Journal of Petroleum Science and Engineering 208 (2022) 109772

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

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

Virtual flow metering of production flow rates of individual wells in oil and
gas platforms through data reconciliation
Gabriel M.P. Andrade a ,∗, Diego Q.F. de Menezes a , Rafael M. Soares a , Tiago S.M. Lemos a ,
Alex F. Teixeira b , Leonardo D. Ribeiro b , Bruno F. Vieira b , José Carlos Pinto a
a
Programa de Engenharia Química, COPPE, Universidade Federal do Rio de Janeiro, Rio de Janeiro, RJ, CEP 21941-972, Brazil
b Centro de Pesquisas Leopoldo Américo Miguez de Mello - CENPES, Petrobras - Petróleo Brasileiro SA, Rio de Janeiro, RJ, CEP 21941-915, Brazil

ARTICLE INFO ABSTRACT

Keywords: In the present work, virtual flow metering (VFM) tools are developed to provide estimates for the individual
Virtual flow metering flow rate measurements of the wells operated simultaneously in typical oil and gas platforms. Data reconcil-
Soft sensor iation (DR) procedures are used to calibrate phenomenological models using available measured profiles of
Data reconciliation
well pressure and temperature, and overall flow rates of gas, oil and water. In order to evaluate the relative
Oil and gas production
importance of key hypotheses, three different phenomenological models were used to describe the platform
operation: (i) a simple single phase fluid flow model, to evaluate whether the DR task can be successfully
accomplished with the available data; (ii) a simple multiphase fluid flow model, in order to evaluate whether
the explicit consideration of the multiphase nature of the problem can affect the obtained DR results; and
(iii) a detailed phenomenological multiphase flow model, to represent accurately the behavior of an oil and
gas production platform. Obtained results show that the proposed DR scheme can be indeed implemented
successfully in the form of a VFM sensor to assign online and in real time the individual production of wells
operated simultaneously in typical oil and gas platforms. The DR-based VFM was able to reconcile total oil
and total water flow rates with average relative deviations of 0.87% and 17% respectively and maximum
deviation of 2.3% for oil flow rates. Besides, the DR-based VFM tool provided insights about the system
behavior and sensoring management, being able to indicate the oil lines and sensors that were subject to
the highest variability.

1. Introduction the incoming flows produced by each well cannot be performed easily,
given the transient and complex multiphase nature of these streams.
The oil and gas production activities are usually grouped into As a matter of fact, characterization of multiphase systems is difficult
three main sectors: upstream, midstream and downstream. While the because properties are heterogeneous and change with the spatial
downstream sector is responsible mainly for refining the product, the coordinates, so that even the appropriate measurements of average
upstream and midstream sectors involve the extraction and transporta- flow properties can be difficult, unreliable and very expensive (Fal-
tion activities, respectively. In particular, it must be considered that cone et al., 2002; Thorn et al., 2012). Nonetheless, it is essential to
the oil and gas extraction is generally performed simultaneously in characterize how much oil and gas a single well produces for several
several different wells, from which the transporting pipelines con- reasons, including payment of royalties, optimization of oil and gas pro-
verge to a common treatment station, equipped with different unit
duction and control of well aging, among others (Bikmukhametov and
operations including three-phase separators, flash drums, cyclones,
Jäschke, 2020; Falcone et al., 2002). Consequently, there are increasing
hydrocyclones, among others. After the appropriate gas and oil pro-
demands for developing individual production metering methods in the
cessing, the product output comprises at least three different phases:
oil and gas industry (Bikmukhametov and Jäschke, 2020).
oil, gas and water (Bikmukhametov and Jäschke, 2020).
Traditional propositions for measuring the individual income flows
Commonly, the properties of each product stream are measured
still find widespread use in the oil and gas industry nowadays, almost
independently, especially because single-phased systems can be charac-
terized more accurately and precisely than multiphase systems (Falcone always based on the so-called ‘‘well tests’’, when a stream produced by
et al., 2002; Thorn et al., 2012). On the other hand, characterization of a single well is deviated to an independent separation line, allowing the

∗ Corresponding author.
E-mail address: [email protected] (G.M.P. Andrade).

https://doi.org/10.1016/j.petrol.2021.109772
Received 29 April 2021; Received in revised form 13 October 2021; Accepted 29 October 2021
Available online 30 November 2021
0920-4105/© 2021 Elsevier B.V. All rights reserved.
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 1. Virtual flow metering scheme.

characterization of the oil, gas and water product streams (Makogon, fact, the phenomenological model equations generally involve variables
2019). The wells are characterized alternately, according to a pro- that have never been measured but can be inferred with the information
grammed characterization plan. Although feasible, this strategy does provided by the other measured variables. Otherwise, as the empirical
not necessarily produce good results because the dynamics of the flow approach is intrinsically attached to the process data used to adjust
can be significantly disturbed by the operation (Bikmukhametov and the model, its use can be unfeasible when the process data are scarce
Jäschke, 2020; Falcone et al., 2002; Thorn et al., 2012). Moreover, and when the variable of interest is nor measured (Bikmukhametov and
it is necessary to wait for the whole testing separation line to reach Jäschke, 2020; Makogon, 2019).
the steady state condition, which can be time consuming and subject It must also be considered that process data are obtained with
to time constraints associated with the necessity to characterize all sensors that are subject to systematic or spurious measurement errors,
wells in a reasonable time interval (Idsøo et al., 2014; Makogon, so that proper statistical treatment of acquired data can be fundamental
2019). It is certain that multiphase sensors can be coupled to the to avoid poor fitting of the model and consequently wrong model
production pipelines, although the operation costs can be excessive and predictions (Prata et al., 2009). de Menezes et al. (2020) discussed
the instrument calibration frequency will probably be very high in this important procedures for the best performance of monitoring based on
case (Makogon, 2019; Patel et al., 2014). (DR), which includes steps of data pre-treatment and characterization;
Based on this scenario, the use of virtual flow sensors has been DR with the support of mass balance equations; gross errors detection
strongly encouraged in the oil and gas industry. In the analyzed prob- to identify failures and malfunctions; and observation of unmeasured
lem, virtual flow metering (VFM) consists in estimating individual variables, through a virtual sensor. The data characterization step is
production flow rates based on mathematical models that can be ad- fundamental to understand the dynamic nature of the data and the
justed using the available overall production data available as described process, assisting in the best selection of the model that supports the
in Fig. 1 (Rasmussen, 2004; Toskey, 2012). Measured data commonly
DR steps, systematic deviation detection, and virtual sensor. However,
used to provide information about the well are temperature and pres-
data treatment and model fitting can be performed simultaneously, in
sure profiles, with sensors normally placed at the (PDG), X-tree (TPT)
the context of Data Reconciliation and Parameter Estimation (DRPE)
and downstream and upstream positions of the choke valve (Carbone,
procedures (Prata et al., 2010). In short, DR consists in ‘‘correcting ’’
2007; Monks et al., 2019). Process data commonly used to provide
the measured process data in order to maximize the likelihood of
information about the production process includes flow rates of gas,
obtaining the available measurements and simultaneously satisfy the
oil and water streams, with sensors placed at the discharge of the
model constraints (Prata et al., 2009). Despite the importance of DR for
separation process (Bikmukhametov and Jäschke, 2020; Seman et al.,
on-line real-time applications, DR procedures have been rarely applied
2021).
in oil and gas production facilities and even less applied to build
Two different approaches have been frequently used to build math-
VFM sensors for characterization of individual well production (Câmara
ematical models to perform VFM, based either on phenomenological
or empirical modeling approaches. Phenomenological models are usu- et al., 2017; de Menezes et al., 2020; Soares et al., 2011).
ally more complex, attempting to describe the process dynamics with In order to provide estimates of the individual well production,
equations that present physical meaning, normally based on mass and a traditional approach consists in fitting a process model to process
energy balances and flow equations, in order to simulate the observed pressure, temperature and flow rate data acquired during the ‘‘well
phenomena as accurately as possible and to provide more reliable test ’’ and afterwards using the fitted model to evaluate the well per-
interpolations and extrapolations of process data. On the other hand, formance. Model fitting can be performed with the help of standard
empirical models can also be used to fit available process data, relating parameter estimation procedures usually through manipulation of head
input variables, like measured temperatures and pressures, with out- loss coefficients or through definition of a correction factor (or bias)
put variables, like measured flow rates (Bikmukhametov and Jäschke, that is used to correct the model predictions (Udofia et al., 2012). Some
2020). Empirical data-driven or machine learning based models can important problems can be associated with this traditional approach: (i)
often be built more easily, using simpler numerical procedures and it disregards the effects of the combined operation, as the many wells
presenting faster computational performances, which favor the devel- operated at the platform can affect the behavior of each other; (ii) it
opment and implementation of real-time applications, although also assumes that the conditions of the test are similar to the real operation
providing less reliable extrapolations of the observed data (Hastie et al., conditions, although process equipment and pipelines are different; (iii)
2009). it assumes that the flow conditions remain constant between successive
It is important to observe that phenomenological modeling allows tests, which is mostly untrue and frequently denied by the successive
the VFM to be extended beyond the intervals of measured variable val- tests.
ues used to estimate the model, because obtained models are intrinsi- An alternative way to assign the individual well production consists
cally related to the phenomena, not to the data set itself. As a matter of of using DR procedures to reconcile the available profiles of well

2
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

temperature and pressure, and overall flow rates of oil, gas and water 2.1.2. Simplified multiphase model
within the framework of a process model developed to describe the The simplified multiphase model is an obvious extension of the
streams of each phase leaving the reservoir. In this case, available mea- simplified single-phase model and was used to evaluate whether the
surements must be ‘‘corrected’’ and model parameters must be adjusted explicit consideration of multiple phases can affect the quality of the
·
to maximize the level of adherence of the model predictions to process obtained DR results. Assuming that 𝑀 𝑡,𝑘 is measured the total mass flow
data and simultaneously satisfy the model constraints. Assuming that ·
rate of phase 𝑘; 𝑀 𝑖,𝑘 is the individual mass flow rate of phase 𝑘 in well
this task can be performed successfully, then the product flow rates of
𝑖 (not measured); 𝛥𝑃𝑖,𝑗 is the measured pressure drop of segment 𝑗 in
individual wells can be obtained as surrogate variables, once they are
well 𝑖; 𝐾𝑖,𝑗 is the pressure drop constant of segment 𝑗 in well 𝑖; 𝑤𝑘,𝑖 and
predicted by the available model.
𝜌𝑘,𝑖 are, the mass fraction and density of phase 𝑘 in well 𝑖, respectively;
Based on the previous paragraphs, VFM tools are built here for
with 𝑖 = 1, … , 𝑁𝑃 , 𝑗 = {𝑃 𝐷𝐺, 𝑇 𝑃 𝑇 , 𝑐ℎ𝑜𝑘𝑒} and 𝑘 = {𝑤𝑎𝑡𝑒𝑟, 𝑜𝑖𝑙, 𝑔𝑎𝑠},
the first time to provide the production of the many individual wells
then it becomes possible to write:
that are operated simultaneously in oil and gas platforms, using a phe-
nomenological flow model as a reference and available measurements · ∑
𝑁𝑃 ·
of the well pressure profiles, well temperature profiles and overall 𝑀 𝑡,𝑘 = 𝑀 𝑖,𝑘 (3)
𝑖=1
product flow rates of gas, oil and water streams to calibrate the model [ ]
with help of DR procedures.
· 2 ∑ 𝑤𝑘,𝑖
𝛥𝑃𝑖,𝑗 = 𝐾𝑖,𝑗 𝑀 𝑖,𝑘 (4)
In order to evaluate the relative importance of the many involved 𝑘
𝜌𝑘,𝑖
phenomena, three different phenomenological models are used here to ∑
describe the platform operation: (i) a simple single phase fluid flow 𝑤𝑘,𝑖 = 1 (5)
𝑘
model is used to evaluate whether the DR task can be successfully
accomplished with the available data; (ii) a simple multiphase fluid As in the previous case, it is assumed that the head loss constants
flow model is used to evaluate whether the explicit consideration of the 𝐾𝑖,𝑗 are known. Besides, it is also assumed that the densities of the
multiphase nature of the problem can affect the obtained DR results; phases and the mass fractions of the distinct phases are also known,
and (iii) a detailed phenomenological multiphase flow model is used which can in many cases be achieved through techniques of thermo-
to represent the actual behavior of an oil and gas production platform. dynamic characterization of the phase properties, characterization of
Besides, different combinations of available process data were investi- the product properties or estimation during the DR procedure. The
gated to characterize the minimum information required for successful multi phase nature of the analyzed flow is considered explicitly in the
implementation of the proposed DR scheme. Finally, it is shown that formulation in order to analyze whether the multi phase nature can
the proposed DR scheme can be implemented successfully in the form affect the obtained DR results. For example, in this case the number
of a VFM sensor to assign the production of the many individual wells of measured outputs is increased and Eq. (4) explicitly indicates that
that are operated simultaneously in typical oil and gas platforms on line the head loss constants depend significantly on the mass fractions and
and in real time, as originally devised. characteristics of the distinct phases.

2. Theoretical framework 2.1.3. Phenomenological multiphase model

The third analyzed model is a phenomenological steady-state mul-
In the present section, the different analyzed models are presented tiphase flow simulator called MARLIM II (Seman et al., 2020), which
and the real industrial oil and gas production process is described. is used currently to perform flow analyzes in commercial oil and
Moreover, the DR strategy is discussed highlighting the proposed VFM gas production facilities. MARLIM II, an in-house well simulator by
features and the characteristics of the adopted numerical procedure. (Petrobras), calculates the pressure drops along individual wells by di-
viding the flow system into segments that can be modeled individually
2.1. Modeling with different available well-known flow correlations that consider the
local thermodynamic states such as pressures, temperatures and com-
2.1.1. Simplified single-phase model positions; local volumetric flow rates and geometrical features of the
The simplified single-phase model was proposed to evaluate whether pipelines. MARLIM II can also handle the external feeding of gas and
the DR task could be accomplished successfully with the available data. liquid streams, as in the case of the continuous gas lift operation (Seman
The model comprises essentially a quadratic correlation between flow et al., 2020). As shown in the next sections, the precision of simulations
rate and pressure drop used to generate the pressure profiles in the performed with MARLIM II can be better than 1% for flow rates and
individual wells, which is parameterized by a head loss constant, as pressure values.
usual (White, 2011). Assuming that 𝑄𝑡 is the measured total product MARLIM II is used here as the reference phenomenological model
flow rate of the process, 𝑄𝑖 is the flow rate of the individual well 𝑖 (not for purposes of data reconciliation. MARLIM II is a proprietary com-
measured), 𝛥𝑃𝑖,𝑗 is the measured pressure drop in segment 𝑗 of well 𝑖 mercial software that provides information about oil flowrates, pressure
and 𝐾𝑖,𝑗 is the pressure drop constant for the segment 𝑗 of well 𝑖; with profiles and temperature profiles along the production lines, when the
𝑖 = 1, … , 𝑁𝑃 and 𝑗 = {𝑃 𝐷𝐺, 𝑇 𝑃 𝑇 , 𝑐ℎ𝑜𝑘𝑒}, then it is possible to write: geometrical characteristics of the pipelines, the rheological features of
the fluid and observed pressure drop between two measured pressure
∑
𝑁𝑃
𝑄𝑡 = 𝑄𝑖 (1) nodes (usually, the first one is placed at the gas lift injection point and
𝑖=1 the second one is placed just before the choke valve) are provided as
𝛥𝑃𝑖,𝑗 = 𝐾𝑖,𝑗 𝑄2𝑖 (2) model inputs, although other pressure nodes can certainly be consid-
ered. Then, the software solves mass, energy and momentum balance
It is assumed in this case that the head loss constants 𝐾𝑖,𝑗 are known equations to calculate individual flowrates, pressure and temperature
and remain constant during the operation. This set of parameters car- profiles as outputs, as described elsewhere (Ribeiro et al., 2009). How-
ries the geometrical features of the pipelines and the physical properties ever, some important features of the phenomenological model must be
of the flowing material and constitute the essential phenomenological highlighted:
model information. The multiphase nature of the analyzed flow is
neglected here in order to analyze whether the available information, • The geometrical features of the production pipelines can be pro-
i.e. pressure profiles in the individual wells and overall product flow vided in detail by the user, including the possibility to divide the
rate, is sufficient to assure the successful implementation of the DR production pipelines into segments with different characteristics.
scheme. Common structural features of each pipeline segment include

3
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

angle, length, internal diameter, roughness, conductivity, among 2.3. Data reconciliation and parameter estimation
others. These features are used to calculate local heat transfer co-
efficients and head loss coefficients with help of well-established The DRPE problem is similar to a typical parameter estimation prob-
multi-phase constitutive correlations available in the literature for lem or a DR problem. For a parameter estimation problem the objective
different flow regimes, pressures and temperatures; function must be minimized through manipulation of the adjustable
• Each pipeline segment can be described by different fluid flow model parameters, and the deviations between model predictions and
correlations, including the ones originally presented by Poettman measured variables must be weighted by variances of respective mea-
and Carpenter (1952), Orkiszewski (1967), Hagedorn and Brown sured variables. On the other hand, a DR problem the objective function
(1965), Duns and Ros (1963), among others; must be minimized through manipulation of the model input variable,
• In order to perform the calculations, some characteristic features and the deviations between adjusted variables and measured variables
of the reservoir must be specified, including the composition must be weighted by variances of respective variables. Must be in mind,
of the inlet well stream and the external temperature profile. when the degree of freedom is satisfied for the data set and model,
The compositions can be provided in different forms, including the solution of minimization problem is possible. Therefore, in a DRPE
definition of inlet RGO and BSW values with respective thermo- problem the input variables and parameters are manipulated jointly to
dynamic characteristics (which can be parameterized in terms of provide the point of minimum of the analyzed optimization problem,
the API degree), or definition of oil stream compositions as mix- while simultaneously satisfying the constraints given by the model
tures of lumps with different molar masses and thermodynamic equations (de Menezes, 2021; Prata, 2009). Usually, the objective func-
characteristics; tion is the maximum likelihood estimator derived from the Gaussian
• Gas and liquid viscosities are calculated locally with help of distribution, presenting the well-known quadratic form, although other
constitutive correlations, as functions of the local temperature, objective functions derived of the robust statistic can also be utilized, as
pressure and composition values; presented by de Menezes et al. (2021) in a review article of 48 robust
• The injection of external gas and liquid feed streams into different M-estimators.
positions of the well can be informed and used for computation Different approaches can be used to perform the solution of the
of flowrates, head losses and temperature profiles; DR problem, depending on how the optimization problem, the model,
• Valves can be placed at different pipeline segments and are and the objective function are structured (de Menezes, 2021; Prata
modeled as recommended by the respective manufacturers, with et al., 2010; Prata et al., 2009). In the present work, two methods
characteristic discharge factors computed as functions of the local were used: successive linearization (Knepper and Gorman, 1980) and
temperature, pressure, flowrates and compositions; the sequential optimization approach (Prata et al., 2010). The first
• Finally, an additional constitutive model (Thornhill-Craver (Craft method is based on the successive linearization of the original nonlinear
et al., 1962), TUALP (Bertovic, 1995) or WinEads (Winkler and set of equations, with the iterative solution of the obtained linear
Eads, 1991)) must be selected when the gas lift flowrate must be formulation of the original optimization problem. The second method
computed. comprises two steps: (i) first, the numerical algorithm solves the model
equations, using suitable guesses for the model inputs and adjustable
2.2. Structure of the production platform model parameters; then (ii) the objective function is evaluated and the
optimizer is used to provide new sets of guesses, in order to minimize
The production platform described here belongs to Petrobras and the objective function.
comprises fifteen wells that have been operating simultaneously at an
offshore platform of the Campos Basin for approximately one decade. 2.3.1. General DR formulation
This platform has two separation trains, containing several pieces of
The Maximum Likelihood formulation was introduced by Fisher,
equipment, such as cyclones, hydrocyclones and three-phase separa-
although some of the basic ideas had been previously discussed by
tors, and one separation test line, which is used periodically to evaluate
Daniel Bernoulli and Gauss, and is the procedure used most often to ob-
individual performances of each particular well. In order to implement
tain estimators for data regression problems in different (experimental,
the DR-based virtual sensor, stored production data were collected
laboratory or industrial) environments (de Menezes et al., 2021).
during a period of two months, between the months of October and
Initially, it is necessary to assume that the measured data is com-
November of 2019. Although no steady state detection was performed,
posed of a true value and an error, where 𝑧𝑖 = 𝑧̂ 𝑖 + 𝜀𝑖 , 𝑧𝑖 are the
the DR structure itself is able to produce insights about the process’
measured variables, 𝑧̂ 𝑖 are the reconciled variables (estimated true
current state, as discussed further on. After careful analysis of the
values), and 𝜀𝑖 is a white Gaussian noise error, related to the pop-
available instrumentation, the final data set used for reconciliation
ulation of sampled measures and distributed according with the 𝑓𝑖
purposes contained the following variables for each of the analyzed
probability function. It is also assumed that the random errors 𝜀𝑖 of
wells of the platform:
the instrumentation are not correlated with the reconciled (estimated
• Pressure and feed flow rate of the gas lift injection stream; true) values (𝑧̂ 𝑖 ), making the measurements values independent of the
• Pressure and temperature at the PDG; measuring errors. Thus, the possible effects of heteroscedasticity of
• Pressure and temperature at the TPT; measurements are neglected. Thereby, the joint probability functions
• Pressure and temperature prior to the choke valve; 𝑓 of the observed measures can be calculated as the product of the
• Pressure, temperature and flow rate of the product gas input and individual probability functions 𝑓𝑖 (𝜀𝑖 ). Thus, the maximum likelihood
output streams at the platform; estimate can be defined as the one that maximizes the probability
• Pressure, temperature and flow rate of the product oil output of observation of the experimental data 𝑓 , according to Eq. (6) (de
stream at the platform; Menezes et al., 2021; de Menezes, 2021).
• Pressure, temperature and flow rate of the product water output ∏
𝑛

stream at the platform; max 𝑓 = max 𝑓𝑖 (6)

𝒛̂ 𝒛̂
𝑖=1
• Overall features of the independent well production streams,
including (BSW) and (GOR), characterized during the well tests. Assuming that the probability distribution can be described by a
Gaussian distribution and the errors of the measured variables are
The structure of the production platform is represented schemati- uncorrelated, and therefore with null covariance (𝑽 = 𝑑𝑖𝑎𝑔[𝝈 2 ]), the
cally in Fig. 2. principle of maximum likelihood for the Normal distribution (Weighted

4
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 2. Schematic representation of the structure of the production platform.

Least Squares estimator — WLS) can be applied, as shown in Eqs. (7) The linearization of Eq. (15) in relation to the two groups of
to (9). variables (𝒛̂ and 𝒖) generates Eq. (17).
{ [ ]}
∏𝑛
1 1 (𝑧𝑖 − 𝑧̂ 𝑖 )
2
𝑨𝑧̂ 𝒛̂ + 𝑨𝑢 𝒖 = 𝒃 (17)
max √ exp − (7)
𝒛̂
𝑖=1 𝜎𝑖 2𝜋 2 𝜎𝑖2
{ [ ]} wherein
∑𝑛
1 (𝑧𝑖 − 𝑧̂ 𝑖 )
2 𝜕ℎ1 𝜕ℎ1 𝜕ℎ1
min − ln exp − (8) ⎡ ⋯ ⎤
𝜕 𝑧̂ 1 𝜕 𝑧̂ 2 𝜕 𝑧̂ 𝑚
𝒛̂
𝑖=1
2 𝜎2 ⎢ ⎥
𝑖
( 𝑖 𝑖) ⎢ ⎥
𝜕ℎ2 𝜕ℎ2 𝜕ℎ2
⋯
∑𝑛
1 (𝑧𝑖 − 𝑧̂ 𝑖 )
2
𝑨𝑧̂ = 𝛁𝒛̂ 𝒉 𝒛̂ , 𝒖 = ⎢
𝑖 𝑇 𝜕 𝑧̂ 1 𝜕 𝑧̂ 2 𝜕 𝑧̂ 𝑚 ⎥ (18)
min
2
(9) ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
𝒛̂
𝑖=1
2 𝜎 ⎢ ⎥
𝑖
⎢ 𝜕ℎ𝑘 𝜕ℎ𝑘
⋯
𝜕ℎ𝑘 ⎥
where 𝜎𝑖2 is the variance of the measurement errors of the measured ⎣ 𝜕 𝑧̂ 1 𝜕 𝑧̂ 2 𝜕 𝑧̂ 𝑚 ⎦
𝜕ℎ1 𝜕ℎ1 𝜕ℎ1
variable 𝑧𝑖 . Finally, the problem can be represented in the form: ⎡ ⋯ ⎤
𝜕𝑢1 𝜕𝑢2 𝜕𝑢𝑙
⎢ ⎥
1
( 𝑖 𝑖) ⎢ ⎥
𝜕ℎ2 𝜕ℎ2 𝜕ℎ2
̂ 𝑇 𝑽 −1 [𝒛 − 𝒛]
𝒛̂ = arg min [𝒛 − 𝒛] ̂ (10) ⋯
𝒛̂ 2 𝑨𝑢 = 𝛁𝒖 𝒉 𝒛̂ , 𝒖 = ⎢
𝑖 𝑇 𝜕𝑢1 𝜕𝑢2 𝜕𝑢𝑙 ⎥ (19)
subject to: ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ ⎥
⎢ 𝜕ℎ𝑘 𝜕ℎ𝑘
⋯
𝜕ℎ𝑘 ⎥
̂ 𝒖) = 𝟎
𝒉(𝒛, (11) ⎣ 𝜕𝑢1 𝜕𝑢2 𝜕𝑢𝑙 ⎦
( )
̂ 𝒖) ≥ 𝟎
𝒈(𝒛, (12) 𝒃𝑖 = 𝑨𝑧𝑖̂ 𝒛̂ 𝑖 + 𝑨𝑖𝑢 𝒖𝑖 − 𝒉 𝒛̂ 𝑖 , 𝒖𝑖 (20)
𝐿 𝑈
𝒛̂ ≤ 𝒛̂ ≤ 𝒛̂ (13) where 𝑘 is the number of model constraints, 𝑚 is the number of
𝒖𝐿 ≤ 𝒖 ≤ 𝒖𝑈 (14) measured variables, 𝑛 is the number of unmeasured variables and 𝑖 is
the 𝑖th iteration of the method.
where 𝒛̂ is the vector of the reconciled variables; 𝒛 is the vector of the In order to eliminate one of these matrices, the QR decomposition
measured variables; 𝑽 is the variance matrix of measurement errors; method must be used to uncouple the unmeasured variables from
𝒖 is the vector of the unmeasured variables; 𝒉 is the vector of linear measured variables, as described by Romagnoli and Sanchez (1999).
or nonlinear algebraic constraint equations; 𝒈 is the vector of the After algebraic manipulation of the linearized constraints, the final
inequalities of linear or nonlinear algebraic constraints; 𝒛̂ 𝑈 , 𝒛̂ 𝐿 are the solution can be given in the form of Eqs. (21) and (22).
upper and lower parameter vectors of the reconciled variables 𝒛; ̂ and
𝒖𝑈 , 𝒖𝐿 are the upper and lower parameter vectors of the unmeasured ( 𝑇 )𝑇 [( 𝑇 ) ( )𝑇 ]−1 [( 𝑇 ) ]
𝒛̂ = 𝒛 − 𝑽 𝑸𝑢2 𝑨𝑧̂ 𝑸𝑢2 𝑨𝑧̂ 𝑽 𝑸𝑇𝑢2 𝑨𝑧̂ 𝑸𝑢2 𝑨𝑧̂ 𝒛 − 𝑸𝑇𝑢2 𝒃
variables 𝒖.
(21)
2.3.2. Successive linearization
The successive linearization approach is commonly applied to DR
𝒖1 = 𝑹−1 𝑸𝑇 𝒃 − 𝑹−1
𝑢1 𝑢1
𝑸𝑇 𝑨 𝒛̂ − 𝑹−1
𝑢1 𝑢1 𝑧̂
𝑹 𝒖
𝑢1 𝑢2 2
(22)
problems where the constraint model is formed by bi-linear or non-
linear equations (Knepper and Gorman, 1980). Let us assume that any where
non-linear model is represented by 𝒉 (𝒘) and that its variables 𝒘 can
be divided into two groups of variables, measured (𝒛)
̂ and unmeasured 𝑨 𝑢 = 𝑸𝑢 𝑹 𝑢 (23)
(𝒖): 𝑨𝑢(𝑘×𝑙) 𝛱𝑢(𝑙×𝑙) = 𝑸𝑢(𝑘×𝑘) 𝑹𝑢(𝑘×𝑙) (24)
[ ]
𝒉 (𝒘) = 𝟎 (15) 𝑸𝑢(𝑘×𝑘) = 𝑸𝑢1(𝑘×𝑙) 𝑸𝑢2(𝑘×𝑘−𝑙) (25)
[ ] [ ]
𝒛̂ 𝑹𝑢1(𝑟×𝑟) 𝑹𝑢2(𝑟×𝑙−𝑟)
𝒘= (16) 𝑹𝑢(𝑘×𝑙) = (26)
𝒖 𝟎1(𝑘−𝑟×𝑟) 𝟎2(𝑘−𝑟×𝑙−𝑟)

5
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 3. Schematic representation of the sequential DR scheme.

The problem is reconcilable if 𝑟 = 𝑙 = 𝑟𝑎𝑛𝑘(𝑨𝑢 ) = 𝑟𝑎𝑛𝑘(𝑹𝑢1 ). The 2.3.4. Methodology with real data
procedure must be performed iteratively until attainment of a specified To apply DR-based VFM, the following sequential steps must be
tolerance. carried out: 1 - Data acquisition; 2 - Data pre-treatment; 3 - Data
characterization; 4 - Data reconciliation; 5 - Estimation of unmeasured
variables and 6 - Analysis of the estimated variables (de Menezes et al.,
2.3.3. Sequential optimization 2020).
The implemented sequential approach, also known as the feasible Data acquisition constitutes the first step of the proposed procedure,
path algorithm, solves the RD problem in two steps. At the internal when the available process information is collected with help of the
step, model equations are solved for suitable initial estimates of the available instrumentation and oil characterization procedures. Then,
input variables and adjustable parameters, using algorithms that can data pre-treatment must be performed prior to use, being essential to
provide numerical solutions for nonlinear equations, such as the stan- ensure the robustness and correct utilization of mathematical and statis-
dard Newton–Raphson procedure. At the external step, the objective tical procedures. For instance, text inputs should be replaced by actual
function is minimized with help of gradient-based optimization proce- numbers, gross measured values should be filtered, process constraint
dures, which can provide new sets of parameter estimates for a new violations should be corrected, among other important consistency
iteration, while convergence criteria are not satisfied. Therefore, the tests. Finally, data characterization must be performed to evaluate the
model represented by mass balances, energy balances, and empirical behavior of the measured variables (including computation of means,
constitutive equations is always solved at each numerical iteration, variances, and variable correlations, for instance) and the type of oper-
guaranteeing the calculation of feasible solutions for the optimization ation regime (stationary or transient). These preliminary analyzes can
problem (Ma et al., 2020; Câmara et al., 2017), as represented in be important to enhance the performances of model calculations and
of the optimization procedure, and to minimize the risks of obtaining
Fig. 3. Different optimization algorithms can be used to minimize
meaningless numerical results.
the objective function, although the most common ones are based on
The data reconciliation step should be ideally preceded by a formal
gradient descent functions that evaluate the directions that lead to the
Classification of Variables step, in order to characterize the observabil-
point of minimum to provide new guesses (Koziel and Yang, 2011).
ity of the available mathematical model and process measurements, as
On the other hand, stochastic algorithms can also be used to handle
the lack of degrees of freedom can prevent the correct use of DR-based
the optimization task, making use of statistical methods and extensive
schemes. When the system is observable, then the reconciliation step
sampling of the parameter space (Koziel and Yang, 2011).
can be used to perform the minimization of measurement deviations
The optimization process is executed continuously during the DR (between actual measurements and model outputs) in a statistically
procedure until a defined tolerance criterion is satisfied. According to coherent manner, also allowing the determination of unmeasured vari-
this approach, the model can be separated from the optimization algo- ables with help of the model. This way, the DR-based scheme behaves
rithm, which communicate to each other when the objective function as a soft sensor and allows the implementation of a digital twin,
is calculated; this way, the model constraints can always be satisfied. A which in the specific considered case is a Virtual Flow Measurement
possible disadvantage of the method is the high computational demand system. Finally, the estimated variables must be critically evaluated and
when complex phenomenological models are considered (Câmara et al., monitored by operators and engineers, as also performed with other
2017). standard process sensors.

6
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Table 1
‘‘True’’ values of head loss constants and individual flow rate measurements, and respective calculated pressure drops for the five-well DR
problem.
𝐾𝑃 𝐷𝐺 𝐾𝑇 𝑃 𝑇 𝐾𝑐ℎ𝑜𝑘𝑒 𝑄𝑖 𝛥𝑃𝑃 𝐷𝐺 𝛥𝑃𝑇 𝑃 𝑇 𝛥𝑃𝑐ℎ𝑜𝑘𝑒
Pa h Pa h Pa h km3
Well∕Unit
m3
(normal) m3
(normal) Nm3
(normal) h
(normal) kPa kPa kPa
1 0.72 0.69 0.67 70.19 3 551.58 3 417.62 3 312.82
2 0.43 0.83 0.04 23.38 234.66 455.99 22.99
3 0.73 0.49 0.55 66.37 3 230.62 2 157.35 2 404.45
4 0.61 0.55 0.75 62.54 2 402.14 2 154.51 2 946.81
5 0.69 0.59 0.34 24.51 412.64 352.87 202.43

2.4. Hardware and software Table 2

Variances of ‘‘measured’’ reconciled variables using the single-phase model.

The algorithms were written in Python with an integrated exe- 𝛥𝑃𝑃 𝐷𝐺 𝛥𝑃𝑇 𝑃 𝑇 𝛥𝑃𝑐ℎ𝑜𝑘𝑒 𝑄𝑇
km3
cutable phenomenological simulator (MARLIM II). All evaluations were Well∕Unit kPa kPa kPa h
(normal)
performed in a laptop with the following configurations: 1 28.974 32.815 38.511 –
2 0.123 0.632 0.002 –
• Processor Intel i7 8th Generation CPU @ 2.20 GHz hexa-core; 3 24.377 14.413 21.082 –
• 16 GB RAM; 4 15.907 15.823 34.813 –
• 128 GB SSD. 5 0.458 0.391 0.169 –
– – – – 0.051

3. Results

First the simplified single-phase model was used to perform the DR the on-line and real-time actual industrial application is supposed to
task using simulated corrupted data to evaluate whether the DR task be carried out iteratively, so that the previously obtained estimated
could be accomplished successfully with the available data. The DR values are expected to be available and to be close to the new estimated
task was solved considering the existence of 5 wells and 15 wells, in values, if the sampling frequency is not too low and is compatible with
order to evaluate the effect of the number of wells on the obtained nu- the process dynamics. The diagonal covariance matrix was defined after
merical performances. Then the simplified multiphase model was used characterization of the variance spectrum procedure (Feital and Pinto,
to evaluate whether the explicit consideration of multiple phases could 2015) of ‘‘measured’’ data, using the first elements of each spectrum as
affect the quality of the obtained DR results. Once more, the DR task references, as shown in Table 2.
was solved considering the existence of 5 wells and 15 wells, in order At this point, it must be emphasized that convergence problems
to evaluate the effect of the number of wells on the obtained numerical are possible in the proposed nonlinear iterative procedures due to
performances. Finally, the detailed phenomenological multiphase flow
occurrence of gross measurement errors, inappropriate estimation of
model was used to perform the DR task, using a real data set collected
measurement variances or use of poor initial guesses. Nevertheless,
from an existing commercial oil and gas platform.
convergence problems are not likely to occur during steady state op-
erations, as estimates obtained after one iteration constitute excellent
3.1. Simplified single-phase model
initial guesses for initialization of the next iterative step, as said previ-
ously. Besides, gross error measurements can be detected and removed
The simplified single-phase model was used to perform the DR
before the initialization of the iterative scheme. It must be observed,
task, using simulated corrupted data and the successive linearization
though, that problems related to flow assurance can cause the fast
method. In order to illustrate and discuss the particularities of the
and continuous drift of the measured data, removing the process from
proposed DR procedure, only the DR study performed with five wells is
steady state operation and indicating the inadequacy of the proposed
discussed here. Information about the DR study performed with fifteen
steady state reconciliation scheme. As a matter of fact, the occurrence
wells can be found in the supplementary material section. First, values
of convergence problems can alarm the onset of undesired flow patterns
of individual flow rates and head loss constants were randomly gener-
ated with Gaussian distributions with means and standard deviations and should not be neglected by the analyst.
defined as reported in the literature (flow rates km3 ∕h (normal)]: mean The obtained results showed that, in order to carry out the DR task,
= 50, std = 15; head loss constant [ Pa.h∕m3 (normal)]: mean = 0.5, std one must necessarily define the 𝐾𝑖,𝑗 head loss parameters beforehand.
= 0.1; where ‘normal’ represents the flow measurement taken under If any of the model parameters is not known, the problem becomes
normal conditions, that is, at a pressure of 1 atm and a temperature of unobservable, as indicated by the rank of the 𝑄 and 𝑅 matrices used
273.15 K (Bikmukhametov and Jäschke, 2020; White, 2011). Table 1 in Eqs. (22)–(23). This particular result is not affected by the number
shows the values used for numerical analyzes. Then, sampling windows of wells connected to the platform, indicating that the simultaneous
containing 500 points were generated for pressure drops and total estimation of adjustable model parameters during the DR task can lead
flow rates. The model constraints were used to generate the average to numerical difficulties in the analyzed data set scenario. This result is
sample values, which were contaminated with Gaussian noises with very important because it indicates that the fundamental characteristics
standard deviations of 5% and 2% of the average values, respectively. of the production network must be specified with the best possible
The standard deviation values were defined considering that pressure accuracy and that possible adjustment of network properties must be
sensors in oil and gas production platforms are often more subject to handled with care during the DR analyzes.
measurement noise than flow rate sensors, as flow rate measurements Nevertheless, as one can see in Figs. 4 and 5, the DR task could
are also used for regulatory purposes (Bikmukhametov and Jäschke, be performed successfully to evaluate the production flow rates for
2020). each individual well, based on the available pressure drop and total
Regarding the numerical optimization scheme, ‘‘measured’’ pressure flow rate measurements, when 5 wells were considered. Individual flow
drop values and individual ‘‘true’’ flow rate values were used as initial rates were properly estimated, as shown in Fig. 4, with fluctuation
guesses for numerical computations, considering the relative tolerance of estimated values around the unknown ‘‘true’’ values, as desired
of 10−8 as convergence criterion. Although during the real DR appli- and expected. The observed deviations between the simulated and
cation the individual ‘‘true’’ flow rate values are not known a priori, reconciled data were due to the random Gaussian noise applied to

7
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 4. Reconciled total flow rates and estimated individual flow rates for 5 wells using the simplified single-phase model.

the simulated data, as one might expect in real operations. Similar variables. First, initial average values were defined. The specific mass
behavior was observed when 15 wells were considered and can be of each phase, GOR, BSW, volumes of oil and water produced by each
found in the supplementary material. However, as one might expect, independent well and pressure drop values (𝛥𝑃𝑃 𝐷𝐺 , 𝛥𝑃𝑇 𝑃 𝑇 , 𝛥𝑃𝑐ℎ𝑜𝑘𝑒 )
the computational time increased when additional wells were added were defined based on common values reported for real applica-
to the DR structure. For example, in the 15-well DR problem, it took tions (Bikmukhametov and Jäschke, 2020; White, 2011). With these
1.90 s for the algorithm to converge and provide the results, which variables and using the simplified multiphase model as reference, it
was approximately two times longer than the 0.86 s needed for 5- was possible to calculate the volumetric flow rates of the gas stream,
well DR problem to converge. In both cases, however, the computation the mass flow rates of each independent phase, the total mass flow rate
times were extremely small, making possible the implementation of the and the mass fractions of each phase. All defined ‘‘true’’ variable values
numerical procedures on line and in real time. are displayed in Tables 3 and 4.
Subsequently, oil, water and gas mass fractions and PDG, TPT and
3.2. Simplified multiphase model choke head loss values were generated for each well by sampling
from Gaussian distributions centered at the previously reported average
The simplified multiphase model was used to perform the DR values and considering standard deviations of 5% in respect to the
task, using simulated corrupted data and the successive linearization averages. Finally, head loss constants were calculated with the model.
method. The dataset was created as described in the previous section; The final set of variables used to represent the process are presented in
however, this model requires the handling of additional number of Table 5.

8
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 5. Reconciled individual pressure drops for 5 wells using the simplified single-phase model.

Table 3 Table 4
Initial average values for process variables in the multiphase model. Initial average values of well features in the multiphase model.
Specific mass Mass flow Volumetric Mass fraction Well feature Value Unit
[kg∕m3 ] rate [kg∕h] flow rates
GOR 122 –
[m3 ∕h]
BSW 102 –
Oil 884.00 442 000.00 500.00 0.64 𝛥𝑃𝑃 𝐷𝐺 3 000 kPa
Gas 0.75 49 650.00 66 200.00 0.07 𝛥𝑃𝑇 𝑃 𝑇 5 000 kPa
Water 999.00 199 980.00 200.00 0.29 𝛥𝑃𝑐ℎ𝑜𝑘𝑒 1 400 kPa

As explained and performed in the previous section, ‘‘measured’’ of 10−8 was defined to characterize the convergence of the numeri-
cal procedure. The diagonal variance matrix of measured values was
pressure drop values and ‘‘true’’ individual flow rates values were used
characterized as described in the previous section and is presented in
as first guesses in the numerical procedure and the relative tolerance Table 6. Then, sampling windows containing 500 points were generated

9
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Table 5
‘‘True’’ mass fraction and head loss constant values for each analyzed well in the 5-well DR multiphase problem.
𝑤𝑔𝑎𝑠 𝑤𝑜𝑖𝑙 𝑤𝑤𝑎𝑡𝑒𝑟 𝐾𝑃 𝐷𝐺 𝐾𝑇 𝑃 𝑇 𝐾𝑐ℎ𝑜𝑘𝑒 𝑀𝑖 𝛥𝑃𝑃 𝐷𝐺 𝛥𝑃𝑇 𝑃 𝑇 𝛥𝑃𝑐ℎ𝑜𝑘𝑒
Pa h Pa h Pa h kg
Well∕Unit – – – m3
(normal) m3
(normal) m3
(normal) h
kPa kPa kPa
1 0.066 0.601 0.333 6.72 × 10−8 1.11 × 10−7 3.05 × 10−8 695 857 2 867.15 4 712.49 1 353.06
2 0.066 0.644 0.290 7.08 × 10−8 1.24 × 10−7 3.15 × 10−8 696 566 3 023.74 5 276.92 1 394.86
3 0.069 0.595 0.336 7.00 × 10−8 1.26 × 10−7 3.41 × 10−8 682 646 2 987.00 5 369.13 1 513.00
4 0.067 0.600 0.333 6.79 × 10−8 1.21 × 10−7 3.10 × 10−8 692 486 2 896.61 5 138.55 1 376.82
5 0.068 0.651 0.281 7.10 × 10−8 1.14 × 10−7 3.24 × 10−8 686 096 3 031.55 4 860.84 1 438.54

Table 6 3.3. Phenomenological multiphase flow model

Variances of ‘‘measured’’ reconciled variables using the multiphase model.
𝛥𝑃𝑃 𝐷𝐺 𝛥𝑃𝑇 𝑃 𝑇 𝛥𝑃𝑐ℎ𝑜𝑘𝑒 𝑤𝑔𝑎𝑠 𝑤𝑜𝑖𝑙 𝑀𝑇 Based on the previous DR analyzes, it must be emphasized that, in
kg 2
Well∕Unit (kPa)2 (kPa)2 (kPa)2 – – ( h
) order to carry out the appropriate estimation of the individual flow
1 919.49 2 284.19 216.66 2.10 × 10−7 2.11 × 10−5 – rates of the wells, the structure of the production platform must be
2 875.43 2 014.67 269.87 3.05 × 10−7 1.62 × 10−5 – carefully described and some critical properties of the flow must be
3 586.99 3 877.51 243.72 2.08 × 10−7 1.19 × 10−5 – defined, such as the overall phase compositions, represented by GOR
4 884.08 4 885.35 240.50 1.66 × 10−7 1.30 × 10−5 –
5 1 211.26 3 303.76 141.90 1.87 × 10−7 3.27 × 10−5 –
and BSW. Based on these preliminary results, the phenomenological
– – – – – – 2.87 × 108 multiphase flow model, available in MARLIM II, was used to perform
the virtual flow metering of individual flow rates of 15 wells during
the actual operation of a platform through data reconciliation and
using the sequential optimization approach. It must be emphasized
for pressure drops, total flow rates and oil and gas mass fractions. The that the sequential optimization approach is indicated when derivate
model constraints were used to generate the average sample values, calculations are complex and computationally costly, since it separates
which were contaminated with Gaussian noises with standard devia- the model evaluation from the optimization algorithm (Câmara et al.,
tions of 5% for pressure drop values, 2% for flow rate values and 3% 2017).
for mass fractions.
For evaluation of the numerical performance of the proposed DR-
Obtained results showed that, in order to carry out the DR task,
based VFM, the following data set was made available to describe the
one must necessarily define the 𝐾𝑖,𝑗 head loss parameters beforehand,
operation of the platform: gas lift flow rates for each independent well
as described in the previous section. Moreover, the structure of the DR
(𝑄𝑔𝑎𝑠,𝑖 ); the pressures of the independent gas lift streams (𝑃𝑔𝑎𝑠,𝑖 ); the
problem also requires the a priori definition of the oil and gas mass
choke pressures of the independent wells (𝑃𝑐ℎ𝑜𝑘𝑒,𝑖 ); the total water flow
fractions of the individual wells; otherwise, as observed previously, the
problem becomes unobservable, as indicated by the rank of the 𝑄 and rate (𝑄𝑊 ); and the total oil flow rate (𝑄𝑂 ). This structure is similar
𝑅 matrices used in Eqs. (21) and (22). Based on Eq. (4), this might to the ones used in the first two evaluations, when pressure data and
already be expected, because the phase properties and composition total flow measurements were used to perform the DR-based VFM
modify mainly the absolute value of the head loss parameters, but do monitoring. However, instead of using several pressure data points,
not change the structural properties of the model equations. Once more, MARLIM II requires only two boundary values as input variables to
this particular result is not affected by the number of wells connected to estimate the individual product flow rates and respective pressure drops
the platform. This result is very important because it indicates that the for each well. Preferably, the individual flow rates of the gas lift streams
fundamental characteristics of the production network and the overall (𝑄𝑔𝑎𝑠,𝑖 ) should be provided, but this information was not available for
gas and water contents of the product streams must be specified with some wells. Therefore, the pressure of the gas lift stream (𝑃𝑔𝑎𝑠,𝑖 ) was
the best possible accuracy and that possible adjustment of network used as input when 𝑄𝑔𝑎𝑠,𝑖 was not available, using the flow model for
properties and gas and water contents of the product streams must be computation of 𝑄𝑔𝑎𝑠,𝑖 . Moreover, the overall gas and water contents of
handled with care during the DR analyzes. streams produced by each independent well were assumed to be known,
Nevertheless, as one can see in Figs. 6–8, the DR task could be described in terms of BSW and GOR, as recommended in the previous
performed successfully to evaluate the production flow rates for each sections and characterized through well tests.
well, based on the available pressure drop and total flow rate measure- An actual operation window of two hours with a sampling rate of
ments, when 5 wells were considered. As also observed in the previous five minutes, resulting in 24 sampling points, was selected to calcu-
case, individual flow rates were estimated properly, as shown in Fig. 6, late the measurement variances of the reconciled variables through
with fluctuation of estimated values around the unknown ‘‘true’’ values, calculation of the variance spectra, as described before (Feital and
as desired and expected. As explained before, the observed deviations Pinto, 2015). The first elements of the analyzed variance spectra were
between the simulated and reconciled data were due to the random selected to represent the measurement signals; however, in the case of
Gaussian noise applied to the simulated data, as one might expect in
𝑃𝑐ℎ𝑜𝑘𝑒,𝑖 , this calculated variance was too high, apparently indicating the
real operations. Similar behavior was observed when 15 wells were
occurrence of measurement problems at the platform. For this reason,
considered and can be found in the supplementary material. These
the variance value used in this case to perform the DR calculations was
results reinforce the hypothesis that the DR procedure can indeed
arbitrarily made equal to 10% of the calculated variance; otherwise,
constitute a good alternative to perform the virtual flow metering
𝑃𝑐ℎ𝑜𝑘𝑒,𝑖 values would be mostly neglected during model fitting. The
of flow rates of the products obtained from independent oil wells.
However, as one might expect, the computational time increased again variances used to perform the DR-based VFM monitoring are shown in
when additional wells were added to the DR structure. As observed Table 7. Calculated variance spectra are presented in the supplementary
in the numerical study, the computational time increased from 1.6 to material.
4.4 s when the number of wells was increased from 5 to 15. When Given the practical necessity to provide reconciled data every two
compared to the numerical performance of the single-phase model, hours, data were reconciled every two hours, assuming constant flow
the computational times increased for the multi-phase model, because behavior during this operation time window. Although the previous
of the higher dimension of the proposed DR problem. However, the sections indicated the successful DR-based VFM monitoring of indepen-
computation times were once more extremely small, making possible dent product flow rates of production wells, the initial results obtained
the implementation of the numerical procedures on line and in real with the phenomenological model and the actual industrial data were
time. not satisfactory. As one can see in Fig. 9, a constant positive offset

10
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 6. Reconciled total flow rates and estimated individual flow rates for 5 wells using the simplified multiphase model.

could be observed between reconciled and measured total water flow each individual run of the proposed DR procedure: (i) the pressure
rate, resulting in average 78% deviation, and oil flow rate, resulting in correction factors (𝑑𝑃𝑖 ), defined as numbers that multiply the head
average 5.7% deviation, showing that the proposed numerical structure losses calculated with the flow model and compensate for possible
was inefficient to allow the appropriate description of the platform inaccuracies of the head loss predictions and of the production network;
performance. (ii) gas/oil ratio (𝐺𝑂𝑅𝑖 ), which compensate for inaccuracies of the
For all the reasons presented in the previous paragraphs, a modified gas fraction characterization and possible transient behavior of the
approach was implemented here to allow the correction of the model production network; (iii) and water cut (𝐵𝑆𝑊𝑖 ) fractions of the oil
predictions and simultaneously avoid the formal estimation of the stream, which compensate for inaccuracies of the water fraction char-
model parameters and prevent the previously reported observability acterization and possible transient behavior of the production network.
problems, as described in the previous sections. In order to adjust the Basically, random combinations of these parameters were generated,
well features and overcome the numerical issues, some structural and according to the uniform distribution and within narrow search regions
compositional parameters were varied inside small and well-defined of ±20% around the nominal values: the measured data for 𝐺𝑂𝑅𝑖 and
search regions, centered at the nominal structural and compositional 𝐵𝑆𝑊𝑖 and 1.0 for 𝑑𝑃𝑖 . Then, the original DR problem was solved for
values. In particular, the following parameters were estimated during each random combination of these data. The DR problem was regarded

11
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 7. Reconciled individual pressure drops for 5 wells using the simplified multiphase model.

as successful when 99.7% of the reconciled oil flow rates were placed adjustment of BSW values, which should be clearly re-evaluated at
inside the confidence region centered at the measured data and built plant site, as observed measurements suggest that the BSW values of the
with ±3 standard deviations around the center. produced oil were significantly smaller than the ones used nominally.
The obtained results showed that it is indeed possible to use DR- Regarding computational time, the majority of DR evaluations took
based procedures for implementation of VFM using complex flow mod- less than two hours to finish and, therefore, are indicated to perform
els as references in real applications. Fig. 9 shows the good adjustment VFM in a real time. Some improvements may be indicated such as
of the overall oil flow rate, with errors always smaller than 2.33%, with model optimization to diminish model computational time and parallel
average value of 0.87%. Although the average errors for the overall computing. At some level, parallelization was already applied to this
water flow rate were equal to 17%, obtained data were clearly more ac- work since each model run for each well was assigned to a differ-
curate that the ones observed in Fig. 10, indicating the appropriateness ent computational process. However, it is possible to achieve smaller
of the proposed corrections. It is important to highlight that deviations computational time paralleling the optimization algorithm.
of the calculated water flow rates were not regarded as very important, Fig. 11 shows that many corrected 𝑑𝑃𝑖 values reached the upper
because adjustment of oil flow rates is much more important at the oil bound of the parameter search region. Opposite behavior can be seen
platform and because water flow rates can be easily corrected through in Figs. 12 and 13 for both 𝐺𝑂𝑅𝑖 and 𝐵𝑆𝑊𝑖 estimation. This clearly

12
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 8. Reconciled oil and gas mass fractions and estimated water fraction for 5 wells using the simplified multiphase model.

indicates that the model predictions must be corrected to compensate able to assign the individual unmeasured oil flow rates for each oil
for inaccuracies of the flow model computations and of the representa- well, based on the available pressure and flow rate measurements,
tion of the production network structure. It can be seen in Fig. 14 that
constituting a powerful tool to indicate changes in production and
the reconciled pressure values are usually very close to the measured
values, although some small positive biases can be observed in some identify operation failures and faults in the production structure. Fig. 13
cases, as observed in well 2, 3 and 14, indicating that some additional also shows that the operation of the wells can be quite heterogeneous,
𝑑𝑃𝑖 corrections can still be still necessary. However, these discrepancies
as indirectly shown by pressure profiles. Particularly, low flow rates
may also indicate the occurrence of sensor wearing, instrument miscal-
ibration or development of flow instability and unusual variations of can be associated with low pressure drops and vice-versa. Fig. 13 can
the gas concentrations and compositions. This result emphasizes how be obviously utilized for programming of well test sequences, as low
powerful DR is for VFM since it can provide vital information about
oil production can possibly indicate operation problems that must be
system characteristics and sensoring problems.
Fig. 15 constitutes one of the most import results of the presented corrected in the well. Additional reconciliation results for 𝑃𝑔𝑎𝑠,𝑖∕𝑄𝑔𝑎𝑠,𝑖 and
DR-based VFM method as it illustrates how the DR-based VFM was 𝑄𝑤,𝑖 can be seen in the complementary material.

13
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 9. Reconciled overall oil and water flow rates without evaluation of structural and compositional deviations and using MARLIM II.

Fig. 10. Reconciled overall oil and water flow rates with evaluation of structural and compositional deviations and using MARLIM II.

Table 7 4. Conclusions
Variances of reconciled variables for the first DR evaluation using the
phenomenological multiphase model.
Virtual flow metering (VFM) tools were developed and implemented
Well 𝑄𝑔𝑎𝑠 𝑃𝑐ℎ𝑜𝑘𝑒 𝑃𝑔𝑎𝑠
in the present work for the first time to provide the individual flow
1 – 0.145 0.021
2 1 180 891.9 0.283 –
rate measurements of the many wells that are operated simultaneously
3 – 2.266 0.075 in typical oil and gas platforms, using phenomenological flow models
4 – 0.717 0.115 as references and available well pressure profiles, well temperature
5 – 0.475 0.017 profiles and overall product flow rates of gas, oil and water streams to
6 238 781 0.145 –
calibrate the model with help of data reconciliation (DR) procedures.
7 – 0.038 0.064
8 – 0.024 0.084 Initially, simplified flow models were used to analyze the effect of
9 – 0.230 0.067 different DR structures on the obtained results. Then, a more complex
10 12 251.2 0.077 – phenomenological flow model (MARLIM II) was used to evaluate the
11 – 0.096 0.137
performance of the proposed DR-based VFM tool to represent a real
12 – 0.013 0.074
13 3.03 × 109 0.003 – production field.
14 7 746.5 2.298 – Two simplified models were developed, assuming the single-phase
15 – 0.003 0.067 and the multiphase nature of the oil stream. The evaluation of the
numerical performances of both models indicated that the fundamental
characteristics of the production network and the overall gas and water
contents of the product streams must be specified with the best possible
accuracy and that possible adjustment of network properties and gas
and water contents of the product streams must be handled with care

14
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 11. Estimated 𝑑𝑃𝑖 values for each well using MARLIM II.

during the DR analyzes, as this may lead to observability problems. It the higher dimension of the proposed DR problem. However, compu-
was also shown that the quality of the obtained results was not affected tation times associated with the simplified models were always small,
making possible the implementation of the numerical procedures on
by the number of wells connected to the platform, indicating that the
line and in real time.
proposed DR-based scheme can handle a large number of independent
Based on the previous results, MARLIM II was used to perform
wells connected simultaneously to a single production platform. When
the DR of actual production data, being possible to implement the
compared to the numerical performance of the single-phase model, the VFM tool successfully, to achieve good representation of the available
computational times increased for the multi-phase model, because of data and to assign the individual production flow rate values for each

15
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 12. Estimated 𝐵𝑆𝑊𝑖 values for each well using MARLIM II.

independent oil well. However, in order to do that, some structural and fractions of the oil stream, which compensate for inaccuracies of the
compositional parameters were corrected simultaneously during the DR water fraction characterization and possible transient behavior of the
procedure: (i) pressure correction factors (𝑑𝑃𝑖 ), defined as numbers production network.
that multiply the head losses calculated with the flow model and The method provided relative errors not greater than 2.33% for oil
compensate for possible inaccuracies of the head loss predictions and total flow rate with average error of 0.87%. Also, it was possible to
of the production network; (ii) gas/oil ratio (𝐺𝑂𝑅𝑖 ), which compensate get some insights about sensoring management and system behavior.
for inaccuracies of the gas fraction characterization and possible tran- For example, wells 2, 3 and 14 displayed greater discrepancies be-
sient behavior of the production network; (iii) and water cut (𝐵𝑆𝑊𝑖 ) tween reconciled and measured choke pressure than the other wells,

16
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 13. Estimated 𝐺𝑂𝑅𝑖 values for each well using MARLIM II.

indicating that these sensors were not well calibrated or were subject and faults in the production structure, being therefore also useful for
programming of well test sequences.
to wearing; or even that the system flow was experiencing some sort of
instability. Particularly, the proposed DR-based VFM was able to assign List of Abbreviations

the individual unmeasured oil flow rates for each oil well, based on the BSW basic sediments and water
DR Data Reconciliation
available pressure and flow rate measurements, constituting a powerful
DRPE Data Reconciliation and Parameter Estimation
tool to indicate changes in production and identify operation failures GOR gas oil ratio

17
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 14. Reconciled choke pressures using MARLIM II.

PDG permanent downhole gauge 𝑑𝑃 Pressure correction factor

Petrobras Petróleo Brasileiro S.A. 𝑓 Probability function
TPT temperature and pressure transmitter 𝐾 Pressure drop constant [Pa.h∕m3 (normal)]
VMF Virtual Flow Metering 𝑃 Pressure [kPa]
WLS Weighted Least Squares
𝑄 Flow rate [km3 ∕h (normal)]
List of Symbols [units] 𝑤 Mass fraction
𝒈 Vector of the inequalities of linear or nonlinear algebraic constraints 𝜀 White Gaussian noise
𝒉 is the vector of linear or nonlinear algebraic constraint equations 𝜌 Specific mass [kg∕m3 ]
𝒖 is the vector of the unmeasured variables 𝜎 Standard deviation
𝑽 Covariance matrix 𝑧̂ Reconciled variable
Ṁ Mass flow rate [kg∕h] 𝑧 Measured variable

18
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

Fig. 15. Individual oil flow rates using MARLIM II.

CRediT authorship contribution statement Declaration of competing interest

Gabriel M.P. Andrade: Conceptualization, Methodology, Software, The authors declare that they have no known competing finan-
Validation, Formal analysis, Investigation, Data curation, Writing – cial interests or personal relationships that could have appeared to
original draft, Writing – review & editing, Visualization. Diego Q.F. influence the work reported in this paper.
de Menezes: Conceptualization, Methodology, Software, Validation,
Acknowledgments
Formal analysis, Investigation, Data curation, Writing – original draft,
Writing – review & editing, Visualization. Rafael M. Soares: Con-
This work was supported by the Petrobras – Petróleo Brasileiro
ceptualization, Methodology, Software, Investigation, Data curation,
SA (Grant number: 2018/00120-0); Conselho Nacional de Desenvolvi-
Writing – review & editing, Visualization. Tiago S.M. Lemos: Concep- mento Científico e Tecnológico (CNPq); Coordenação de Aperfeiçoa-
tualization, Writing – review & editing, Supervision, Project adminis- mento de Pessoal de Nível Superior (CAPES) and Fundação Carlos Cha-
tration, Visualization. Alex F. Teixeira: Conceptualization, Resources, gas Filho de Apoio à Pesquisa do Estado do Rio de Janeiro (FAPERJ).
Writing – review & editing. Leonardo D. Ribeiro: Conceptualization,
Resources. Bruno F. Vieira: Conceptualization, Project administration, Appendix A. Supplementary data
Resources, Funding acquisition. José Carlos Pinto: Conceptualization,
Methodology, Writing – original draft, Writing – review & editing, Supplementary material related to this article can be found online
Project administration, Funding acquisition, Resources. at https://doi.org/10.1016/j.petrol.2021.109772.

19
G.M.P. Andrade et al. Journal of Petroleum Science and Engineering 208 (2022) 109772

References Makogon, T.Y., 2019. Handbook of Multiphase Flow Assurance. Gulf Professional
Publishing.
Bertovic, D., 1995. A unified model for gas-lift valve performance incorporating Monks, J.S., Campos, M.C.M.D.M., Pellanda, P.C., 2019. Inferência da Vazão de
temperature effects. (Ph.D. thesis). University of Tulsa. Produção em Poços de Petróleo. Revista Militar de CiÊNcia E Tecnologia 36 (1).
Bikmukhametov, T., Jäschke, J., 2020. First principles and machine learning virtual Orkiszewski, J., 1967. Predicting two-phase pressure drops in vertical pipe. J. Pet.
flow metering: a literature review. J. Pet. Sci. Eng. 184, 106487. Technol. 19 (06), 829–838.
Câmara, M.M., Soares, R.M., Feital, T., Anzai, T.K., Diehl, F.C., Thompson, P.H., Patel, P., Odden, H., Djoric, B., Garner, R.D., Vea, H.K., 2014. Model based multiphase
Pinto, J.C., 2017. Numerical aspects of data reconciliation in industrial applications. metering and production allocation. In: Offshore Technology Conference Asia, All
Processes 5 (4), 56. Days, URL https://doi.org/10.4043/25457-MS, OTC-25457-MS.
Carbone, L.C., 2007. Metodologia para Medição Virtual de Vazão e Otimização da Poettman, F.H., Carpenter, P.G., 1952. The multiphase flow of gas, oil, and water
Produção. III Seminório de Elevação Artificial e Escoamento, Búzios, RJ, Brasil. through vertical flow strings with application to the design of gas-lift installations.
Craft, B., Holden, W., Graves, E., 1962. Well design: Drilling and production. In: Drilling and Production Practice. OnePetro.
Prentice-Hall, ISBN: 9780139500220, URL https://books.google.com.br/books?id= Prata, D.M., 2009. Reconciliação robusta de dados para monitoramento em tempo real.
flwZAQAAIAAJ. (Ph.D. thesis). COPPE - Universidade Federal do Rio de Janeiro, Rio de Janeiro,
de Menezes, D.Q.F., 2021. Monitoramento de processos de separação baseado em RJ.
reconciliação de dadoso. (Ph.D. thesis). COPPE - Universidade Federal do Rio de Prata, D.M., Lima, E.L., Pinto, J.C., 2009. Nonlinear dynamic data reconciliation in real
Janeiro, Rio de Janeiro, RJ. time in actual processes. Comput. Aided Chem. Eng. 27, 47–54.
de Menezes, D.Q.F., de Sá, M.C.C., Fontoura, T.B., Anzai, T.K., Diehl, F.C., Thomp- Prata, D.M., Schwaab, M., Lima, E.L., Pinto, J.C., 2010. Simultaneous robust data
son, P.H., Pinto, J.C., 2020. Modeling of spiral wound membranes for gas reconciliation and gross error detection through particle swarm optimization for
separations—Part II: Data reconciliation for online monitoring. Processes 8 (9), an industrial polypropylene reactor. Chem. Eng. Sci. 65 (17), 4943–4954.
1035, URL https://doi.org/10.3390/pr8091035. Rasmussen, Å., 2004. Field Applications of Model-Based Multiphase Flow Computing.
de Menezes, D.Q.F., Prata, D.M., Secchi, A.R., Pinto, J.C., 2021. A review North Sea flow measurment workshop, St Andrews, UK.
on robust M-estimators for regression analysis. Comput. Chem. Eng. (ISSN: Ribeiro, C.A.I., Madureira, D.F., Carvalho, A.S., 2009. Guia de Utilização: MARLIM,
0098-1354) 147, 107254, URL https://www.sciencedirect.com/science/article/pii/ Vol. II, 20094.
S0098135421000326. Romagnoli, J.A., Sanchez, M.C., 1999. Data Processing and Reconciliation for Chemical
Duns, H., Ros, N., 1963. Vertical flow of gas and liquid mixtures in wells. In: 6th World Process Operations, Vol. 2. Academic Press.
Petroleum Congress. OnePetro. Seman, L.O., Giuliani, C.M., Camponogara, E., Müller, E.R., Vieira, B.F., Miy-
Falcone, G., Hewitt, G., Alimonti, C., Harrison, B., 2002. Multiphase flow metering: atake, L.K., Medeiros, A.G., 2021. Tuning of oil well models with production data
Current trends and future developments. J. Pet. Technol. (ISSN: 0149-2136) 54 reconciliation. Comput. Chem. Eng. 145, 107179.
(04), 77–84, URL https://doi.org/10.2118/74689-JPT. Seman, L.O., Miyatake, L.K., Camponogara, E., Giuliani, C.M., Vieira, B.F., 2020.
Feital, T., Pinto, J.C., 2015. Use of variance spectra for in-line validation of process Derivative-free parameter tuning for a well multiphase flow simulator. J. Pet. Sci.
measurements in continuous processes. Can. J. Chem. Eng. 93 (8), 1426–1437. Eng. 192, 107288.
Hagedorn, A.R., Brown, K.E., 1965. Experimental study of pressure gradients occurring Soares, M., Machado, F., Guimaraes, A., Amaral, M.M., Pinto, J.C., 2011. Real-
during continuous two-phase flow in small-diameter vertical conduits. J. Pet. time monitoring and parameter estimation of the emulsion polymerization of
Technol. 17 (04), 475–484. carboxylated styrene/butadiene latexes. Polym. Eng. Sci. 51 (10), 1919–1932.
Hastie, T., Tibshirani, R., Friedman, J., 2009. The Elements of Statistical Learning: Data Thorn, R., Johansen, G.A., Hjertaker, B.T., 2012. Three-phase flow measurement in the
Mining, Inference, and Prediction. Springer Science & Business Media. petroleum industry. Meas. Sci. Technol. 24 (1), 012003.
Idsøo, E.S., Sperle, I.L., Aasheim, R., Wold, M.S., 2014. Automatic subsea deduction well Toskey, E.D., 2012. Improvements to deepwater subsea measurements RPSEA program:
testing for increased accuracy and reduced test time. In: Abu Dhabi International Evaluation of flow modelling. In: OTC Offshore Technology Conference, All Days,
Petroleum Exhibition and Conference, Day 1 Mon, November 10, 2014, URL URL https://doi.org/10.4043/23314-MS, OTC-23314-MS.
https://doi.org/10.2118/171840-MS, D011S006R001. Udofia, E., Akporuno, M., Vandenberg, F., Beijer, V., Oguntimehen, G., Oni, O., 2012.
Knepper, J.C., Gorman, J.W., 1980. Statistical analysis of constrained data sets. AIChE Advances in production allocation: Bonga field experience. In: SPE Intelligent
J. 26 (2), 260–264. Energy International Conference and Exhibition, All Days, URL https://doi.org/10.
Koziel, S., Yang, X.-S., 2011. Computational Optimization, Methods and Algorithms, 2118/150450-MS, SPE-150450-MS.
Vol. 356. Springer. White, F.M., 2011. Fluid Mechanics. McGraw-Hill.
Ma, Y., McLaughlan, M., Zhang, N., Li, J., 2020. Novel feasible path optimisation Winkler, H., Eads, P., 1991. Applying the basic performance concepts of single-element
algorithms using steady-state and/or pseudo-transient simulations. Comput. Chem. unbalanced gas-lift valves for installation design. In: Producing Solutions for the
Eng. 143, 107058. 90’s. pp. 43–58.

20