You are on page 1of 16

Assessment and sensitivity analysis of multiphase flow

pipeline models against experimental and field data


P Dhoorjaty, D Erickson, R Kowta, M Mai, P Parthasarathy
Wood Group, USA

ABSTRACT

Transient multiphase pipeline simulators are seeing increasing application for live
computation of oil and gas production pipeline and process flows. This study considers
the performance of Wood Group’s Virtuoso tool and three other commercially available
simulators used widely in the industry, on an extended set of field data from a large gas-
condensate pipeline with three-phase flow. A steep escarpment strongly influences the
dynamics of pipeline pressure, and liquid content and outlet rates; production rates span
the region of transition between separated and slug flow regimes. ‘Out-of-the-box’
model predictions of pressure drop can differ from measured values by 20% or more.
While model predictions of liquid holdup agree reasonably well  both mutually and
against experimental data  for small-diameter pipes at gentle inclination, predictions
vary by 50% or more under field conditions. The extent of scatter in predicted holdup is
a large fraction of available slug catcher capacity. Analysis indicates that pressure and
liquid volume predictions are sensitive to flow regime determination, specifically via the
interfacial friction factor and the holdup at which slugging initiates; pressure prediction
error can be reduced by 50% through moderate adjustment of associated modeling
parameters. These aggregate assessments of model performance suggest that field-
specific model tuning remains necessary to achieve the level of prediction accuracy
demanded by online systems used for liquids management, forecasting, and pipeline
integrity monitoring.

SYMBOLS AND ABBREVIATIONS

d days S0 Virtuoso simulator


D diameter S1 commercial simulator 1
f friction factor S2 commercial simulator 2
Fr Froude number S3 commercial simulator 3
g gravitational acceleration t time
j superficial velocity T temperature
MEG mono-ethylene glycol vs slip velocity
P pressure  holdup fraction
PVT thermodynamic data  pipe inclination
Subscripts
expt experimental l liquid
g gas mix mixture
i interfacial pred model predicted

© BHR Group 18 MPT 2017 209


1 INTRODUCTION

Oil and gas production processes commonly involve the simultaneous transport of
multiple phases: oil, gas, aqueous, and other phases such as sand and hydrates. Efficient
design and safe operation of pipelines and processing facilities requires good knowledge
of the thermodynamics and fluid dynamics of multiphase flow. Fluid dynamical
phenomena of multiphase flow are complex, involving a variety of flow patterns, but
ultimately govern parameters of practical interest, such as the pressure drop, in situ
liquid content, and liquid outlet rates and cumulative volumes under transient or even
nominally steady-state operating conditions. The problem has, as such, received much
attention and over the last three decades sufficient progress has been made that transient
multiphase flow calculations are routinely used during pipeline and facilities design.
Confidence in the reliability of multiphase pipeline simulation has indeed grown to the
point that there is an increasing demand in the industry for advisory systems that perform
live simulation of the field (‘online systems’). Common uses of the technology include
virtual well metering; pipeline monitoring for hard-to-measure quantities such as liquid
holdup and fluid temperature profile; forecasting of cool-down and no-touch times upon
shutdown, or of process response to planned operational events; pig campaign planning;
hydrate blockage detection; and leak detection and location. While reasonably
conservative model prediction may be acceptable, or even desirable, at the design stage,
viable application to online systems places more stringent demands on model robustness
and accuracy. For example, a 10% over-prediction of the pressure drop can easily cause
an online system to miss a significant hydrate blockage event; conversely a 10% under-
prediction can prevent the detection of a major leak.

This study seeks an assessment of current multiphase simulation capabilities under flow
conditions typical of the field. Multiple commercial simulators are now available.
Three such tools are considered in this study, referred to herein as S1, S2, and S3. S1
and S3 are the same product, with S3 being a later version that has presumably benefited
from further model refinement. S2 is an independent tool from a different vendor. All
three are currently in common use in the flow assurance industry. Additionally, this
study considers Wood Group’s own ‘Virtuoso’ model, which is referred to herein as S0.
Virtuoso has seen extensive application over the last two decades to online systems
covering a wide range of pipeline size, fluid type (gas-condensate and black oil), and
operating pressure and temperature (1,2,3). Virtuoso’s installation base makes available
a valuable library of field data.

Field data – while ‘uncontrolled’ in comparison to experiment and sometimes requiring


careful validation and post-processing – access conditions hard to achieve in the
laboratory. For example, pipeline diameters are often larger and liquid loadings are
often much smaller than in experiment. Certain aspects of modeling – for instance the
flux of liquids transported as droplets in the gas phase – are sensitive to pipe diameter
scale up. Whereas experiments usually yield steady-state or time-averaged results, field
data allow study of model response to drastic transients (examples in this study show
how prediction during rate transients affects subsequent prediction errors over extended
periods of time). Long field pipeline lengths, and associated pressure and temperature
changes, elevate the importance of fluid density variation and mass transfer, and
challenge the mixture velocity invariance approximation made in some theoretical
models, such as for slug flow. Field pipelines can also exhibit extended spans of
counter-current flow, which can affect model prediction of interfacial wave properties
and transition to slug flow. Nevertheless, data from experiments remain (for their

210 © BHR Group 18 MPT 2017


certainty) the bedrock of model validation and tuning. This study therefore considers
data from both experiments and field.

The remainder of this paper is organized as follows. Section 2 provides a brief overview
of the Virtuoso model, relevant to subsequent discussion. Section 3 compares model
performance against a set of experimental data for relatively small-diameter pipes at
gentle inclination. Section 4 considers model performance against field data from a
particular large gas-condensate pipeline: specifically, model predictions are presented for
pressure, liquid holdup, and liquid cumulative volumes during production rate ramp-ups.
Section 5 discusses likely contributors to prediction error; an example of model ‘tuning’
is presented that can reduce the error and it is argued why such tuning may be
unavoidable given current levels of modeling knowledge. Section 6 contains concluding
remarks.

2 THE VIRTUOSO MODEL

The Virtuoso pipeline model solves one-dimensional equations for mass, momentum,
and energy balances. Mass and momentum balances are calculated separately for each
phase and the energy balance is combined for all phases. A pipeline network is
discretized into branches and sections thereof; geometric parameters (inclination,
diameter, section length, wall properties, etc.) can vary section to section.
Thermodynamic properties (‘PVT’) are usually obtained from look-up tables, but it is
possible to calculate these within the pipeline solver using specialized models when
necessary (e.g. for gas density under high-pressure conditions or for liquid viscosity
where stable oil-water emulsions occur).

Multiphase flow calculation involves, at least in the context of one-dimensional models,


the determination of the applicable flow pattern. For near-horizontal flow (i.e. the angle
range applicable to most pipelines but not to wells) Virtuoso distinguishes stratified,
stratified-wavy, annular, slug, and fully dispersed flow regimes. Regime prediction is
based on well-known theories (4). Stratified, stratified-wavy, and annular flow regimes
are treated using full momentum balances for each phase. Dispersed flow regimes are
calculated using drift-flux models. The slug regime is treated as a combination of
separated and dispersed flow, with the contribution of each being determined through
regime and slug properties calculations.

Liquid droplets can enter the gas phase through entrainment. Droplet entrainment and
deposition balances set the droplet radial concentration profile (5,6). When both oil and
aqueous phases are present, either may be fully or partially dispersed into the other. The
degree of dispersion is based on balancing available turbulent energy in the continuous
liquid phase with the amount of surface energy required to generate stable droplets of the
dispersed liquid phase (7). A number of closure relationships are needed to fully solve
the problem  for gas-liquid interfacial friction (8), for slug bubble velocity (9), for
critical gas-liquid slip at which entrainment commences (5), and so on. Baseline closure
relationships are based on these results in the open literature, but these have been further
tuned to improve agreement with available experimental and field data.

This broad outline of the modeling approach in Virtuoso is unlikely to be qualitatively


different from those of the other simulators considered here. Conversely, the various
simulators are likely to differ in the details of the closure relationships and more
specifically in the tunings that have been performed. The scatter in prediction results

© BHR Group 18 MPT 2017 211


seen in the following therefore likely reflects not just the natural scatter in measurement
(in repeatability or reproducibility) but also the degree of uncertainty yet remaining in
our understanding of multiphase flow. This is particularly true of the disagreement
between S1 and S3, which are the same software product.

3 COMPARISON WITH EXPERIMENTAL DATA

This section examines the performance of the various models against some experimental
data. These data are all for two-phase flow, for small diameter (D) pipes, ranging from
D = 0.1m to 0.2m, and at inclinations (θ) ranging from -1 degree (down-sloping) to +3
degrees.

Figure 1 shows parity plots of liquid holdup (αl) prediction in various models for a set of
experimental data in a D = 0.1m pipe. Here θ ranges from 0 to +3 degrees. Gas
superficial velocity (jg) ranges from 0.5 to 2.5 m/s and the no-slip holdup, jl / ( jg + jl),
from 2×10-3 to 0.07, where jl is the liquid superficial velocity. Regime determination
results (not shown) are similar in all the models. Virtuoso (S0), in particular, predicts
stratified or annular flow for most of the low holdup data and slug flow for some of the
higher holdup data. S0 and S2 over-predict liquid content at low holdups. S1 and S3
show a more centred error distribution, but with larger scatter than S0, as quantified by
the standard deviation of the relative holdup error, (αl,pred - αl,expt) / αl,expt.

0.6 0.6
(a) S0 (b) Simulator S1

0.4 0.4
Model holdup

Model holdup

+/- 20% +/- 20%

0.2 0.2

Mean of relative error = 0.148 Mean of relative error = 0.136


Std dev of relative error = 0.212 Std dev of relative error = 0.421

0 0
0 0.2 0.4 0.6 0 0.2 0.4 0.6
Expt holdup Expt holdup

0.6 0.6
(c) Simulator S2 (d) Simulator S3

0.4 0.4
Model holdup

Model holdup

+/- 20% +/- 20%

0.2 0.2

Mean of relative error = 0.353 Mean of relative error = 0.084


Std dev of relative error = 0.519 Std dev of relative error = 0.383

0 0
0 0.2 0.4 0.6 0 0.2 0.4 0.6
Expt holdup Expt holdup

Figure 1: Parity plots of αl for two-phase flow in D = 0.1m pipe at θ up to +3


degrees, for various models. Means and standard deviations are shown for
relative holdup error. Lines are relative errors of -20%, 0%, and +20%.

212 © BHR Group 18 MPT 2017


In any model, it can be seen that the scatter is the largest at intermediate holdups. This is
the region of transition from stratified-wavy to slug flow, and modeling uncertainty is
large here compared to separated flow (stratified or annular) or slug flow. It can be
shown that the average holdup in fully developed slug flow is independent of the details
of slug unit structure (10), being a function of the superficial velocities, gas bubble
velocity, and gas void fraction in the liquid slug (and independent of slug length, bubble
length, and film thickness). Further, other experimental data show that aeration of the
slug body is negligible when the Froude number (Fr2 ≡ jmix2 / gD) is less than 2 (8). Fr’s
here for the high holdup data are less than 1. Consequently, αl in the slug regime can be
well-predicted by drift-flux correlations, and the good prediction by models is not too
surprising. Holdup prediction in the intermediate range is a more challenging problem,
as models need to capture details of the regime transition process; this is considered
further in later discussion (Section 5).

A second aspect of importance to holdup prediction in the transition region from


stratified to slug flow is the high sensitivity to θ at small up-sloping inclinations. Figure
2(a) shows holdup predictions in a D = 0.1m pipe at various θ, with jg = 1 m/s and jl =
0.01 m/s everywhere. All models predict stratified flow at lower angles and slug flow at
the highest angle; the S0 model differs slightly in predicting a drop in αl after the slug
regime is attained (positive dαl/dθ is recovered with a further increase in angle, not
shown in the figure). The slope dαl/dθ is primarily a function of the friction factor fi at
the gas-liquid interface. The ratio of interfacial to gas-wall friction factors, fi / fg, can be
in excess of 15 and depends on the properties of the interfacial waves, beyond a critical
slip velocity at which the Kelvin-Helmholtz instability occurs (11,12). It would not be
surprising if fi in all models has been tuned to datasets of this type, as has been done in
S0, since there is a fair amount of variation in experimental correlations and no one
correlation has been found to be entirely satisfactory, in our experience, across the range
of data available from experiments and fields. The surface tension is relevant during the
transition to slug flow as well. Experiments are often conducted with kerosene or diesel,
whose surface tension and viscosity can be significantly different from light condensates
and black oils seen in the field. It appears that no general theory applicable over a range
of flow conditions and fluid properties is as yet available (11). The scatter in model fi’s
is further examined in Section 5.

Figure 2(b-d) show histograms of absolute holdup error in the models for a data set
(containing over 2,000 points) at D = 0.2m and θ in the range of -1 to +1 degrees. The
smallest mean and standard deviation errors are seen in down-sloping flow. This is not
too surprising since the flow is largely in the smooth stratified regime, where modeling
uncertainty is predominantly due to fi. Larger scatter occurs, for reasons discussed
earlier, in up-sloping flow; the quality of predictions for horizontal flow is intermediate
to up- and down-sloping results. The S0 model presents a more centred error distribution
for horizontal flow, but this is in our experience largely a matter of model tuning.
Indeed, from a design perspective it can be desirable to avoid the under-prediction of
holdup; this may explain the mean errors in simulators S1, S2, and S3, which primarily
target design application.

© BHR Group 18 MPT 2017 213


0.4 30
(a) Holdup vs. angle (b) Horizontal pipe S0
S1
S2
Mean Std Dev
0.3 S0 0.003 0.029 S3
S1 0.019 0.029
20 S2 0.020 0.032
S3 0.023 0.031

PDF
Holdup

0.2

10
Data
0.1 S0
S1
S2
S3
0 0
-0.25 0 0.25 0.5 0.75 1 1.25 -0.2 -0.1 0 0.1 0.2
Angle (deg) Holdup error: model - data

30 30
(c) Upwards pipe S0 (d) Downwards pipe S0
S1 S1
S2 S2
Mean Std Dev Mean Std Dev
S0 -0.006 0.062 S3 S0 -0.015 0.021 S3
S1 0.005 0.061 S1 -0.007 0.023
20 S2 0.001 0.061 20 S2 -0.009 0.023
S3 0.008 0.062 S3 -0.003 0.027
PDF

PDF
10 10

0 0
-0.2 -0.1 0 0.1 0.2 -0.2 -0.1 0 0.1 0.2
Holdup error: model - data Holdup error: model - data

Figure 2: θ-dependence of model αl: (a) data and predictions for αl(θ) in D = 0.1m
pipe; (b-d) histograms of absolute holdup error in horizontal, up-sloping (up to
+1 degree), and down-sloping (down to -1 degree) D = 0.2m pipes, respectively.

4 COMPARISON WITH FIELD DATA

Field data of three-phase flow, from an existing Virtuoso (S0) installation in a large
offshore gas-condensate field, are considered next. This online system provides various
functions: virtual well metering, pipeline monitoring, multiple specialized forecasts, and
monitoring for leaks and restrictions. The pipeline network analysed here is an
independent part of larger network and is chosen for its interesting multiphase flow
dynamics.

Figure 3(a) is an overview of the pipeline network and relevant onshore facilities.
Offshore production from two locations (points ‘A’ and ‘B’) is flowed through two short
flow lines and gathered into a long, production export line that runs from point ‘C’ to an
onshore plant. The production fluid contains gas, condensate, and water phases; mono-
ethylene glycol (MEG) is continually injected at both production points for hydrate
mitigation. (All multiphase models studied here treat MEG as part of the aqueous
phase.) Pipe diameter of the ‘C to Plant’ pipeline varies, but is everywhere greater than
0.7m. Physical multiphase flow meters at points ‘A’ and ‘B’ provide the flow rate data
to drive the models. Pressure and temperature measurements are available at points ‘A’,
‘B’, and ‘C’. Onshore reception facilities include a fairly large slug-catcher,
instrumented with pressure, temperature, and level measurements. Liquids from the slug
catcher and the knock-out drum are flowed through a heater and a pressure letdown
station into a three-phase separator. The most upstream onshore liquid flow rate
measurements are at the outlet legs of this separator, which is instrumented with total
liquid and oil-water interface level measurements.

214 © BHR Group 18 MPT 2017


Figure 3(b) shows the elevation profiles of the gathering lines and the export pipeline1;
gathering lines’ length and elevation change are much smaller than the export line’s.
Many of the interesting dynamics of this pipeline are driven by the presence of steep
escarpment in the first half of the export pipeline. Pipe angles in this region approach
+10 degrees. Slugging occurs at sufficiently high gas rates and manifests as liquid
surges at the onshore plant during gas rate ramp up events.

10
(b) Pipeline elevation profiles

3
Normalized elevation x 10
6

2
A to C
B to C
0
C to Plant

0 0.2 0.4 0.6 0.8 1


Normalized distance

Figure 3: System overview: (a) pipeline network and facilities showing liquids
handling and relevant measurement points, and (b) pipeline elevation profiles.

Field data have been extracted over two time periods (sets 1 and 2), with each set
comprising about 100 days of continuous operation. Set 1 is during early life of the field
when overall production rates are relatively low. Set 2 is during mature production when
the production rates are higher. Figure 4 shows the data for gas production rate and point
‘C’ pressure. It can be seen that the rate and pressure are higher in set 2. Both sets
feature multiple operational transients. A few of the production ramp-up events (denoted
as ‘1.1’ and ‘2.1’—‘2.3’) are identified in the figure and analysed in this study for liquid
cumulative volumes.

1.5 2.5 1.5 2.5


(a) Data set 1 Production rate Production rate (b) Data set 2
Pressure Pressure event 2.2
1.2 1.2
Normalized pressure at point C

event 2.1 Normalized pressure at point C


Normalized Gas Std Vol Rate

Normalized Gas Std Vol Rate

event 2.3
2 2
event 1.1
0.9 0.9

0.6 0.6
1.5 1.5

0.3 0.3

0 1 0 1
0 25 50 75 100 0 25 50 75 100
Time (days) Time (days)

Figure 4: Measured (normalized) gas production rate and point ‘C’ pressure in
field data sets analysed here, during (a) early (set 1) and (b) mature (set 2)
production. Ramp-up events identified here are analysed for cumulative volumes.

1
Field data are normalized everywhere in this paper. Pipeline lengths and elevations are normalized
by the length of the export pipeline. Liquid volumes are normalized by slug catcher capacity.
Liquid volume rates are presented in units of normalized volume per hour. Gas production rates are
normalized by a nominal design rate, and pressures are normalized by a nominal design pressure at
the slug catcher.

© BHR Group 18 MPT 2017 215


The models configured in S0 and each of S1—S3 are identical with regards to network
layout, pipeline geometric and wall properties, ambient temperature, and pipeline
elevation profiles. PVT look-up tables for each simulator contain the same
thermodynamic data. These data have been tuned to the observed liquid loadings in the
field and the known production fluid composition. The boundary conditions for each
model are (i) the measured gas, condensate, production water, and MEG injection rates
at pipeline network inlet points ‘A’ and ‘B’; (ii) measured temperatures at points ‘A’ and
‘B’; and (iii) the measured slug catcher pressure at the plant. Predicted model pressures,
holdups, and outlet flow rates are directly output from each simulator.

Figure 5(a) compares measured and predicted pressures at point ‘C’ for set 1. All
models are in excellent agreement with measurement until time t = 20d. This period of
data follows immediately after a pigging campaign, so that there is very little liquid in
the pipeline at t = 10d. During this initial period liquid builds up at the base of the
escarpment (near export line inlet) as the gas rate is insufficient to transport liquids up
the slope. Note that the normalized gas rate (Figure 4a) is maintained at a low value of
0.3 during this period. Around t = 20d the production rate is ramped up from 0.3 to 0.7.
This results in the gradual transport of liquid up the escarpment. Trends of cumulative
liquid volume at the outlet show no change until approximately t = 30d, in all models.

The pressure rise (see Figure 4a) observed during this process has contributions from
both frictional and gravitational components. In the range of gas rates from 0.3—0.7, the
hydrostatic contribution to total pipeline pressure drop is, correspondingly, in the range
of 50-30%, almost of all it obtaining at the escarpment. All models predict a liquid
holdup fraction in the range of 10-30%. Model predictions and measurement start to
diverge shortly after the gas rate has been ramped up. The first divergence is seen in the
t = 20d to t = 30d period. Herein all models under-predict field pressure drop by
approximately 20%; see Figure 5(b), which plots the relative error in pipeline pressure
drop, i.e. (PC, pred  PC, meas) / (PC, meas  Pplant, meas). Examination of model holdup profiles
during this period shows that oil-water slippage causes only oil to be transported up the
escarpment, while the aqueous phase remains at the base. The interpretation from field
data is that the aqueous phase is more easily transported uphill in reality. Aqueous-phase
transport up the escarpment commences in the models at t  30d. Predicted pressures
then come into better agreement with measurement.

2 50
(a) Pred. and meas. press.: set 1 Field (b) Model pressure error: set 1 S0
S0 S1
S1 S2
S2 S3
1.8 S3 25
Relative pressure error (%)
Normalized pressure

1.6 0

1.4 -25

1.2 -50
20 40 60 80 100 20 40 60 80 100
Time (days) Time (days)

Figure 5: Set 1 pressure results at point ‘C’: (a) measured and predicted pressures,
and (b) relative pressure error for each simulator.

216 © BHR Group 18 MPT 2017


Subsequently, the models differ in the amount of liquids that are carried over the top of
the escarpment and towards the plant. S2 and S3, and to a lesser extent S1, over-predict
the pressure during the t = 30d to 40d period. The relatively high production rate of 0.7
is maintained (with a short pause) until slightly after t = 40d. Larger disagreements are
seen in S1S3 pressures during production turn-down and the extended period of low
production rate (t  45d to 70d), with substantial pressure under-prediction (> 20%; see
Figure 5b), due to discharging too much liquid over the escarpment. Liquid builds up in
the pipeline during this period and all simulators predict zero pipeline outlet liquid rates.
A subsequent ‘clean-up’ operation (t  75d) removes these liquids from the pipeline.
Following this transient, simulator S3 pressures line up with measurement better than
before, whereas simulators S1 and S2 continue to under-predict the pressure, this time by
around 20% (t > 80d). S0 model pressures agree better with measurement throughout
these transients.

Pipeline liquid volume predictions by various models for set 1 are shown in Figure 6(a).
Predictions diverge starting from t = 20d, which is when liquids first arrive at plant
following the production ramp-up. It is clear that simulators S2 and S3 retain more
liquids at and upstream of the escarpment during this initial transient, as their holdup
predictions are much larger than S0 or S1. S2 and S3 holdups remain the highest through
the simulation, although interestingly the change in their holdups during the clean-up
operation (t  75d) agree relatively well with S1 and S0 (see later, in discussion related
to Figure 9). The larger holdups in S3 are consistent with the higher pressures predicted
by that model during the t = 30d to 40d period, when the liquids are on the escarpment.
(S3 holdups are larger during the subsequent low-production period as well, but the
liquids are now at the base of the escarpment and S3 pressures are smaller than
measured.) The scatter in model predictions for water holdup fraction (volume-
integrated αw / αl), shown in Figure 6(b), is smaller, although S3 predicts the highest
water fractions as well; all models, except S2, predict that most of liquids in the pipeline
are aqueous after the initial transient during t = 20d to 30d. S0 results are intermediate to
S1 and S3 for both total holdup and water fraction.

1
(a) Model liquid holdups: set 1 S0 (b) Aqueous holdup fraction, set 1 S0
S1 S1
S2 S2
S3 0.8 S3
Normalized total liquid holdup

3
Aqueous holdup fraction

0.6

0.4

1
0.2

0 0
20 40 60 80 100 20 40 60 80 100
Time (days) Time (days)

Figure 6: Set 1 model holdup predictions from S0 and simulators S1—S3:


(a) total liquid holdup, and (b) water holdup fraction.

Figure 7 presents results for pressure prediction and pressure drop errors for set 2. At the
higher gas rate, there is tendency in most models (S0, S2 and S3) to over-predict the
pressure (Figure 7a); S1 shows a more centred error distribution but the scatter in
pressure error is similar to those of the other models (Figure 7b and also Figure 12d,
discussed later). There are several production rate transients in this data set (marked in

© BHR Group 18 MPT 2017 217


Figure 4), and gas rate ramp-ups and turn-downs are associated with sloshing of the
liquids up and down the escarpment. Sometimes the gas rate is sufficiently high to
evacuate the built-up liquids from the pipeline, as seen from episodes of sharp holdup
decrease in Figure 8(a). As in set 1, the predicted pressure at point ‘C’ is dependent
upon the amount of liquids retained on the escarpment; if a model evacuates too much
liquid the period of pressure under-prediction that follows the transient can be quite
prolonged (e.g. Figure 5a).

Figure 8(b) shows results for pipe water volume as a fraction of liquid volume in each
model. The variation in total holdup correlates to the amount of excess αw, i.e. a model
predicting a higher total holdup also predicts a higher water fraction. S3 is a notable
outlier, with the difference between its holdup and that of S1 exceeding 50% of slug
catcher capacity, during low rate periods. S0 prediction is again intermediate to these
two models, whereas S2 agrees more with S1.

50
Field (a) Pred. and meas. press.: set 2 (b) Model pressure error, set 2
2.4 S0
S1
S2
S3 25
Relative pressure error (%)

2
Normalized pressure

1.6
-25
S0
S1
S2
S3
1.2 -50
0 20 40 60 80 20 40 60 80 100
Time (days) Time (days)

Figure 7: Set 2 pressure results at point ‘C’: (a) measured and predicted pressures,
and (b) relative pressure error for each simulator.

1.5 1
(a) Model liquid holdup: set 2 S0 (b) Aqueous holdup fraction, set 2 S0
S1 S1
S2 S2
1.2 S3 0.8 S3
Normalized total liquid holdup

Aqueous holdup fraction

0.9 0.6

0.6 0.4

0.3 0.2

0 0
20 40 60 80 100 20 40 60 80 100
Time (days) Time (days)

Figure 8: Set 2 model holdup predictions from S0 and simulators S1—S3:


(a) total liquid holdup, and (b) water holdup fraction.

An analysis has been conducted of liquid cumulative volumes during specific ramp-up
events (marked in Figure 4). The models terminate at the pipeline outlet (slug catcher
inlet). Since field processes downstream of pipeline outlet are not simulated, a direct
comparison of measured and predicted liquid cumulative volumes is not possible.
Available measurements are used instead to infer the pipeline outlet liquid rate occurring
in the field, obtained as follows. Starting with the measured aqueous and condensate rate
at the outlet of the three-phase separator and separator liquid level measurement, the inlet

218 © BHR Group 18 MPT 2017


liquid rate to the separator is calculated using the known level to volume relationship of
the vessel. This involves the time-derivative of the level, which is computed after fitting
the level measurement time-series to a cubic spline. Next, the calculated separator inlet
liquid rate is flashed from separator pressure and temperature (P, T) conditions to the (P,
T) conditions at slug catcher outlet (note that a heater and pressure letdown station
intervene). The composition used in this flash calculation is that of the liquid obtaining
at the slug catcher (P, T). Slug catcher P and T are quite constant during field operation,
so that compositional error introduced due to lag effects is expected to be small.
Finally, the calculated slug catcher outlet liquid volume rate is used in combination with
the slug catcher level measurement and the known level to volume relationship to obtain
the inferred pipeline outlet liquid rate in the field. Time-integration of this rate over
periods of interest yields the cumulative volume delivered into the slug-catcher. There
are uncertainties associated with field measurement and the calculation procedure. In
addition to these, there are (relatively small but potentially persistent) unmetered flows
entering the three-phase separator: from a MEG injection point just upstream of the slug
catcher and from secondary process flows that also drain into the separator.

100 0.1 4
(a) Liquids measurements: event 1.1 S0
S1
3.5 S2
80 0.08 S3
Normalized total liquid holdup

3
Normalized volume rate

2.5
60 0.06
Level (%)

SC level
Sep. level
Cond. rate
2
Wat. rate
40 0.04
1.5

1
20 0.02
0.5
(b) Model holdups: event 1.1
0 0 0
72 74 76 78 72 74 76 78
Time (days) Time (days)

0.3 2
(d) Surge volumes: event 1.1
(c) Model pipeline liquid rates

0.25
1.5
Normalized liq. volume rate

Normalized liq volume

0.2
Field
S0
0.15 S1 1
S2
S3
0.1
0.5 S0
S1
0.05
S2
S3
Field
0 0
72 74 76 78 72 74 76 78
Time (days) Time (days)

Figure 9: An example of field cumulative volume estimation: (a) separator liquid


flow rate and separator and slug catcher level measurements; (b) model holdup
transient during the transient; (c) model predictions of outlet liquid flow rate;
and (d) predicted and inferred liquid cumulative volumes.

Figure 9 shows a liquid surge event from data set 1, occurring during the gas rate ramp-
up around t = 70d. The event has a clear signature in the slug catcher level trends
(Figure 9a) and is associated with a sharp holdup decrease in all models (Figure 9b).
Liquid outlet rates predicted by all models are zero both before and after the transient
(Figure 9c) – thus this event provides a good check of the cumulative volume calculation
since liquids accumulated in the pipeline during a prolonged production turndown arrive

© BHR Group 18 MPT 2017 219


at once at the slug catcher. The inferred cumulative volume for the field is plotted
alongside model predictions in Figure 9(d). The models agree well with each other, as
they would be expected to. The discrepancy between liquid volume inference and
predictions is in the range of 10-20%, and this number can be viewed as the level of
overall uncertainty in the calculation. It is noteworthy that the cumulative volumes of S2
and S3 are close to that of S0 and S1 even though the total holdups in S2 and S3 are
larger throughout the transient. Such a holdup difference could be very significant,
though, for pig forecasting.

A similar analysis has been carried out for three other transient events in data set 2; these
are associated with production ramp-ups at t  35d, 60d, and 90d (see Figure 4). Table 1
shows the results. Model cumulative volumes are generally within the uncertainty level
of the inference calculation for S0, S1 and S3, although the persistently high predictions
of S3 (for event 2.1 in particular) suggests that the holdup in that model is too large.

5 A DISCUSSION OF MODEL SENSITIVITY

The results presented so far show that relatively large disagreements in prediction versus
measurement, as well as between predictions of different models, are associated with the
transition from stratified to slug flow2. The sensitivity of model predictions to details of
how related phenomena are modelled is discussed in the following.

Table 1: Predicted and inferred normalized cumulative volumes for selected events.

Time Cumulative liquid volume Change in holdup


Event
(d) Field S0 S1 S2 S3 S0 S1 S2 S3
1.1 72-78 1.6 1.7 1.7 1.9 1.8 1.3 1.3 1.4 1.2
2.1 34-36 1.0 1.1 0.9 0.5 1.4 0.3 0.1 0.1 0.5
2.2 59-65 3.3 2.7 2.5 1.4 3.0 0.7 0.5 0.4 0.8
2.3 91-97 2.3 2.2 2.1 1.4 2.2 0.6 0.5 0.2 0.5

The interfacial friction factor (fi) determines separated flow gas-liquid slip, and affects
the onset and development of interfacial waves. In up-sloping pipes, there is a range of
gas velocity wherein a small difference in fi can cause a relatively large change in the
predicted holdup (and thence the pressure drop via the hydrostatic contribution). Figure
10(a) plots a calculation for jl(αl) in two-phase flow, using the correlations for fi in (8),
for various jg. The curves can be viewed as the predicted outlet liquid rate in a section of
the pipe. At steady state this rate should match the inlet jl (when there is no mass
transfer); the figure shows an inlet jl = 0.01 m/s for illustration. At the highest jg of 5.5
m/s, there is only one low-holdup solution, and likewise at the lowest jg of 5.0 m/s. At
the intermediate jg, however, three solutions appear. When holdup is small, the
gravitational force on the film can be overcome by interfacial friction at low jg. The

2
The field data comparisons of the previous section suggest that models also differ in the extent to
which the aqueous phase is mobilized, but a consideration of this problem is outside the scope of
this paper.

220 © BHR Group 18 MPT 2017


influence of gravity increases with increasing holdup and film velocity drops, leading to
the intermediate holdup solution. With progressively higher holdups, the film falls
backwards (jl is negative), until the constriction of gas passage in the pipe increases gas
velocity, increasing the interfacial friction and, finally, causing the film to flow forwards
again – which leads to the high holdup solution. The intermediate holdup solution is
unstable, since djl / dαl < 0, but the two extreme holdup solutions are realizable. Thus
there is, with varying jg, a bifurcation effect3. Note that these are steady-state solutions.
In transient flow, it is likely that the predicted holdup would be some average of the two
roots. Nevertheless, this predicted holdup would remain sensitive to jg, since the
bifurcation is unavoidable.

Models for fi in commercial simulators are proprietary details, but it is possible to infer
the degree of variation across simulators by examining the steady-state holdup curves in
multiple-root regions of the parameter space. Figure 10(b) shows steady state holdup
curves from various simulators. All display a jg-region of sharp transition, but there is
considerable variation in where this occurs. The higher the ‘critical’ jg, the smaller is the
modelled friction factor. It can be estimated, for example, that the friction factor
variation between S2 and S3 is more than 25%. There is experimental evidence that the
interfacial friction factor increases with the onset of large, ‘roll’ waves, and can be
correlated to wave height. Models are likely to have been tuned to distinct sets of
experimental or field data. Further, the liquid-wall friction, as well, is affected by the
presence of interfacial waves, and this friction factor could also be based on varying
correlations.

0.04 0.5
(a) (b) S0
S1
S2
0.4
j g = 5.5 S3
0.02
Superficial liquid velocity

Holdup fraction

0.3

0
0.2

-0.02
0.1
j g = 5.0
j g = 5.25
0
-0.04
2 3 4 5 6 7
0 0.1 0.2 0.3 0.4
Holdup fraction Superficial gas velocity

Figure 10: (a) jl in stratified flow as a function of αl for a D = 0.7m, θ = +5 degree


pipe, at various jg; the horizontal line represents an inlet jl = 0.01 m/s and
intersections with this line represent steady-state solutions. (b) Steady-state
holdup curve in various models in the region of back-flowing liquid film.

3
The condition of multiple roots has been proposed as a predictor of slug flow, with the highest
stable holdup corresponding to the slug region and the lowest stable holdup to the film region – but
the phenomenon can occur in smooth stratified flow as well. Slugging can require wave growth,
which is dependent on parameters (e.g. surface tension) that do not affect smooth stratified flow.
One proposed model (13) examines roots of a drift-flux jl(αl, jg)  jl,in function, but the validity of a
drift-flux model in predicting a stratified flow root is questionable.

© BHR Group 18 MPT 2017 221


A second aspect of modeling likely to be governed by tuning parameters is the threshold
for the onset of slug flow. The applicability of drift-flux models in fully developed slug
flow is well accepted, and supported by experimental data. Figure 11(a) plots
experimental data against two calculations of jl – one from a stratified flow model
incorporating a wave-dependent interfacial friction factor and the second from a drift
flux model, with the slip velocity vs  (1 - αl)2 where the proportionality constant is
related to the gas-bubble velocity. It is evident that data at low holdup are well-predicted
by the stratified flow calculation and at high holdup by the drift-flux calculation.
Regime transition occurs over a narrow range of holdup and is likely to be a complex,
intrinsically unsteady phenomenon. Models are likely to determine the transition
threshold based on tunings. Figure 11(b) shows transition curves (for a particular set of
parameters) in the various simulators, in terms of a function of jl that is 0 in stratified
flow and 1 in fully developed slug flow. Also shown are two criteria for transition to
slugging: the necessary onset of Kelvin-Helmholtz waves (4) and a critical holdup
threshold criterion for slug stability (11). The models select a slug transition threshold
somewhat larger than theory, and this selection is likely to be the result of tunings
against specific data.

1.2
(a) (b)
1
(j l,model - j l,strat ) / (j l,df - j l,strat )

0.9 K-H waves


Superficial liq velocity

Data
S0
0.6 0.5 S1
Slug stability S2
S3

0.3

Stratified flow model 0


Drift-flux model
Data
0
0 0.1 0.2 0.3 0.4 0 0.2 0.4 0.6
Holdup Holdup

Figure 11: Transition to slug flow: (a) Data and superficial liquid velocity curves
calculated using representative stratified flow and drift-flux models. (b) Data and
slug transition models as a function of holdup in a pipe at +1 degrees.

It is therefore of interest to determine overall model sensitivity to these parameters.


Figure 12 shows how S0 predictions change when the interfacial friction factor and slug
transition criteria are adjusted by 25% to favour transition to slug flow. S0-predicted
holdups, which were intermediate originally to S1 and S3 predictions are brought into
near-coincidence with S1 (Figure 12a; a short time-window is shown for clarity of
plotting but is representative). S0’s over-prediction of measured pressure is reduced by
50% for set 2, and overall prediction of set 1 pressure deteriorates only marginally. This
exercise is not meant as a demonstration of fully tuning a model, but merely as an
illustration of model sensitivity to these aspects of flow physics.

222 © BHR Group 18 MPT 2017


1.2 2.5
S0, baseline (b) Set 2: Model and field pressures
(a) Set 2, model holdups
S1
S0, adjusted
Normalized total liquid holdup
0.9 2.25

Normalized pressure
Measured
S0, baseline
0.6 2 S1
S0, adjusted

0.3 1.75

0 1.5
40 50 60 70 80 70 72 74 76
Time (days) Time (days)

25 25
S0, baseline S0, baseline
(c) Set 1: pressure error distribution. (d) Set 2: pressure error distribution.
S1 S1
S2 S2
20 S3 20 MEAN STD DEV S3
S0, adjusted S0, baseline 0.108 0.067 S0, adjusted
S1 0.009 0.076
MEAN STD DEV
S2 0.067 0.073
S0, baseline -0.077 0.066
15 15 S3 0.074 0.101
S1 -0.150 0.360
S0, adjusted 0.053 0.056
PDF

PDF
S2 -0.169 0.162
S3 -0.063 0.427
10 S0, adjusted -0.084 0.077 10

5 5

0 0
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
P model - P field P model - P data

Figure 12: ‘Adjusted’ S0 model results: (a) αl for set 2; (b) point ‘C’ pressure; (c)
pressure error distribution in set 1; and (d) pressure error distribution in set 2.

6 CONCLUDING REMARKS

The comparison of various multiphase simulators against experiment and field data
considered here shows a scatter in the results in the range of 10% or more. While this
level of accuracy may be considered adequate for qualifying a model, the error has real
consequences for the viability of an online system. A 10% error in pressure prediction,
for instance, can translate to an absolute error of 10 bar or more, which is too large for
the viability of an online leak-detection or restriction-detection system. The scatter in
model holdup predictions is a large fraction of field liquids handling capacity.

It is reasonable to assume the tunings applied internally in different commercial


simulators have been derived from distinct data sets. It is not our argument here that any
one simulator is superior to others in general. Rather, the point is that there are yet
intrinsic limitations to the accuracy of available correlations  such as for interfacial
friction, slug transition mechanisms, and liquid-wall and liquid-liquid friction factors 
that yield the levels of prediction scatter observed herein. Indeed, variation between S1
and S3 (which are different versions of the same tool) can be viewed as indicative of the
levels of prediction uncertainty. Further, modeling uncertainties in the field are
compounded by PVT uncertainties, such as liquid-gas and liquid-liquid surface tension,
that can have significant impact on wave properties, for example. Finally, field
prediction is often particularly sensitive to slug transition. The ‘take-off’ point in the
overall holdup curve αl(jg) of a pipeline (where dαl/djg becomes large with decreasing jg)
is closely related to the appearance of film backflow; pipelines often operate in this
region, since larger gas velocities cause too much frictional pressure drop and smaller
gas velocities lead to excessive holdup.

© BHR Group 18 MPT 2017 223


For these reasons it seems to remain necessary to tune flow assurance models to field-
specific data. Wood Group’s experience with field systems shows that viable online
simulation tools are indeed achievable with comprehensive validation and analysis of
field data followed by judicious tuning. Flow assurance engineers working with design
tools are well advised to consider the sensitivity of results to modeling parameters that
are intrinsically uncertain. It is recommended that simulation tools expose such
parameters affecting internal model calculations to facilitate these sensitivity analyses.

REFERENCES

(1) Parthasarathy, P. & Mai, M. (2016) ‘Bridging the gap between the design world
and online, real-time, dynamic simulation world’, Offshore Technology
Conference, Houston, OTC-27232-MS.
(2) Mai, M. C. & Erickson, D. D. (2012) ‘Twenty years of practical observations using
online real-time dynamic simulation in the upstream O&G industry’, SPE
Advanced Technology Workshop, Bangkok.
(3) Parthasarathy, P. V., Haldipur, P. L. & Erickson, D. D. (2007) ‘Real-time pipeline
monitoring’, in Pipeline and Gas Technology, May 2007 issue, pp. 22-25.
(4) Shoham, O. (2006) ‘Mechanistic modeling of gas-liquid two-phase flow in pipes’,
SPE.
(5) Pan, L. & Hanratty, T.J. (2002) ‘Correlation of entrainment for annular flow in
horizontal pipes’ Int. J. Mult. Flow 28, pp. 385-408.
(6) Skartlien, R. et al. (2011) ‘Simultaneous entrainment of oil and water droplets in
high Reynolds number gas turbulence in horizontal pipe flow’ Int. J. Mult. Flow
37, pp. 1282-1293.
(7) Brauner, N. (2001) ‘The prediction of dispersed flow boundaries in liquid-liquid
and gas-liquid systems’ Int. J. Mult. Flow 27, pp. 885-910.
(8) Tzotzi, C. & Andritsos, N. (2013) ‘Interfacial shear stress in wavy stratified gas-
liquid flow in horizontal pipes’ Int. J. Mult. Flow 54, pp. 43-54.
(9) Bendiksen, K. H. (1984) ‘An experimental investigation of the motion of long
bubbles in inclined tubes’ Int. J. Mult. Flow 10, pp. 467-483.
(10) Taitel, Y. & Barnea, D. (1990) ‘Two-phase slug flow’ Adv. Heat Trans. 20, pp. 83-
132.
(11) Hurlburt, E.T. & Hanratty, T.J. (2002) ‘Prediction of the transition from stratified
to slug and plug flow for long pipes’ Int. J. Mult. Flow 28, pp. 707-729.
(12) Fabre, J. (2002) ‘Modelling of stratified gas-liquid flow’ in Modelling and control
of two-phase flow phenomena, Int. Cent. Mech. Sci., Udine Italy.
(13) Danielson, T. J. (2011) ‘A simple model for hydrodynamic slug flow’ Offshore
Tech. Conf., Houston TX (21255-PP).

224 © BHR Group 18 MPT 2017

You might also like