Sunteți pe pagina 1din 12

SPE 68420

Using an Artificial Neural Network to Develop a Wall-Loss Model for Coiled Tubing
Fracturing Operations

Keith Rispler, SPE, Joanne McNichol, SPE, Kevin Matiasz, SPE, Halliburton Energy Services, Inc., and Mark
Rheinlander, SPE, Quality Tubing

Copyright 2001, Society of Petroleum Engineers, Inc. Operating within the safe working envelope of the CT is
This paper was prepared for presentation at the SPE/ICoTA Coiled Tubing Roundtable held in critical to the success of these operations. Wall loss from ero-
Houston, Texas, 7-8 March 2001.
sion affects operational CT parameters, and previously identi-
This paper was selected for presentation by an SPE Program Committee following review of fied patterns of erosion1 required a better understanding of this
information contained in an abstract submitted by the author(s). Contents of the paper, as presented,
have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the phenomenon. Three of the 35 strings used during CT fracturing
author(s). The material, as presented, does not necessarily reflect any position of the Society of
Petroleum Engineers, its officers, or members. Papers presented at SPE meetings are subject to operations performed in the year 2000 were evaluated for wall
publication review by Editorial Committees of the Society of Petroleum Engineers. Permission to loss. Magnetic wall-loss detection was performed with Hall-
copy is restricted to an abstract of not more than 300 words. Illustrations may not be copied. The
abstract should contain conspicuous acknowledgment of where and by whom the paper is presented. effects sensors. The strings were measured at 82-ft (25-m) inter-
Write Librarian, SPE, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A, fax 01-972-952-9435.
vals and at 30° intervals around the circumference of the string.
The data from one string clearly indicate maximum erosion loss
on the outer radius of the pipe and minimal wear on the inner
Abstract radius of the pipe (Table 1). Wall thickness for all three strings
Coiled tubing (CT) erosion can occur during CT fracturing is plotted in Figs. 1, 2, and 3. The wall loss from the tests
operations. Resulting wall loss and fatigue can limit CT life and demonstrates a repetitive pattern similar to those previously
prevent safe wellsite operations. An artificial neural network identified.1 Pipe thickness affects critical CT operating param-
(ANN) has been successfully developed for accurately predict- eters, including fatigue, internal pressure capacity, and tensile/
ing wall loss resulting from erosion. compressive loading. A method for accurately predicting ero-
This paper presents a case history in which ANN technol- sion could help operators use traditional CT pipe-management
ogy was used to successfully manage tubing strings during CT simulators to effectively manage CT fracturing strings.
fracturing operations. During these operations, wall loss affects
CT pressure ratings, tensile strength, and fatigue, all of which are Using an ANN to Develop a Wall-Loss Model
critical performance parameters used for determining CT life and Several technical papers have been written about the develop-
identifying a safe operating envelope. An ANN can predict ment of ANNs. Neural nets have been given many different names,
erosional wall loss and quantify critical performance parameters including black boxes, empirical models, universal approximators,
for specific applications. and parallel models. The basic function of an ANN is to map data
from one multidimensional space to another. Fig. 4 shows an
Introduction example of a simple ANN with three inputs, two hidden layers with
Hydraulic fracture-stimulation treatments performed through two neurons in each layer, and one output. The example has 12
CT provide a cost-effective method of stimulating wells in which unknown weights. The weights must be determined so that the
producing intervals have multiple stringers. This technique is inputs can be properly mapped to the corresponding outputs.
being successfully applied in many areas and is being used on a ANNs are constructed of neurons organized in layers. An
daily basis in the shallow gas fields of southern Alberta. Fractur- ANN can consist of a single layer or multiple layers, and each
ing operations in these shallow gas fields typically require a layer consists of one or more neurons. Artificial neurons receive,
2 3/8-in. (0.203-in.) QT-900 tubing string that is 2,625 to 2,789 ft consider, and calculate input from other sources. This informa-
(800 to 850 m) in length. The zones targeted for stimulation are tion is then displayed by an algorithmic process or transfer
located at depths of 984 to 2,362 ft (300 to 720 m). Typical function. The number of layers, neurons in each layer, and
treatments involve three to eight stages in which a total of connecting transfer functions depend on the problem that must
110,231 to 220,462 lb (50 to 100 tonnes) of proppant is pumped. be solved.
In addition, most treatments involve the use of a gas-assist with Data from both the CT fracturing-service treatments and CT
either carbon dioxide or nitrogen added to the stimulation fluid. wall-measurement tests were compiled for analysis. All data

References at the end of the paper.


USING AN ARTIFICIAL NEURAL NETWORK TO DEVELOP A WALL-LOSS
2 SPE 68420
MODEL FOR COILED TUBING FRACTURING OPERATIONS

collected were initially used as “input” for building a neural net tric tubing can be determined from a uniform wall-thickness
model. The “output” was the wear measurements. After each calculation based on 80% of the predicted wall loss of uniformly
neural-net operation, data were correlated for determining the eroded tubing.
model’s ability to accurately predict wall reduction. Parameter Pressure ratings estimated by the CT simulator4 (including
responses were charted for determining the data fields that varying amounts of accumulated fatigue) have also been plotted
affected pipe wear. These interpretations were used for deter- in Fig. 11. A pressure rating has also been calculated from
mining the parameters that most affected pipe erosion: measured average wall thickness and tensile strength from the
• core-end measurement (Fig. 5) samples tested to failure. These curves show a good correlation
• total volume of sand pumped through a given core (Fig. 6) based on average thickness values. The calculated curve repre-
These parameters were used as the input data in a neural sents samples with fatigue ranges from the curves calculated by
network developed for String 13358_B. The correlation of pre- the CT simulator. Pressure-testing all samples to 10,000 psi
dicted wear vs. actual wear had an R2 match of 0.96 for the string further validated these curves.
(Fig. 7). The predicted and actual wear values were very similar Cumulative fatigue for the three strings tested has been
(Fig. 8). Numerous variations were run, and this model was calculated by the CT simulator4 (Figs. 12, 13, and 14). These
determined to be capable of accurately predicting erosion on the values include progressive wall-reduction factors. However,
other strings. This model never significantly underestimated this model cannot easily simulate nonuniform wall loss, and all
erosion (Figs. 9 and 10). In addition, this model was capable of strings were modeled assuming uniform wall loss along the
using data from strings used the previous year to accurately entire length and circumference of the string. Sections from the
predict erosion. maximum erosion points of these strings were cycled to failure
for estimating the CT simulator’s4 ability to accurately calculate
fatigue and verify that a correct application or safety factor is
Correlating Critical Performance Parameters to Wall
being used. The tests were performed on a 72-in. radius goose-
Loss
neck, but actual field operations included a reel with a 110-in.
Yield- and tensile-strength tests were performed on the string
diameter and a gooseneck with a 96-in. radius, resulting in a
sections exhibiting maximum erosion (Table 2). These data dem-
predicted fatigue variance of approximately 6% at the 750-psi
onstrate that a trend of yield strength compared to average wall
test pressure. These tests indicated that the CT simulator4 accu-
thickness was observed. The samples used in these tests exhib-
rately calculated fatigue for these strings and allowed an appro-
ited significant cyclic softening; the load-displacement traces
priate safety factor (Fig. 15).
exhibited significant curvature during the elastic-to-plastic tran-
sition.
Surprisingly, the yield strength exhibited during the tests Managing Pipe with the ANN Model
correlated more strongly to wall loss than to predicted fatigue Before ANNs were used for predicting wall loss, the primary
(Table 3). However, both wall loss and fatigue were shown to method of managing CT fracturing strings was to use a CT
impact the tensile strength of the pipe. These data appear to fatigue simulator as the single determining factor. The intent of
support combined-loading calculations derived from average managing these fracturing strings is to calculate the fatigue
wall-thickness values run in a CT simulator. The data collected throughout the life of the CT string by relating travel path across
from these strings indicate that the average wall thickness (12 the gooseneck and reel so that safe working parameters can be
points) of the pipe in the most highly eroded sections is 4% maintained. The wall reduction per treatment was based on
greater than the average of nominal and minimum wall thickness. previous measurements1 and entered as a minimum wall thick-
Therefore, an average of nominal and minimum wall thickness ness.
appears to be an appropriate calculation for predicting tensile The addition of the ANN model complements the fatigue
strength. simulator by providing a predicted input for fatigue calculations
High treating pressures are typical during CT fracturing and a safe operating envelope for combined tubing loading. The
operations, and screenouts and pressure spikes can occur under ANN model predicts wall reduction by correlating the volume of
these conditions. Therefore, operators must understand the sand pumped with the corresponding string positions. Wall
pressure ratings of CT fracturing strings. Fig. 11 shows that the reduction and associated fatigue calculations are then used as
internal yield-strength capacity of 2 3/8-in. (0.203-in.) eccentri- input for the CT simulator.4 The ANN model uses this input to
cally eroded pipe with a wall thickness of 0.12 to 0.16 in. is determine a maximum wall reduction that can be used in the
750 psi greater than that of uniformly eroded pipe with a similar fatigue simulator. The fatigue simulator then applies this reduc-
ID and wall thickness. For new pipe with a wall thickness of tion to the entire length and circumference of the string. How-
0.133 in. and an internal yield pressure of 10,000 psi, a 0.015-in. or ever, string measurements indicate that most string erosion
1,000-psi variance exists, which corresponds to a minimum wall- occurs on the reel after the core wrap and on the outside curve
thickness variance of approximately 20% for eccentrically and of the pipe. Therefore, applying the maximum wall reduction to
uniformly eroded tubing. Therefore, the wall thickness for eccen- the entire length and circumference of the string may be conser-
SPE 68420 K. RISPLER, J. MCNICHOL, K. MATIASZ, M. RHEINLANDER 3

vative. Perhaps a more accurate approach would be to adjust the used in similar applications. Because the treatments pumped
predicted wall thickness to an average wall thickness based on through the three CT strings had limited variability, factors such
nominal and minimum thicknesses for calculation of critical as pump rate did not appear to have a significant effect on wall
parameters. loss. Consequently, this model is limited in scope and may only
The wall thickness of a CT fracturing string should allow the be applicable to treatments that fit within the window of inves-
string to withstand pressures and tensile stresses encountered tigation.
during normal fracturing operations. Often, working pressures The nondestructive and destructive tests that were per-
reach 40 to 45 MPa, and sudden screenout conditions can formed on these strings indicate that using an effective wall
produce pressure spikes as high as 70 MPa. In addition, situa- thickness calculated from wall-loss predications from the ANN
tions arise during which a significant tensile strength is required model can accurately predict CT pipe parameters.
to either allow the BHA to be moved or to be able to part the shear
assembly if the BHA cannot be moved. Therefore, CT should References
have a minimum wall thickness capable withstanding pressure 1. Kazakov, S. and Rispler, K.: “Optimizing and Managing Coiled
spikes or increased tension without experiencing catastrophic Tubing Frac Strings,” paper SPE 60747 presented at the 2000
failure. The objective of managing a CT fracturing string is to SPE/IcoTA Coiled Tubing Roundtable, Houston, April 5-6.
complete as many wells as possible within safe operating param- 2. Coiled Tubing Technical Catalog, Quality Tubing Inc., Houston
eters. Effective string management depends on accurately pre- (1997).
dicting wall-loss erosion and resulting fatigue. 3. Engineering Design, Second Edition, Faupel, J.H. and Fisher,
F.E., John Wiley and Sons, New York (1981).
4. Cerberus Coiled Tubing Modeling Software Version 5.0 with
Conclusion Halliburton Fatigue Module. [CD-ROM] Conroe, Texas: Coiled
The ANN developed from analysis of three CT strings can be Tubing Engineering Services (1999).
used to determine wall loss resulting from erosion for CT strings
USING AN ARTIFICIAL NEURAL NETWORK TO DEVELOP A WALL-LOSS
4 SPE 68420
MODEL FOR COILED TUBING FRACTURING OPERATIONS

Table 1 String 1117A UT Thickness Survey


Meters 1 2 3 4 5 6 7 8 9 10 11 12
0 0.196 0.198 0.198 0.199 0.2 0.203 0.204 0.214 0.205 0.206 0.205 0.2
12.5 0.188 0.194 0.196 0.2 0.202 0.209 0.214 0.223 0.208 0.203 0.202 0.198
37.5 0.176 0.188 0.198 0.21 0.216 0.218 0.215 0.22 0.21 0.194 0.191 0.182
62.5 0.168 0.17 0.18 0.197 0.207 0.214 0.214 0.213 0.202 0.192 0.181 0.175
87.5 0.165 0.172 0.185 0.2 0.207 0.22 0.214 0.212 0.206 0.202 0.192 0.18
112.5 0.162 0.168 0.189 0.197 0.204 0.215 0.212 0.213 0.202 0.2 0.191 0.18
137.5 0.142 0.163 0.192 0.202 0.206 0.209 0.212 0.206 0.202 0.19 0.173 0.187
162.5 0.147 0.149 0.17 0.189 0.195 0.2 0.204 0.191 0.19 0.185 0.161 0.16
187.5 0.161 0.165 0.17 0.189 0.2 0.202 0.202 0.205 0.2 0.198 0.185 0.175
212.5 0.146 0.149 0.151 0.189 0.2 0.202 0.202 0.195 0.185 0.177 0.165 0.155
237.5 0.148 0.152 0.171 0.191 0.198 0.198 0.2 0.195 0.192 0.18 0.16 0.152
262.5 0.152 0.175 0.189 0.189 0.201 0.201 0.201 0.195 0.19 0.185 0.18 0.163
287.5 0.154 0.169 0.19 0.19 0.202 0.202 0.202 0.192 0.19 0.186 0.178 0.157
312.5 0.166 0.174 0.184 0.195 0.2 0.204 0.206 0.192 0.19 0.186 0.171 0.166
337.5 0.176 0.184 0.19 0.19 0.198 0.2 0.201 0.201 0.204 0.195 0.184 0.174
362.5 0.186 0.188 0.195 0.197 0.202 0.205 0.204 0.203 0.2 0.198 0.193 0.186
387.5 0.196 0.194 0.194 0.197 0.2 0.203 0.206 0.193 0.196 0.198 0.196 0.196
412.5 0.197 0.196 0.2 0.203 0.2 0.206 0.21 0.193 0.193 0.197 0.195 0.196
437.5 0.198 0.198 0.2 0.2 0.204 0.201 0.2 0.195 0.192 0.197 0.198 0.199
462.5 0.2 0.2 0.197 0.2 0.199 0.204 0.209 0.208 0.204 0.202 0.201 0.201
487.5 0.199 0.197 0.197 0.2 0.198 0.202 0.202 0.208 0.202 0.202 0.2 0.2
512.5 0.196 0.2 0.198 0.196 0.198 0.198 0.195 0.193 0.196 0.197 0.198 0.197
537.5 0.2 0.2 0.2 0.2 0.199 0.2 0.197 0.198 0.193 0.201 0.202 0.202
562.5 0.197 0.2 0.2 0.2 0.201 0.2 0.193 0.195 0.199 0.2 0.198 0.198
587.5 0.177 0.18 0.2 0.2 0.2 0.199 0.195 0.2 0.201 0.202 0.203 0.197
612.5 0.198 0.198 0.198 0.2 0.201 0.194 0.205 0.2 0.195 0.196 0.198 0.198
637.5 0.198 0.198 0.197 0.198 0.198 0.196 0.195 0.2 0.205 0.202 0.2 0.2
SPE 68420 K. RISPLER, J. MCNICHOL, K. MATIASZ, M. RHEINLANDER 5

Table 2 Yield and Tensile Strength Data


Tensile
Avg.Min Wall Avg. Wall Avg. OD, Area Yield Strength, Yield Strength,
Sample Strength
(in.) (in.) (in.) (in.2) psi (0.2% Offset) psi (1.0% Offset)
(psi)
A-1 0.149 0.194 2.391 1.339 84,800 94,400 100,600
A-2 0.169 0.189 2.38 1.301 84,600 95,800 103,500
A-3 0.159 0.193 2.374 1.322 82,500 93,200 101,900
B-1 0.134 0.176 2.384 1.221 78,900 89,900 97,600
B-2 0.132 0.181 2.381 1.251 75,800 93,100 95,400
B-3 0.122 0.18 2.392 1.251 79,600 88,500 95,400
C-1 0.155 0.186 2.382 1.283 81,100 80,000 99,600
C-2 0.157 0.189 2.388 1.306 79,100 89,100 98,700
C-3 0.155 0.182 2.392 1.264 81,200 93,000 100,900

Table 3 Fatigue, Wall Thickness, and Tensile Strength


String Avg. Fatigue Avg. Wall Thickness Avg. Tensile Strength
13358_A 12% 0.159 83966
13358_B 31% 0.129 78100
13358_ 31% 0.155 80466

Fig. 1—String 13358 wall thickness.


USING AN ARTIFICIAL NEURAL NETWORK TO DEVELOP A WALL-LOSS
6 SPE 68420
MODEL FOR COILED TUBING FRACTURING OPERATIONS

Fig. 2—String 13358_B wall thickness.

Fig. 3—String 13358_C wall thickness.


SPE 68420 K. RISPLER, J. MCNICHOL, K. MATIASZ, M. RHEINLANDER 7

Fig. 4—Simple artificial neural network (ANN).

Fig. 5—Neural response curve.


USING AN ARTIFICIAL NEURAL NETWORK TO DEVELOP A WALL-LOSS
8 SPE 68420
MODEL FOR COILED TUBING FRACTURING OPERATIONS

Fig. 6—Neural response curve.

Fig. 7—Actual wear vs. ANN predicted wear for Model 13358_B C&T 96.
SPE 68420 K. RISPLER, J. MCNICHOL, K. MATIASZ, M. RHEINLANDER 9

Fig. 8—Model 13358_B C&T 96.

Fig. 9—Model 13358_ C&T 96, predicted 13358_A vs. actual 13358_A.
USING AN ARTIFICIAL NEURAL NETWORK TO DEVELOP A WALL-LOSS
10 SPE 68420
MODEL FOR COILED TUBING FRACTURING OPERATIONS

Fig. 10—Model 13358_B C&T 96, predicted 13358_C vs. actual 13358_C.

Fig. 11—CT Yield and burst pressures for eroded QT-900 2.38 x 0.203w.
SPE 68420 K. RISPLER, J. MCNICHOL, K. MATIASZ, M. RHEINLANDER 11

Fig. 12—String 13358 2 3/8-in. 0.203 wall fatigue.

Fig. 13—String 13358_1 2 3/8-in. 0.203 predicted fatigue.


USING AN ARTIFICIAL NEURAL NETWORK TO DEVELOP A WALL-LOSS
12 SPE 68420
MODEL FOR COILED TUBING FRACTURING OPERATIONS

Fig. 14—String 13358_B 2 3/8 -in. 0.203 wall fatigue.

Fig. 15—Fatigue, simulated vs. test.

S-ar putea să vă placă și