IOP Conference Series: Earth and Environmental Science

PAPER • OPEN ACCESS

Correlation between numerical simulations and measurements to assess

uncertainties: a case study on a hydroelectric runner
To cite this article: M Gagnon et al 2019 IOP Conf. Ser.: Earth Environ. Sci. 240 072005

View the article online for updates and enhancements.

This content was downloaded from IP address 179.61.188.36 on 27/03/2019 at 17:30

29th IAHR Symposium on Hydraulic Machinery and Systems IOP Publishing
IOP Conf. Series: Earth and Environmental Science 240 (2019) 072005 doi:10.1088/1755-1315/240/7/072005

Correlation between numerical simulations and

measurements to assess uncertainties: a case study on a
hydroelectric runner

M Gagnon1, A Immarigeon2, J Nicolle1 and J-F Morissette1

1
Institut de recherche d’Hydro-Québec (IREQ), Varennes, QC, J3X 1S1, Canada
2
Hydro-Québec, Montréal, QC, H2Z 1A4, Canada

[email protected]

Abstract. For fatigue reliability analysis, we need an estimate of the structure response and
uncertainty at critical locations where often direct observation from a measurement campaign
cannot be obtained. At these critical locations, two types of information are required: static
amplitudes and dynamic ranges. In this work, we assume a linear correlation between
simulated and measured values to estimate the information needed at critical locations on the
runner blades. The proposed methodology is illustrated using the data from a six-blade
propeller runner. The methodology enables one to easily transpose information from observed
values to a critical location and assess uncertainty for fatigue reliability analysis.

1. Introduction
One of the goals of numerical simulation is to accurately predict the response of a system prior to
empirical evidence from direct observations or representative experiments. To use these results,
decision makers need to be confident that the results obtained have enough precision to support their
actions. This confidence can often only be gained a posteriori through observations; hence the need
for validation using observed data [1]. However, even if the numerical model is valid within the
precision needed at the measured locations, for fatigue analysis [2], we still need to estimate the
response and uncertainty levels at critical locations where we do not have experimental data.
This paper presents how to perform such estimation for static and dynamic strains in a six-blade
hydraulic turbine propeller by combining Computational Fluid Dynamic (CFD), Finite Element
Analysis (FEA) and experimental results. The fatigue analysis requires that the static and dynamic
responses be estimated for every steady operating condition of interest. Figure 1 illustrates how the
response is decomposed in static amplitude and dynamic range. We define the static amplitude as the
mean value of the response and the dynamic range corresponds to the maximum peak-to-peak interval
over the considered time interval in steady operating condition.

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
29th IAHR Symposium on Hydraulic Machinery and Systems IOP Publishing
IOP Conf. Series: Earth and Environmental Science 240 (2019) 072005 doi:10.1088/1755-1315/240/7/072005

Figure 1. Static amplitude and dynamic range.

These values are estimated using numerical models during the design phase of the turbine and need
to be validated during commissioning through field measurements. To validate the stress amplitudes
used in fatigue analysis, a first difficulty resides in the choice of measurement locations. It is often
impossible to position a strain gauge exactly where the maximal values should be observed. Hence, a
transposition method is needed to estimate expected values and uncertainty levels at these desired
critical locations. In this study, to transpose the information at the critical location, we assume that
simulation results and measurements are linearly correlated which enables one to estimate both the
expected value and its uncertainty.
The study is structured as follows: first, the proposed correlation methodology is presented and
then our case study is detailed by reviewing numerical simulation results followed by observed data
from the measurement campaign. Finally, the transposition results and their implications are
discussed. The unit studied is a six-blade propeller turbine that has previously been referred to by the
authors in others studies [3-5]

2. Correlation methodology
Numerical simulations are an approximation of the reality of interest. In our case, this approximation
needs to answer the following engineering question: are the maximum values at critical locations (i.e.
hotspots) acceptable to warrant reliable operation during the expected life of the runner? This should
hold true if numerical simulation and measurements are well correlated. In our case, a linear
correlation between numerical results and measurements is assumed. This statement implies the
following model function:
𝑦 = 𝛽0 + 𝛽1 𝑥 (1)
where 𝑥 is the simulation result and 𝑦 is the measured value which are the two variables of interest.
Ideally, one would expect an intercept 𝛽0 = 0 and a slope 𝛽1 = 1. However, because we expect that
the model might have bias either constant 𝛽0 ≠ 0 or non-constant 𝛽1 ≠ 1, the regression problem
becomes:
𝑦𝑖 = 𝑓𝑖 (𝑥𝑖 + 𝛿𝑖 ; 𝛽) − 𝜖𝑖 (2)
where the 𝑦𝑖 are a function of the 𝑥𝑖 and the unknown parameters 𝛽 subjected to error δ𝑖 and 𝜖𝑖 . This
means that we expect error 𝜖𝑖 in measured values and δ𝑖 in the simulation results. Such problem can
be solved using Orthogonal Distance Regression (ODR) using the following minimization:
𝑚𝑖𝑛𝛽,𝛿,𝜖 ∑𝑛𝑖=1( 𝜖𝑖 2 + 𝛿𝑖 2 ) (3)
where 𝑛 is the number of 𝑥𝑖 and y𝑖 available. Furthermore, to account for unequal uncertainty across
the 𝑥𝑖 and y𝑖 , weights in the form of w𝜖𝑖 and w𝛿𝑖 are introduced as follows:

2
29th IAHR Symposium on Hydraulic Machinery and Systems IOP Publishing
IOP Conf. Series: Earth and Environmental Science 240 (2019) 072005 doi:10.1088/1755-1315/240/7/072005

𝑚𝑖𝑛𝛽,𝛿,𝜖 ∑𝑛𝑖=1( 𝑤𝜖𝑖 𝜖𝑖 2 + 𝑤𝛿𝑖 𝛿𝑖 2 ) (4)

This enables us to account simultaneously for error and uncertainty in both the numerical results
and measured values. In this study, we used the software package ODRPACK [6] to estimate the
model parameters 𝛽𝑖 using weighted orthogonal distance regression. Afterwards, the confidence
interval was constructed using the 𝛽𝑖 , the covariance matrix 𝐶 and the appropriate Student’s
distribution values [7]. The confidence interval was obtained as follows:
𝑗=𝑛 𝜕𝑓 𝜕𝑓
𝑦̂ ± 𝑡𝛼⁄2,𝜈 √𝜒𝜐2 ∑𝑗=0 ∑𝑘=𝑛
𝑘=0 𝜕𝛽 𝐶𝑗𝑘 (5)
𝑗 𝜕𝛽𝑘

Where 𝜒𝜐2 is the reduced Chi square and 𝑡𝛼⁄2,𝜈 the Student's t-distribution with 𝜐 degree of freedom
at 𝛼 significance level. For any given x value, this interval is interpreted as the probability that the true
value of the function should be within its limits (1 − 𝛼)% of the time.
An example using randomly generated values is presented in Figure 2. In this example, randomly
generated error and uncertainty for values between 0 and 5 have been used. Using the confidence
interval around the estimated linear model shown in blue area, we can obtain, for any interval
representing the uncertainty around a value 𝑥, the uncertainty interval for the corresponding 𝑦 value.
Notice, the black line representing the relation 𝑦 = 𝑥, which was included to help visualize the biases
in the estimated linear model. In this example, we observe that for 𝑥 = 6 ± 0.25, we obtain the
interval [5.18, 7.32] in red given a 95% confidence interval. This approach can be used to estimate the
expected uncertainty interval for any values as long as we accept the assumptions made initially by
using a linear model to correlate both datasets while neglecting the spatial relation between values.
Due to the limited number of sensors, spatial information is difficult to take into account using spatial
correlation. Accounting for spatial correlation would give a better assessment of uncertainty by using
distances between measured locations and the location of interest.

Figure 2. Weighted orthogonal distance regression

example.

3. Numerical simulation results

Four different operating conditions for this propeller turbine have been simulated with the use of
commercial CFD solver Ansys CFX. They consist in four stable regimes, namely: speed-no-load
(SNL), part load (60% opening), best efficiency (85% opening), and full load (100% opening).
Two configuration setups were used for the different simulations (Figure 3). A complete machine
setup with the entire spiral casing, distributor, runner and draft tube was used for three operating
points at stable regime with the advanced SAS (Scale-Adaptive Simulation) turbulence model to get as
much dynamic content as possible. The speed-no-load condition was modelled without the spiral
casing but with the entire distributor, air injection, and also used the SAS turbulence model.

3
29th IAHR Symposium on Hydraulic Machinery and Systems IOP Publishing
IOP Conf. Series: Earth and Environmental Science 240 (2019) 072005 doi:10.1088/1755-1315/240/7/072005

Those simulations enabled us to get the pressure loads. FEA structural analyses were then
conducted in Ansys Mechanical with the full-runner geometry meshed with 1,34M nodes, refined at
the strain gauges locations shown in Figure 4. Quasi-static analyses were used with two pressure loads
per operating point to determine the dynamic range. Those loads correspond to the instants with
maximum and minimal recorded torque on a specific blade. Dynamic amplification was not simulated
because it was not detected during measurements. Static values were simply the arithmetic mean of the
maximum and minimum values for each location and operating point.

(a) (b)

Figure 3. CFD setup for pressure loading. (a) Full machine setup for steady operating
condition, (b) 360 model used for speed-no-load.

Figure 4. FEA model for structural analysis, the trailing

edge measurements location E, F and G are visible.

In order to account for uncertainties, an arbitrary 10% of the maximum simulated value for a given
operating condition was used and applied at every location on the blade. The same methodology was
used for both the static and dynamic estimates.

4. Field measurement data

The runner was instrumented with strain gauges at seven locations replicated on two different blades.
Figure 5a presents the location of the strain gauges and an example of the measured data at location A
for a complete start-to-stop cycle is presented in Figure 5b. The locations A, B, C and D were

4
29th IAHR Symposium on Hydraulic Machinery and Systems IOP Publishing
IOP Conf. Series: Earth and Environmental Science 240 (2019) 072005 doi:10.1088/1755-1315/240/7/072005

instrumented with strain gauges rosettes while E, F and G were monitored with unidirectional strain
gauges.

(a) (b)

Figure 5. Strain gauge measurement. (a) Location of strain gauges. (b) Measured data at location A.

The static amplitude and dynamic range for maximum wicket gate opening (full load), 85% wicket
gate opening (best efficiency), 60% wicket gate opening (part-load) and speed-no-load (SNL)
conditions were obtained from the measured data. The uncertainty intervals used to weight the static
amplitudes include:
 Differences between the zeros before and after the considered operating condition
 Differences between similar measured operating conditions
 Unaccounted uncertainty sources with 10% of the maximum static stress
For dynamic ranges, the uncertainty intervals used to weight the data include:
 Differences between similar measured operating conditions
 Unaccounted uncertainty sources with 10% of the maximum dynamic range
The zeros were verified before and after each measurement record while the unit was at rest. Notice
that the unaccounted uncertainty sources include more than only variations in parameters like
geometry, material properties, loading, etc. They also include epistemic uncertainty due to lack of
knowledge and modeling assumptions.

5. Correlation and expected maximum values

The correlation results and uncertainty intervals expected at the location of the maximum deformation
on the model (which is in the vicinity of measurement locations A and B in Figure 5a) are presented in
Figure 6 and 7. In each figure, 95% confidence bands are presented in blue and the interval at the
maximum location is shown in red. The static amplitude results showed in Figure 6 exhibit good
correlations with small confidence bands for all operating conditions except SNL. However, it can be
noticed that for the SNL condition all the values are grouped together and have relatively small values
which might explain the wide confidence band observed when projected to higher value. This also
highlights the difficulty in choosing proper measurement locations on a runner in order to obtain
useful data for a wide range of operating conditions.

5
29th IAHR Symposium on Hydraulic Machinery and Systems IOP Publishing
IOP Conf. Series: Earth and Environmental Science 240 (2019) 072005 doi:10.1088/1755-1315/240/7/072005

(a) Maximum opening (b) 85% opening

(c) 60% opening (d) SNL

Figure 6. Correlation results for static amplitude.

For the dynamic range, Figure 7 shows that none of the conditions studied exhibit good linear
correlation. For every studied condition, we observe large constant biases 𝛽0 and non-constant biases
𝛽1 . Furthermore, there is a large dispersion that translates into a large 95% confidence band. Hence,
larger uncertainty intervals are estimated at the maximum stress location. These large biases and
dispersion indicate that our numerical model is not well representative of the measured values which
might warrant for larger unaccounted uncertainty sources. Nonetheless, as such, this generates large
uncertainty intervals for maximum dynamic stresses that could be decreased by improving the
numerical model.

6
29th IAHR Symposium on Hydraulic Machinery and Systems IOP Publishing
IOP Conf. Series: Earth and Environmental Science 240 (2019) 072005 doi:10.1088/1755-1315/240/7/072005

(a) Maximum opening (b) 85% opening

(c) 60% opening (d) SNL

Figure 7. Correlation results for dynamic range.

6. Discussions
The principal advantage of the proposed methodology is that it can bring added value to numerical
analysis and measurements by accounting for errors and sources of uncertainty in the various inputs to
quantify overall uncertainties. The outcomes of the regression are estimates of the intercept 𝛽0 , slope
𝛽1 and confidence bands to account for the dispersion. Using these, we can obtain estimates for the
expected value and uncertainty at any location. In this case, the value obtained shows good overall
behavior as it properly rises with uncertainty and dispersion. However, since we neglect spatial
correlation between the locations of interest, uncertainty might not be properly estimated. Also, we
have observed that additional care is required if we want to use the correlation results for validation
because rather than having just one metric to consider, we now have three (constant bias 𝛽0 , non-
constant bias 𝛽1 , and dispersion). Furthermore, linear correlation results for different runners cannot
be directly compared because the number of measured locations and their locations are often different.
Nonetheless, the proposed methodology proves to be a good qualitative tool to assess the validity of
numerical model.

7. Conclusions
The objective of this paper was to propose a methodology to correlate the measurement data and
simulation results in order to estimate the critical values on runner blades. With the proposed
methodology, we have shown that it is possible to account for errors in measured data and in
simulation results while simultaneously weighting the data to account for uncertainty using weighted
orthogonal distance regression. The results of our study case show better correlations and lower

7
29th IAHR Symposium on Hydraulic Machinery and Systems IOP Publishing
IOP Conf. Series: Earth and Environmental Science 240 (2019) 072005 doi:10.1088/1755-1315/240/7/072005

uncertainty for static amplitude than for dynamic fluctuations. These results were to be expected since
dynamic ranges are more difficult to obtain using numerical simulation. The methodology proves to be
a good qualitative tool to assess the validity of these simulation results and helps highlight the model
deficiencies. Notice that having three metrics to take into account at the same time (constant bias 𝛽0 ,
non-constant bias 𝛽1 and dispersion) makes the correlation results not well suited for quantitative
model comparison [8]. Nonetheless, the proposed methodology, with its simplicity, enables the analyst
to easily transpose the observed value to the critical location and assess uncertainty for fatigue
reliability analysis. Other studies are needed to obtain a methodology that could at the same time be
used for quantitative model comparison, account for spatial correlation and estimate expected values
with uncertainties at critical locations.

References
[1] ASME 2006 Guide for Verification and Validation in Computational Solid Mechanics
vol PTC 60/V&V 10
[2] Gagnon M, Tahan A, Bocher P and Thibault D 2013 A probabilistic model for the onset of High
Cycle Fatigue (HCF) crack propagation: Application to hydroelectric turbine runner
International Journal of Fatigue 47 300–307
[3] Gagnon M, Jobidon N, Lawrence M and Larouche D 2014 Optimization of turbine startup:
Some experimental results from a propeller runner. 27th IAHR symposium on hydraulic
machinery and systems, 22-26 of September 2014, Montreal, Canada
[4] Diagne I, Gagnon M and Tahan A 2016 Modeling the dynamic behavior of turbine runner
blades during transients using indirect measurements 28th IAHR symposium on hydraulic
machinery and systems, 4-8 of July 2016, Grenoble, France
[5] J Nicolle and J-F Morissette 2016 Simulation of air admission in a propeller hydroturbine
during transient events 28th IAHR symposium on hydraulic machinery and systems, 4-8 of
July 2016, Grenoble, France
[6] Boggs P T, Byrd R H, Rogers J E and Schnabel R B 1992 User’s Reference Guide for
ODRPACK Version 2.01 Software for Weighted Orthogonal Distance Regression National
Institute of Standards and Technology Gaithersburg, MD 20899
[7] Wolberg J 2006 Data Analysis Using the Method of Least Squares: Extracting the Most
Information from Experiments Springer-Verlag Berlin Heidelberg
[8] Chatenet Q, Tahan A, Gagnon M and Chamberland-Lauzon J 2016 Numerical model validation
using experimental data: Application of the area metric on a Francis runner 28th IAHR
symposium on hydraulic machinery and systems, 4-8 of July 2016, Grenoble, France