Journal of Petroleum Science and Engineering 208 (2022) 109762

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

journal homepage: www.elsevier.com/locate/petrol

A comparative study on the performance of multiphase flow models against

offshore field production data
Gabriela Souza Chaves a, *, Hamidreza Karami b, Virgilio Jose Martins Ferreira Filho a,
Bruno Ferreira Vieira c
a
Federal University of Rio de Janeiro, Brazil
b
University of Oklahoma, USA
c
Petrobras R&D Center, Brazil

A R T I C L E I N F O A B S T R A C T

Keywords: In the oil industry, multiphase flow models are commonly used to simulate the flow from reservoir to the pro­
Multiphase flow duction unit. Flow simulations help in the production monitoring and optimization to support the decision-
Production optimization making process. Despite the need for accurate simulations to support production operations, multiphase flow
Offshore data analysis
models are mostly developed using laboratory data and rarely validated under field-scale conditions. To improve
Flow assurance
the understanding of multiphase flow models in field conditions, this work evaluates the performance of a large
Flow models
set of models for a dataset composed of two production units with 20 producing wells and 865 measured pro­
duction points. The wells are divided and analyzed in two segments: the surface flowline and the wellbore. The
evaluation compares the models’ performances using various statistical parameters and trending charts. The
evaluation analyzes the impacts of production parameters, well geometry, predicted flow pattern, and pressure
drop components on the model performances. Overall, Ansari’s model presented the best performance with
13.8% absolute average percent error for the flowline segment and 14.1% error for the wellbore segment. In
addition, Gray can also generate good results for most but not all of the tested wellbores. From the production
parameters evaluation, higher flow rate tests resulted in more stable outputs. Ansari and Beggs models showed a
trend of increasing errors for wells with larger horizontal wellbore lengths. Also, Beggs model showed larger
errors when the transition flow pattern was predicted. Finally, both Ansari and Beggs models showed larger
overprediction errors for tests with higher pressure losses in the flowline segment, while Gray showed smaller
errors for these cases. The conclusions can be used to select the best models given the production system’s flow
parameters, or be incorporated in systems to increase the accuracy.

1. Introduction Jäschke, 2020), optimization (Chaves et al., 2020a), surveillance (Pou­

lisse et al., 2006), and monitoring (Lunde et al., 2013). This wide
The oil and gas production system is responsible to transport the application encourages researchers to map and understand the use and
hydrocarbon mixture from the reservoir to the market place, passing accuracy of various empirical and mechanistic models.
through the well, flowlines, and risers. This mixture, composed of oil, Until the 1980s, multiphase flow was mostly modeled by empirical
water, and gas, is a multiphase fluid. Multiphase flow models play a key correlations due to its complex nature (Ansari et al., 1994). Empirical
role in simulating the production system. Field monitoring, manage­ correlations were developed for a specific range of variables and con­
ment, and surveillance are examples of production operations that de­ ditions used in the experiment process. Beggs and Brill (1973) and
mand the flow behavior simulation. This is what makes multiphase flow Dukler et al. (1964) are the two most commonly used empirical corre­
models a vital part of production monitoring and optimization. lations to compute the pressure drop in horizontal flows. For vertical and
Multiphase flow models are implemented to simulate the flow sharply inclined flows Hagedorn and Brown (1965), Duns and Ros
behavior in many problems faced in the industry, including well allo­ (1963), Orkiszewski (1967), Aziz et al. (1972), and Gray (1978) are the
cation (Rashid, 2010), virtual flow metering (Bikmukhametov and most commonly used empirical correlations (Shoham, 2006).

* Corresponding author.
E-mail address: [email protected] (G.S. Chaves).

https://doi.org/10.1016/j.petrol.2021.109762
Received 16 April 2021; Received in revised form 6 October 2021; Accepted 29 October 2021
Available online 2 November 2021
0920-4105/© 2021 Elsevier B.V. All rights reserved.
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Beggs and Brill (1973) conducted their experiments in a 90-ft long pattern is predicted, pressure drop calculations differ for each flow
acrylic pipe using air and water as fluids, a small pipes diameter (1.0 and pattern.
1.5 inches), and angles from − 90◦ to +90◦ . Although the experiments Hasan and Kabir (1988) presented a mechanistic model to predict
were done for horizontal and vertical inclinations, only horizontal flow pressure drop in vertical wells considering bubbly, slug, churn, and
patterns were mapped, resulting in a correlation more suitable for hor­ annular flow patterns. Similar to most other mechanistic models, churn
izontal flow. flow is not independently modeled due to its chaotic and irregular na­
Dukler et al. (1964) focused their work on frictional pressure drop ture, and instead treated similar to slug flow in pressure drop
and proposed an approach based on the principles of similarity. The calculations.
correlation results were satisfactory; however, it was dependent on Xiao et al. (1990) studied the previous models for horizontal and
holdup results, which were not discussed in this work. near-horizontal pipelines, and proposed a comprehensive mechanistic
Hagedorn and Brown (1965) performed their experiments in a approach. The model can be considered as a modified version of Taitel
1500-ft deep vertical experimental well using air and four types of crude and Dukler’s work. Their model predicted flow patterns and flow
oils, and collected 360 data points with three sizes of tubing (1.0, 1.25, characteristics for stratified, intermittent, annular, and dispersed bubble
and 1.5 inches). The liquid holdups were not collected, and the resulting flow. The model was statistically evaluated against three data bases,
correlation did not distinguish the flow pattern. A pseudo-holdup including experimental and field pipelines. Similarly, Ansari et al.
concept was introduced and applied in developing the correlation. (1994) studied the previous works and formulated a comprehensive
Duns and Ros (1963) developed their correlation using a 185-ft long model for vertical and near-vertical tubulars. The model is partially
vertical flow pipe with 1.26–5.60 inches diameter. The multiphase flow based on Taitel at al.‘s work, and differs between bubble, slug, and
was represented by air for the gas phase and water or crude oil for the annular flow patterns. Details about these models can be found in the
liquid phase, constructing an overall of around 4000 tests and data corresponding articles or books like Shoham (2006) and Al-Safran and
points. Brill (2017).
Orkiszewski (1967) developed a correlation by checking previous Mechanistic models are based on the physical nature of the flow,
works with field data and mixing methods according to flow pattern making them applicable to a wide range of flow variables. This advan­
calculations. As a result, this correlation uses Duns and Ros (1963) to tage over the empirical models is undeniable. However, mechanistic
obtain the slug-mist flow transition boundary and to compute the mist models present inherent numerical discontinuities. The numerical dis­
flow pressure drop. Orkiszewski (1967) also used Griffith and Wallis continuities are derived from multiple roots at the convergence process
(1961) to obtain the transition boundary between bubble and slug flow, (often faced in calculating holdup depending on numerical techniques)
and to calculate the bubble flow pressure drop. As their contribution, and sudden transitions in flow regimes. These discontinuities represent a
Orkiszewski proposed a calculation method for slug flow. challenge for the industry, especially in cases where the model is
Gray (1978) developed a correlation that presents satisfactory per­ incorporated into systems to solve larger problems (Shippen and Bailey,
formance for high gas content wells. This is an empirical correlation, 2012). The tradeoff between continuous simplified models and robust
developed using 108 vertical well test data from gas and condensate models with more accuracy needs to be considered by the industry.
fluids. Pressure drops were not dependent on flow pattern, and In the recent years, a new era of multiphase modeling has started
single-phase flow was assumed in this correlation. The authors affirmed with unified mechanistic models. These models use the same combined
the correlation is suitable for flow velocities below 50 ft/s, pipe di­ momentum balance equation for the liquid-gas flow, regardless of the
ameters smaller than 3.5 inches, condensate/liquid ratios below 50 predicted flow pattern. This approach helps provide a continuous and
bbl/mmscf, and water/liquid ratios below 5 bbl/mmscf. more accurate solution for the pressure drop. Most of these models use
Aziz et al. (1972) represent the transition to mechanistic modeling. slug flow as the default flow pattern and predict the transitions from this
In this work, the flow pattern transitions were evaluated using physical pattern. One of the earliest attempts for this approach was done by
concepts. However, only the two commonly encountered flow patterns Bendiksen et al. (1991) using a wide range of available experimental
in oil wells, bubble and slug, are considered. Aziz et al. (1972) marked a data. They developed a unified model to predict pressure drop of
period called awakening years (Shippen and Bailey, 2012). Researchers multiphase flow for all flow patterns and inclination angles using the
faced difficulties on correlations’ extrapolations for sizes of pipes, fluids concept of minimum slippage. This model and its later modifications
types, and high flow rates. Authors started to propose analytical models were used as the foundation for a popular software that has been
representing the physical behavior of the flow and the mechanistic packaged and marketed as OLGA™ over the years.
models appeared in the literature. At the same time, computers were Zhang et al. (2003) developed a two-phase flow model, commonly
getting faster and more affordable to solve complex numerical problems. known as TUFFP unified model, capable of predicting flow character­
Mechanistic models are based on the combined momentum balance istics for all flow patterns and inclination angles. The model uses slug
equation, in which flow patterns are defined by transitions boundaries, flow as the default flow pattern, and transition boundaries are predicted
and independent models are developed for each flow pattern. Flow using the relative lengths of liquid film and slug body regions. The model
pattern prediction is dependent on variables like gas and liquid flow was benchmarked and validated against a wide range of field and
rates, fluid properties, pipe diameter, and angulation. Mechanistic experimental data points.
models combine these variables using a physical basis and generate a Most of the aforementioned published works use laboratory test
coordinate system to predict the relationship between the variables and data. However, in the field the truth is different. Most experimental
the transitions boundary. A large number of coordinate systems can be facilities have low-pressure conditions, with model oil or water as the
created, which may be dimensional or dimensionless. The main differ­ liquid and compressed air as the gas phase. In addition, the pipe incli­
ence between empirical and mechanistic models is that the first one nation angle is either fully horizontal or vertical. A field production
generates the maps using data and the second one using physical re­ system however, is combined of many long sections with oscillating
lationships. As a result, the mechanistic models are ideally not con­ inclination angles, sometimes varying pipe diameters, and changing
strained to a limited range of flow parameters. pressure resulting in varying fluid characteristic over the field lifetime.
Taitel and Dukler (1976) developed one of the earliest mechanistic According to Shippen and Bailey (2012), validating models under
models, predicting flow boundaries in horizontal two-phase flow. The field-scale conditions is a research area that still requires further study in
study only evaluated horizontal flow, but it was continued later by Taitel the multiphase area. In recent years, Waltrich et al. (2019) investigated
et al. (1980), who developed a model for vertical flow. Barnea (1987) the performance of multiphase flow models focused on the
studied the effects of pipe inclinations, releasing a mechanistic model wort-case-discharge cases and affirmed comparisons between models
that includes the inclination angles. For all these models, after the flow are rare in the open literature. Waltrich et al. (2019) highlighted the

2
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Table 1 Table 2
Average well information for the two production units. Data collection summary.
Production Unit A Production Unit B Production Production
Unit A Unit B
Number of wells 9 11
Fluid API 18.6 20.0 Period Average per well 7 years and 4 3 years and 6
Gas Specific Gravity 0.648 0.657 months months
Flowline TVD (m) 1660 1744 Number of tests Total 580 285
MD (m) 7125 5108 Average per well 64 26
d (in.) 6 7 Measured Water Cut [%] 19 10
Wellbore TVD (m) 1360 1284 parameters - GOR [m3/m3] 155 217
MD (m) 1472 1507 average Liquid flow rate 1793 937
d (in.) 5.4 5.4 [m3/day]
Reservoir Temperature (◦ C) 59.8 58 QGL [m3/day] 328,000 148,000
Pressure (kPa) 28,439 26,870 Pressure at surface - 1666 2254
Psup [kPa]
Pressure at wellhead 10,983 8041
- Pwh [kPa]
Pressure at 18,437 12,749
bottomhole - Pbh
[kPa]

well test can take a few hours to a few days, and the result is an average
of the measurements taken. The production test is executed sporadically,
depending on the well behavior or regulatory agencies’ demands. The
average frequency is a month for this case.
The water, gas, and oil production rates are measured after separa­
tion in platform facilities. Those values are used to obtain water cut
(water-liquid ratio - WC) and the gas-oil ratio (GOR). The gas lift in­
jection flow rate (QGL) is a controlled variable, and it is also presented in
the well test. Pressures and temperatures are taken by transmitters
Fig. 1. A sample well geometry for the field. localized at three points: surface platform (abbreviated as Psup and Tsup),
wellhead (abbreviated as Pwh and Twh), and bottom hole (abbreviated as
importance of studies mapping specific conditions where the models Pbh and Tbh).
present higher errors, such as large-diameter pipes, high gas/liquid ra­ The measurement process and equipment have some inherent un­
tios, and high flow rates. certainties that lead to uncertainties in the test results. Some data judged
This work aims to improve our understanding of multiphase flow by as unreliable were removed from the dataset. A summary of the testing
evaluating the performance of a large group of multiphase flow models parameters and averages of the resulting data is presented in Table 2.
compared to a field dataset. The dataset is composed of two production Production unit A provides a longer duration of testing with more tests
units from an offshore field, comprising 20 wells with period of pro­ per well and higher average production rates and pressures. The com­
duction from 3 to 11 years and an overall of 865 measured production bination of these two units with multiple wells helps give us a wide
points. range of data at varying flow rates and pressures. One example of the
well test data plotted over time for a well in production unit A is given in
2. Field data structure Fig. 2. A natural decline in production rate and increase in water cut can
be observed over the years.
The performances of multiphase flow models are benchmarked in
comparison to a large dataset. This dataset includes the data from two 3. Methodology
production units in an offshore oil field. Production unit A is composed
of 9 wells and production unit B has 11 wells. All the wells are satellite This work aims to study multiphase flow model performances against
wells, i.e. no subsea manifolds are used and wells are directly routed to real field production data. The objective is not only to compare the
the production unit. The artificial lift method used for all of these wells is performance among the most used models, but also to map the most
gas lift. Table 1 presents a summary of some basic information for these suitable ranges of operating conditions, where a given model’s errors are
production units. The true vertical depth (TVD) is the depth measured minimized.
vertically from the wellhead of each well, while the measured depth To perform the analysis each producing well is divided into two
(MD) represents the total drilled length of the well. The flowline rep­ segments: flowline and wellbore, as shown in Fig. 1. The multiphase
resents the production line that connects the wellhead to the production flow models considered are Beggs and Brill (Beggs and Brill (1973),
unit. Fig. 1 shows a typical well design, where the flowline and the Hagedorn and Brown (1965), Orkiszewski (1967), Duns and Ros (1963),
wellbore segments can be observed, as well as the typical inclination Gray (1978), Aziz (Aziz et al., 1972), Mukherjee (Mukherjee and Brill,
angles, and the gas lift injection point. The wellbore segments are mostly 1985), Dukler (Dukler et al., 1964), Xiao (Xiao et al., 1990), and Ansari
near-vertical, while the flowlines are composed of long near-horizontal (Ansari et al., 1994). In utilized datasets, the wellbore segments have a
pipes followed by vertical risers. This pattern is similar for all wells. The vertical/near-vertical design, and the flowline segments have a
numbers in the table show the average values for the wells in the given near-horizontal and a near-vertical part. Combining this information
production unit. with the specific inclination angles suitable for each model, the models
Production well tests are a common practice in the oil industry to were tested as follows in Table 3 for each segment. Each model formu­
track a well’s behavior over its life. Production units are equipped with a lation can be found in the references mentions in Table 3. Overall, 9
test facility shared among all the wells, and one well is connected to the models were tested for the wellbore section, and 10 models were tested
facility at a time to perform the test. During the well test, production for flowline section. In addition, combinations of various model pairs
data are measured a couple of times after reaching the steady state. The were tested for the flowline section depending on the inclination angle.

3
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Fig. 2. Well test data example, including WC, GOR, temperatures, pressures and rates with time.

Table 3 Table 4
Multiphase flow models application. Used correlations.

Models Application: Variable Available correlations/

reference
Wellbore Beggs and Brill (1973) The model applied for the full wellbore
Hagedorn and Brown segment Black-oil Oil Formation Volume Standing (1947)
(1965) Factor Vazquez and Beggs (1977)
Orkiszewski (1967) Glaso (1980)
Duns and Ros (1963) Petrosky and Farshad (1998)
Gray (1978) Gas Solubility Ratio Vazquez and Beggs (1977)
Aziz et al. (1972) Standing (1947)
Mukherjee and Brill (1985) Glaso (1980)
Dukler et al. (1964) Petrosky and Farshad (1998)
Ansari et al. (1994) Water Formation Volume McCain (1990)
Flowline Beggs and Brill (1973) The model applied for the full flowline Factor
Hagedorn and Brown segment Fluid Properties Gas Compressibility Factor Kareem (2014)
(1965) Gas Critical Pressure Standing (1981)
Orkiszewski (1967) Gas Critical Temperature Standing (1981)
Duns and Ros (1963) Gas Viscosity Lee et al. (1966)
Gray (1978) Water Viscosity McCain (1991)
Aziz et al. (1972) Dead Oil Viscosity Beggs and Robinson (1975)
Mukherjee and Brill (1985) Standing (1981)
Dukler et al. (1964) Glaso (1980)
Ansari et al. (1994) Labedi (1992)
Xiao et al. (1990) Chew and Connally (1959)
Beggs and Brill (1973) + First model applied to the horizontal/ Saturated Oil Viscosity Beggs and Robinson (1975)
Hagedorn and Brown near-horizontal parts, while second Petrosky and Farshad (1995)
(1965) model applied to the vertical/near- Kartoatmodjo and Schmidt
Xiao et al. (1990) + Ansari vertical parts (1991)
et al. (1994) Standing (1981)
Beggs and Brill (1973) + ( Thermal Sagar et al. (1991)
Orkiszewski (1967) Exchange
Beggs and Brill (1973) +
Gray (1978)
Beggs and Brill (1973) + Calculations were coded and executed within the structure of SELMA
Aziz et al. (1972) (Chaves et al., 2020b), a proprietary steady-state multiphase flow soft­
Beggs and Brill (1973) + ware. The fluid mixture was characterized by a black-oil model using
Duns and Ros (1963)
models available in the literature. The oil formation volume factor (Bo)
and gas solubility ratio (RS) were evaluated by models chosen using PVT
An inclination angle of 20◦ was adopted to switch between the data among the ones listed in Table 4. The fluid properties were updated
near-horizontal and near-vertical flow models. The models of Duns and along the flow path for each thermodynamic state, as also listed in
Ros, Mukherjee and Dukler presented larger errors, and are therefore not Table 4. The oil viscosity is an important property available from PVT
included in the evaluation, which focuses only on the better performing data, thus the best performing correlation among the ones listed against
models. the PVT data was selected. The oil-water emulsion viscosity was also

4
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

properties calculations included at Table 4. The calculations over the

segments are numerically integrated. Within each pipe segment, a
trial-and-error solution process is used to compute the output pressure.
This output pressure is used as the input for the next segment. Multi­
phase flow models are used to calculate the pressure gradient (dp/dL)
within each segment, using the flow conditions as input. Fig. 3 shows a
scheme of the marching algorithm procedure, where n is the number of
segments.

3.1. Model evaluation procedure

The model evaluations are carried out based on the calculated

pressure drops in the flowline and wellbore segments. The measured
temperature data for surface, wellhead, and bottomhole points are used
to adjust the temperature drop calculations and obtain more accurate
results. The producing flow rate and one pressure point are used as in­
puts to the multiphase flow model, while the other pressure point is
calculated for the evaluation. For the flowline segment, the surface
platform pressure (Psup) is used as the input, and the wellhead pressure
(Pwh) is calculated. For the wellbore segment, the wellhead pressure
(Pwh) is the input, and the bottomhole pressure (Pbh) is calculated.
The pressure drop evaluation is based on the comparison between
the measured pressure drop ( − ΔPM ) and the calculated value ( − ΔPC )
for all available data in wellbore or flowline. Three statistical parameters
are used in the evaluation: the average percent error (ε1 ), the absolute
average percent error (ε2 ), and the percent standard deviation (ε3 ).
These three parameters are based on the percent relative error (er )
described in Equation (1), and they are expressed in % in Equations (2)–
(4), respectively. The n value is the total number of analyzed data used
in model evaluation.
− ΔPC − ( − ΔPM )
er = ∙100 (1)
− ΔPM
Fig. 3. Marching algorithm schematic. ( )
1 ∑
n
ε1 = er,i (2)
n
treated with a proprietary exponential model, provided by the operator. i=1

These fluid characteristics are then used to calculate the in-situ flow ( )
∑
n ⃒
rates and properties for the liquid and gas phases. The outputs of this ε2 =
1 ⃒
⃒er,i ⃒ (3)
segment get implemented as inputs for the multiphase flow models. n i=1
The cumulative pressure drops in wellbores and flowlines are
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
computed using what is called in “marching algorithm” (Brill and ∑
n
( )2
er,i − ε1
Mukherjee, 1999). To apply this algorithm, the production system is ε3 = (4)
n− 1
divided into multiple segments, with one segment border as the starting i=1

inlet. The flow model calculations are performed for each pipe segment. The absolute average percent error (ε2) is the most relevant statistical
For this purpose, pressure and temperature are assumed constant at each parameter to evaluate the models. Opposite to the average percent error
segment to compute the flow conditions using the black-oil and fluids (ε1), the positive and the negative values do not cancel out in ε2

Table 5
Overall evaluation.
Flowline Segment Wellbore Segment

Model No of Data ε1 (%) ε2 (%) ε3 (%) Model No of Data ε1 (%) ε2 (%) ε3 (%)

Ansari 858 5.6 13.8 18.1 Ansari 836 − 0.2 14.0 18.2
Xiao_Ansari 848 6.3 14.9 19.8 Gray 842 − 12.5 15.2 11.4
Xiao 760 − 1.5 15.5 21.4 Hagedorn 816 3.2 18.1 27.9
Beggs_Gray 827 − 11.1 18.7 20.7 Beggs 858 6.6 19.4 29.6
Beggs 835 15.2 19.8 22.0 Orkiszewski 838 − 1.8 20.6 29.7
Gray 858 − 18.5 20.0 18.9 Aziz 858 − 14.3 21.1 20.4
Beggs_Duns 833 22.1 30.1 38.1 Duns 858 41.9 42.0 41.1
Beggs_Aziz 833 7.5 31.3 35.7 Total 858
Duns 850 14.0 31.8 43.1
Beggs_Orkiszewski 823 34.7 36.5 23.5
Beggs_Hagedorn 825 41.2 42.9 23.5
Hagedorn 854 42.4 44.4 25.7
Aziz 865 24.4 50.7 57.5
Orkiszewski 825 49.5 50.8 28.7
Total 865

5
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Fig. 4. Overall dataset boxplot.

calculation. On the other hand, this property of ε1 is important to indi­ with regards to their overall error values. Then, a clustering study is
cate the agreement between the predicted values and the measured data. presented, and existing trends in model performances are presented
Positive values denote over-prediction, while negative values denote considering various operational and multiphase flow parameters.
under-prediction.
Boxplot graphics are used to assist with the visualization of the er­ 4.1. Overall evaluation
rors. Boxplot is a commonly used graphic that shows 5 data metrics
(minimum, first quartile, median, third quartile, and maximum), the The overall evaluation includes the entire dataset, presented in
outliers, and the data symmetry. The lower and upper box borders are Table 5. Models are sorted in ε2 ascending order. Ansari’s model pre­
the first and third quartile, respectively. The first quartile represents a sents the best overall performance considering the absolute average
point where 25% of the data are lower than the quartile. Similarly, the error for both inspected segments, 13.8% for flowline segment and
third quartile is the point where 75% of the data are lower than the 14.1% for the wellbore segment. In the flowline segment, Ansari’s model
quartile. The line inside the box is the median or the second quartile, shows the best performance considering all three statistical parameters.
with 50% of the data below and the other 50% of the data above this In the wellbore segment, Ansari presents the best absolute average error
line. Lastly, the whiskers are the “minimum” (Equation (5)) and and average error, but not the smallest standard deviation. However, the
“maximum” (Equation (6)) values excluding the outliers. The outliers differences in data are not sufficiently high to invalidate the model
are defined as the data points out of the “minimum” and “maximum” performance.
range. IQR in equations (5) and (6) is the interquartile range, which The number of data, shown in Table 5, implies the number of field
includes the range between the first to the third quartile, that encom­ data, where the model has successfully converged. This number can be
passes 50% of the data points. compared with the total data, 865 for the flowline and 858 for the
′′Minimum′′ = Q1 − 1.5 ∗ IQR (5) wellbore segment. The number of calculated data for Ansari model is
among the highest for both flowline and wellbore segments. It must be
′′Maximum′′ = Q3 + 1.5 ∗ IQR (6) mentioned that mechanistic models have relatively more complex
convergence algorithms, which can lead to a greater rate of convergence
4. Evaluation failure. In contrast, correlations like Aziz converge for all cases. How­
ever, their accuracies may be a bigger question mark.
The objective of this section is to evaluate the performances of the The combinations of two models for the flowline segment intend to
aforementioned multiphase models against the available field data. decrease the error by applying the most suitable correlation according to
First, an overall evaluation is presented, where models are compared the pipe inclination. In this respect, it was expected that the combined
models would result in smaller errors. This was the case for all the

Table 6
Production unit evaluation for the flowline section.
Flowline Segment

Production Unit A Production Unit B

Model No of Data ε1 (%) ε2 (%) ε3 (%) Model No of Data ε1 (%) ε2 (%) ε3 (%)

Gray 573 − 9.7 12.0 13.6 Ansari 282 − 1.6 12.3 16.4
Xiao 531 3.5 12.6 19.1 Xiao_Ansari 282 − 2.3 12.8 16.7
Beggs_Gray 575 − 2.0 12.9 15.7 Beggs 255 7.0 16.8 31.5
Ansari 576 9.2 14.6 17.8 Xiao 229 − 13.1 22.4 21.9
Xiao_Ansari 565 10.5 15.9 19.9 Beggs_Orkiszewski 252 26.2 29.3 26.5
Beggs 580 18.7 21.2 14.8 Hagedorn 285 25.7 30.5 25.7
Beggs_Duns 580 15.4 21.9 25.1 Beggs_Hagedorn 253 26.3 30.8 24.9
Duns 580 3.8 23.0 28.2 Beggs_Gray 252 − 31.9 32.1 14.6
Beggs_Aziz 580 24.9 29.0 24.6 Gray 285 − 36.1 36.2 15.4
Beggs_Orkiszewski 571 38.4 39.6 21.1 Beggs_Aziz 253 − 32.4 36.6 23.0
Beggs_Hagedorn 572 47.7 48.2 19.5 Aziz 285 − 32.3 41.6 31.6
Hagedorn 569 50.8 51.4 21.2 Orkiszewski 269 40.3 42.7 32.4
Orkiszewski 556 54.0 54.8 25.6 Beggs_Duns 253 37.3 48.8 54.9
Aziz 580 52.3 55.1 45.7 Duns 270 35.8 50.7 58.7
Total 580 Total 285

6
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Fig. 5. Production unit clustering boxplot for flowline section.

Table 7
Production unit evaluation for the wellbore section.
Wellbore Segment

Production Unit A Production Unit B

Model N◦ of Data ε1 (%) ε2 (%) ε3 (%) Model N◦ of Data ε1 (%) ε2 (%) ε3 (%)

Ansari 563 − 9.0 10.7 7.9 Gray 275 − 5.6 12.1 13.8
Beggs 575 − 7.8 10.7 8.7 Ansari 273 18.0 20.9 19.8
Hagedorn 544 − 8.6 11.7 11.0 Aziz 283 − 1.0 21.2 30.0
Orkiszewski 565 − 15.4 16.3 8.5 Orkiszewski 273 26.2 29.4 37.3
Gray 567 − 15.9 16.7 8.2 Hagedorn 272 26.7 30.8 35.6
Aziz 575 − 20.8 21.0 7.2 Beggs 283 35.9 37.1 35.0
Duns 575 25.3 25.3 25.0 Duns 283 75.7 76.0 46.3
Total 575 Total 283

models, except for Ansari’s model. near-vertical risers. The pressure drops are mainly dictated by the
Fig. 4 shows the boxplot for the top 5 correlations in each section. gravitational losses in the riser. First, looking at production unit A, Gray
The boxplot presents the error distribution and comparison of the top 5 presents the best performance with 12.0% absolute average error, high
models. The box sizes confirm the table results, showing the best results number of successful convergences (573 from 580) and small standard
for Ansari and Gray (only in the wellbore). It can be observed that the deviation. The top 3 models present similar absolute average percent
median value (line inside the box) and the majority of the error distri­ errors. However, Xiao model has a lower convergence number (531 from
bution are negative for the wellbore segment. This means that the 580) and a higher standard deviation (19.1%), confirmed by the pres­
models generally tend to underpredict the pressure drop. However, the ence of outliers in the boxplot in Fig. 5.
outliers seem to mostly be in the overpredicting side. Analyzing the production unit B, Ansari presents the best perfor­
mance considering not only the absolute average error (12.3%), but also
4.2. Clustering evaluation the number of successful convergences (282 from 285) and standard
deviation (16.4%). Xiao and Ansari models present similar results in all
Models may perform differently depending on the operational pa­ three metrics. It is interesting to highlight that the combination of Beggs
rameters or flow characteristics. A clustering evaluation is presented and Orkiszewski models results in a great improvement compared to the
here to study these effects. The cluster is divided by production unit. use of single Orkiszewski model. Also, the combination of Beggs and
Table 6 shows the model performances for the flowline segments of the Hagedorn models performs worse than both individual models, which is
two production units. It needs to be mentioned that the flowline seg­ not expected. More data may be required to fully verify the observations
ments are combinations of the near-horizontal seabed pipelines and of this evaluation. The differences between both platforms’ evaluations

Fig. 6. Production unit clustering boxplot for wellbore section.

7
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Fig. 7. Production parameters analysis.

can be due to numerous reasons. Some overall differences between the models perform similarly. The standard deviations in the wellbore
platforms’ data include: wells from platform A have longer MD and segment are notably small, showing less outliers. At the production unit
smaller diameter at the flowline section; the GOR from platform B is B, Gray presents the best performance considering all the metrics.
larger, but gas lift injection rates are larger for platform A; platform A
encompasses a longer time of production, which results in increasing
4.3. Parametric evaluation
WC trends, while platform B only has lower values of WC; platform A
wells have higher liquid flow rates.
The objective of this section is to find patterns in the models’ per­
Table 7 shows the same clustering evaluation in the production unit
formances with regards to flow parameters. The evaluation is divided
level for the wellbore segment. The wellbore segments contain mainly
into four parts: production parameters, well geometry, flow pattern
near-vertical tubulars with gravity dominated pressure losses. The re­
prediction, and frictional/gravitational pressure drop prediction. The
sults differ from the previous results. At the production unit A, Ansari
first two parts are related to inputs; while the last two parts are linked to
and Beggs perform similarly, with absolute average percent errors of
the calculation outputs. Although this analysis was performed for all
10.7% and similar other metrics. The boxplot in Fig. 6 also shows how
models, only the results of Ansari, Beggs, and Gray models are shown.

8
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Fig. 8. Well geometry analysis.

The reason was that these models showed better performances as pre­ parameters to perform the analysis. It must be emphasized that, some
sented earlier. other flow variables are not investigated in this work. For instance, the
oil API gravity and the tubing diameter play big roles in pressure drop
4.3.1. Production parameters calculations. However, the used datasets do not present significant
Production parameters include the flow rates and fluid properties changes in these parameters to make comparisons meaningful. The WC
varying over time. They are obtained from the well test data described is selected to analyze the effects of produced water quantity. The gas
earlier. Water cut, liquid flow rate, and gas flow rate are the selected flow rate includes the GOR and the injected gas lift rate, corresponding

Fig. 9. Flow pattern prediction analysis.

9
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Fig. 10. Pressure drop components analysis.

to the full amount of gas flowing in the tubing. The liquid flow rate is length. This length is zero for the case of a fully vertical well, and simply
also evaluated to study the model performances at varying well shows a measure of how deviated the well is.
productivities. As observed, the horizontal wellbore length is directly proportional
Fig. 7 shows the errors of Ansari, Beggs and Gray models with respect to the errors for Ansari, and Beggs models. In Ansari and Beggs models,
to WC, gas rate and liquid rate. Each data point shows the percent bigger horizontal lengths result in increasing error trends. This shows
relative error (er) of the model for a specific well and time compared to that the model performances are not as good for horizontal wells with
the conducted production test. Results from flowline and wellbore seg­ long horizontals. At Gray’s model plot, this correlation is not present.
ments are all shown and distinguished by color, blue for flowline data
and orange for wellbore data. The data scatter is noticeable in all the 4.3.3. Flow pattern prediction
plots with no clear trends. However, high errors associated with over­ Flow pattern prediction is a crucial step in the pressure drop calcu­
prediction are encountered mostly in cases with low liquid flow rates lation process. All mechanistic models and most of the empirical cor­
(less than 1000 m3/day) and/or low gas flow rates (less than 300,000 relations use flow pattern distinction as the first step before the pressure
m3/day) for Ansari and Beggs models. Low flow rates do not always drop fractions are computed. The flow pattern decision changes the
induce high errors, but high errors occur more often in this group. This pressure drop calculation procedure. To study the flow pattern predic­
can be interpreted as higher flow rates resulting in more stable outputs. tion repercussions for each model, the main flow pattern that was pre­
Similar behavior is seen for the Gray model, but with model under­ dicted to exist over the majority of the wellbore or flowline length was
prediction. The highest errors associated with underprediction for Gray outlined. This flow pattern was plotted with respect to the percent
are encountered in low liquid rates and/or low gas rates. relative error of the model. This means that, more than one flow pattern
may be encountered over the length of the segment, but only the pre­
4.3.2. Well geometry dominant flow pattern is labeled to represent the data point in this
The geometry of a pipe or tubing plays an important role in the flow analysis.
mechanism. Pipe inclination affects the flow patterns, and consequently, Fig. 9 shows the analysis for Beggs and Ansari models for wellbore
the pressure drop calculation. Empirical multiphase flow correlations and flowline segments. Gray’s model does not distinguish flow patterns,
are commonly developed based on a specific pipe orientation (hori­ so it is not considered in this analysis. Beggs model reveals an interesting
zontal or vertical) and adapted for a range of angles. For a field scale observation, as most of the data with errors higher than 100% show
tubing however, the angle varies frequently over the line, which makes transition flow patterns. Particularly for the wellbore segment, it can be
the investigation of its effects challenging. In order to study the geom­ affirmed that all of the data derived from this flow pattern are over­
etry effects, the pressure drop prediction errors were plotted against the predicted. The transition flow pattern in Beggs models corresponds to a
vertical and the horizontal lengths of the wellbore segment, the hori­ transition between stratified and slug flow patterns. This is a range that
zontal lengths plots are shown in Fig. 8. No trends were observed for the has been poorly understood and is being studied more extensively in the
vertical wellbore lengths and neither for the flowline segment. The recent literature. On the other hand, a more detailed analysis is needed
vertical wellbore length shows the true vertical depth (TVD) of the well. to derive conclusions for Ansari’s model.
The horizontal wellbore length is the horizontal projection of the total

10
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

4.3.4. Frictional and gravitational pressure drop Acknowledgments

The pressure gradient of flow in a tubular is composed of three terms,
hydrostatic or gravitational, frictional and accelerational. Considering This research was supported in part by Conselho Nacional de
the velocities encountered in oil and gas production lines, the accel­ Desenvolvimento Científico e Tecnológico - Brasil (CNPq) and Coor­
erational term is commonly negligible. The other two terms are calcu­ denação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil
lated based on the methodology of the proposed mechanistic or (CAPES).
empirical models and combined to provide the flow’s pressure gradient.
In vertical tubulars, the gravitational component is dominant, while the References
horizontal parts are frictional dominated. Even though the production
line structure is a mix of varying inclinations, the dominance of the Al-Safran, E., Brill, J.P., 2017. Applied Multiphase Flow in Pipes and Flow Assurance.
SPE.
gravitational term in the wellbore segment is still noticeable. The fric­ Ansari, A.M., Sylvester, N.D., Sarica, C., Shoham, O., Brill, J.P., 1994. A comprehensive
tional component becomes more significant in the flowline segment. mechanistic model for upward two-phase flow in wellbores. SPE Prod. Facil. 9,
Fig. 10 shows a comparison of the gravitational and frictional pres­ 143–152. https://doi.org/10.2118/20630-PA.
Aziz, K., Govier, G.W., Fogarasi, M., 1972. Pressure drop in wells producing oil and gas.
sure drop values for the wellbore and flowline segments, using Ansari, J. Can. Pet. Technol. 11, 38–48. https://doi.org/10.2118/72-03-04.
Beggs and Gray models. The results are plotted with respect to the Barnea, D., 1987. A unified model for predicting transitions for the whole pipe
relative percent error in model prediction. The wellbore segment does inclinations. Int. J. Multiphas. Flow 13, 1–12.
Beggs, D.H., Brill, J.P., 1973. A study of two-phase flow in inclined pipes. J. Petrol.
not show a clear correlation between the error and the two pressure drop Technol. https://doi.org/10.2118/4007-PA.
components. However, it is interesting to observe the smaller values of Beggs, H.D., Robinson, J.R., 1975. Estimating the viscosity of crude oil systems. JPT (J.
frictional component compared to the gravitational component. In the Pharm. Technol.) 1140–1141.
Bendiksen, K.H., Maines, D., Moe, R., Nuland, S., 1991. The dynamic two-fluid model
flowline segment, the prediction errors increase with increases in fric­
OLGA: theory and application. SPE Prod. Oper. 6, 171–180.
tional and gravitational losses. Although this correlation is positive for Bikmukhametov, T., Jäschke, J., 2020. First principles and machine learning virtual flow
the three models, each one has its own peculiarity. Ansari and Beggs metering: a literature Review. J. Petrol. Sci. Eng. 184, 106487. https://doi.org/
present larger errors for larger pressure losses. For Beggs model, this 10.1016/j.petrol.2019.106487.
Brill, J.P., Mukherjee, H., 1999. Multiphase Flow in Wells, first ed. Society of Petroleum
correlation is stronger with larger than 100% errors for tests with hy­ Engineers.
drostatic pressure losses above 10,000 kPa. On the other hand, Gray Chaves, G.S., Monteiro, D.D., Duque, M.C., Ferreira Filho, V.M., Baioco, J.S., Vieira, B.F.,
underpredicts the pressure drop at cases with lower pressure drops, but 2020a. Short-term production optimization under water-cut uncertainty. SPE J.
1–21.
the accuracy increases greatly at higher pressure drops. Chaves, G.S., Monteiro, D.D., Jardim, T.S., Ferreira Filho, V.M., Baioco, J.S., 2020b.
Selma-pmot : a production management optimization tool coupled to multiphase
5. Conclusions flow simulator. In: Rio Oil & Gas Expo and Conference 2020. https://doi.org/
10.48072/2525-7579.rog.2020.440 ©.
Chew, J.-N., Connally, C.A., 1959. A viscosity correlation for gas-saturated crude oils.
Data from two offshore platforms over several years of production Trans. AIME 216, 23–25. https://doi.org/10.2118/1092-g.
were used to evaluate the performances of the most commonly used Dukler, A.E., Wicks, M., Cleveland, R.G., 1964. Frictional pressure drop in two-phase
flow: B. An approach through similarity analysis. AIChE J. 10, 44–51. https://doi.
multiphase models. Overall, Ansari’s model presented the best perfor­ org/10.1002/aic.690100118.
mance with 13.8% absolute average percent error for the flowline Duns, H., Ros, N.C.J., 1963. Vertical flow of gas and liquid mixtures in wells. 6th World
segment and 14.1% error for the wellbore segment. Looking closer, Gray Pet. Congr 451–465.
Glaso, O., 1980. Generalized pressure-volume-temperature correlations. J. Petrol.
can also generate good results for most but not all of the tested well­
Technol. 32, 785–795. https://doi.org/10.2118/8016-pa.
bores. From the production parameters evaluation, higher flow rate tests Gray, H., 1978. Vertical flow correlation in gas wells. User’s man. In: API 14B Surf.
resulted in more stable outputs. Control. Subsurf. Saf. Valve Sizing Comput. Program, second ed. Appendix B) June.
A well geometry analysis was conducted to analyze the effects of well Griffith, P., Wallis, G.B., 1961. Two-Phase slug flow. J. Heat Tran. 83, 307–318. https://
doi.org/10.1115/1.3682268.
deviation on model performances. Ansari and Beggs’s models showed Hagedorn, A., Brown, K., 1965. Experimental study of pressure gradients occurring
increasing errors for wells with larger horizontal wellbore lengths. Also, during continuous two-phase flow in small-diameter vertical conduits. J. Petrol.
Beggs model showed larger errors when the transition flow pattern was Technol. 17 https://doi.org/10.2118/940-PA.
Hasan, A.R., Kabir, C.S., 1988. Study of multiphase flow behavior in vertical wells. SPE
predicted. Finally, both Ansari and Beggs models showed larger over­ Prod. Eng. 3, 263–272. https://doi.org/10.2118/15138-pa.
prediction errors for tests with higher pressure losses in the flowline Kareem, L.A., 2014. Z Factor: implicit correlation, convergence problem and pseudo-
segment, while Gray showed smaller errors for these cases. The results reduced compressibility. In: 38th Niger. Annu. Int. Conf. Exhib. NAICE 2014 -
Africa’s Energy Corridor Oppor. Oil Gas Value Maximization through Integr, vol. 1.
represent how the laboratory models respond on field-scale application Glob. Approach, pp. 324–344. https://doi.org/10.2118/172373-ms.
that encompass a complex fluid composition combined with inclination Kartoatmodjo, R.S.T., Schmidt, Z., 1991. New correlations for crude oil physical
angles oscillating along the flow. These conclusions are valuable to the properties. In: Soc. Pet. Eng. AIME. SPE.
Labedi, R., 1992. Improved correlations for predicting the viscosity of light crudes.
industry, and can be used to select the best models given the production J. Petrol. Sci. Eng. 8, 221–234. https://doi.org/10.1016/0920-4105(92)90035-Y.
system’s flow parameters. Lee, A.L., Gonzalez, M.H., Eakin, B.E., 1966. The viscosity of natural gases. J. Petrol.
Technol. 18, 997–1000. https://doi.org/10.2118/1340-pa.
Lunde, G.G., Rudrum, G., Angelo, P., Holmas, K., Setyadi, G.R., 2013. Ormen lange flow
Credit author statement
assurance system (FAS) - online flow assurance monitoring and advice. Offshore
Technol. Conf. https://doi.org/10.4043/24297-ms.
Gabriela Souza Chaves: Conceptualization, Methodology, Soft­ McCain, W.D., 1991. Reservoir-fluid property correlations-state of the art. SPE Reservoir
Eng. 6, 266–272. https://doi.org/10.2118/18571-pa.
ware, Formal analysis, Investigation, Writing - Original Draft, Visuali­
McCain, W.D., 1990. The Properties of Petroleum Fluids, second ed. PennWell Books.
zation. Hamidreza Karami: Conceptualization, Methodology, Mukherjee, H., Brill, J.P., 1985. Pressure drop correlations for inclined two-phase flow.
Validation, Writing - Review & Editing, Supervision, Project adminis­ J. Energy Resour. Technol. 107, 549–554. https://doi.org/10.1115/1.3231233.
tration. Virgilio Jose Martins Ferreira Filho: Conceptualization, Orkiszewski, J., 1967. Predicting two-phase pressure drops in vertical pipe. J. Petrol.
Technol. 19, 829–838. https://doi.org/10.2118/1546-pa.
Project administration, Funding acquisition. Bruno Ferreira Vieira: Petrosky, G.E., Farshad, F.F., 1998. Pressure-volume-temperature correlations for gulf of
Funding acquisition. Mexico crude oils. SPE Reservoir Eval. Eng. 249–258. https://doi.org/10.2118/
29468-ms, 1998.
Petrosky, G.E., Farshad, F.F., 1995. Viscosity correlations for gulf of Mexico crude oils.
Declaration of competing interest In: Soc. Pet. Eng. - SPE Prod. Oper. Symp. 1995, POS 1995 1995-April, pp. 249–258.
https://doi.org/10.2118/29468-ms.
The authors declare that they have no known competing financial Poulisse, H., Van Overschee, P., Briers, J., Moncur, C., Goh, K.C., 2006. Continuous well
production flow monitoring and surveillance. SPE Intell. Energy Conf. Exhib. 2,
interests or personal relationships that could have appeared to influence 514–519. https://doi.org/10.2523/99963-ms, 2006.
the work reported in this paper.

11
G.S. Chaves et al. Journal of Petroleum Science and Engineering 208 (2022) 109762

Rashid, K., 2010. Optimal allocation procedure for gas-lift optimization. Ind. Eng. Chem. Taitel, Y., Dukler, A.E., 1976. A model for predicting flow regime transitions in
Res. 49, 2286–2294. https://doi.org/10.1021/ie900867r. horizontal and near horizontal gas-liquid flow. AIChE J. 22, 47–55. https://doi.org/
Sagar, R., Doty, D.R., Schmidt, Z., 1991. Predicting temperature profiles in a flowing 10.1002/aic.690220105.
well. SPE Prod. Eng. https://doi.org/10.2118/19702-pa. Vazquez, M., Beggs, H.D., 1977. Correlations for fluid physical property prediction. In:
Shippen, M., Bailey, W.J., 2012. Steady-state multiphase flow - past, present, and future, Proc. - SPE Annu. Tech. Conf. Exhib. 1977-Octob. https://doi.org/10.2118/6719-pa.
with a perspective on flow assurance. Energy Fuels. https://doi.org/10.1021/ Waltrich, P.J., Capovilla, M.S., Lee, W., De Sousa, P.C., Zulqarnain, M., Hughes, R.,
ef300301s. Tyagi, M., Williams, W., Kam, S., Archer, A., Singh, J., Nguyen, H., Duhon, J.,
Shoham, O., 2006. Mechanistic modeling of gas-liquid two-phase flow in pipes. Soc. Pet. Griffith, C., 2019. Experimental evaluation of wellbore flow models applied to worst-
Eng. case-discharge calculations for oil wells. SPE Drill. Complet. 34, 315–333. https://
Standing, M.B., 1981. Volumetric and Phase Behavior of Oil Field Hydrocarbo Systems. doi.org/10.2118/184444-PA.
Soc. Pet. Eng. AIME. Xiao, J.J., Shoham, O., Brill, J.P., 1990. Comprehensive mechanistic model for two-phase
Standing, M.B., 1947. A pressure-volume-temperature correlation for mixtures of flow in pipelines. In: Proc. - SPE Annu. Tech. Conf. Exhib. Pi, pp. 167–180.
California oils and gases. In: Drilling and Production Practice. American Petroleum Zhang, H.Q., Qian, W., Sarica, C., Brill, J.P., 2003. A unified mechanistic model for slug
Institute, New York, New York, pp. 275–287. liquid holdup and transition between slug and dispersed bubble flows. Int. J.
Taitel, Y., Bornea, D., Dukler, A.E., 1980. Modelling flow pattern transitions for steady Multiphas. Flow 29, 97–107. https://doi.org/10.1016/S0301-9322(02)00111-8.
upward gas-liquid flow in vertical tubes. AIChE J. 26, 345–354. https://doi.org/
10.1002/aic.690260304.

12