OTC 24981

ANN Powered Virtual Well Testing

A. Aggarwal, S. Agarwal, Indian School of Mines

Copyright 2014, Offshore Technology Conference

This paper was prepared for presentation at the Offshore Technology Conference Asia
held in Kuala Lumpur, Malaysia, 2528 March 2014.

Introduction

This paper was selected for presentation by an OTC program committee following review
of information contained in an abstract submitted by the author(s). Contents of the paper
have not been reviewed by the Offshore Technology Conference and are subject to
correction by the author(s). The material does not necessarily reflect any position of the
Offshore Technology Conference, its officers, or members. Electronic reproduction,
distribution, or storage of any part of this paper without the written consent of the Offshore
Technology Conference is prohibited. Permission to reproduce in print is restricted to an
abstract of not more than 300 words; illustrations may not be copied. The abstract must
contain conspicuous acknowledgment of OTC copyright.

Due to the complex structures and heterogeneous media,

of oil and gas reservoirs, characterizing reservoirs
precisely is a herculean task. Petroleum Engineers solve
this challenge by acquiring and analyzing in depth
reservoir information, which is crucial for the study of the
reservoir performance (Vaferi et al., 2011). Well Testing
or Pressure Transient Testing has proved to be a powerful
reservoir characterization tool to study such complex
media (Muskat, 1937). Pressure testing is conducted by
recording the well bore bottom hole pressure responses
which are created as result of induced flow disturbances.
The most general test methods are 1) By creating a
pressure drawdown in the wellbore by producing the well
at a constant rate, after keeping it in shut-in condition for
a set period (Figure 1); 2) By developing a pressure
build-up in the wellbore by shutting-in at the bottom hole,
due to which formation fluids cannot flow into the
wellbore. Thus, measured flow rates and pressures during
these tests can provide sufficient information for the
characterization of the tested well (Matthews et al., 1967;
Earlougher et al., 1977).

Abstract
Due to the drying up of old oil fields throughout the
globe, the age of easy oil is over and the newly discovered
fields have reservoirs with complex heterogeneous media.
The reservoir parameters are identified indirectly by
correctly interpreting well test model which is recognized
by the feature of pressure derivative curves. Well testing
involves creation of disturbance in fluid flow by injecting
liquids and simultaneously recording the pressure
transient data. Lost production, equipment and personnel
costs turn well testing as highly cost intensive job making
it difficult to cover all the important wells in a particular
field. But with the advent of artificial neural networks
(ANN) it is now possible to generate synthetic pressure
transient data. This technique provides a basis to leach out
detailed information from the available pressure transient
data and it doesnt eradicate the need for actual well tests.
This technique can also prove to be very vital in cases
where equipment breakdown may have taken place and
full set of data couldnt be availed. This simulated well
testing involves training of a neural network from
pressure transient data obtained from designated wells in
the field, which has the potential to generate pressure
transient responses at other well sites where no well test
has been conducted.
In this paper a 3 layer multi-layer perceptron (MLP) Time
Delay Neural Network - NARX model has been designed
working on resilient backpopagation algorithm for
training. Cubic Spline Interpolation has been used from
enriching the data before feeding it to NARX model. A
simulated example which highlights the efficiency of
NARX model in attaining accurate synthetic pressure
transient data has been discussed. The neural network is
successful in predicting well test interpretation model.
The ANN thus produces expeditious and reliable synthetic
data which has the potential to revamp the industry.

Well testing enables accurate determination of cash flow

of the well by obtaining comprehensive reservoir
description. The core purpose of well testing lies in the
determination of the fluid production capability of a
formation and incurring the inherent reason for
productivity of well. A meticulously designed and
performed well test operation can provide accurate facts
and figures about formation permeability, average
conductivities, extent of wellbore damage or stimulation,
reservoir pressure, and extended testing may deliver
information regarding underlying geological barriers
(faults, pinch-outs, etc.), reservoir boundaries and
heterogeneities (Matthews et al., 1967; Earlougher et al.,
1977). On the same note Da Prat et al. in 1992 have
classified well test objectives into short term and long
term objectives. Obtaining the reservoir description in the
vicinity of the well bore, from analysis of gathered data,
are categorized as short term objective. While, long term
objectives of well test are to analyze gathered data for
obtaining complete description of the whole reservoir.
However, huge production time losses, manpower and
equipment costs turn it into a cost intensive affair
(Dakshindas, 1999). Since its introduction to the

OTC-24981-MS

petroleum engineering industry around 1937 by ground

water hydrology scientists (Gringarten, 2008), plethora of
novel technologies have been introduced both for data
acquisition and analysis. The introduction of electronic
pressure gauges was a great step ahead in enhancing
obtaining of reservoir description, and continuously
upgraded versions of these are being developed to meet
current challenges. Moreover, for expeditious data
acquisition and interpretation, and in depth rigorous
insight into the reservoir, industry anticipates close
integration of advanced microprocessors and innovative
computational techniques.

comprising of smaller units termed as neuron, as these

seek to imitate the brain microstructure (Noriega, 2005).
The multilayer perceptron, a widely applied example of
artificial neural networks, is a programming epitome that
was developed around half a century back by W.S.
McCulloch and W. Pitts in 1943. The neural network is an
intensely parallel, distributive, adaptive, non-arithmetic
and non-digital system, which can solve problems ranging
from pattern recognition, to innovative symbolic
manipulation

Theory

The neural network attempts to model functioning of

biological nervous system or human brain. The brain is
made up of huge parallel interconnected processing units.
These processing units or neurons are electrically
excitable, which through electrical and chemical signals
transmits and process information. Majorly dendrites,
soma and axon form the neuron body. A neuron receives
the input information from other neurons in the network
connected to it at synapses through dendrites. The
received information is then processed at the soma, which
integrates it over time and space, the generated output is
activated depending on the input. The synapses located at
the end of axon transmit the output signal to connected
dendrites. This forms an intensely complex parallel
computer.

This paper features a unique approach for synthetic

pressure transient data generation by application of Cubic
Spline Interpolation technique and Artificial Neural
Networks (ANN). The main objective of the work
presented is to formulate an artificial neural network
which has the ability to forecast transient pressure
responses without any need for an actual well test. The
data thus generated can be analyzed using traditional well
test analysis methods for reservoir characterization. In a
nutshell, this is made possible by training the neural
network using pressure transient data available from
proximate wells, flow parameters values and reservoir
characteristics, and then the error is reduced significantly
by retraining the network and testing the network
efficiency by comparing the network results to the data
available from actually conducted well tests.
Cubic Spline Interpolation
Data points are interpolated by using cubic spline curve
fitting (CSCF) technique. This can be realized by
arranging the available data into a table [p i,qi], i [0,n],
thus providing with n intervals for n+1 control points. The
cubic spline curve is a continuous piecewise third order
polynomial, satisfying all the input values (Figure 2).
Each polynomials second derivative is usually set to zero
at the endpoints, as this offers a boundary condition which
makes the system of n-1 equations complete. Also this is
not the only possible option, as other boundary conditions
can also be used. Thus a "natural" cubic spline is
produced that results to an elementary tridiagonal system,
which is computed to deliver coefficients of the
polynomials.
The main purpose of using CSCF technique is to increase
data points between minimum and maximum, as with
higher amount of input data points better results through
ANN can be obtained. Moreover its high accuracy of
estimation and capacity to produce seamless curves makes
it a popular interpolation technique.

Biological Basis

Similarly, an artificial neural network is a composite

architecture of soft computing based neurons integrated
into numerous parallel layers which are connected to all
the neurons of preceding and succeeding layers by the
means of weights. A variety of developed neurons and
network specifications can be programmed to interpret,
recognize and retrieve patterns to solve optimization
problems and clear noise from input data (Kumar, 2012).
Network Mechanics
The neural network has two types of learning processes,
supervised and un-supervised. Supervised learning or
training process is the crux of ANN mechanics, initially
random weight values (between -1 to 1) are assigned to
network connections, furthermore the network analyses
input data to estimate weights of the connections between
successive neuron layers (Figure 3). The data is input to
the first layer in which each neuron multiplies the data
with the associated weight and is summed with a bias.
The result is then fed to an activation function, or transfer
function, of the neuron which determines if the result is
above the threshold or not, to instruct the neuron for
transmitting data. This output to activation function is
calculated by

Neural Networks
Neural Networks are a form of massive parallel
distribution of biologically inspired processing units

The output values and the input values are compared for
error and then accordingly weights are updated for

OTC-24981-MS

minimising error. This process is repeated a number of

times (epochs) to achieve precise weight values. Now, the
network output is tested against the input data sets for
measuring the prediction quality, which if acceptable can
be applied to generate data, for environments whose
outputs are unknown.
Background
These are being applied to counter a large variety of
challenges from simple pattern-recognition task, to
advance symbolic manipulation (Noriega, 2005). Neural
networks have gained ground in geophysical application
and well test interpretation in last couple of decades.
Their effective application is to provide remarkably
precise solution to solve a range of problems like well-log
analysis (Huang et al., 1991), seismic deconvolution
(Wang et al., 1992; CalderonMacias et al., 1997),
waveform recognition and rst-break picking (Murat et
al., 1992; McCormack et al., 1993); for electromagnetic
(Poulton et al., 1992), magnetotelluric (Zhang et al.,
1997), and seismic inversion purposes (Roth et al., 1994;
Langer et al., 1996; CalderonMacias et al., 1998);
event classication (Dowla et al. 1990; Romeo, 1994),
zone identification (White et al., 1995), trace editing
(McCormack et al., 1993) and for shear-wave splitting
(Dai et al., 1994), in geophysics domain (Van der Bann et
al., 2000), while making significant contribution to the
well test interpretation domain by enabling permeability
prediction (Singh et al., 2005), reservoir model
identification (Vaferi et al., 2011), well test model
recognition (Sung et al., 1996) and many more.

NARX Neural Network Model

Instead of conventional dynamic networks that are either
feedforward networks or focused networks, with only
input layer dynamics. In this, fully connected feedback
connections are enclosed in numerous layers of a
recurrent dynamic network which is Nonlinear
Autoregressive with eXogenous inputs (NARX recurrent
models). Contrary to other recurrent networks, feedback
comes only from output neuron instead from hidden
states, it is modeled with a tapped delay line (embedded
memory) which is clubbed to a delayed feeding line from
the output, second tapped delay line. This type of limited
viewed on part of the input series is referred to as time
window (Diaconescu, 2008). A NARX model is
formulated by the following equation
y(t) = f(y(t-1), y(t-2),.., y(t-ny), u(t-1), u(t-2),.., u(t-nu))
where, y(t) and u(t) denotes input and output signal of the
network at time t, ny and nu signifies input and output
order, while f represent the mapping performed by
Multilayer Perceptron (Siegelmann et al., 1997). The
value of dependent output y(t) is reverted to earlier output
values and values of eXogenous input (Figure 5). This
real-time feeding of output to the network is called as
parallel architecture, which results in more accurate
training and due its purely feedforward architecture, static
backpropagation can be used.
Application Methodology

However, not much application of this Artificial

Intelligence (AI) technology has been observed in the
generation of synthetic well test Pressure Transient Data
(PTD) and rigorous utilization of advanced neural
network functions, training algorithms and advanced
computational techniques in this specific field have not
been witnessed.

The steps incurred in throughout this process are

explained below:

Time Delay Neural Networks (TDNN)

2.

The designing of neural networks can be understood by

classifying them into two categories dynamic and static.
Static networks are comparatively simple with neither
feedback elements nor delays. On the other hand, in case
of dynamic networks, the output generated is governed by
current inputs, previous inputs, outputs and network
states. However, dynamic networks can be trained by the
same algorithms used by static networks but due to
complex nature of error surfaces computing gradients is
more intensive. Moreover in this study NARX model has
been implemented which is a type of dynamic network or
more specifically Time Delay Neural Network (TDNN)
(Figure 4).

1.

3.

4.
5.

6.

Candidate Identification: A set of wells are chosen,

which are producing from a particular formation
(zone). Segregate the wells where prediction is to be
done.
Information Gathering: Pressure Testing is
performed on the selected zone at all picked well
locations and PTD is recorded. Other vital
information about the test and reservoir is also
acquired.
Data Enrichment: NARX works best when highly
dense data sets are fed to it. Thus according to the
type of cures produced by the specific pressure
testing procedure followed, data interpolation
technique is chosen and applied.
Normalization: The input data for the NARX is
scaled down to the range of 0 to 1.
Network Training: The network is configured with
appropriate parameters and then input data is fed to
the network for training it.
Network Testing: Performance of the network is
validated against known data sets, if the results are

7.

OTC-24981-MS

unsatisfactory then the network parameters are

tweaked and the NARX is retrained.
Prediction: Utilize the trained NARX for simulating
data for unknown outputs.

To minimize error and produce the best pressure transient

predictions, the whole process was tested numerous times
with varied network models and interpolation techniques.
Due to the monotonically increasing nature of pressure
curve generated using PTD the best suited technologies
are identified was Cubic Spline Curve Fitting (CSCF) and
NARX neural model (Figure 6).
Case Study
This case deals with simulator generated pressure buildup test data considering of five active producing wells for
an infinitely large homogenous field (Figure 7). The data
was generated using analytical simulator, which is based
on principle of superposition and infinite acting line
source solution. All the five wells considered had
identical shut-in and production time, and different flow
rates. The wells are shut-in at 215 hours and the data is
recorded for 67 hours.
The ANN is trained from pressure responses of four wells
and the data for well 5 is predicted. Complete data set for
each well is prepared for training the NARX network.
Each wells pressure transient data, (tp+t)/t, modified
inter-well distances, and flow rates along with their
functional links are used as inputs for the network
(Figure). The output from the network is pressure which
is generated by also taking into account the interference
effects from the proximate wells. Modified distance is
formulated to be:
Distmod = QW1 * Dist(W1 W2)/QW2
After taking into account modified distance and functional
links better prediction results were observed but with
increased CPU usage and network training time.
Similarly, initially 60 data values were available, which
by using Cubic Spline Interpolation were increased to 89.
Resilient backpropagation algorithm was used to update
bias values and connection weights during training
because it enables optimizing the magnitude of weight
change by reducing it in case weights are oscillating and
increasing it when for several iterations weight changes
continuously in the same direction. Sigmoid or
squashing functions were used as transfer functions for
all the neurons at every layer, as the derivative of sigmoid
function can be swiftly calculated which is needed to be
backpropagated to calculate error. It is defined by:
F(i) = (1+e-i)-1
Where, F(i) is the function output and i is the input. The
performance of the network was judged on the basis of

mean square error (MSE). The input set was divided using
interleaved indices. The applied NARX model is a three
layer network with 18 input layer neurons, 25 neurons in
hidden layer and 1 output layer neuron. The time delay
configured for this particular problem is 1:2, thus
producing 87 outputs for 89 inputs.
A good match is observed is observed between the
network and the simulator output (Figure 8-11),
confirming that NARX has the potential to predict well
test pressure responses accurately. The prediction
performed by the network for well 5 is also remarkable
considering the complexity of the problem (Figure 12).
With increased available data from more number of wells
the network will deliver more precise predictions but the
wells must be chosen meticulously to ensure inclusion of
interference effects, flow rate effects, shut-in and
production effects, boundaries and heterogeneity effects.
But the risk that network might get over-trained also
escalates with increased number of training wells.
Conclusions
In this paper a unique synthetic pressure transient data
generation has been introduced. The discussed simulated
field case study has justified the approach and delivered
recommendations on how to increase the accuracy of the
prediction. The precision of the network remarkably
improves when neural network is exposed to variety of
data from more number of strategically selected wells.
Incorporation of more data functional links, multiple
inputs from the field after rigorous iterative testing can
increase reliability of the network.
The NARX model has been able to analyze and interpret
the response almost perfectly which is proved by the
fashion in which the output curve closely traces the
simulator curve. Moreover, for cases where due to some
downhole equipment failure test couldnt be completed or
some data is lost, by using this we can complete the data.
This doesnt eradicate the need for actual well tests, but it
remarkably curbs the frequency of actual tests when
clubbed with tactical well test pattern planning. Using
NARX technology more informed well tests can be
designed by extracting more information from the
available data, thus reflecting enormous potential to
revamp the petroleum industry by delivering expeditious
and reliable solutions.
Acknowledgements
The authors are indebted to officials from Schlumberger,
ONGC, Reservoil and Dept. of Petroleum Engineering,
Indian School of Mines Dhanbad for their constant
support during this research. The authors are also obliged
to Mr. Mohit Punjabi, Mr. Swapnil Gupta, Mr. Ajay
Singh, Mr. Paras Goel and Mr. Praveen Pushkar, all from
Indian School of Mines Dhanbad, for always stimulating
and rejuvenating us.

OTC-24981-MS

References
1.

2.

3.

4.
5.

6.

7.
8.

9.

10.

11.
12.

13.

14.

15.
16.

17.
18.

Bartels, R. H. Beatty, J. C. and Barsky, B. A. 1998.

Hermite and Cubic Spline Interpolation, An Introduction to
Splines for Use in Computer Graphics and Geometric
Modelling, Morgan Kaufmann, San Francisco, CA, Chapter
3: 9-17.
Beale, M.H. Hagan, M.T. and Demuth, H.B. 2013. Neural
Network ToolboxTM Users Guide, The MathWorks Inc.,
MA, USA.
Bertolini, A.C. Booth, R.J. Morton, K.L. and Fitzpatrick,
A.J. 2013. Design of Objective Function for Interference
Well Testing, OTC 24513, Offshore Technology
Conference Brazil, Rio de Janeiro, 29-31 October.
Burden, R. L. Faires, J. D. and Reynolds, A. C. 1997,
Numerical Analysis, Brooks/Cole, Boston, MA. 6: 120-121.
Calder onMacas, C. Sen, M. K. and Stoffa, P. L. 1997.
Hopeld neural networks, and mean eld annealing for
seismic deconvolution and multiple attenuation,
Geophysics: 9921002.
Calder onMacas, C. Sen, M. K. and Stoffa, P. L. 1998.
Automatic NMO correction and velocity estimation by a
feedforward neural network, Geophysics: 1696-1707.
Dai, H. and MacBeth, C. 1994. Split shear-wave analysis
using an articial neural network? First Break: 605613.
Dakshindas, S.S. Ertekin, T. and Grader, A.S. 1999. Virtual
Well Testing, SPE 5745, SPE Eastern Regional Meeting,
Charleston, West Virginia, 21-22 October.
Diaconnescu, E. 2008. The use of NARX Neural Networks
to predict Chaotic Time Series, WSEAS Transactions of
Computer Research.
Dowla, F. U. Taylor, S. R. and Anderson, R. W. 1990.
Seismic discrimination with articial neural networks:
Preliminary results with regional spectral data, Bull. Seis.
Soc. Am.: 13461373.
Earlougher, R.C. Jr. 1997. Advances in Well Test Analysis,
Monograph Series, SPE, Dallas.
Gringarten, A.C. 2008. From Straight Lines to
Deconvolution: The Evolution of the State of the Art in
Well Test Analysis, SPE 102079, SPE Reservoir
Evaluation & Engineering, SPE.
Huang, S. C. and Huang, Y. F. 1991. Bounds on the
number of hidden neurons in multilayer perceptrons, IEEE
Trans. Neur. Networks: 4755.
Kumar, A. 2012. Artificial Neural Network as a Tool for
Reservoir Characterization and its Application in the
Petroleum Engineering, OTC 22967, Offshore Technology
Conference, Houston, 30 April 3 May.
Lee, J. 1982. Well Testing, SPE, New York, USA.
Lin, T. Giles, L. Horne, B.G. and Kung, S.Y. 1998. A
Delay Damage Model Selection Algorithm for NARX
Neural Networks, University of Maryland Technical Report
Matthews, C.S and Russell, D.G. 1967, Pressure Buildup
and Flow Tests in Wells, Monograph Series, SPE, Dallas.
McCormack, M. D. Zaucha, D. E. and Dushek, D. W.
1993. Firstbreak refraction event picking and seismic data
trace editing using neural networks, Geophysics: 6778.

19. McCulloch, W.S. and Pitts, W. 1943. A Logical Calculus

of the Ideas Immanent in Nervous Activity, Bulletin of
Mathematical Biophysics, 5.
20. Mohaghegh, S. Arefi, R. Ameri, S. and Rose, D. 1994.
Design and Development of an Artificial Neural Network
for Estimation of Formation Permeability, SPE 28237, SPE
Petroleum Computer Conference, Dallas. 31 July 3 Aug.
21. Murat, M. E. and Rudman, A. J. 1992. Automated rst
arrival picking: A neural network approach, Geophysics
Prosp: 587604.
22. Muskat, M. 1937. Flow of Homogeneous Fluids Through
Porous Media, McGraw Hill, New York.
23. Noriega, L. 2005. Multilayer Perceptron Tutorial, School
of Computing, Staffordshire University, Staffordshire.
24. Poulton, M. M. Sternberg, B. K. and Glass, C. E. 1992.
Location of subsurface targets in geophysical data using
neural networks, Geophysics: 15341544.
25. Roth, G. and Tarantola, A. 1994. Neural networks and
inversion of seismic data, J. Geophys. Res.: 67536768.
26. Ramey, H.J. Jr. 1993. Advances in Practical Well-Test
Analysis, SPE 20592, Journal of Petroleum Technology,
SPE, USA.
27. Romeo, G. 1994. Seismic signals detection and
classication using articial neural networks, Annali di
Geosica: 343353.
28. Siegelmann, H.T. Horne, B.G. and Giles, C.L. 1997.
Computational Capabilities of Recurrent NARX Neural
Networks, IEEE Transactions on Systems, Man and
Cybernetics Part B: Cybernetics, 27 (2): 1083-4419.
29. Singh, S. 2005. Permeability Prediction Using Artificial
Neural Network (ANN): A Case Study of Uinta Basin,
SPE-99286-STU, SPE ATCE, Dallas, 9 10 October.
30. Sung, W. Inhang, Y. Seunghoon, R. and Heungjun, P.
1996. Development of the HT-BP Neural Network System
for the Identification of a Well-Test Interpretation Model,
SPE 30974, SPE Eastern Regional Meeting, Morgantown,
17-21 September.
31. Vaferi, B. Eslamloueyan, R. and Ayatollahi, S. 2011.
Automatic recognition of oil reservoir models from well
testing data by using multi-layer perceptron networks,
Journal of Petroleum Science and Engineering, Elsevier
B.V., 77 : 254-262.
32. Van der Baan, M. and Jutten, C. 2000. Neural networks in
geophysical applications, Geophysics, SEA, 65 (4): 10321047.
33. Wang, L.X. and Mendel, J. M. 1992. Adaptive minimum
prediction error deconvolution and source wavelet
estimation using Hopeld neural networks, Geophysics:
670679.
34. White, A.C. Molnar, D. Aminian, K. Mohaghegh, S.
Ameri, S. and Esposito, P. 1995. The Application of ANN
for Zone Identification in a Complex Reservoir, SPE
30977, SPE Eastern Regional Conference & Exhibition,
Morgantown, West Virginia, 19 21 September.
35. Zhang, Y. and Paulson, K. V. 1997. Magnetotelluric
inversion using regularized Hopeld neural networks,
Geophysics Prosp.: 725743.

OTC-24981-MS

Figure 1: A Schematic diagram of drawdown test of a homogeneous reservoir with innite acting boundary (Vaferi et al. 2011)

(a)

(b)

Figure 2: (a) Three polynomials making up a cubic spline; (b) Input curve for cubic spline interpolation

OTC-24981-MS

Figure 3: Mathematical neuron (Baan et al., 2000)

Figure 4: A fully connected recurrent Time Delay neural network (Siegelmann et al., 1997)

Figure 5: A NARX neural network with ny = 2, nu = 3 and H= 4 (Lin et al., 1998)

Figure 6: Overall flowchart for synthetic pressure transient data generation system

OTC-24981-MS

Figure 7: Case evaluating infinitely large field with 5 wells

5100

Pressure (psia)

4900
4700
4500
4300

actual

4100

ann

3900
3700
3500
210

220

230

240

250

260

270

280

290

Time (hr)
Figure 8: Comparison of simulator and ANN data for Training Well 1

5100

Pressure (psia)

4900
4700
4500
4300

actual

4100

ann

3900
3700
3500
210

220

230

240

250

260

270

Time (hr)
Figure 9: Comparison of simulator and ANN data for Training Well 2

280

OTC-24981-MS

5100

Pressure (psia)

4900
4700
4500
4300

actual

4100

ann

3900
3700
3500
210

220

230

240

250

260

270

280

290

Time (hr)
Figure 10: Comparison of simulator and ANN data for Training Well 3

5100

Pressure (psia)

4900
4700
4500
4300

actual

4100

ann

3900
3700
3500
210

220

230

240

250

260

270

280

290

Time (hr)
Figure 11: Comparison of simulator and ANN data for Training Well 4

5100

Pressure (psi2a)

4900
4700
4500
4300
4100

ann

3900
3700
3500
210

220

230

240

250

260

270

Time (hr)
Figure 12: NARX model prediction result for Well 5

280

290