Bistrian 2015 AMM

JID: APM
ARTICLE IN PRESS [m3Gsc;November 21, 2015;8:44]
Applied Mathematical Modelling 000 (2015) 1–22
Contents lists available at ScienceDirect
Applied Mathematical Modelling

journal homepage: www.elsevier.com/locate/apm
Weighted proper orthogonal decomposition of the swirling flow

exiting the hydraulic turbine runner
D.A. Bistrian a,∗, R.F. Susan-Resiga b
a
Electrical Engineering and Industrial Informatics Department, ”Politehnica” University of Timisoara, 331128 Hunedoara, Romania
b
Hydraulic Machinery Department, ”Politehnica” University of Timisoara, Bvd. Mihai Viteazu 1, RO-300222 Timisoara, Romania
a r t i c l e i n f o a b s t r a c t
Article history: In this paper we propose a framework for orthogonal decomposition of swirling flows applied
Received 13 May 2015 to problems originating from turbomachines, where dynamics of the swirling flow in the draft
Revised 30 October 2015
cone is strongly influenced by the turbine discharge coefficient. A weighted proper orthogonal
Accepted 10 November 2015
decomposition method (wPOD) is proposed for analyzing the evolution of the swirling flow
Available online xxx
exiting the hydraulic turbine runner as the turbine discharge is modified. The chief idea is
Keywords: that through orthogonal decomposition one can better identify the leading modes of the axial
Reduced order modeling and circumferential velocity profiles perturbation with respect to a simple base flow. More-
Proper orthogonal decomposition over, it is expected that only one mode is actually responsible for the stability loss. The key
Radial basis functions innovation introduced in this paper resides in identification of two perturbation quantities
Unsteady swirling flows having vanishing integrals on the computational domain. By applying the weighted POD on
these perturbation quantities, the property of vanishing integral is conserved for each individ-
ual mode. As a result, the POD representation of the velocity field is achieved with a number
of modes significantly lower compared with other classic techniques. The efficiency of the
reduced order model developed in this paper is tested by comparing the computed solution
with the experimentally measured profiles. In addition, a qualitative analysis of the reduced
order model by its correlation coefficient and root mean squared error (RMSE) is performed.
© 2015 Elsevier Inc. All rights reserved.
1. Introduction
In the field of modern hydraulic turbines, despite the advent and maturation of high-performance computing, Direct Numer-
ical Simulations (DNS) solvers remain so computationally intensive that they cannot be performed as often as needed (see for
reference the work of Buron et al. [1], Guenette et al. [2], Nicolet et al. [3], Ruprecht [4], for turbomachinery problem examples).
Alternative numerical techniques of modeling the fluid-structure interaction in turbomachinery have been proposed, each
having its advantages and disadvantages. For example, a methodology underlying the calculation of axisymmetrical turbulent
flow in the diffuser cone of the hydraulic Francis turbine was presented by Resiga and his collaborators [5] and the comparison
with Laser Doppler Anemometry measurements [6] confirms the correctness of the numerical results. Tridon et al. [7] pro-
posed a novel flow control technique focused on discharge imbalance mitigation in Francis turbine draft-tube bays. Furthermore,
accurate numerical simulations are dealing with very high dimensional systems. To overcome these limitations, solution algo-
rithms based on spectral methods have been proposed in our previous investigations [8,9] and a reduced order model based on a
∗
Corresponding author. Tel.: +400761365811.
E-mail address: diana.bistrian@fih.upt.ro, diaalina@yahoo.com (D.A. Bistrian).
http://dx.doi.org/10.1016/j.apm.2015.11.015
S0307-904X(15)00729-5/© 2015 Elsevier Inc. All rights reserved.
Please cite this article as: D.A. Bistrian, R.F. Susan-Resiga, Weighted proper orthogonal decomposition of the swirling flow
exiting the hydraulic turbine runner, Applied Mathematical Modelling (2015), http://dx.doi.org/10.1016/j.apm.2015.11.015
JID: APM
2 D.A. Bistrian, R.F. Susan-Resiga / Applied Mathematical Modelling 000 (2015) 1–22
Nomenclature
r radial coordinate
rw dimensionless radial distance to the cone wall
q dimensionless discharge coefficient
m dimensionless flux of moment of momentum
v(aq) (r), v(uq) (r) axial and circumferential velocity profiles
v(aq) , v(uq) base flow

v(aq) ,
v(uq) perturbation part
(q)
f (q) (r) ≡ rva (r) perturbation quantity
(q) (q)
g(q) (r) ≡ r2
va (r) vu (r) perturbation quantity
N number of experimentally measured velocity profiles
n number of grid points in r− direction
w , = 1, . . ., N the weights of the snapshot matrix
φ j (r), j = 1, . . ., N the POD basis functions
(q)
a j , j = 1, . . ., N the POD coefficients
p number of the retained POD modes
( q)
uPOD (r) the reduced order solution
( q)
RMSEu the root mean square error between the experimental data u and its reduced order solution
mapped onto the computational domain
(q)
Cu the correlation coefficient between the experimental data u and its reduced order solution
λ j , j = 1, . . . , N eigenvalues of POD decomposition

Ep = N
j=p+1 λ2j projection error of order p in terms of Frobenius norm
List of abbreviations
POD proper orthogonal decomposition
wPOD weighted proper orthogonal decomposition
DNS direct numerical simulations
SPIV stereoscopic particle image velocimetry
ROM reduced order model
BEP best efficiency point
EVD eigenvalue decomposition
CVT centroidal Voronoi tesselation
EVD-wPOD weighted POD algorithm based on eigenvalue decomposition
EVD-POD classic POD algorithm based on eigenvalue decomposition
RBF Radial basis function
wPOD-RBF model reduced order model based on weighted POD and RBF interpolation
RMSE the root mean square error
two-level differential quadrature algorithm was derived in [10] to study interactions between roughness structures and concen-
trated vortices.
Evaluation of the hydraulic turbine efficiency for the whole range of admissible discharge, is possible in the early stage of
design by the standard experimental investigation on a turbine model. The real challenge is to determine the velocity profiles
exiting the runner without a priori knowledge of the runner design, aiming to develop further solutions to stabilize the flow in
the discharge cone. Two directions for investigating the velocity profiles cohabit in the recent studies. Most papers existing in
the literature consider theoretical swirl configurations. For instance, Wang and Rusak [11] provide a new study of the axisym-
metric vortex breakdown phenomenon, based on a thorough investigation of the axisymmetric unsteady Euler equations. Their
study is limited to inviscid and axisymmetric swirling flows and does not consider the interaction with the vortex breakdown
phenomenon. In a previous work, Resiga and his coworkers proposed a suitable analytical representation of the swirling flow as
a superposition of three distinct vortices, taking the discharge coefficient as independent variable [12]. In the opposite manner,
investigators consider velocity profiles fitted by experimentally measured data. The averaged velocity field inside the inter-blade
channel of a propeller turbine runner measured using a stereoscopic particle image velocimetry (SPIV) technique is addressed
by Aeschlimann [13] and his coworkers. A particle image velocimetry (PIV) system is used by Iliescu et al. [14] to investigate the
flow velocity field in the case of a developing cavitation vortex at the outlet of a Francis turbine runner. Both DNS numerical
simulation based on theoretical characterization of swirling flow and performing the experimental measurements in situ are not
JID: APM
D.A. Bistrian, R.F. Susan-Resiga / Applied Mathematical Modelling 000 (2015) 1–22 3
appropriate solutions for investigating hydraulic machines due to the tremendous effort required in computational resources
and materials.
Among the most significant reduced-order modeling techniques in fluid mechanics is the Proper Orthogonal Decomposition
(POD), which is also known in the literature as Karhunen–Loève expansion and principal component analysis. POD has its incep-
tion in statistical analysis and is rooted in the work of independent investigators like Kosambi [15] in 1943, Loéve [16,17] in 1945,
Karhunen [18] in 1946 and Obukhov [19,20] in 1954. Lumley [21] was the first to apply this method to identify coherent struc-
tures in dynamical systems in 1967. Most informative on the application of proper orthogonal decomposition to fluid dynamics
is the book authored by Holmes, Lumley and Berkooz [22].
The tendency of using POD is growing, although the application of this technique is primarily limited to flows whose coherent
structures can be hierarchically ranked by their energy content. The range of applications for POD reduced order modeling is
substantial, being successfully applied in fluid mechanics (Caiazzo et al. [23], Stefanescu et al. [24], Luo et al. [25]), turbulent
flows (Wang et al. [26], Walton et al. [27], Osth et al. [28]) and oceanography (see for reference the work of Du et al. [29], Fang
et al. [30], Alekseev and Navon [31]). Recently, in computational fluid dynamics of swirling jets, POD has gained special attention.
Markovich et al. [32] employed POD for a comparative analysis of low and high swirl confined flames and jets. Oberleithner
et al. [33] investigated the three-dimensional coherent structures in a swirling jet undergoing vortex breakdown using proper
orthogonal decomposition and dynamic mode decomposition.
Our purpose is to enhance POD to make it an efficient practical tool for numerical investigation of swirling flow exiting the
runner of Francis hydraulic turbine. We propose the application of the method of Proper Orthogonal Decomposition to obtain
a low dimensional approximate description of the radial profiles of both the axial and circumferential velocity components of
the swirling flow exiting the hydraulic turbine runner, considering a set of snapshots generated at various values of a turbine
parameter. These low dimensional approximate descriptions are used as basis functions for a reduced order model (ROM). In
our approach the snapshots are the radial distribution of the radial and circumferential velocities experimentally measured at
the inlet of the discharge cone for a set of discharge coefficients of the hydraulic turbine. The issue is that, in the state space, the
distribution of snapshots can be far from uniform, meaning that in some regions of the state space the snapshots are concentrated
while in other regions they are non-uniformly distributed. The irregular distribution of snapshots can bias the resulting reduced
POD basis. Based on the method of snapshots or ”strobes”, developed by Sirovich [34], we extend the standard POD approach by
incorporating a technique to obtain an efficient reduced POD basis from recorded snapshots, which are not in general uniformly
distributed in the state space. In our approach, the basis vectors are computed through the eigenvalue decomposition (EVD) of a
weighted correlation snapshot matrix, where each column is defined as a snapshot vector multiplied by a weighting coefficient.
Our procedure is a development of the ideas initiated in [35] for parameter-dependent nonturbulent flows and exemplified also
in [36]. We refer to this approach as weighted POD or wPOD.
In the classic approach of POD [18–20], the choice and number of basis functions is critical to the accuracy of the ROM. The
POD method establishes the basis functions, whilst the number of basis functions in the ROM model is based on the desired
accuracy. The work performed in this paper focuses on achieving a quality solution of the POD model of the swirling flow exiting
the turbomachinery runner independently of the retained number of the POD basis. The key innovation introduced in this paper
resides in identification of two perturbation quantities having vanishing integrals on the computational domain. By applying the
weighted POD on these perturbation quantities, the property of vanishing integral is conserved for each individual mode. As a
result, the POD representation of the velocity field is achieved with a number of modes significantly lower than with other classic
techniques.
The proposed framework outlines steps towards hydrodynamic stability analysis of flow in modern hydraulic turbines with
potential applications, where computational efficiency remains an important consideration. Investigation of the hydrodynamic
stability of the leading modes can offer an insight into hydrodynamic stability of the flow in the discharge cone.
There are other two very important advantages that result when a reduced order model is employed instead of a full DNS
simulation. First, the method proposed in this paper extends the limitation of experimental measurements and allows the com-
putation of the swirling flow in any operating regime starting from a set of few measured profiles. Second and by far the most
important advantage is that, compared with the marching computational approach in which terms are evaluated on the fine
mesh at each time step to solve the complete Navier-Stokes equations, we introduce an improved POD technique to obtain a
reduced order model, so that the problem of computing the velocity field is reduced in dimensionality.
The remainder of this article is organized as follows. Identification of flow quantities for an efficient orthogonal decomposition
is presented in Section 2. The principles governing the Proper Orthogonal Decomposition are discussed in detail in Section 3 that
includes also an improved algorithm for computing the weighted proper orthogonal modes for identification of the swirling flow
exiting the runner of hydraulic turbine. In particular, we discuss the properties of the new model. We devote the Section 4 to
revealing the numerical results estimated for the hydraulic turbine within the full operating range. The results are compared to
existing measurements for radial velocity profiles and a rigorous error analysis for the ROM model is performed. Summary and
conclusions are drawn in the final section.
2. Identification of flow quantities for an efficient orthogonal decomposition
Due to their adaptability to a wide range of site conditions in terms of head and discharge [37], Francis turbines cover the
largest part of the installed hydropower capacity in the world, owing their name to the British-American engineer James B.
Francis in 1840. A cross-section of the device is presented in Fig. 1. The flow enters through a volute or scroll casing which is
JID: APM
Fig. 1. Meridional cross-section of the runner with the velocity profiles measured in the discharge cone downstream section S.
designed to evenly distribute the flow around the periphery of the inlet guide vane. As the inlet guide vanes increase the angular
momentum of the fluid, the turbine rotor turns the flow from the radial to the axial direction. The draft tube is the machine
component where the flow exiting the runner is decelerated, where the excess of kinetic energy is converted into static pressure.
Several simplified assumptions are made on the swirling flow at the runner outlet, consistent with the flow conditions in
Francis turbines. We assume a steady, axisymmetric, parallel flow, of an inviscid and incompressible fluid. In order to derive the
mathematical model of the swirling flow in the draft tube, we express the physical quantities in their dimensionless form using
a reference radius Rref , chosen as the runner outlet radius and a reference velocity Vre f = Rre f , where represents the runner
angular velocity, as shown in Fig. 1. Consequently, by introducing the dimensionless velocity v = V/(Rre f ), we are concerned
with the axial-circumferential velocity profile downstream the runner, corresponding to the radial section S = { r | r ∈ [0, rw ]},
where rw = Rw /Rre f is the dimensionless radial distance to the cone wall and has the value 1.063 on our test bench (see Fig. 1).
The integral quantities that define a swirling flow are, according to the fundamental equation of turbomachinery,
rw
q≡ v(aq) (r)2rdr, (1)
0
rw
m≡ v(aq) (r) rv(uq) 2rdr, (2)
0
representing the discharge coefficient and the flux of angular momentum, respectively, in dimensionless form. Eqs. (1) and (2)
define a relationship m(q) which is an integral blueprint of the turbine operation.
Let us consider the swirling flow exiting a hydraulic turbine runner and further ingested by the draft tube, as measured (or
computed) axial and circumferential velocity profiles
v(aq) (r), v(uq) (r). (3)

The operating point is indicated by the discharge coefficient q, which varies within a possibly wide range around the best effi-
ciency point.
More often, the hydraulic turbine has to operate far away from the best efficiency point (BEP), which is the operating regime
when the efficiency ”hill chart” [12] presents a peak efficiency. The corresponding operating range is defined by flow rates which
are higher or lower than the value producing the highest efficiency of the turbine. In order to derive the mathematical model of
the velocity components, both axial and circumferential velocity profiles are measured in the survey section S downstream the
runner, using the Laser Doppler Velocimetry technique [5,12], at several operating regimes covering both the overload, the BEP
and the partial load of Francis turbine.
The integrals in relations (1) and (2) are computed applying a Newton–Cotes rule and the corresponding values are given in
Table 1, for the considered operating regimes. In Fig. 2 the dependence of the dimensionless moment of momentum flux on the
discharge coefficient is illustrated.
We intend in this paper to introduce a novel mathematical framework based on Proper Orthogonal Decomposition to identify
a reduced order model for estimating the axial and circumferential velocities in the computational domain S as the turbine is
operated within the range [q1 , q7 ]. The reduced order model presented in this paper aims at a-priori computation of the swirling
flow exiting the runner, in the early design stages of the hydraulic turbine. The novelty introduced herein resides in identification
JID: APM
0.06
Moment of momentum flux coefficient

0.05 m(q)
0.04
0.03
0.02
0.01
0.28 0.30 0.32 0.34 0.36 0.38 0.40 0.42 0.44
Discharge coefficient q
Fig. 2. Dimensionless moment of momentum flux coefficient m(q) versus the discharge coefficient.
Table 1
Turbine coefficients (q, m) computed from experimental data.
Operating Discharge coefficient Moment of momentum

regime q flux coefficient m
110.7%BEP q1 = 0.440220 m1 = 0.016394

105%BEP q2 = 0.412889 m2 = 0.025731
102.5%BEP q3 = 0.406330 m3 = 0.029602
100%BEP q4 = 0.393281 m4 = 0.032020
97.4%BEP q5 = 0.385796 m5 = 0.034291
91.9%BEP q6 = 0.368836 m6 = 0.043610
71.4%BEP q7 = 0.286725 m7 = 0.055262
of two quantities having vanishing integrals on the computational domain. By applying the orthogonal decomposition on these
perturbation quantities, the property of vanishing integral is conserved for each individual mode, respectively. This strategy
compensates for the loss of accuracy when the POD series is truncated.
v(aq) ,
First we consider the velocities expressed as a juxtaposition of a base flow (v(aq) , v(uq) ) and a perturbation part ( v(uq) ). The
flow perturbation with respect to the base flow is defined as
v(aq) (r) = v(aq) (r) − v(aq) and

v(uq) (r) = v(uq) (r) − v(uq) (r). (4)
A quick analysis of the velocity profiles shape (3) suggests that a simple base flow can be considered, in the form of a constant
axial base flow and a parabolic shape for the circumferential base flow,
v(aq) = constant, (5)

(q)
vu (r) = −α r + β r . 2
(6)
The base flow (5) and (6) should also possess the integral properties (1) and (2), i.e.
rw
v(aq) 2rdr = q, (7)
0
rw
v(aq) (r) rv(uq) 2rdr = m. (8)
0
The integral condition (7) immediately leads to

q
v(aq) = . (9)
rw 2
JID: APM
Table 2
The axial velocity of the base flow (5) and the parameters of the
circumferential velocity of the base flow (6) computed from ex-
perimental data.
Operating regime va α β
110.7%BEP 0.363818 0.307257 0.430320

105%BEP 0.341230 0.132564 0.276666
102.5%BEP 0.335810 −0.031209 0.107664
100%BEP 0.325026 −0.063327 0.084679
97.4%BEP 0.318840 −0.012586 0.157451
91.9%BEP 0.304823 0.074726 0.298761
71.4%BEP 0.236963 −0.161386 0.125566
From (1) and (7) it follows that the axial velocity perturbation has a vanishing discharge,
rw
v(aq) (r)2rdr = 0,
∀q, (10)
0
by construction.
To determine the base flow for the circumferential velocity, we introduce the flow quantity
v(aq) (r)
g(q) (r) = r2 v(uq) (r), (11)
and we determine the base flow v(uq) (r) such that the condition
rw
g(q) (r)dr ≡ 0, ∀q, (12)
0
holds. Eq. (12) states that the flow quantity g(q) (r) has a null integral on the computational domain and, hence it has a null
moment of momentum flux
rw
v(aq) (r)
2r2 v(uq) (r)dr ≡ 0, ∀q. (13)
0
Introducing the relation (4) in (13) leads to

rw rw
v(uq) (r)r2 dr = v(uq) (r)r2 dr. (14)
0 0
The system of Eqs. (8) and (14) allows the computation of the two parameters α and β for the base flow circumferential
velocity profile (6),
⎧ rw rw
⎪ m
⎨−α r3 v(aq) (r)dr + β r4 v(aq) (r)dr =
0 0 2
rw . (15)
⎪
⎩−α r 4
r 5
( )
+β
w w
= r vu (r)dr
2 q
4 5 0
Introducing the notation

rw rw rw
r3 v(aq) (r)dr = I1(q) , r4 v(aq) (r)dr = I2(q) , r2 v(uq) (r)dr = I3(q) ,
0 0 0
it follows from (15) that the parameters of the circumferential base flow are
2 mrw 5 − 10I2(q) I3(q) 5 mrw 4 − 8I1(q) I3(q)

α= · , β= · . (16)
rw 4 5I2(q) − 4I1(q) rw 2rw 4 5I(q) − 4I(q) rw
2 1
The axial velocity of the base flow (5) and the parameters of the circumferential velocity of the base flow (6) are given in Table 2,
for the considered operating regimes.
Examining the quantities involved in Eqs. (10) and (12) it seems that the orthogonal decomposition should be applied to the
following two quantities
v(aq) (r),
f (q) (r) ≡ r (17)
v(aq) (r)
g(q) (r) ≡ r2 v(uq) (r), (18)
where both perturbation quantities (17) and (18) have vanishing integrals,
rw rw
f (q) (r)dr = 0, and g(q) (r)dr = 0, ∀q. (19)
0 0
JID: APM
The next figures present the experimental data achieved for the swirling flow in the draft tube at section S for the considered
(q) (q)
operating regimes. The measured velocity profiles (va , vu ) and, respectively, the axial and circumferential velocity profiles of
the base flow (v(aq) , v(uq) ) are depicted in Fig. 3. These measurements are used to develop the reduced order model that predicts
the inlet velocity components for the operating regimes for which the experimental measurements are not available. We will
further detail the results in the section dedicated to numerical experiments.
3. Reduced order model formulation
3.1. Proper Orthogonal Decomposition (POD) and Weighted Proper Orthogonal Decomposition (wPOD)
The Proper Orthogonal Decomposition (POD), represents at the moment state-of-the-art for many model reduction problems.
The strong point of POD is that it can be applied to non-linear partial differential equations, especially for smooth systems in
which the energetics can be characterized by the first few modes. Reduced order modeling based on POD has recently received
an increasing amount of attention to study the dynamics of the flows in different applications. According to Lumley [38] the
method of POD is an alternative to the usual Fourier decomposition and yields an optimal set of basis functions in the sense that
no other decomposition of the same order captures an equivalent amount of kinetic energy. The POD and its variants are also
known as Karhunen–Loève expansions in feature selection and signal processing, empirical orthogonal functions in atmospheric
science or principal component analysis in statistics. Recently, POD was successfully applied in weather and climate modeling
[39,40], in data assimilation [41,42] and also in inverse problems [43,44].
In the field of turbomachinery, the development of accurate and reliable low dimensional models represents an extremely
important task. The aim of this section is twofold: first we introduce a weighted proper orthogonal decomposition of the pertur-
bation quantities and second, we derive a reduced order model based on radial basis function interpolation.
Suppose that S = (0, rw ) ⊂ R represents the computational domain, we consider the space L2 (S) of square integrable func-
tions on S

L2 (S) = φ : S → R |φ|2 dr < ∞ , (20)
S
be a Hilbert space endowed with the inner product

φi (r), φ j (r) L2 (S) = φi (r) φ j (r) dr f or φi , φ j ∈ L2 (S), (21)
S

and the induced norm φ L2 (S) =
φ , φL2 (S) for φ ∈ L2 (S).
We concentrate in this section on approximating the components of the perturbation quantities (f(q) (r), g(q) (r)) previously
obtained over the computational domain S (see Fig. 1) as a finite sum of form

p
u(q) (r) ≈ a(jq) φ j (r), (22)
j=1
expecting that this approximation becomes exact as p → +∞. To simplify the presentation, we denote any perturbation quantity
by u(q) (r). A solution to the approximation problem (22)is given by the spectral methods [8] when orthogonal polynomials like
Chebyshev polynomials [9] or Legendre polynomials are used for the basis functions φ j (r). In Proper Orthogonal Decomposition
approach, we seek orthonormal basis functions, i.e.

φi (r), φ j (r) L2 (S) = δi j , (23)
where δ ij is the Kronecker delta symbol

1 if i= j
δi j = (24)
0 if i = j
and the coefficients aj are computed by projecting the velocity field onto the POD modes

a(jq) = u(q) (r), φ j (r) . (25)
L2 (S)
The POD problem reduces to find the subspace X = span{φ1 , φ2 , . . ., φ p } spanned by the sequence of orthonormal functions φ j (r)
such that the p-approximation of u(q) (r) is the best possible in the least square sense. The approximation problem (22) is then
equivalent to the following constrained minimization problem
2
p
(q)
(q)
min u ( r ) − u (r), φ j (r) φ j (r) dr (26)
φ1 ,φ2 ,...,φ p S 2
j=1 L (S )

s.t. φi , φ j 2 = δi j , 1 ≤ i ≤ j ≤ p.
L (S )
JID: APM
Fig. 3. The experimentally measured velocity profiles and the axial and circumferential velocity profiles of the base flow for the considered operating regimes.
JID: APM
For the step size h = rw /(n − 1) let us introduce a spatial grid in S by
ri = (i − 1)h f or i = 1, . . ., n. (27)
Given the discharge coefficients q1 , . . ., qN , we consider the following ensemble of snapshots

U := u(q1 ) (r), . . ., u(qN ) (r) , (28)
as n realizations of each u(q) (r) in locations r1 , r2 , . . ., rn on the computational domain S. p

The approximation problem (26) is then equivalent to finding an orthonormal basis φ j (r) with p ≤ min {n, N} solving
j=1
2
N

p
(q )
(q )
min u (r ) − u (r ), φ j (r) φ j (r) . (29)
φ1 ,φ2 ,...,φ p
=1
j=1
2
L (S )
The projection error of order p in the Frobenius norm is given by

N

E p = I − T U = λ2j , (30)
F
j=p+1
where p represents the dimension of the subspace on which all the state vectors live, λj are the singular values of the snapshot
matrix, I stands for identity matrix of order p, is the column matrix of POD basis vectors truncated at its first p columns and
T stands for its transpose.
For the turbine investigated in this paper, the experimental data is not obtained as a function of time, rather it is derived as
a function of the coefficient of discharge that varies in a non-uniform fashion. Considering that the velocity field is measured for
the set of discharge coefficients of the hydraulic turbine presented in Table 1, it is obvious that in the state space, the distribution
of snapshots is far from uniform, being clustered for overload and non-uniformly distributed at partial load. This irregular dis-
tribution of snapshots can bias the resulting reduced POD basis. Specifically, under the problem assumptions, given a snapshot
set, for any number of basis (q)
functions POD minimizes the average of the square distance between the snapshots u (r) and the
projections u(q) (r), φ j (r) 2 φ j (r). This concern occurs in problems where the snapshots depend on some controls or parame-
L (S)
ters. To overcome this, Du and Gunzburger [45]introduced the centroidal Voronoi tesselation (CVT) to generate a reduced-order
basis from the generators of a CVT of snapshots. CVT is a clustering technique which for discrete data sets is known as k-means
clustering. In this approach, the generators of a CVT of snapshots are the centroids of the clusters formed from the snapshot set.
It was then a widely used method for data compression and it was employed also in the work of Burkardt et al. [46] for complex
systems and Du et al. [47] where more efficient algorithms for constructing CVTs are discussed. Although CVT was described as
a model and method for optimal point distributions of regions or sets of discrete data, this technique is useful when it is applied
to large data sets, contributing to an optimal reduction of the model by clustering the snapshots.
For the problem we are investigating, only seven sets of snapshots are experimentally provided according to Table 1, thus the
clustering is not an appropriate solution. We address the issue of irregular distribution of snapshots by employing a weighted
POD technique exemplified also by Christensen et al. [35] for parameter-dependent non turbulent flows and Carlberg and Farhat
[36] for model reduction of static systems.
In our approach, the basis vectors are computed applying a slightly modified POD on a weighted snapshot matrix, where each
column is defined as a snapshot vector multiplied by a weighting coefficient. We refer to this approach as weighted POD or wPOD
[41,48,49].
We consider a weighting function as the Gaussian function
2
γi j = exp −qi − q j /2σ 2 , (31)
where σ controls the kernel length. The element γ ij of the weighting matrix has a large value when qi and qj are close. Then we
compute the snapshot weights in the form

N
γ j
w = , = 1, . . ., N. (32)
N
j=1
A weighted snapshot matrix can be defined as

Uw := w1 u(q1 ) (r), w2 u(q2 ) (r), . . ., wN u(qN ) (r) , (33)
where 0 < w ≤ 1 for each w .
JID: APM
The wPOD basis vectors can be determined by the following sequence of constrained, weighted least-squares problem
2

N p
(q )
(q )
min w u ( r ) − u (r ), φ j (r) φ j (r) (34)
φ1 ,φ2 ,...,φ p
=1
j=1
2
L (S)

s.t. φi , φ j L2 (S) = δi j , 1 ≤ i ≤ j ≤ p.
In matrix notation, the wPOD method constructs a basis matrix w that minimizes the Frobenius norm of the difference
between Uw and its projection w w T U w . The projection error of order pof weighted snapshot matrix Uw onto the wPOD basis
in Frobenius norm is given by

N

E p = I − w w U
w T w
= λ2j , (35)
F
j=p+1
where p represents the dimension of the subspace on which all the state vectors live, λj are the singular values of the weighted
snapshot matrix, I stands for identity matrix of order p, w is the column matrix of wPOD basis vectors truncated at its first p
columns and ’T’ stands for its transpose.
We have two choices to determine the wPOD basis vectors. The first one is achieved by employing an eigenvalue decomposi-
tion of a correlation matrix. In the second approach, a singular value decomposition is applied directly to the weighted snapshot
matrix. The eigenvalue decomposition does have strong connections with singular value decomposition and more details can be
found in [50–52]. If the snapshot matrix is symmetric and positive definite, then its eigenvalues coincide with its singular values.
We adopt the first approach in this paper and we present hereinafter the algorithm for the weighted proper orthogonal
decomposition. We consider the data stored in the weighted snapshot matrix U w given by (33).
Algorithm: Weighted POD algorithm based on eigenvalue decomposition (EVD-wPOD)
(1:) Precompute the weighted snapshot matrix

Uw := w1 u(q1 ) (r), w2 u(q2 ) (r), . . ., wN u(qN ) (r) .
(2:) Calculate the empirical correlation matrix
T

N
C = U w (U w ) / wi , (36)
i=1
where (U w )T represents the transpose of the weighted snapshot matrix.

(3:) Compute the eigenvalue decomposition
CV = DV, (37)
where D = diag(λ1 , . . ., λN ) ∈ RN×N
is the matrix of eigenvalues and V ∈ RN×N
is the column matrix of eigenvectors.
(4:) Sort the eigenvalues in descending order λ1 > λ2 > · · · > λN and also the corresponding eigenvectors.
(5:) Determine the orthonormal basis vectors φ ∈ w :
U w v
φ = , = 1, . . ., N, (38)
U w v 2
T
where v = v1 , . . ., vN
∈ V denotes the eigenvector to C with corresponding eigenvalue λ and · 2 stands for Euclidean
norm.
The columns of snapshot matrix U are given by
1
N
u(q ) = φi , w u(q ) Rn φi . (39)
w
i=1
A point of controversy in the community is whether the POD should be applied to the snapshot matrix itself or to its fluctuat-
ing part obtained after subtracting the ”ensemble” mean. In most previous studies in the literature the second approach has been
adopted. In our context, the perturbation quantities (f(q) (r), g(q) (r)) determined in Section 2 have null averages over the compu-
r r
tational domain, i.e. 0 w f (q) (r)dr = 0 w g(q) (r)dr =0, ∀q, therefore the wPOD method has been applied without subtracting the
ensemble mean. Centering the snapshots in our case is thus equivalent to introducing an extra correlation. The key innovation
introduced in this study is that the presented approach is able to preserve the property of a null integral on every wPOD mode.
Introducing an extra correlation like subtracting the ensemble mean will alter this property.
The question of subtracting or not the ensemble snapshots mean is addressed also by Chatterjee [50], where two numerical
examples are provided. Moreover, information on applying the POD efficiently can be found in the excellent paper of Aubry [53]
and Aubry et al. [54].
JID: APM
3.2. Identification of the reduced order modelby RBF interpolation
The radial basis function (RBF) method has undergone rapid progress as an active tool of mathematical interpolation. Since
it was introduced by Rolland Hardy in 1970 [55] for applications in cartography, the interpolation of scattered data of various
types based on RBFs has been demonstrated to be useful in many application domains like image processing [56], domain de-
composition [57] or more recently for reduced order modeling of unsteady fluid flows [58,59]. A detailed description of this topic
for various problems may be found in Micchelli [60], Fornberg et al. [61] and Driscoll and Fornberg [62]. Investigations upon
accuracy and stability of RBF based interpolation may be found in Fornberg and Wright [63], Fasshauer [64] and Chenoweth [65].
The algorithm previously described allows the identification of a reduced order model of the form

p
u(POD
q )
(r) = a(jq ) φ j (r), = 1, . . . , N, (40)
j=1
(q )
in which φ j (r) represents the wPOD basis, the coefficients a j are computed by projecting the snapshots onto the wPOD modes
and p is the truncation order. In the following, we employ the method of RBF interpolation, as a generalization of Hardy’s multi-
(q)
quadric and inverse multiquadric method [66], for numerical interpolation of the coefficients a j for q ∈ [q1 , qN ].
Considering the determined wPOD coefficients as a set of distinct nodes {xi }i=1 ⊂ R2 and a set of function values { fi }i=1 ⊂ R,
p×N p×N
the problem reduces to find an interpolant s : R2 → R such that
s(xi ) = fi f or i = 1, . . ., p × N, (41)
where N is the number of operating regimes for which experimentally measured data are available and p is the number of
(q )
retained wPOD modes. Note that we use the notation fi = a j , j = 1, . . ., p, = 1, . . ., N for scattered points values and x =
(x, y) ∈ {1, . . ., p} × [q1 , qN ] for scattered points coordinates.
Considering BL2 R2 the Beppo-Levi space [67] of distributions on R2 with square integrable second derivatives, equipped
with the rotation invariant semi-norm
2 2 2
∂ 2 s(x) ∂ 2 s(x) ∂ 2 s(x)
s(x) 2 = + +2 dx, (42)
R2 ∂ x2 ∂ y2 ∂ x∂ y
we seek the smoothest interpolant surface in the affine space

SBL = s ∈ BL2 R2 | s(xi ) = fi , i = 1, . . ., p × N , (43)
i.e.,
s(x) = arg min s(x) 2 .

(44)
s∈SBL
The semi-norm (42) measures the energy or ”smoothness” of the surface interpolant s, such that interpolants having a small
semi-norm are considered smoother than those having a large semi-norm. Following Duchon [68] and Green and Silverman
[69], the solution to the problem (44) is a function of the form

p×N
s(x) = c0 + c1 x +
βi k( x − xi 2 ), (45)
i=1
where k is a real valued function defined on the kernel k ∈ K : R p×N × R p×N → R, · 2 is the Euclidian distance between the
points x and xi , the coefficients βi ∈ R are constant real numbers and P (x) = c0 + c1 x is a global polynomial function, usually
considered of small degree. The points xi are referred as centers of the Radial Basis Functions k(r) = K (x, xi ), where the variable
r stands for x − xi 2 . In our approach we use the so called thinplatekernel k(r) = r2 ln (r + 1).
Ensuring that the interpolation surface lies in the Beppo-Levi space s ∈ BL2 R2 implies the following side conditions

p×N

p×N

p×N
βi = xi βi = yi βi = 0, (46)
i=1 i=1 i=1
and the constraints

p×N
βi P (xi ) = 0. (47)
i=1
Considering that {p1 , p2 } represents a basis for the polynomial P and {c0 , c1 } are the coefficients that give the polynomial P (x)
in terms of this basis, the interpolation conditions (41) with the side conditions (46) and constraints (47) lead to the following
JID: APM
linear system to be solved for the coefficients that specify the RBF

K P β f
= , (48)
PT 0 c 0
T
where Ki j = k xi − x j , i, j = 1, . . ., p × N, Pi j = p j (xi ), i = 1, . . ., p × N, j = 1, 2, β = β1 , . . ., β p×N , c = (c0 , c1 )T , f =
T 2
f1 , . . ., f p×N . The zeros in (48)denote matrices or vectors of appropriate dimensions and ‘T’ stands for the transpose of a
matrix or vector. Solving the linear system (48) determines the constant coefficients β and the polynomial coefficients c and
hence the interpolant surface s(x).
The methodology presented herein leads to the following linear model (denoted in the following wPOD-RBF model) for esti-
mation of the perturbation quantities (f(q) (r), g(q) (r)) ≡ u(q) (r) for any operating regime q ∈ [q1 , qN ]

p
u(POD
)
a(j ) φ j (r), a(j ) = x ∈ {1, . . ., p} × [q1 , qN ],
q q q
(r ) = s(x), (49)
j=1
( q)
where a j are the interpolated coefficients, φ j (r) are the wPOD basis functions, p represents the number of the wPOD basis
functions retained for the reduced order model and q denotes any value of the discharge coefficient in the interval [q1 , qN ].
4. Analysis of the reduced order model
4.1. Validation of the reduced order model and numerical results
The weighted POD methodology presented in the previous section was employed to construct a POD representation of the
difference between the original experimental data and its crude approximation as a base flow. Considering the set of N = 7
snapshots presented in Fig. 3, representing the axial and circumferential velocity profiles of the swirling flow experimentally
measured at the inlet of the draft tube (see Fig. 1), we employ the weighted POD algorithm described in Section 3 combined with
RBF interpolation to obtain a mathematical model of the velocity field.
The key innovation introduced in this paper resides in the manner of application of the POD method. Usually, POD practition-
ers apply the POD directly onto the experimental or numerical data as snapshots. We improved the classical methodology and
we focused on achieving a quality solution of the mathematical model of the swirling flow exiting the turbine runner minimiz-
v(aq) (r) and g(q) (r) ≡ r2
ing the retained number of the POD basis. First we identified the flow quantities f (q) (r) ≡ r v(aq) (r)
v(uq) (r)
having vanishing integrals on the computational domain. By applying the weighted POD methodology presented in the paper
on these perturbation quantities, the property (19) of vanishing integral is conserved for each individual mode. The mathemat-
ical demonstration of this property within the presented approach is given in Appendix (Proposition 1). The advantage of this
approach is that the POD representation of the velocity field is achieved with a number of modes significantly lower than in the
classic manner.
The projection error is given by the root mean square error (RMSE) which is calculated by

n 2
(q) 1 (q)
RMSEu = uPOD (ri ) − u(q) (ri ) , q = q1 , . . . , q7 . (50)
n
i=1
(q)
In this expression, uPOD and u(q) denote the reduced order solution (mapped onto the computational domain) and experimental
data at the node ri , respectively, and n represents the number of nodes on the computational domain. The correlation coefficients
defined below are used as additional metrics to estimate the quality of the reduced order model

(q) H (q) 2
uPOD (r) · u (r)
Cu( ) =
q 2
(q) H (q) , q = q1 , . . ., q7 , (51)
uPOD (r) · uPOD (r) u (r) · u(q) (r)2
(q ) H
2
( q)
where u(q) represents the measured data, uPOD is the estimated solution obtained with the wPOD-RBF model, ( · ) represents the
Hermitian inner product, · 2 stands for the Euclidean norm and H denotes the conjugate transpose.
In order to select the number of the retained modes for estimating the perturbation quantities (f(q) (r), g(q) (r)) in each operating
regime, we compare the solution root mean square errors, as well as the correlation coefficients of the reduced order model. We
determine the minimum number of the POD modes for each quantity (f(q) (r), g(q) (r)) ≡ u(q) (r) such that the conditions
RMSEu( ) ≤ 10−2 , Cu( ) ≥ 0.6

q q
(52)
hold.
The resulting number of POD modes and corresponding error estimates are provided in Table 3 for f(q) (r) and, respectively in
Table 4 for g(q) (r).
The first three orthonormal modes computed with the improved algorithm presented herein are depicted in Fig. 4 for modal
decomposition of the perturbation quantities f(q) (r) and g(q) (r), respectively.
JID: APM
Table 3
The retained number of POD modes and corresponding error estimates for
(q)
fPOD (r).
RMSE (f ) C (f )
q q
Operating regime Number of POD modes
110.7%BEP p=3 0.00288213 0.987

105%BEP p=2 0.00239951 0.986
102.5%BEP p=2 0.00207512 0.988
100%BEP p=2 0.00207134 0.973
97.4%BEP p=2 0.00437007 0.925
91.9%BEP p=3 0.00281274 0.933
71.4%BEP p=2 0.00349582 0.994
Table 4
The retained number of POD modes and corresponding error estimates for
g(POD
q)
(r).
RMSEg( ) Cg( )
q q
Operating regime Number of POD modes
110.7%BEP p=2 0.00003674 0.998

105%BEP p=3 0.00008106 0.988
102.5%BEP p=3 0.00007240 0.961
100%BEP p=3 0.00011817 0.601
97.4%BEP p=3 0.00008394 0.921
91.9%BEP p=3 0.00008193 0.777
71.4%BEP p=3 0.00004519 0.997
Fig. 4. The first three orthonormal modes obtained by weighted orthogonal decomposition of perturbation quantities (a) f(q) (r) and (b) g(q) (r).
(q) (q)
The coefficients of the reduced order models to estimate the perturbation quantities fPOD (r) and gPOD (r) in the investigated
operating regimes are listed in Table 5 and, respectively in Table 6.
The evolution of the coefficients corresponding to the two leading modes in the window of the discharge coefficient is illus-
trated in Fig. 5.
Analyzing the evolution of the two leading coefficients as the discharge coefficient varies from part load values to overload
values, one can observe that the second coefficient has a soft oscillating trajectory and does not influence the stability of the
second mode of the perturbations. On the contrary, the coefficient of the first mode varies sharply. In consequence, the first
mode can exhibit an unstable evolution as the turbine is operated at variable discharge, thus it can constitute the source of
the hydrodynamic instability of the swirling flow in the discharge cone. The methodology proposed in this paper is a viable
technique in conjunction with the linear stability analysis available in the community, to investigate the hydrodynamic stability
of swirling flows in turbomachines. This topic will be addressed carefully in our future work. Recent studies have been performed
by Oberleithner et al. [70] on the control of global modes in swirling jet experiments and linear stability analysis was applied to
a swirl-stabilized combustor flow [71]. They employed POD analysis of the PIV data for identification of dominant structures and
the related phase information.
JID: APM
Table 5
(q)
The POD coefficients for reduced order model fPOD (r).
Operating regime POD coefficients
110.7%BEP a1 = −0.119568 a2 = −0.022246 a3 = 0.024935

105%BEP a1 = −0.098378 a2 = −0.019184 −
102.5%BEP a1 = −0.085314 a2 = −0.033356 −
100%BEP a1 = −0.055463 a2 = −0.022710 −
97.4%BEP a1 = −0.064220 a2 = −0.036969 −
91.9%BEP a1 = 0.013616 a2 = −0.028783 a3 = −0.039180
71.4%BEP a1 = 0.218372 a2 = −0.049121 −
Table 6
The POD coefficients for reduced order model g(POD
q)
(r).
Operating regime POD coefficients
110.7%BEP b1 = −0.004917 b2 = 0.001389 −

105%BEP b1 = −0.003557 b2 = −0.000081 b3 = −0.000339
102.5%BEP b1 = −0.001419 b2 = −0.000582 b3 = −0.000808
100%BEP b1 = 0.514443 × 10−3 b2 = 0.049435 × 10−3 b3 = −0.445149 × 10−3
97.4%BEP b1 = −0.948931 × 10−3 b2 = −0.656887 × 10−3 b3 = −0.753368 × 10−3
91.9%BEP b1 = −0.285490 × 10−3 b2 = −0.649175 × 10−3 b3 = −0.188055 × 10−3
71.4%BEP b1 = 0.003981 b2 = 0.001233 b3 = −0.000817
Fig. 5. Evolution of the two leading coefficients obtained by weighted orthogonal decomposition of perturbation quantities (a) f(q) (r) and (b) g(q) (r).
(q) (q)
The coefficients of the reduced order models fPOD (r) and gPOD (r) have been estimated for a discharge coefficient by interpo-
lating the POD computed coefficients using radial basis functions (RBF). They are depicted in Fig. 6, respectively, where circles
represent the POD computed coefficients.
The validity of the methodology introduced in this paper is checked by comparing how the reduced order model assesses the
velocity field in parallel with experimentally measured velocity profiles. The reconstruction of the perturbation quantities f(q) (r)
and g(q) (r) employing the reduced order model having the POD modes depicted in Fig. 4 and the coefficients listed in Table 5 and,
respectively in Table 6, is illustrated in Fig. 7.
After orthogonal decomposition of the perturbation quantities f(q) (r) and g(q) (r), we obtained the velocity field of the pertur-
bations by postprocessing, i.e.
(q)
1 (q) 1 gPOD (r)
v(aqPOD
)
(r) = f (r), v(uqPOD
)
(r) = . (53)
r POD 2
r v(aq) (r)
The singularities have been treated by a local approximation in Taylor series and retaining the first three terms to give an
accurate approximation to the local behavior of the solution. Comparisons between the experimentally measured velocity per-
turbations and their profiles estimated by the reduced order model developed in this study (denoted as wPOD-RBF model) are
presented in Fig. 8.
The RMSE errors given by the wPOD-RBF model areillustrated
in Fig. 9a. For both axial and circumferential velocity profiles
estimation, the reduced order model error is of order O 10−2 . A comparison of the correlation coefficients of the reduced order
JID: APM
(q)
Fig. 6. The coefficients of the reduced order models (a) fPOD (r) and (b) g(POD
q)
(r) obtained by radial basis functions (RBF) interpolation. Circles represent the POD
computed coefficients.
Table 7
Projection errors (eigenvalues decay) of order p, E p = N
j=p+1 λ2j , employing different
algorithms: weighted POD on perturbation quantities f(r), g(r) (EVD-wPOD), classic POD
on perturbation quantities f(r), g(r) (EVD-POD), POD applied directly on velocity snap-
shots va (r), vu (r)(direct POD).
Projection order EVD-wPOD on f(r) EVD-POD on f(r) Direct POD on va (r)
p=2 0.000391 0.002742 0.007044

p=3 9.633177 × 10−5 0.000674 0.0009004
p=4 2.862232 × 10−5 0.0002005 0.000616
p=5 9.409749 × 10−6 6.591246 × 10−5 0.000297
p=6 4.357300 × 10−6 3.051659 × 10−5 10−8
Projection order EVD-wPOD on g(r) EVD-POD on g(r) Direct POD on vu (r)
p=2 3.334562 × 10−7 2.340860 × 10−6 0.012686

p=3 9.494396 × 10−8 6.651764 × 10−7 0.001488
p=4 3.522649 × 10−8 2.467201 × 10−7 0.000774
p=5 2.002421 × 10−8 1.402727 × 10−7 0.000386
p=6 3.422557 × 10−9 2.396784 × 10−8 10−8
model, estimating the velocity profiles is provided in Fig. 9b. The values of the correlation coefficients are greater than 94%, 87%,
for axial and circumferential velocity profiles, respectively, and confirm the validity of the reduced order model.
The numerical results support the conclusions drawn from the theoretical error estimates, that the reduced order model
developed in the present research can accurately predict the velocity field with a minimum number of POD modes.
4.2. Qualitative analysis of the reduced order model
The aim of this section is to illustrate the computational efficiency of the proposed weighted POD algorithm applied to the
flow quantities f (q) (r) ≡ rv(aq) (r) and g(q) (r) ≡ r2
v(aq) (r)
v(uq) (r) having vanishing integrals on the computational domain.
The weighted orthogonal decomposition developed in this paper is based on the eigenvalue decomposition of the correlation
matrix (approach denoted EVD-wPOD). In this section, we compare the proposed method with two different techniques. The
first one is the classic approach in which the irregularly distributed snapshots of f(q) (r) and g(q) (r) are not weighted (approach
denoted by EVD-POD). The second technique to be evaluated is obtained by applying POD on velocity fields directly, named in
the following as direct POD method.
N
The projection error is directly related to the truncated eigenvalues {λi }i=p+1 of POD decomposition. The projection errors (or
eigenvalues decay) given by the classic POD algorithm and the weighted POD algorithm developed in the present approach, re-
spectively given by Eqs. (30) and (35) are presented in comparison in Table 7, where p represents the dimension of the subspace
on which all the state vectors live. To highlight the benefit obtained by applying the orthogonal decomposition on the perturba-
tion quantities f(q) (r) and g(q) (r) and not on the velocity snapshots directly, we present in Table 7 the projection error obtained
by employing the algorithm developed in this paper (EVD-wPOD) in comparison with the projection error obtained by applying
POD on velocity fields (direct POD). The logarithmic plots of the eigenvalues computed with the aforementioned approaches are
illustrated in Fig. 10.
JID: APM
Fig. 7. Measured perturbations quantities f(q) (r) and g(q) (r) (circles) and the reconstructed perturbations using the truncated wPOD-RBF model (present research),
employing the POD modes depicted in Fig. 4 and the coefficients listed in Tables 5 and 6, respectively.
JID: APM
Fig. 8. Measured axial and circumferential perturbations of the velocity profiles (circles) and the reconstructed perturbations of the velocity field using the
truncated wPOD-RBF model (present research), employing the POD modes depicted in Fig. 4 and the coefficients listed in Tables 5 and 6, respectively.
JID: APM
Fig. 9. (a) RMSE of the measured axial and circumferential perturbations of the velocity profiles vs. the estimated perturbations of the velocity profiles using the
wPOD-RBF model (present research). (b) Correlation coefficients of the reduced order model (present research).
Fig. 10. Logarithmic representation of the eigenvalues, computed with weighted POD on perturbation quantities f(r), g(r) (EVD-wPOD), classic POD on pertur-
bation quantities f(r), g(r) (EVD-POD), POD applied directly on velocity snapshots va (r), vu (r) (direct POD).
It is evident from Table 7 that the weighted orthogonal decomposition is more efficient than the classic POD, in term of
diminishing the projection error. By applying a weighted POD based on eigenvalue decomposition approach (EVD-wPOD) on
irregularly distributed snapshots, as we proceed in this study, the mathematical model reduces to the first two or three leading
modes. To the problem investigated here, obtaining a projection error less than O(10−3 ) by applying POD on velocity snapshots
directly, requires a model that involves at least six modes. On the contrary, the EVD-wPOD algorithm developed in this survey,
applied on the perturbation quantities f(r) and g(r), leads to a reduced
order
model that involves a maximum number of three
leading modes, with a projection error of order O(10−5 ) for f(r) and O 10−8 for g(r), respectively.
The approach assessed in this paper is seen to be a very reliable alternative to the rival techniques, ensuring an optimal
truncation of the POD bases to the first few modes. In our approach, the dominant modes are extracted and a reduced order
model of the flow dynamics is developed based on two or three leading POD modes.
5. Summary and conclusions
In this paper we have proposed a framework for orthogonal decomposition of swirling flows applied to parameter depen-
dent problems. The investigated problem originates from turbomachines, where dynamics of swirling flow in the draft cone is
strongly influenced by the turbine discharge coefficient. We made several simplified assumptions on the swirl at runner outlet,
consistent with the flow conditions in the Francis turbine. We assume a steady, axisymmetric, parallel flow, of an inviscid and
incompressible fluid. Several key innovations have been introduced in this paper:
JID: APM
• In the state space, the distribution of snapshots is far from uniform, being clustered for overload and non-uniformly dis-
tributed at partial load. The irregular distribution of snapshots hinder the effectiveness of POD. To overcome this, we have
implemented a weighted proper orthogonal decomposition (wPOD) technique, where each column of the snapshot matrix is
defined as a snapshot vector multiplied by a weighting coefficient. After performing a rigorous error analysis we emphasized
the computational performances of the weighted POD method over the classic POD method for the investigated problem.
• The weighted POD methodology presented herein was employed to construct a POD representation of the difference between
the original experimental data and its crude approximation as a base flow. Considering the set of only seven snapshots,
representing the axial and circumferential velocity profiles of the swirling flow, experimentally measured at the inlet of the
draft tube, we first identify the representation of the base flow.
• The key innovation introduced in this paper resides in the manner of application of the weighted POD method. We im-
proved the classical methodology of applying the POD directly on the velocity snapshots and we identified the flow quanti-
v(aq) (r) and g(q) (r) ≡ r2
ties f (q) (r) ≡ r v(aq) (r)
v(uq) (r) having vanishing integrals on the computational domain. By applying the
weighted POD methodology presented in the paper on these perturbation quantities, the property of vanishing integral is
conserved for each individual mode. We showed that the advantage of this approach is that the POD representation of the
velocity field is achieved with a number of modes significantly lower comparing with different classic techniques.
• This property of our model has been analytically demonstrated and the efficiency of the proposed technique over the appli-
cation of POD directly on the velocity snapshots, so far performed in the literature, has been numerically proved.
• Combining the weighted proper orthogonal decomposition with the radial basis function (RBF) interpolation we have derived
a reduced order model for estimating the velocity profiles at the inlet of the discharge cone of the hydraulic turbine. Analyzing
the evolution of the leading modes coefficients is an instance of apprehending the hydrodynamic stability of the flow in the
discharge cone, thus this paper outlines steps towards hydrodynamic stability analysis of flow in modern hydraulic turbines
with potential applications.
We emphasized the excellent behavior of the reduced order model developed in this paper by comparing the computed
solution with the experimentally measured profiles and we found a close agreement. In addition, we performed a qualitative
analysis of the reduced order model by its correlation coefficient and root mean squared error (RMSE).
The numerical results support the conclusions drawn from the theoretical error estimates, that the reduced order model
developed in the present research can accurately predict the velocity field dependent on the discharge coefficient. The values of
the correlation coefficients greater than 94%, 87%, for axial and circumferential velocity profiles respectively, confirm the validity
of the reduced order model. Within the proposed methodology, the dominant modes are extracted and a reduced order model
of the flow dynamics is developed based on two or three leading POD modes. We defer to our future research the application
of the proposed methodology in conjunction with the linear stability analysis of the leading modes, for hydrodynamic stability
investigation of swirling flows in turbomachines.
Acknowledgments
The first author acknowledges the partially support of strategic grant POSDRU/159/1.5/S/137070 (2014) of the Ministry of
National Education, Romania, co-financed by the European Social Fund–Investing in People, within the Sectorial Operational
Program Human Resources Development 2007–2013. The second author was partially supported by a grant of the Romanian
Ministry of Education and Scientific Research CNCS-UEFISCDI PN-II-ID-PCE-2012-4-0634. The authors would like to thank the
anonymous reviewers who greatly contributed in improving the presentation of the paper.
Appendix

Proposition 1. Let U := u(q1 ) (r), . . . , u(qN ) (r) be a data collection with null average of each snapshot on the computational domain,
i.e.

u(q) (r)dS = 0, ∀q. (54)
S

N
( q)
Let u(q) (r) = a j φ j (r) be the proper orthogonal decomposition of data in U. Then the orthogonal modes preserve a null average
j=1
on the computational domain

φ j (r)dS = 0, j = 1, . . ., N. (55)
S

Proof. Let us denote by Fj = S φ j (r)dS, j = 1, . . ., N the averages of the proper orthogonal modes over the computational domain.
By associating the POD model with condition (54) yields

N
a(j ) φ j (r)dr = 0, ∀q.
q
(56)
S
j=1
JID: APM
In system formulation, relation (56) reads

⎧ (1) (1) (1)
⎪
⎪a1(2) F1 + a2(2) F2 + · · · + aN(2) FN = 0,
⎪
⎨a F1 + a F2 + · · · + a FN = 0,
1 2 N
. (57)
⎪
⎪..
⎪
⎩ (N)
a1 F1 + a(2 ) F2 + · · · + a(N ) FN = 0.
N N

(q)
Let us denote by A = a j , j, q = 1, . . ., N, the system matrix of coefficients and
T T
a1 = a(1 ) , a(1 ) , . . ., a(1 ) , a2 = a(2 ) , a(2 ) , . . ., a(2 )
1 2 N 1 2 N
, . . .,
T
aN = a(N ) , a(N ) , . . ., a(N )
1 2 N
∈ RN , (58)
the vectors containing the columns of A. Solution to the system (57)is nontrivial, since the coefficient matrix A generated by POD
is singular.
Let us denote by C (A) = span{a1 , . . ., aN } the column space of A which is equal to the span of the vectors ai , i = 1, . . ., N. Then
the system (57) yields the relation
a1 F1 + a2 F2 + · · · + aN FN = 0. (59)
To find the solution to the linear system (57) we test in the following the linear independence of vectors ai , i = 1, . . ., N.
Employing the POD algorithm described in Section 3, the coefficients of proper orthogonal decomposition are computed by
projecting the snapshot data onto the POD modes, i.e.
A = ( )T U, (60)

where = φ1 ... φN is the matrix of orthogonal modes and A is the matrix of coefficients computed by the POD
algorithm.
It follows from (60) that
(A)T = U T , (61)
which is equivalent with
A = U T . (62)
Let us denote by

C (U ) = span{u1 , . . ., uN }, C ( ) = span φ1 , . . ., φN ,
the column space of snapshot matrix U and POD modes matrix , respectively. Then
⎛ ⎞
u1 φ1 u1 φ2 ... u1 φN
⎜ u2 φ1 u2 φ2 ... u2 φN ⎟
A=⎜ ⎟, (63)
⎝ .. .. .. .. ⎠
. . . .
uN φ1 uN φ2 ... uN φN
which means that
T T
a1 = u1 φ1 , . . . uN φ1 , . . ., aN = u1 φN , . . .uN φN . (64)
We used the canonical inner product in RN . Relation (59)is then equivalent with
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
u1 φ1 u 1 φ2 u1 φN
⎜ u2 φ1 ⎟ ⎜ u 2 φ2 ⎟ ⎜ u2 φN ⎟
F1 ⎜ ⎟ + F2 ⎜ ⎟ + · · · + FN ⎜ ⎟ = 0, (65)
⎝ .. ⎠ ⎝ .. ⎠ ⎝ .. ⎠
. . .
uN φ1 uN φ2 uN φN
JID: APM
which in system formulation yields

⎧
⎪ F1 u1 φ1 + F2 u1 φ2 + · · · + FN u1 φN = 0,
⎪
⎨ F1 u2 φ1 + F2 u2 φ2 + · · · + FN u2 φN = 0,
.. (66)
⎪
⎪
⎩ .
F1 uN φ1 + F2 uN φ2 + · · · + FN uN φN = 0.
If we multiply each jth equation with vector φ j , j = 1, . . ., N and sum the resulting equations, it goes for the orthonormality
property of POD vectors that
F1 u1 + F2 u2 + · · · FN uN = 0. (67)
Thus the problem reduced to the linearly independence of the set of the snapshot vectors, which is achieved by their construction.
It follows that F1 = F2 = · · · = FN = 0.
References
[1] J.D. Buron, S. Houde, R. Lestriez, C. Deschenes, Application of the non-linear harmonic method to study the rotor-stator interaction in Francis-99 test case,
J. Phys.: Conf. Ser. 579 (2015) 1–12. Article number 012013.
[2] V. Guenette, S. Houde, G.D. Ciocan, G. Dumas, J. Huang, C. Deschenes, Numerical prediction of a bulb turbine performance hill chart through RANS simula-
tions, IOP Conf. Ser. Earth Environ. Sci. 15 (2012) 1–8. Article number 032007.
[3] C. Nicolet, A. Zobeiri, P. Maruzewski, F. Avellan, On the upper part load vortex rope in Francis turbine: experimental investigation, IOP Conf. Ser. Earth
Environ. Sci. 12 (2010) 1–10. Article number 012053.
[4] A. Ruprecht, Unsteady flow simulation in hydraulic machinery, Task Q. 6 (1) (2002) 187–208.
[5] R. Susan-Resiga, S. Muntean, F. Avellan, I. Anton, Mathematical modelling of swirling flow in hydraulic turbines for the full operating range, Appl. Math.
Modell. 35 (2011) 4759–4773.
[6] R. Susan-Resiga, S. Muntean, T. Ciocan, E. Joubarne, P. Leroy, L. Bornard, Influence of the velocity field at the inlet of a Francis turbine draft tube on perfor-
mance over an operating range, IOP Conf. Ser. Earth Environ. Sci. 15 (2012) 1–8. Article number 032008.
[7] S. Tridon, S. Barre, G. Ciocan, C. Segoufin, P. Leroy, Discharge imbalance mitigation in Francis turbine draft-tube bays, J. Fluids Eng. 134 (4) (2012) 1–8. Article
number 041102.
[8] D.A. Bistrian, Mathematical and numerical treatment of instabilities of non-axisymmetric confined vortices under the Dirichlet boundary conditions, Appl.
Math. Modell. 37 (6) (2013) 3993–4006.
[9] D.A. Bistrian, A solution of the parabolized Navier–Stokes stability model in discrete space by two-directional differential quadrature and application to
swirl intense flows, Comput. Math. Appl. 68 (2014) 197–211.
[10] D.A. Bistrian, Parabolized Navier–Stokes model for study the interaction between roughness structures and concentrated vortices, Phys. Fluids 25 (10)
(2013) 1–22. Article number 104103.
[11] S. Wang, Z. Rusak, The dynamics of a swirling flow in a pipe and transition to axisymmetric vortex breakdown, J. Fluid Mech. 340 (1997) 177–223.
[12] R. Susan-Resiga, G. Ciocan, I. Anton, F. Avellan, Analysis of the swirling flow downstream a Francis turbine runner, ASME J. Fluids Eng. 128 (2006) 177–189.
[13] V. Aeschlimann, S. Beaulieu, S. Houde, G.D. Ciocan, C. Deschanes, Inter-blade flow analysis of a propeller turbine runner using stereoscopic PIV, Eur. J. Mech.
– B/Fluids 42 (2013) 121–128.
[14] M.S. Iliescu, G.D. Ciocan, F. Avellan, Analysis of the cavitating draft tube vortex in a Francis turbine using particle image velocimetry measurements in
two-phase flow, J. Fluids Eng. 130 (2008) 1–10. Article number 021105.
[15] D.D. Kosambi, Statistics in function space, J. Indian Math. Soc. 7 (1943) 76–88.
[16] M. Loève, Fonctions Aleatoires du Second Ordre, Comptes Rendus de l’Académie des Sciences, Paris, 1945. 220.
[17] M. Loève, Probability Theory, Van Nostrand, 1955.
[18] K. Karhunen, Zur Spektraltheorie Stochastischer Prozesse, Annales Academiae Scientiarum Fennicae, 1946. 37.
[19] A.M. Obukhov, On the distribution of energy in the spectrum of turbulent flow, Izvestiia Akademii Nauk SSSR. Ser. Geogr. Geophys. 4–5 (1941) 435–466.
[20] A.M. Obukhov, Statistical description of continous fields, Trudy Geofizicheskogo Instituta, Akademiya Nauk SSSR 24 (1954) 3–42.
[21] J.L. Lumley, Atmospheric turbulence and radio wave propagation, J. Comput. Chem. 23 (13) (1967) 1236–1243.
[22] P. Holmes, J.L. Lumley, G. Berkooz, Turbulence, Coherent Structures, Dynamical Systems and Symmetry, Cambridge University Press, 1996.
[23] A. Caiazzo, T. Iliescu, V. John, S. Schyschlowa, A numerical investigation of velocity-pressure reduced order models for incompressible flows, J. Comput. Phys.
259 (2014) 598–616.
[24] R. Stefanescu, A. Sandu, I.M. Navon, Comparison of POD reduced order strategies for the nonlinear 2d shallow water equations, Int. J. Numer. Methods Fluids
76 (8) (2014) 497–521.
[25] Z. Luo, H. Li, P. Sun, J. Gao, A reduced-order finite difference extrapolation algorithm based on POD technique for the non-stationary Navier–Stokes equations,
Appl. Math. Modell. 37 (2013) 5464–5473.
[26] Z. Wang, I. Akhtar, J. Borggaard, T. Iliescu, Proper orthogonal decomposition closure models for turbulent flows: a numerical comparison, Comput. Methods
Appl. Mech. Eng. 237-240 (2012) 10–26.
[27] S. Walton, O. Hassan, K. Morgan, Reduced order modelling for unsteady fluid flow using proper orthogonal decomposition and radial basis functions, Appl.
Math. Modell. 37 (2013) 8930–8945.
[28] J. Osth, B.R. Noack, S. Krajnovic, D. Barros, J. Boree, On the need for a nonlinear subscale turbulence term in POD models as exemplified for a high-Reynolds-
number flow over an Ahmed body, J. Fluid Mech. 747 (2014) 518–544.
[29] J. Du, F. Fang, C.C. Pain, I.M. Navon, J. Zhu, D. Ham, POD reduced order unstructured mesh modelling applied to 2d and 3d fluid flow, Comput. Math. Appl.
65 (2013) 362–379.
[30] F. Fang, C.C. Pain, I.M. Navon, M.D. Piggott, G.J. Gorman, P. Allison, A.J.H. Goddard, Reduced order modelling of an adaptive mesh ocean model, Int. J. Numer.
Methods Fluids 59 (8) (2009) 827–851.
[31] A.K. Alekseev, I.M. Navon, Numerical control of two-dimensional shock waves in dual solution domain by instant temperature disturbances, Int. J. Numer.
Methods Fluids 71 (2013) 175–184.
[32] D.M. Markovich, S.S. Abdurakipov, L.M. Chikishev, V.M. Dulin, K. Hanjalic, Comparative analysis of low- and high-swirl confined flames and jets by proper
orthogonal and dynamic mode decomposition, Phys. Fluids 26 (2014). Article number 065109.
[33] K. Oberleithner, M. Sieber, C.N. Nayeri, C.O. Paschereit, C. Petz, H.C. Hege, B.R. Noack, I. Wygnanski, Three-dimensional coherent structures in a swirling jet
undergoing vortex breakdown: stability analysis and empirical mode construction, J. Fluid Mech. 679 (2011) 383–414.
[34] L. Sirovich, Turbulence and the dynamics of coherent structures part I: Coherent structures, Q. Appl. Math. XLV (3) (1987) 561–571.
[35] E.A. Christensen, M. Brons, J.N. Sorensen, Evaluation of proper orthogonal decomposition-based decomposition techniques applied to parameter-dependent
nonturbulent flows, SIAM J. Sci. Comput. 21 (4) (2006) 1419–1434.
JID: APM
[36] K. Carlberg, C. Farhat, A low-cost, goal-oriented compact proper orthogonal decomposition basis for model reduction of static systems, Int. J. Numer. Methods
Eng. 86 (3) (2011) 381–402.
[37] G. Ingram, Basic Concepts in Turbomachinery, Grant Ingram and Ventus Publishing ApS, 2009.
[38] J.L. Lumley, Stochastic Tools in Turbulence, Academic Press, New York, 1970.
[39] F. Fang, D. Pavlidis, C.C. Pain, A.G. Buchan, I.M. Navon, Reduced order modelling of an unstructured mesh air pollution model and application in 2d/3d urban
street canyons, Atmos. Environ. 96 (2014) 96–106.
[40] R. Stefanescu, I.M. Navon, POD/DEIM nonlinear model order reduction of an ADI implicit shallow water equations model, J. Comput. Phys. 237 (2013) 95–114.
[41] X. Chen, S. Akella, I.M. Navon, A dual weighted trust-region adaptive POD 4d-var applied to a finite-volume shallow-water equations model on the sphere,
Int. J. Numer. Methods Fluids 68 (3) (2012) 377–402.
[42] T. Iliescu, Z. Wang, Variational multiscale proper orthogonal decomposition: Navier–Stokes equations, Numer. Methods Partial Differ. Equ. 30 (2) (2014)
641–663.
[43] C. Winton, J. Pettway, C.T. Kelley, S. Howington, O.J. Eslinger, Application of proper orthogonal decomposition (POD) to inverse problems in saturated
groundwater flow, Adv. Water Resour. 34 (2011) 1519–1526.
[44] Y.E.A. Eldahab, N.H. Saad, A. Zekry, Enhancing the maximum power point tracking techniques for photovoltaic systems, Renew. Sustain. Energy Rev. 40
(2014) 505–514.
[45] Q. Du, M.D. Gunzburger, Centroidal Voronoi tessellation based proper orthogonal decomposition analysis, Int. Ser. Numer. Math. 143 (2002) 137–150.
[46] J. Burkardt, M. Gunzburger, H.C. Lee, Centroidal Voronoi tessellation-based reduced-order modeling of complex systems, SIAM J. Sci. Comput. 28 (2) (2006)
459–484.
[47] Q. Du, M. Gunzburger, L. Ju, Advances in studies and applications of centroidal Voronoi tessellations, Numer. Math. Theory Methods Appl. 3 (2) (2010)
119–142.
[48] D.N. Daescu, I.M. Navon, A dual-weighted approach to order reduction in 4D-Var data assimilation, Mon. Weather Rev. 136 (3) (2008) 1026–1041.
[49] X. Chen, I.M. Navon, F. Fang, A dual weighted trust-region adaptive POD 4D-Var applied to a finite-element shallow water equations model, Int. J. Numer.
Methods Fluids 65 (5) (2011) 520–541.
[50] A. Chatterjee, An introduction to the proper orthogonal decomposition, Curr. Sci. 78 (7) (2000) 808–817.
[51] K. Kunisch, S. Volkwein, Galerkin proper orthogonal decomposition methods for parabolic problems, Numerische Mathematik 90 (2001) 117–148.
[52] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
[53] N. Aubry, On the hidden beauty of the proper orthogonal decomposition, Theor. Comput. Fluid Dyn. 2 (1991) 339–352.
[54] N. Aubry, R. Guyonnet, R. Lima, Spatio-temporal analysis of complex signals: theory and applications, J. Stat. Phys. 64 (3/4) (1991) 683–739.
[55] R.L. Hardy, Multiquadric equations of topography and other irregular surfaces, J. Geophys. Res. 76 (8) (1970) 1905–1915.
[56] J.C. Carr, R.K. Beatson, J.B. Cherrie, T.J. Mitchell, W.R. Fright, B.C. McCallum, T.R. Evans, Reconstruction and representation of 3D objects with radial basis
functions, ACM SIGGRAPH 2001 (2001) 67–76.
[57] R.K. Beatson, W.A. Light, S. Billings, Fast solution of the radial basis function interpolation equations: domain decomposition methods, SIAM J. Sci. Comput.
22 (5) (2000) 1717–1740.
[58] S. Walton, O. Hassan, K. Morgan, Reduced order modelling for unsteady fluid flow using proper orthogonal decomposition and radial basis functions, Appl.
Math. Modell. 37 (2013) 8930–8945.
[59] D. Xiao, F. Fang, C. Pain, G. Hu, Non-intrusive reduced order modelling of the Navier–Stokes equations based on RBF interpolation, Int. J. Numer. Methods
Fluids 79 (11) (2015) 580–595.
[60] C.A. Micchelli, Interpolation of scattered data: Distance matrices and conditionally positive definite functions, Constr. Approx. 2 (1) (1986) 11–22.
[61] B. Fornberg, T.A. Driscoll, G. Wright, R. Charles, Observations on the behaviour of radial basis function approximations near boundaries, Comput. Math. Appl.
43 (3–5) (2002) 473–490.
[62] T.A. Driscoll, B. Fornberg, Interpolation in the limit of increasingly flat radial basis functions, Comput. Math. Appl. 43 (3-5) (2002) 413–422.
[63] B. Fornberg, G. Wright, Stable computation of multiquadric interpolants for all values of the shape parameter, Comput. Math. Appl. 48 (2004) 853–867.
[64] G.E. Fasshauer, Meshfree Approximation Methods with MATLAB, World Scientific Publishers, 2007.
[65] M.E. Chenoweth, A numerical study of generalized multiquadric radial basis function interpolation, SIAM Undergrad. Res. Online 2 (2009) 58–70.
[66] R.L. Hardy, Theory and applications of the multiquadric-biharmonic method: 20 years of discovery, Comput. Math. Appl. 19 (1990) 163–208.
[67] H. Aikawa, On weighted Beppo Levi functions, Potential Theory, Springer, US, 1988, pp. 1–6.
[68] J. Duchon, Splines minimizing rotation-invariant semi-norms in Sobolev spaces, Constructive Theory of Functions of Several Variables, Lecture Notes in
Mathematics, Springer-Verlag, 1977, pp. 85–100.
[69] P.J. Green, B.W. Silverman, Nonparametric Regression and Generalized Linear Models: A roughness Penalty Approach, 1st, Chapman and Hall/CRC, 1993.
[70] K. Oberleithner, M. Sieber, C.N. Nayeri, C.O. Paschereit, On the control of global modes in swirling jet experiments, J. Phys., Conf. Ser. 318 (2011) 1–19.
Proceedings of 13th European Turbulence Conference (ETC13).
[71] K. Oberleithner, S. Terhaar, L. Rukes, C.O. Paschereit, Why nonuniform density suppresses the precessing vortex core, J. Eng. Gas Turbine Power 135 (2013)
1–9. Article number 121506.

Bistrian 2015 AMM

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bistrian 2015 AMM

Uploaded by

Copyright:

Available Formats

JID: APM

ARTICLE IN PRESS [m3Gsc;November 21, 2015;8:44]

Applied Mathematical Modelling 000 (2015) 1–22

Contents lists available at ScienceDirect

Applied Mathematical Modelling

Weighted proper orthogonal decomposition of the swirling ﬂow

2. Identification of flow quantities for an efficient orthogonal decomposition

v(aq) (r), v(uq) (r). (3)

Moment of momentum flux coefficient

Operating Discharge coeﬃcient Moment of momentum

110.7%BEP q1 = 0.440220 m1 = 0.016394

v(aq) (r) = v(aq) (r) − v(aq) and

v(aq) = constant, (5)

The integral condition (7) immediately leads to

110.7%BEP 0.363818 0.307257 0.430320

Introducing the relation (4) in (13) leads to

Introducing the notation

2 mrw 5 − 10I2(q) I3(q) 5 mrw 4 − 8I1(q) I3(q)

3. Reduced order model formulation

be a Hilbert space endowed with the inner product

where δ ij is the Kronecker delta symbol

For the step size h = rw /(n − 1) let us introduce a spatial grid in S by

Given the discharge coeﬃcients q1 , . . ., qN , we consider the following ensemble of snapshots

as n realizations of each u(q) (r) in locations r1 , r2 , . . ., rn on the computational domain S. p

The projection error of order p in the Frobenius norm is given by

A weighted snapshot matrix can be deﬁned as

where 0 < w ≤ 1 for each w .

Algorithm: Weighted POD algorithm based on eigenvalue decomposition (EVD-wPOD)

(1:) Precompute the weighted snapshot matrix

where (U w )T represents the transpose of the weighted snapshot matrix.

3.2. Identiﬁcation of the reduced order modelby RBF interpolation

the problem reduces to ﬁnd an interpolant s : R2 → R such that

s(x) = arg min s(x) 2 .

and the constraints

4. Analysis of the reduced order model

4.1. Validation of the reduced order model and numerical results

RMSEu( ) ≤ 10−2 , Cu( ) ≥ 0.6

110.7%BEP p=3 0.00288213 0.987

110.7%BEP p=2 0.00003674 0.998

110.7%BEP a1 = −0.119568 a2 = −0.022246 a3 = 0.024935

110.7%BEP b1 = −0.004917 b2 = 0.001389 −

Projection order EVD-wPOD on f(r) EVD-POD on f(r) Direct POD on va (r)

p=2 0.000391 0.002742 0.007044

p=2 3.334562 × 10−7 2.340860 × 10−6 0.012686

4.2. Qualitative analysis of the reduced order model

5. Summary and conclusions

In system formulation, relation (56) reads

which is equivalent with

which in system formulation yields

You might also like

as n realizations of each u(q) (r) in locations r1 , r2 , . . ., rn on the computational domain S. p