which was originally developed in a
finite difference context and is also used in the MAC method of Harlow and Welch [ 151. This has been extended
to the Finite Element Method by Donea et al. [16] and Ramaswamy et al. [17]. In the present study, the
algorithm has been extended to the solution of turbulent flow in a complex geometry.
In the Eulerian velocity correction method the solution at each time step is obtained through three stages.

Step I: Cdculation of provisional velocities. The provisional velocities are here calculated from the
momentum equations by dropping the pressure terms. Since these explicitly advanced velocities do not
necessarily satisfy the continuity equation, they are called provisional velocities and shown starred.

(4)
P.K. Mqji, G. Biswas I Comput. Methods Appl. Mech. Engrg. 167 (1998) 167-190 171

The above equation is used to calculate the provisional velocities u*, u* and IV*.
Step 2: Solution of pressure equation. The pressure term, which was initially ignored is taken into account
now. Thus, the momentum equations (Eq. (2)) become

au ;lp au,*
l= _
(37 z+-G
(5)
/
Expanding the above equation in time domain the following equation is obtained.
II + I
,z+ %
u,
I
=u,*-AT- ax, .
(6)

Taking the divergence of Eq. (6) and imposing the condition (au, /ax,)“+’ = 0, we get the Poisson equation

the right-hand side (RHS) of which can be computed from the provisional velocities, and the equation then
solved with proper boundary conditions to obtain the pressure at the n + 1 time step.
Step 3: Velocity correction. The velocities are corrected at the next time step by updating the provisional
velocities using the evaluated pressure and Eq. (6) to calculate the velocities CL”+‘, u “+I and We+‘. One can
interpret the role of pressure in incompressible flow as a projection operator which projects an arbitrary vector
field into a divergence free vector field.
The salient features of the above-mentioned solution procedure may be summarized as
(1) Start with initial conditions (u,),; p,,; n = 0.
(2) Calculate the provisional velocities, u,*, using Eq. (4).
(3) Calculate the pressure p”+’ using Eq. (7).
(4) Calculate the corrected velocities u;+’ from Eq. (7).
(5) Update values LI;~u~+‘, p”tp”+‘; ntn + 1.
Steps 2 to 5 are repeated till a steady state condition is reached.

4. Weighted residual method

Let n (geometry of interest) be a bounded region in !)I’ with the piecewise smooth boundary r Let x = {x,},
i = 1,2, 3 denote the vector of spatial coordinates of a generic point in R and let r denote the time value in the
interval I = [0, 01. Also consider n’ as the outward normal vector to r, and <, and c, are non-overlapping
sections of the boundary r which satisfy the following:

<ur;,=r (8)

GUG=!J (9)

The superimposed bar in Eq. (8) represents set closure and 8 in Eq. (9) denotes the empty set.
Now our focus is on the finite element formulation of the problem defined by Eqs. (4), (5) and (7) along with
the following initial and boundary conditions

l Initial and boundary conditions for velocities:

u,(x, 0) =(Q,
“,(X, 7) = g, ; kf’xE q
V7E(O,qJ, 7,,>0 (10)
ii%, = h, ; VXEI;

t/7E(O,7,,1, q,>o
172 P.K. Mqji, G. Biswas I Comput. Mrthodc Appl. Mech. En~rg. 167 (1998) 167- 190

which simply indicate that Dirichlet boundary conditions for velocity are specified on $, and Neumann
conditions on I;, and initial conditions are specified for the velocity.

l Initial and boundary conditions for pressure:

P(X? 0) = PI,
/?(x,r)=s; tlxE4,
v 7 E (0, T,], 7 > 0 (11)
;.Vp = b ; vxq
VTE(O,7,,], 7>0

which means that Dirichlet boundary conditions for pressure are specified on 4, and Neumann conditions on $,,
and the initial pressure field is specified. The quantities g,, h,, s, b are specified functions of x, J, z and r. The
initial quantities (u,),, and p. are the specified functions of x, J, z.
We now write the weak formulation (Weighted Residual Method) of Eqs. (4), (5) and (7).

5. Weak formulation

Let S and V be the vector spaces of the admissible trial functions and test functions respectively which may
be defined as

S = (G(0) 1CDE (H’(R))‘, CD= CDqon rz} (12)

where,

rz = 4 when velocities are specified

r,T = I;, when pressure is specified

@ = {“,> PI (13)
Qx = {g,, s>
V={WIWEH’(Q), W=Oon rz}

where Sobolev space H ‘(0) is the space of the functions that together with their first partial derivatives are
square integrable in R and (H’(fi))4 is the product space of H’(0)
Multiplying Eq. (4) by W, and integrating over 0 yields

(14)

where

Application of Green’s theorem to the diffusion term of above equation together with W= 0 on 1; and
~3~1,
/ax, n, = &,/an = h, (Eq. lo), the above equation reduces to

(15)

Now, multiplying Eq. (5) by W and integrating over R yields


P.K. Maji, G. Biswas I Comput. Methods Appl. Mech. Engrg. 167 (1998) 167-190 173

(16)

Similarly, multiplying Eq. (7) by W and integrating over R yields

do=0 (17)
I
On application of Green’s theorem to the first term of the above equation together with W= 0 on I;,, the above
equation reduces to

(18)

6. Finite element formulation

In order to find a discrete solution of Eqs. (15), (16) and (18) we assume R is discretized in n, hexahedral
elements such that

fizza
C=,
II<, (19)

n ne=0
?=I
where R’ denotes the interior domain of an element. Let r’ be boundary of 0’. Finally, define the ‘interior
boundary’ $,,, as the following
II(I

J,, = u r’ - r (20)
<I=I
Let (u:, p”) be member of 5” and W” be member of V” where S” and V” are finite dimensional subspaces of the
trial (S) and test (V) function spaces, respectively (i.e. S” C S and Vh C V) and defined as

S” = {(u:, ph) E (C’(n))“, U: = g, and pl’ = s on r,*}

vh = {w” E c”(l2), wl’ = 0 on r,g


We assume that this C”(n) functions are typical finite element shape (trial) functions of Lagrange type which
will also belong to H ‘(0) functional space. Now, finite element methods can be formulated by requiring the
discrete solution of, ph to satisfy the weak form of Eqs. (15), (16) and (18).
Thus, Eq. (15) becomes

(21)

In a similar manner, Eq. (16) becomes

a(uh)*
Wh++f-& Wh$$df2. (22)
.I

Following a similar approach, also invoking Eq. (1 l), Eq. (18) becomes
P.K. Muji. G. BISWLY I Cornput. Mrthds Appl. Mrch. Engrg. 167 (1998) 167-I90

(23)

Let the finite set {N,} represent a basis for S”, while {W,} be the basis for V”. The functions Nk, W, are associated
with the node k of the finite element mesh. The discrete solution u:, p’ can be approximated within each
element as a linear combination of the trial functions as follows:

‘I,,
(24)

P” = ,!F, (p(r)),J,,, = {N%(r)1

where n,, is the total number of nodes in each element and {N} represents a column matrix (of dimension 1 X n,,)
with elements N, and {N}7 represents transpose of {N}. The curly brackets { } enclosing some quantity indicate
that the corresponding nodal quantities are arranged node-wise in column vector form.
With these approximations of U: and p” and also discretizing in the time domain together with the concept of
mass lumping, Eq. (21) can be written as

Here and elsewhere, the symbol C3 indicates that the quantity following the symbol is independent of the
integrating variables (see e.g. LHS of above equation). Also, [WI represents a diagonal matrix and can be
expressed as:

WI = diaglw,, W,, . , Y:,!,l

The simplification due to mass lumping is now obvious, and [W] can be trivially inverted and the nodal
acceleration can be separately obtained for each node from Eq. (25).
Similarly, Eq. (22) can be discretized in the time domain

Following a similar approach, Eq. (23) can be written as

(27)

The trial functions {N} are piecewise trilinear and can be written in terms of the transformed local coordinates 5,
r], i (Fig. 5) as

=(I -0~1 -v)(l -5)/s, NZ=(1+5)(1-$(1-~)/8


N,=(l -t5)(1 -n)(l +[)/s, NJ = (1 - <)(I -~)(l + i)/s
(28)
Ns=(l-5)(1+17)(1-1)/S, N6=(t+5)(1+7i9(1-0/8
N,=(1+5)(1+71)(1+~)/8, N,=(l-5)(1+77)(1+63/s
P.K. Maji, G. Biswas I Comput. Methods Appl. Mrch. Engrg. 167 (1998) 167-190 175

Fig. 5. Transformation of a hexahedral element into a cube.

The eight noded hexahedral elements are mapped on a 2 X 2 X 2 cube (Fig. 5) using isoparametric
transformation.
The x, location of any point within the element can be evaluated from the equation

(29)

where, (x,),,~ are the nodal values of x,.


In this formulation, the test functions for the velocity correction equation (25) and pressure correction
equation (27) are the same as the trial functions {N}. However, for Differently Weighted Petrov Galerkin
(DWPG) schemes, the test functions for Eq. (25) are formed by adding a perturbation term to the trial function
and the same will be discussed separately in a subsequent section.
Before going to the next section, the tinal form of velocity correction equation (Eq. (26)) and pressure
Poisson equation (Eqn. (27)) are required to be mentioned.

l Velocity correction equations


The velocity components are corrected using Eq. (26) in the following form

9 Pressure Poisson equation


The pressure correction Eq. (27) may be written as

(31)

where r:, is a boundary of the element 1 such that r: E <,., and U, ri = $.

7. Differently Weighted Petrov-Galerkin (DWPG) based formulation

The numerical scheme based on standard Galerkin finite element method for convection-diffusion problems
produces non-physical oscillatory solution when convection dominates over diffusion. In order to avoid this
problem, alternative weighted residual methods are proposed which are known as Petrov-Galerkin methods. In
the standard Petrov-Galerkin formulation, artificial diffusion is introduced by adding modification functions to
the test functions which are either one or two degrees higher than the basis functions. Christe et al. [ 181
introduced upwinding on linear basis elements by adding a quadratic modification to the standard linear test
functions. Heinrich et al. [19] extended such upwinding to two-dimensional linear elements that used test
176 P.K. Maji. G. Rises I Comput. Methods Appl, Mech. Engrg. 167 (1998) 167-190

functions as tensor products of appropriate one-dimensional functions. Heinrich and Zienkiewicz [20] and
Christie and Mitchell [21] generalized the method using higher-order trial functions. To improve the accuracy of
the time dependent convection dominated problem, Dick [22] has introduced an additional symmetric cubic
perturbation to the quadratic test functions of Christie et al. [IS]. Westerink et al. [23] have indicated that use of
Cubic Petrov-Galerkin (N + 2 degree Petrov-Galerkin) method, which uses test functions two degrees higher
than the trial functions, yields noteworthy improvements of both spatial and temporal accuracy. However, for
nonlinear problems some amount of artificial dissipation may be required when physical dissipation is not
sufficient for convergence. In this case, a small amount of N + I degree upwinding together with N + 2 degree
upwinding may be appropriate for linear elements to make a stable algorithm. Bouloutas and Celia 1241 have
expressed similar views about the cubic Petrov-Galerkin formulation. Their numerical results indicate that
accuracy of Cubic Petrov-Galerkin is comparable to the Galerkin method using quadratic basis functions, at the
cost of much less computational time.
In all the above-mentioned Petrov-Galerkin techniques, the test (weighting) functions applied to the time
derivative is the same as the weighting functions applied to the spatial derivatives. Cardle [25] has used a
modified Petrov-Galerkin scheme for a one-dimensional problem in which the weighting functions applied to
the time derivative are different from those applied to the spatial derivatives. He used the above method to avoid
numerical instability due to the non-symmetrical weighting of the time derivative term in the standard
Petrov-Galerkin method which uses quadratic or a combination of quadratic and cubic test functions.
In the above cited work, Cardle [25] uses symmetrical cubic weighting functions for the time derivative and
non-symmetrical quadratic weighting function for the spatial derivatives. For linear elements, the quadratic bias
influences both the mass and convective matrices, while the cubic bias influences the mass matrices only. Hence,
for linear elements. the cubic modification of the test function affects only a time dependent problem.
Furthermore, the cubic upwind contribution to elemental mass matrices vanish in the case of a lumped mass
formulation.
In the present study, we use linear weighting functions for the time derivative (since we have used the concept
of mass lumping) and quadratic weighting functions for the spatial derivatives. The de&led finite element
discretization of Eq. (4) based on the Differently Weighted Petrov-Galerkin (DWPG) scheme is described
below.
In this formulation, the test functions for the time derivative &A*/&) are the same as trial functions, i.e.

{WI = {N) (32)


The spatial derivatives are upwind biased using the quadratic test functions defined by

(33)

where ( W,<, Wzc), (W, ‘I, W,,,) and (W, i, Wz5) are one-dimensional quadratic test functions corresponding to 5, r]
and 5 direction, respectively. These functions are defined as

w,,=& +a,,, -5X1 +o

w~,=;(1+5)+&(1 +X1+5,

(34)
P.K. Maji, G. B&was I Comput. Methods Appl. Mech. Engrg. 167 (1998) 167-190 177

The coefficients o+, 4, I+ determine the amount of quadratic upwind bias. In one-dimensional steady-state
case it is possible to choose above parameters such that the exact solution is obtained at the nodes. Christie et al.
[ 181 showed that for the one-dimensional case, the optimal (Y is given by

1
cyopt = coth(Pe) - me (35)

where Pe = (U A.x)/~v is the cell Peclet number, AX is the grid spacing and u is the velocity. However, for the
transient case, the truncation errors can not be eliminated and thus there is no precise definition of an optimal (Y.
In our study, we use (Y= Cr (Courant number) as used by Cradle [25]. Because the amplification function of the
numerical solution in three-dimensions is the tensor product of the one-dimensional forms, Taylor series
expansion in the three-dimensional case yields the following rules

+ = Cr, , av = Cr, , q = Cri

where.

“z Al- u Ar u< Ar
Cr,=h, Cr,=*, Cr, = 7
f 0 i
are local Courant numbers in 5, 7, 6 direction, respectively.
Here,
-9 +
u< = eg.u ; % = eg.u ; UC= &l

+++
where umt vectors e,, e,, ej, and element characteristic length h,, hn, h, are defined in Fig. 6. In the present
study we assume that the advective movement on the elemental scale is locally one-dimensional, i.e.

ff$ = cYT= ai_ = Ly= c,

where

C, = Courant number

cuJ!!k<L_+
1+$ (36)
r

Peclet n~m,“@
V

II4 = Ju2+ u2+ iv2 and h = Ih,l + Ih21+ Ih,l (37)


The parameters Ilull, and h are to be evalu%ed2t tm+eelement center (5 = 77= 5 = 0).
Here, h,, h,, h, are the projections of r5, r?, ri vectors in the direction of the normalised local velocity
++-+
vector u. The vectors rE, r,,, rc are in the 5, v, 5 directions, respectively, and can be expressed as

C=2&<

;r: = 2&Z
+ *
ri = 2fcei
178 P.K. Mqji. G. B~SWNS I Com/mt. Mrrhods Appl. Mrch. En,cy,q. It57 (1998) lh7- 190

c=o Planc
q=0 Plant

q =0 Plane k=O Plane

Fig. 6. Typical eight-noded hexahedral element

--f--f
where, et, e,, and < are the unit normal vectors in the 5, 7 and 5 directions, respectively and ,f<,f, and & are
computed from the following expressions

The projection II,, h2 and h, may be expressed as

(38)

The quantity h, can be considered to be a characteristic length of the element in the direction of the flow. The
DWPG based finite element formulation for Eq. (25) is written as

where ri, is a boundary of the element 1 such that r:, E I;, and U, ri? = KS.
P.K. Maji, G. Biswm I Cotput. Methods Appi. Mrch. Engrg. 167 (1998) 167-190 179

Eq. (39) can be rewritten as

i coef~(q)~ = -g, (conv U, + diff u, + porous u,)~ + c (sx,)~ (40)


P= I I

where

coef f, = Nk do

Now, Eq. (40) can be written as

(41)

where ny are the elements which have 9 as common node point and k are the local nodes of elements ny
corresponding to global node q and X (t der u,), is the summation of all the quantities on the right-hand side of
Eq. (41). Then, at each global node q, the contribution from the neighbouring elements will yield

g coef & = C coef ,f,


“Y
(42)
(gtderU,)y= -c ttderu,),
“Y

and finally we get

(C-+Y)x = (;c’q

(43)
( gt der ui L,
(u;), = (u:‘), + AT
g coef
fy
which allows us to compute the provisional velocities at the new time step in term of the quantities at the
previous time steps value.

8. Evaluation of final pressure

In deriving Eq. (7), the divergence au8 /axi)“” has been equated to zero. However, in many simulations, this
may be quite large in some locations compared to other locations due to presence of singularities, grid
distribution, etc. This prevents the evaluation of a correct pressure field. In our simulations we have observed
that maximum residue in some locations is five to ten times more than the r.m.s. residue (typically less than
10m5 has been chosen). This non-zero divergence can spread into the computational domain. Kawamura and
Kuwahara [26] have stated that this divergence term must be included in the Poisson equation in order to
1x0 P.K. Maji. G. Biswu I Cornput. Mrthods Appi. Mwh. Engrg. 167 (1998) 167-190

evaluate the pressure field correctly. In the present simulations, the following pressure Poisson equation, derived
directly from the Navier-Stokes equations is solved along with the pressure boundary conditions (Eq. I I) after
the steady-state velocity field has been obtained:

where

The right-hand side of the above equation is being evaluated using the steady-state velocity fields. Following
a similar approach used for Eq. (7), the Galerkin’s weighted residual formulation of the above Eq. (44) can be
written as

{N}R”’ ’ da + c /,_, {N}6 dT - O(k) (45)


/ y

where the last term (neglected) corresponds to the viscous contribution and the second last due to the Neumann
boundary conditions (Eq. 1I) and the residue at nth and (n + 1)th levels are given below:

9. Evaluation of the Cartesian derivatives

The Cartesian derivatives of any arbitrary function q5 are evaluated in the following way

(46)

where the Cartesian derivatives of shape functions are given by


P.K. Maji, G. Biswas I Comput. Methods Appl. Mech. Engrg. 167 (1998) 167-190 181

-I
J (47)

In the above expression, J ’ is the inverse of Jacobian matrix J and defined as

For calculating the Jacobian, all the derivatives with respect to 5, v and [ are readily evaluated as below using
Eq. (28) in conjunction with Eq. (29)

10. The elemental volume and numerical integration

In the earlier section, we have already mentioned about the elemental volume, do. The elemental volume, di2
is given by

da = (JI d‘$ dv dl (48)

where the determinant of the Jacobian, IJI, is the local volume scaling factor.
In terms of normalised curvilinear coordinates, the elemental integrals take a form

(. . .)IJId5 drl dS

The two point Gauss-Legendre quadrature is used to numerically integrate the above equation. Fig. 7 illustrates
the discrete sampling points in each one of the three directions. The location of Gauss points and their weights
are (-l/A, +I/&) and (1, l), respectively. With this quadrature rule, Eq. (3.93) can be expressed as

Fig. 7. Location of Gauss points in 5 direction.


182 P.K. Muji, G. Biswc~c I Comput. Methods Appl. Mech. E~,qrg. 167 (199X) 167-190

(50)

resulting in a sequence of simple algebraic calculations. Here ig, jg and kg refer to the Gauss points in i, j and k
directions, respectively. The quantities a,,, u,~ and a,< are the weights corresponding to Gauss point (in, jg, kg).

11. Grid size, solution procedure and code validation

In the spiral casing, two-dimensional structured grids are generated at 38 different cross-sections along the
flow direction. Three-dimensional mesh has been formed by connecting corresponding nodes. The number of
grids used in each plane are: 15 X 19 nodes for the inlet straight pipe and 15 X 33 for the spiral-distributor
combined sections.
In order to determine the steady state values, the time marching is continued until the maximum discrepancy
of any dependent variable between two consecutive time steps reaches a predefined upper bound; here a value of
IO-’ has been chosen. The time steps are found out from CFL condition. For the present simulation, the
non-dimensional time unit has been of the order of 10 ‘. Typical CPU time required to reach the steady state for
the above grid system (16 290 nodes and 14 224 elements) is about 36 h on a 32MB RAM, 133 MHz,
DEC-3000/400S machine for a Reynolds number of 10h. A solution method based on incomplete Cholesky
decomposition has been used for solving the discretized pressure Poisson equations.
The numerical scheme was tested on a model problem of flow in a lid-driven cubic cavity with unity linear
dimensions. A (25 X 25 X 25) nonuniform grid has been used to solve the problem for a Reynolds number of
400 and 1000. Fig. 8 shows the profiles of u-component of velocity along the vertical centerline and
u-component of velocity along the horizontal centerline on the x - 4’ plane at z = 0.5 of the cubic cavity for a
Reynolds number of 400.
Fig. 9 show similar results for Re = 1000. The results due to present computation show good agreement with
those presented by Ku et al. [27] using pseudospectral method and Kato et al. [28] using GSMAC FEM. Both
the groups have used 25 X 25 X 25 non-uniform grids for Re = 400 and 3 1 X 31 X 31 non-uniform grids for
Re = 1000. The extreme values of the above-mentioned centerlines velocities for Re = 400 are shown in Table 1

DWPG -&A-
SJPG -Q-+
GWRM +33-
OWPG -&-A-
KU et d. -_~t GWRM -GXI-
Koto et 01. -O-O- Ku et 01 _-MC
Koto et 01. -O-O-

0
I
05 I D 02 01 01 CB I
-I -0 I 0

u X

Fig. 8. Velocity profiles on (a) vertical centerline and (b) horizontal centerline of the x ~ v plane at ; = 0.5 of the cubic cavity for Re = 400.
P.K. Maii, G. Biswas I Comput. Methods Appl. Mrch Engrg. 167 (1998) 167-190 183

OB

06

Y V
SUPG +
oWPG e SUPG--o-o_
04
GWRM 0 OWPG -A-h-
Ku at 01 -- GWRM -

K” .t 01. -wt_
00
KatO et 01 -

(b)

-1 -0 s 0 05 I 0 02 04 CL 08 1

u
X

Fig. 9. Velocity profiles on (a) vertical centerline and (b) horizontal centerline of the x - ?‘ plane at z = 0.5 of the cubic cavity for
Re = 1000.

Table 1
Extreme values of centerline velocities on x -r plane at z = 0.5 for Re = 400

Scheme u ,,,,” at x = 0.5, 7 = 0.5 IJ,,,,,, at .y = 0.5, z = 0.5 U,nlX at J = 0.5, t = 0.5

%~” I’ u ,n,n x lJ IlldX x

DWPG -0.2041 0.27241 -0.33813 0.86553 0.17312 0.13441

12. Results and discussion

The present simulation has been performed for Reynolds numbers of 5 X lo5 and lo6 using the following
initial and boundary conditions imposed on different boundaries of the spiral casing (Fig. 10).

Exal boundary
!
“, = - xD/rpv
ny = 0
“z = - m/rgv

Fig. 10. Plan view of the spiral and the boundary conditions. Fig. 1 I. Velocity vectors on the horizontal mid-plane for Re = IOh.
184 P.K. Muji, G. Biswtrs I Cornput. M&hods Appl. Mech. Engrg. 167 (1998) 167-190

Initial and boundary conditions for velocities


At r = 0,
u = u = 0, w = 1 at the inflow plane
IA= u = w = 0 everywhere in the domain except at the inflow plane

At 7>0,
u = u = w = 0 at the solid wall
u = u = 0 and w = 1 at the inflow plane
&A/& = &1/6+z = aw / an = 0 at the exit boundary

Initial and boundary conditions for pressure


At r = 0, p = 0 everywhere in the domain
At 7>0,
i)p/& = 0 at the inflow plane
?&~/an = 0 at the solid wall
p = gauge pressure with respect to local ambient = 0 at the exit boundary
The above-mentioned initial and boundary conditions for velocities and pressure have been described earlier
in a compact parametric form through Eqs. (10) and (11).
Fig. 11 shows the velocity vectors on the horizontal mid-plane of the casing. It can be seen that the uniform
velocity at the entry gradually culminates in a free vortex flow within the casing. Also, the mid-plane velocity
vectors are in qualitative agreement with the results due to Shyy and Vu [4]. It may be mentioned that the
present simulation has been performed using the same values of Darcy’s coefficient, K which varies between
0.30 to 1.20, as reported by Shyy and Vu [4] for the distributor zone. The variation of K along the distributor
height and in the circumferential direction has also been considered.
Fig. 12 depicts the static pressure variation on the same horizontal mid-plane of the casing. The static pressure

Fig. 12. Static pressure distribution on the horizontal mid-plane Fig. 13. Secondary flow at dlfferent cross sections for Re = IO”.
for Re = IO’. (a) %= 0”; (b) %= 90”; (c) %= 180”; (d) %= 240”.
P.K. Maji, G. Biswas I Comput. Methods AppI. Mech. Engrg. 167 (1998) 167-(90 I85

varies mainly along the radial direction rather than along the circumferential direction. The centrifugal force
pui/r is balanced by the radial pressure gradient dp/ dr. With the exception of the region near the end of the
spiral, the pressure is evenly distributed within the casing. A wavy pattern is observed near the end portion. This
is attributed to the obstruction created by the partition plate. The pressure distribution presented by Vu and Shyy
[4] shows wavy pattern near the outside wall of the casing. No such waviness in pressure distribution is
observed in the present computation.
Fig. 13 depicts the behavior of secondary flow at six different cross-stream planes located at 0 = O”, 60”, 90”,
150”, 180” and 240”. Several factors such as the wall curvature, no-slip boundary along the solid wall,
exit-opening on the inner circumference, and continuous reduction in the cross-sectional area along the
circumference of the spiral casing, influence the development of secondary flow. The secondary flow exhibits
strong inward motion close to the top and the bottom wall region together with double spherical vortices in the
core region. This is primarily brought about by an imbalance between the centrifugal force and the radial
pressure gradient. The inward flow near the wall causes transportation of low energy fluid from the boundary
layer to the inner radius of the spiral. This induces boundary layer suction effect. The strength of secondary flow
varies from one cross-section to another. At the beginning (0 = 0”) of the spiral casing (Fig. 13(a)) the flow is
accelerated evenly along the casing height. As we move along the circumferential direction, the formation of
double vortices due to the interaction between curvature effect and viscous effect becomes more prominent
(f3 = 60”, 90” in Fig. 13(b-c)). After a 150” turn (Fig. 13d) the secondary flow continues to persist. However, in
this case the double vortices are somewhat weaker. After a 240” turn (Fig. 13) the radial flow is further
accelerated thereby causing disappearance of the double vortices. The generation of the secondary flow and its
dissemination are corroborated by the observation of Kurokawa and Nagahara [7].
Fig. 14 shows the static pressure contours at two different cross-sections located at 8 = 90”, and 150”. The
static pressure decreases from the outside wall towards the distributor outlet. The total pressure contours on
cross-stream planes for 0 = 90” and 150” are shown in Fig. 15. Total pressure is maximum at the center while
near the wall it is minimum. This distribution agrees with the observation of Kurokawa and Nagahara [7].
Fig. 16 shows the trajectories of the fluid particles within the spiral. The particles are released from three
different points along a vertical line on a cross-plane. Each particle moves from the inlet region to the exit
describing a helical three-dimensional path. Such particle traces help in extracting the dormant features of
three-dimensionality of fluid motion.
So far, we have discussed about the flow structure and our understanding of the flow physics within the spiral.
In order to provide useful input to design and development, a detailed parametric study has been undertaken.
Before we start a discussion in this direction it is necessary to be acquainted with a set of geometrical
terminologies related to the spiral casing. We take a recourse to Fig. 17 which explains some of the important
zones on a cross-stream plane. Distribution of various flow parameters in these zones characterizes the
performance of a spiral casing.
The radial distribution of flow parameters along HH is shown on four different cross-sectional planes located

Fig. 14. Static pressure contour at different cross sections for Re = IO’. (a) 0 = 90”; (b) 0 = 150”
(b)

Fig. IS. Total pressure contour at different cross sections for Re = IO”. (a) H = 90”: (b) H = 1.50”

Fig. 16. Particle traces withm the spiral casing

Vl
I
I I 4 of
I Hydraulic
turbine

------(

\
V\

Fig. 17. Important zones on the cross-stream plane


P.K. Maji, G. Biswas I Comput. Metho& Appl. Mech. Engrg. 167 (1998) 167-190 187

at 0 = 15”, 90”, 150” and 240” in Fig. 18. The static pressure, p, generally decreases towards distributor exit. The
total pressure (p, = p + 0.5~~) initially increases from outside wall to inner radius and then keeps decreasing up
to the stay vane inlet. However, within the distributor region, the total pressure, pt, increases due to flow
acceleration. The radial velocity, uI, is directed towards the outer wall except near the stay vane (r/r,, = 1)
zone. Near the stay vane, it changes the direction and increases rapidly. The fluid continues to accelerate
throughout the distributor zone. This increase corresponds to a rapid decrease in the static pressure or rapid
increase in favorable pressure gradient dp/dr towards the inner radius. It is also seen that a weak outward u,
component is induced at 0 = 90”. Also, a weak inward radial velocity is induced for H 2 1.50” near the outer
radius. The axial velocity, u, is zero along HH on all the cross-stream planes. The tangential velocity, u,,
increases from the outside wall towards the inner radius up to the distributor inlet. Within the distributor, the
tangential velocity becomes smaller while the radial velocity increases. However, after H 3 240” (Fig. 18(d)),
tangential velocity does not decrease much within the distributor which makes the distribution almost uniform. It
may be mentioned that the distribution of flow parameters on a cross-stream plane at 0 = 90” due to present
computation compares favorably with the experimental observations of Shyy and Vu [4].
The distribution of the flow parameters along VV is shown in Fig. 19 for two different cross-sectional planes
located at 0 = 1.5” and 90”. A low total pressure zone near the wall and a high total pressure zone at the center
are seen in all the sections. Distribution pattern of tangential velocity is similar to the total pressure distribution.
The radial velocity distribution at 0 = 15” (Fig. 19(a)) shows that weak outward flow is induced in the center and
relatively stronger inward flow near the wall. Both the inward and outward radial flows are much stronger at
6’ = 90” (Fig. 19(b)).
The axial distribution of flow characteristics at the exit of the guide vane channel (along GG in Fig. 17) is
shown in Fig. 20 for two different angles (0 = 15”, and 90”). Distribution of radial velocity is almost the same in

(b)

Fig. 18. Radial distribution of flow parameters along HH at different cross sections for Re = IOh. (a) 0 = IS”; (b) 19= 90”; (c) 0 = 150”; (d)
8 = 240”.
IX8 P.K. Maji, G. Biswu I Comput. Methods Appl. Mrch. Engrg. 167 (1998) 167p/90

v
-0.71
-Cl., 4 11 0 0.11 0.1 -0.1 .0.*1 0 0.21

‘/2R y/zR
(4 b)
Fig. 19. Distribution of Row parameters along VV at different cross sections for Re = IO” (a) H = IS”; (b) H = 90”.

(4 (b)

Fig. 20. Axial distribution of flow parameters along GG at different cross sectlons for Re = IO’. (a) H = 1.5’; (b) 0 = 90”.

all the sections. However, the peak value of the radial velocity reduces slightly in the circumferential direction.
On the other hand, the peak value of the tangential velocity increases in this direction. This is due to the
acceleration of the main flow along the spiral. Distribution of axial velocity is same in all the sections and the
axial velocity is zero at the mid-point. The static pressure is zero (gauge pressure with respect to local ambient)
at the guide vane exit. The total pressure increases in the circumferential direction due to flow acceleration in the
guide vane channel. It can also be seen from the axial velocity distributions that the flow diverges slightly
towards the axial direction at the exit.
To examine the variation of mean flow parameters along the circumferential direction, the averaged values of
pressure and velocity over the GG-section are plotted in Fig. 21 as a function of angular position. The radial

P I-

Fig. 21. Average flow characteristics on GG as a function of polar angle (0)


P.K. Maji, G. Biswas I Comput. Methods Appl. Mech. Engrg. 167 (1998) 167-190 189

Table 2
Head loss in a spiral casing

Reynolds No. Inlet head Exit head Total head Head loss pel
(kg-m/s) (kg-m/s) loss (kg-m/s) unit mass (mwc”)

500000 49.95 34.69 15.26 0.07 I


1000000 390.93 279.01 I I I .92 0.26

’ mwc = meter water column.

velocity is uniform in general. However, the tangential velocity increases slightly in the circumferential
direction. As a result, the total pressure also increases in the same direction. The distribution of tangential and
radial velocity gives a nearly axisymmetric flow in a spiral casing except for the zones near the inlet and
towards the end of the spiral.
The average total pressure head at the inlet and the exit of the spiral were determined from the computed
results. The loss of total pressure head within the spiral casing was calculated from the two quantities mentioned
above. Table 2 shows the loss of total pressure head for two different Reynolds numbers of interest.

13. Conclusions

The structure of a three-dimensional flow field and the pressure distribution in the spiral casing of a hydraulic
turbine have been obtained using a Differently Weighted Petrov Galerkin (DWPG) technique. The results show
excellent qualitative agreement with the experimental results. Contribution of this investigation is in the
application of an accurate mathematical tool in developing a reliable model. The model has successfully
analysed a three dimensional flow in an extremely complex geometry at high Reynolds numbers. The model
forms a basis of development of an efficient design tool for hydraulic turbines.

References

[I] D. Ulrich, Berechung der laminaren stromung in einem turbinenspiralgehiiuse mittels einer finite-element-formulierung der Navier-
Stokes-Gleichungen in nicht-druckintegrierter form, Ingenieur Archiv (Archive of Appl. Mech.) 56 (1986) 192-200.
[2] T.C.Vu and W. Shyy,Viscous flow analysis as a design tool for hydraulic turbine components, ASME J. Fluid Engrg. I I2 (1990) 5-11.
[3] T.C. Vu and W. Shyy, Navier-Stokes flow analysis for hydraulic turbine draft tubes, ASME J. Fluid Engrg. I I2 (1990) 199-204.
[4] W. Shyy and T.C. Vu, Modeling and computation of flow in a passage with 360-degree turning and multiple airfoils, ASME J. Fluid
Engrg. I I5 (1993) 103-108.
[5] Z. Mrsa, Optimal design of spiral casing tongue and wicket gate angle by decomposition method, Int. J. Numer. Methods Fluids I7
(1993) 995-1002.
[6] W. Tabakoff, B.V. Vittal and B. Wood, Three-dimensional flow measurements in a turbine scroll, ASME J. Engrg. Gas Turbines Power
106 (1984) 516-522.
[7] J. Kurokawa and H. Nagahara, Flow characteristic in spiral casings of water turbines, in: IAHR Symp. 1986, Vol. 2, Paper No. 62,
Montreal, Canada, 1986.
[S] W. Shyy and T.C. Vu, A numerical study of incompressible Navier-Stokes flow through rectilinear and radial cascade of turbine blades,
Comput. Mech. I (1986) 269-279.
[9] W. Shyy and M.E. Braaten, Three-dimensional analysis of the flow in a curved hydraulic turbine draft tube, Int. J. Numer. Methods
Fluids 6 (1986) 861-882.
[IO] T.C.Vu, W. Shyy, M. Braaten and M. Reggio, Recent development in viscous flow analysis for hydraulic turbine components, in: IAHR
Symp. 1986, Vol. 2, Montreal, Canada, 1986.
[I I] T.C. Vu and W. Shyy, Navier-Stokes computation of radial flow turbine distributor, ASME J. Fluid Engrg. I IO (1988) 29-32.
[ 121 W. Shyy and T.C. Vu, On the adoption of velocity variable and grid system for fluid flow computation in curvilinear coordinates, J.
Comput. Phys. 92 (1991) 82-105.
[13] B.E. Launder and D.B. Spalding, The numerical computation of turbulence flows, Comput. Methods Appl. Mech. Engrg. 3 (1974)
269-289.
1141 A.J. Chorin, A numerical method for solving incompressible viscous flow problems, J. Comput. Phys. 2 (1967) 12-26.
[ 151 F.H. Harlow and J.E. Welch, Numerical calculation of time dependent viscous incompressible flow of fluid with free surface, Phys.
Fluids 8 (1965) 2182-2188.
190 P.K. Maji. G. Biswu I Comput. Mrthod.t Appl. Mech. Eqrg. I67 (1998) 167-190

[ 161 J. Donea, S. Giuliani and H. Lava], Finite element solutions of the unsteady Navier-Stokes equations by fractional step method.
Comput. Method Appl. Mech. Engrg. 30 (1982) 53-73.
[ 171 B. Ramaswamy, T.C. Juf and E. Akin, Semi-implicit and explicit finite element schemes for coupled fluid/thermal problems, Int. J.
Numer. Methods Engrg. 34 ( 1992) 675-696.
[ 181I. Christie, D.F. Griffiths, A. Mitchell and O.C. Zienkiewlcz, Finite element methods for second order differential equations with
significant first derivatives, Int. J. Numer. Methods Engrg. IO (1976) 1389-1396.
[19] J.C. Heinrich. P.S. Huyakorn, O.C. Zienkiewicz and A.R. Mitchell, An upwind finite element scheme for two dimensional convective
transport equation, Int. J. Numer. Methods Engrg. I I (1997) 131-143.
[20] J.C. Heinrich and O.C. Zienkiewicz, Quadratic finite element schemes for two-dimensional convective transport problem, Int. J.
Numer. Methods Engrg. 1 I ( 1997) 183 I- 1844.
[21] 1. Christie and A.R. Mitchell. Upwinding of higher order Galerkin method\ in conduction-convection problems, Int. J. Numer.
Methods Engrg. I2 (1978) 1764-1771.
[22] E. Dick, Accurate Petrov-Galerkin methods for transient convective diffusion problems, Int. J. Numer. Methods Engrg. IY (1983)
1425-1433.
1231 J.J. Westerink and D. Shea, Consistent higher degree Petrov-Galerkin methods for the solution of the transient convectton-diffusion
equation, Int. J. Numer. Methods Engrg. 28 (1989) 1077-I 101.
[24] E.T. Bouloutas and M.A. Celia, An improved cubic Petrov-Galerkin method for simulation of transient advection-diffusion processes
in rectangularly decomposable domain, Comput. Methods Appl. Mech. Engrg. 92 (1991) 289-308.
[2S] J.A. Cardle, A modification of the Petrov-Calerkin method for the transient convection-diffusion equation. Int. J. Numer. Methods
Engrg. 3X (199.5) 171-181.
[26] T. Kawamura and K. Kuwahara. Direct simulation of a turbulent inner flow by finite difference method. AIAA-85.0376, 1985.
1271 H.C. Ku. R.S. Hirsh and T.D. Taylor, A pseudospectral method for solution of the three-dimensional incompressible Navier-Stokes
equations. J. Comput. Phys. 70 ( 1987) 439-462.
1281 Y. Kato, H. Kawai and T. Tanahashi. Numerical flow analysis m a cubic cavity by the GSMAC finite-element method, JSME Int.
Journal, Series II. 33(4) (1990) 649-658.

You might also like