Oil-well Flow-Rate Forecasting Using

Auto-Regressive Model
Shreyas Lele,

B. Tech (Information Technology), Software Engineer.

Abstract: Advancements in IIOT devices and analytics based is providing a major help in analysis and
forecasting of oil production from huge oil-well systems. This paper presents our approach where we
apply autoregressive model to time series data for prediction of possible upcoming values. It focuses how
timeseries prediction along with IIOT helped us to analyze past data of flow rate to forecast the futuristic
flow rate values beforehand with good accuracy. The prediction output can be used to estimate the oil
production as well as will help in detecting anomalous values which in turn will provide idea about any
flaws in the oil well or other machinery in it.

Keywords : Flow-rate, ARIMA, IIOT, timeseries prediction

1. Introduction:
The well head controls surface pressure and well
Due to the widespread use of information technology,
many time-series are collected during industrial
operations. Such data can be explored and analyzed production rate. Inductors are generally chosen so that
using many useful tools and methodologies they do not affect the production rate after the line
developed in modern time series analysis. Time series inductance. It is assumed that the flow through the
analysis (TSA) is often associated with the discovery restrictor takes place under critical flow conditions.
and use of patterns (such as periodicity, seasonality, or The outflow from the well can be critical or subcritical.
cycles), and prediction of future values (specifically From the point of view of controlling the production
termed forecasting in the time series context). of solid and fine particles in the well flow, the critical
flow represents the rate of solid material production
The modern oil wells and the systems mounted on in the flow. Maintaining subcritical flow conditions
them requires extensive instrumentation and may be important for sand or fine sedimentation. If the
monitoring. Smart or intelligent wells are flow exceeds this limit (i.e. becomes critical), the
nonconventional well with downhole instrumentation production of sand and fine particles will increase
(sensors, valves and inflow control devices) installed significantly.
on the production tubing. Such wells allow for the
continuous in-situ monitoring of fluid flow rates and
pressures and the periodic adjustment of downhole 2.Literature Survey
valves Approaches including analytics and machine
learning have become vital to digital oil field None of the various published models used to predict
operations. Most of the data from the sensors on oil production rates through wellhead chokes from
different oil wells are collected in form of continuous fluid composition and pressures can be considered as
streams, which require the use of fast real-time data a universal model for all regions. Here, a model is
analysis techniques. Due to the high frequency of provided to predict liquid production-flow rates for
sensor data in the oil and gas industry and its the Reshadat oil field offshore southwest Iran,
timeseries nature, it is inevitable that oil and gas applying a customized genetic optimization algorithm
companies will need to use efficient timeseries (GA) and standard Excel Solver non-linear and
analysis approaches, those that particularly deal with evolutionary optimization algorithms. The dataset of
historical data. 182 records of wellhead choke measurements spans
liquid flow rates from < 100 to 30,000 stock tank
Before drilling, the production passes through a stop barrels/day. Each data record includes measurements
that serves to accelerate the ideal or desired flow of five variables: liquid production rate (QL), wellhead
relative to the pressure waves in the flow line. pressure, choke size, basic sediment and water, and

Electronic copy available at: https://ssrn.com/abstract=3502754

gas–liquid ratio. 70% of the dataset (127 data records) Knowledge of the pressure-rate behavior can be quite
was used for training purposes to establish the helpful in designing and evaluating stimulation
prediction relationships, and 30% of the dataset (55 treatments or any operation that improves flow
data records) was utilized for independently testing efficiency, like the estimation of future performance
the accuracy of the derived relationships as predictive which is required for forecasting and planning
tools. The methodology applying either the purpose. Inflow performance of a reservoir is defined
customized GA or standard Solver optimization as the functional relationship between the flowing
algorithms, demonstrates significant improvements in bottom-hole and the resulting flow rate. It is the rate
QL-prediction accuracy with the lowest APD (− 7.72 to at which fluid will flow towards the wellbore and
− 2.89), AAPD (7.33–8.51), SD (288.77–563.85), MSE depends on the viscosity of the fluid, the permeability
(91,871–316,429), and RMSE (303.1–562.52); and the of the rock, and the driving force. For an oil/gas well
highest R2 (greater than 0.997) compared to six to flow there must be a pressure difference from
previously published liquid flow-rate prediction reservoir to the well-bore at the reservoir depth. If the
models..[1] well-bore pressure is equal to the reservoir pressure,
there can be no inflow. If the well-bore pressure is
In this paper, an integrated model was reviewed and zero, the inflow would be a maximum possible that is
used to calculate the dimensionless oil flow rate and the Absolute Open Flow (AOF). [6]
bottom hole flowing pressures variables. Oil
production data from a commingled smart well were Novel liquid critical-flow rate models, hybridizing an
simulated on a FASTWELL software and their plotted artificial neural network (ANN) with a teaching-
IPR results analyzed to study; the oil inflow rate for a learning-based optimization (TLBO) algorithms,
single-phase flow, oil inflow rate for a multiple phase involving 3 and 6 input variables, demonstrate
flow, Absolute Open flow (AOF), the effect of water improved accuracy compared to nonlinear regression
and gas on oil inflow rate and the effect of anticipated models, traditional ANN models and published
formation damage. The approach used here provides correlations. The improved accuracy of the developed
a quick insight for early isolation of the errors models is assessed statistically using a data set of 113
contributed by gas and water in the analysis of oil wellhead flow tests from oil wells in South Iran (with
inflow performance relationship (IPR) curves.[2] a full data listing included). The ANN-TLBO (6
parameters) developed model is the most accurate,
Smart well technology is one of the most significant yielding the best liquid critical-flow rate predictions
breakthroughs in modern petroleum production for that data set: coefficient of determination of 0.981;
technologies. It allows operators to remotely control root mean square error of 714; average relative error of
production of multilateral wells without intervention, 2.09%; and, average absolute relative error of 6.5% [7]
thereby, optimizing production, and maximizing
recovery and capital-expenditure efficiency while The effect of valve position, inlet flow configuration
minimizing operating costs.[3] Some applications of and mode of operation on production efficiency was
this technology include managing production where studied Horizontal axis in the model with synthetic
there are significant variations among laterals in lower drive. They anticipated control in reactive and
reservoir pressure, productivity, gas/water fractions, proactive mode. They defined reactive control as the
or permeability due to presence of fractures and faults surface production choking or zonal isolation of
[4]. production through closing valves discretely or
continuously. With proactive control they rely on the
In this paper, the aim is to develop new correlation to data provided by the sensors (distributed electrical
predict oil flow rate through chokes with critical flow panels) installed in the well. Based on this data, it has
for Iraqi oil wells. This study shows that there is a strict been shown that a proactive control strategy can be
quantitative connection between three factors: used to avoid it. They used simulated annealing and
upstream pressure, gas-oil ratio and choke size with conjugate gradient optimization algorithms to
oil flow rate at critical flow for one hundred determine the optimal position and control
production tests of Iraqi oil wells. Many forms have configuration of the valve to maximize the
been tried using nonlinear regression analysis to accumulated oil production. They did not update the
obtain the optimum form of correlation that gives valve configuration on time and assumed that they
minimum differences between the calculated and field would start from the beginning production. Both
data. Here, based on Iraqi oil wells data; new methods have approached the same optimum [8]
correlation has been developed for predicting oil flow
rate through chokes at critical flow. The proposed
correlation exhibits more accuracy (only 4.523%
average absolute error) than the existent correlations.
The correlation coefficient of the new correlation is
determined as 0.997[5]

Electronic copy available at: https://ssrn.com/abstract=3502754

3. Training models

ARIMA is a general class of forecasting model which An ARIMA model is one where the time series was
have random trend, random walk, exponential differenced at least once to make it stationary and you
smoothing, and autoregressive models as special
cases. It is suitable for application on time-series data
which can be made stationary by applying
differencing or simple non-linear transformations like
logarithm, deflation etc. combine the AR and the MA terms. So, the equation
becomes:
A time series dataset is stationary if its statistical
properties are all constant over time. A stationary ARIMA model in words:
series has no trend, its variations around its mean have Predicted Yt = Constant + Linear combination Lags of
a constant amplitude, and it wiggles in a consistent Y (up to p lags) + Linear Combination of Lagged
fashion, i.e., its short-term random time patterns forecast errors (up to q lags)
always look the same in a statistical sense.
To predict values of AR order (p) and MA order(q), we
ARIMA stands for Auto Regressive Integrated Moving use two plots:
Average. Here Autoregressive terms (AR) depend on
the lags of the stationarized data, Moving Average Autocorrelation Function (ACF): It is a measure of the
(MA) depends on lags of the forecasted errors and a correlation between the Time series with a lagged
time series which needs to be differenced to be made version of itself. For instance, at lag 5, ACF would
stationary is said to be an "integrated" version of a compare series at time instant ‘t1’…’t2’ with series at
stationary series. instant ‘t1-5’…’t2-5’ (t1-5 and t2 being end points).
A nonseasonal ARIMA model is denoted by ARIMA Partial Autocorrelation Function (PACF): This
(p, d, q), where: measures the correlation between the TS with a lagged
‘p’ is the number of autoregressive terms, ‘d’ is the version of itself but after eliminating the variations
number of non-seasonal differences suitable for already explained by the intervening comparison.
stationarity, ‘q’ is the number of lagged errors in the
prediction equation. For checking the stationarity of the timeseries we use
the following:
A pure Auto Regressive (AR only) model is one
where Yt depends only on its own lags. That is, Yt is a (A) Plotting Rolling statistics: In this test, we plot the
function of the ‘lags of Yt’. moving average and analyze how much variance is
there in the plot.

(B) Dickey-Fuller Test: This is one of the statistical tests

where, $Y{t-1} $ is the lag1 of the series, $\beta1$ is the for checking stationarity. Here the null hypothesis is
that the TS is non-stationary. The test results comprise
coefficient of lag1 that the model estimates and
$\alpha$ is the intercept term, also estimated by the of a Test Statistic and some Critical values for
model. difference confidence levels. If the ‘Test Statistic’ is less
than the ‘Critical Value’, we can reject the null
Likewise, a pure Moving Average (MA only) hypothesis and say that the series is stationary. Now
model is one where Yt depends only on the lagged let’s plot the Rolling Average and visually analyze
whether data is stationary. Also applying dickey fuller
forecast errors.
for hypothesis testing.

Electronic copy available at: https://ssrn.com/abstract=3502754

4.Results and Analysis

Fig1: Flowrate Data after preprocessing

Fig4: ACF and PACF to determine correlation

Fig2: first difference rolling mean and Standard deviation

Fig5: ARIMA model statistics

Fig3: Dickey-Fuller Test Output

Fig6: Actual Data and forecast in terms of log

Electronic copy available at: https://ssrn.com/abstract=3502754

5.Conslusion [5] Dhefaf J. Sadiq - Predication of Oil Flow Rate
through Choke at Critical Flow for Iraqi Oil Wells
For the flow rate time series data which is used in this
case, the ARIMA model of order (0,1,1) is working well [6] Wiggins, M.L., Russell, J.E. and Jennings, J.W.
and is giving satisfactory forecast. The forecast in fig 6 (1992). Analytical Inflow Performance Relationship for
gives the logarithm of the flow rate since we have used ThreePhase Flow in Bounded Reservoirs, SPE paper
the log to remove any existing trend or seasonality in 24055, presented at Western Regional Meeting,
the data. Now when we take exponential of the Bakersfield, CA,
forecast, we can get the actual value of the flow rate. If
the forecast values change abruptly, then that might [7] AbouzarChoubineh, HamzehGhorbani, David A.
indicate some problematic behavior within the oil- Wood, Seyedeh Robab Moosavi, EliasKhalafi,
well, which can be further inspected. ErfanSadatshojaei - Improved predictions of wellhead
choke liquid critical-flow rates: Modelling based on
Thus, using the flow rate values, the above model will hybrid neural network training learning based
help in predicting any future anomalies or problems optimization
which the oil-well might face and thereby will help the
company or customer to take necessary corrective [8] Khargoria, A., Zhang, F., Li, R., and Jalali, Y. (2002).
measures well in advance. Application of Distributed Electrical Measurements
and Inflow Control in Horizontal Wells under Bottom-
As the risk assessment of the assets used in the Oil and Water Drive
Gas industry is of greater importance and is highly
necessary as even a minute fault within the pipelines,
or electronic pumps and similar other machineries
linked to the oil wells, can result hamper the upstream
production of oil which in turn results in heavy losses
for the company in production as well as economic
views.

Taking into consideration the above point, this paper

emphasizes on the approach to forecast and analyze
the risk using the data-streams from multiple sensors
which indicate different parameters associated with
the oil wells and the machineries attached to it, and
thereby alert the on-field engineers about any
futuristic anomalies and malfunctions.

6. References

[1] Hamzeh Ghorbani, David A. Wood, Jamshid

Moghadasi, Abouzar Choubineh, Peyman Abdizadeh
& Nima Mohamadian - Predicting liquid flow-rate
performance through wellhead chokes with genetic
and solver optimizers: an oil field case study

[2] Udoaka Mfon, Godwill, Dulu Agppah - Analysis of

smart oil wells inflow performance relationship
curves

[3] Robinson, M. (2003). Intelligent Well Completions,

SPE Journal of Petroleum Technology

[4] Mubarak, S., Dawood, N. and Salamy, S. (2009).

Lessons Learned from 100 Intelligent Wells Equipped
with Multiple Downhole Valves. SPE 126089-MS
Paper presented at the SPE Saudi Arabia Section
Technical Symposium, AlKhobar, Saudi Arabia

Electronic copy available at: https://ssrn.com/abstract=3502754