A Novel Neural Network Framework For The Prediction of Drilling Rate of Penetration

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/316963557

A Novel Neural Network Framework For The Prediction Of Drilling Rate Of


Penetration

Conference Paper · March 2017

CITATIONS READS

0 71

2 authors:

Saurabh Tewari U.D. Dwivedi


Rajiv Gandhi Institute of Petroleum Technology Rajiv Gandhi Institute of Petroleum Technology
12 PUBLICATIONS   17 CITATIONS    24 PUBLICATIONS   242 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Lithofacies identification View project

All content following this page was uploaded by Saurabh Tewari on 16 May 2017.

The user has requested enhancement of the downloaded file.


APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

A Novel Neural Network Framework For The Prediction Of Drilling Rate Of


Penetration
1Saurabh Tewari, 2Umakant Dhar Dwivedi*
1, 2
Department of Petroleum Engineering, Rajiv Gandhi Institute of Petroleum Technology, Jais,
Uttar Pradesh.
Email:saurabhtewari@gmail.com, udwivedi@rgipt.ac.in
Mobile No-08853946604, 09616467097
Abstract: Drilling rate of penetration is an important concern during drilling operations. Proper
rate of penetration reduces drilling cost and time, particularly enhances drilling efficiency. To
understand the effect of various drilling parameters on the rate of penetration, conventional
Bourgoyne and Young model has been taken into consideration. The unknown coefficients of
Bourgoyne and Young model were normally estimated through multiple regression analysis.
However, in certain conditions multiple regression analysis may results in physically
meaningless value of coefficients making conventional model unreliable. In this paper, optimum
artificial neural network based estimator model has been developed and tested on Kinabalu field
data for the prediction of drilling rate of penetration. The proposed ANN model provided
predictions with a RMSE of 0.39964, a MAPE of 0.084232 and R2 of 0.86174. From the
comparison of the results, it can be concluded that the proposed ANN based rate of penetration
model outperforms the conventional Bourgoyne and Young model for prediction of drilling rate
of penetration.
Keywords: Artificial neural networks, Bourgoyne and Young model, Rate of penetration.
INTRODUCTION
Drilling parameters are adjusted to make the drilling operations faster, economical, effective
in nature and to achieve optimum value of drilling rate of penetration (ROP). Proper ROP
minimizes the likelihood of facing numerous complications [1] [2]. The important parameters
which effects the drilling rate of penetration are depth, bit number, weight on bit/hook load,
rotary speed, tooth wear Reynolds number function, ECD and pore gradient. Unluckily, there is
no perfect correlation found between drilling rate of penetration (ROP) and various drilling
parameters due to their nonlinear and complex relationships [2]. Numerous researchers struggled
to develop a mathematical model for the estimation of ROP. Speer (1958) was first to
demonstrate the empirical relationships of ROP, weight on bit, rotary speed, hydraulic horse
power and formation drillability [8]. Graham and Muench (1959) analytically calculated weight
on bit and rotary speed combinations and mathematical empirical relationship for bit life
expectancy along with drilling rate as function of depth, rotary speed, and bit weight [5].
Bourgoyne and Young’s (BY) in 1999 finally developed a simplified mathematical model which
was accepted and adopted widely. Eren and Ozbayoglu (2011) explained the procedure of
statistical synthesis of the past data for the prediction of ROP in real time [4]. Normally, the
unknown coefficients of BY model were determined by multiple regression (MRA) technique.
APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

Unfortunately, in certain conditions MRA results in physically meaning less values of


coefficients like zero and negative values [2]. To achieve more reasonable and reliable results,
Non-linear least square data fitting with trust-region method was also implemented whose results
were not entirely satisfactory in nature [1]. Bahari et al. (2008) applied genetic algorithm (GA) to
calculate the coefficients of BY [2]. However, GA tuned MRA model does not show excellent
prediction agreement between estimated ROP and actual ROP. Therefore, there is urgent need of
new techniques which can estimate ROP with high degree of accuracy and reliability.
Recently artificial neural network (ANN) has gained popularity in the field of estimation and
pattern recognition because of its accurate prediction and generalization capability. ANN has an
efficient competency of combining different parameters as input to predict parameter as an
output in various conditions [7]. In this paper, extensive study is done to develop optimum ANN
based estimator model for the prediction of drilling ROP. Despite of changing mud properties,
lithology, bit type, hole size conditions, ANN can predict ROP efficiently due to its flexibility in
structural design. A comparative study is also done between conventional BY ROP model and
ANN based ROP model using three statistical parameters i.e. mean absolute percentage error
(MAPE), root mean square error and (RMSE) and coefficient of correlation (R2).

Figure 1. Proposed ANN framework ROP estimator model.


APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

MODEL DEVELOPMENT
Bourgoyne and Young ROP model
Bourgoyne and Young have proposed a simple ROP prediction model for roller cone bit. The
drilling ROP can be given as:
8
 
ROP  Exp  a1   ai xi  (1)
 i 2 
where xi are various functional drilling parameters which effect the ROP and a1 to a8 are
constants (Table 1 & 2). Drilling behavior of a given formation type is modelled through the
selection of constants from a1 to a8 in equation 1. BY determined the value of constants a1 to a8
through the MRA of field data.
Table 1. Functional relationships associated with ROP model [3].
Serial Equations Effect description
No
1 nX1 n is number of data point
2 X  10, 000.0  De Depth
2
3 1 Formation Compaction
X  De 0.69  G  9.0 
3  p 2
4 Differential pressure
X  D e  G  P 
4  p c
5  W W  W   Bit weight and bit diameter
X  ln    4.0    
5
 d  d t  d t 
6 X  ln  N 100  Rotary speed
6
7 X  H Tooth wear
7
8 X  pq 350  d Fluid properties
8 n

The first constant a1 symbolizes the effects of formation strength, bit types, mud types and solid
content which are not included in ROP model. The terms a2x2 and a3x3 represent the effect of
compaction on the drilling rate. The effect of overbalance on penetration rate is represented by
the a4x4 term. The effect of bit weight and rotary speed is included in the terms a5x5 and a6x6.
Remaining terms a7x7 and represents the effect of tooth wear and characterizes the effects of bit
hydraulics on ROP [3]. The functional relationships associated with equation 1 are listed in
Table 1.The coefficients a1 to a8 must be determined for each formation because of their
APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

dependency toward local drilling conditions [3]. BY has also suggested upper and lower for each
of eight coefficients after testing the different formations as shown in Table 2 [3].
Table 2. Recommended upper and lower bound for Bourgoyne and Young ROP model [3].
Coefficients a1 a2 a3 a4 a5 a6 a7 a8
Lower 0.5 0.0000001 0.0000001 0.000001 0.5 0.4 0.3 0.3
Bound
Upper 1.9 0.0005 0.0009 0.001 0.2 0.1 1.5 0.6
Bound

BY has determined the values of eight coefficients by utilizing the concept of MRA and
recommended that minimum thirty data points are required [3]. MRA gives results out of
recommended bounds in certain conditions results in zero or negative values also. Zero or
negative values of coefficients are practically worthless in nature. This method is also limited to
the recommended ranges for drilling parameters and results are not highly reliable in nature.
Therefore, new technique is required which can predict ROP in highly accurate way.

ARTIFICIAL NEURAL NETWORKS


ANNs are based on the neurons existing inside human brain. ANN workflow structure
resembles the nervous system featuring the network of interconnected large number of simple
processing neurons. The modeled neuron is known as node as shown in figure. Each neurons
element takes n input to produce m output given by a function known as activation function.
n
m f( wi xi  b) (2)
i 1
Where xi are the ith input, wi is the weight associated with each input, b is bias and f is activation
function. Every input is associated with its weight and output is calculated as the weighted sum
of its inputs, called the net sum or net which have too passed through the activation function f.
Y=f (net). Commonly used activation functions are logistic sigmoid function, hyperbolic tangent
sigmoid function, purline etc.
Rosenbatt, coined the term perception showed that neuron has learning capability by updating
weights through a simple learning algorithm.

w(t 1)  w(t) w i xi (3)

All the xi which are misclassified by previous iteration is the learning rate and is the error
associated with input. Many nodes combined together to form layers and these layers combine to
form network. The layers are divided into three types i.e. input layer, hidden layer and output
layer. Number of neurons and layers depends upon the complexity of problems. For training
purpose back propagation algorithm is most suitable for multilayer perceptron neural network.
The data is processed through input layer, then to hidden layer and finally to output layer. The
APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

error between the output and target value are measures and passed backward by updating
individual weights and bias neurons layers.

Figure 2. Schematic of a neuron.


This process repeated continuously until the training error reduces to its minimum value. Error is
calculated by mean square error (MSE). The input data is consisting of eight drilling parameters
and ROP as output parameters.

RESULTS AND DISCUSSION


The proposed ANN model is developed and tested on Kinabalu field drilling dataset to build a
comprehensive model [3][6]. Sample of data utilized in the development of ROP estimator
model is shown in Table 4. The developed model requires eight different drilling parameters, for
prediction of ROP. Acquired dataset is randomly divided for training, cross validation and
testing of ANN based ROP model. The training dataset used 70% of database, 15% is provided
for validation and rest 15% for testing phase. The data were normalized between +1 and -1 with
the average values set to zero.
X i - X Min
Xi  (4)
Norm X Max - X Min

where XMin and XMax are minimum and maximum values of the corresponding datasets. Outlier
analysis is also done to detect and remove unwanted noisy data points. In this research work, for
the estimation of ROP, ANN models are developed with eight inputs (BY parameters) and ROP
as a single output. Optimum number of neurons and hidden layers essential for the prediction of
ROP were determined by minimum RMSE value. The different ANN network architectures were
developed and tested on available data and recorded in Table 3. Early stopping was implemented
to avoid the problem of over fitting and improvise the generalization performance of ANN. The
learning and generalization ability of ANN model is evaluated on the basis of three statistical
parameters i.e.
APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

N
1 Tm  OP
MAPE 
N

I 0 Tm
*100 (5)

N
1 2
RMSE 
N
 T
i 1
m  Op  (6)

N
2
 T m  Op 
R2  1 i 1
2 (7)
N N
 1 
 Tm  N  Tm i 
i 1  i 1 i

where Tm and Op is the measured and predicted value of ROP. The 8-20-1 feed forward ANN
architecture is found to be optimum for prediction of ROP with least RMSE (0.39964) and
highest coefficient of correlation R2 (0.86174). The comparison between BY and ANN model is
done in table 5 using three statistical parameters. The test results clearly indicate that ANN has
better prediction performance as compared with conventional Bourgoyne and Young model.
Table 3. Comparison between various neural network models.

S.No. Transfer Function Model MSE


1 Purlin 8-4-1 0.841
2 Logsin 8-6-1 0.7213
3 Tangsig 8-8-1 0.624
4 Purlin 8-10-1 0.647
5 Logsin 8-12-1 0.468
6 Tangsig 8-14-1 0.4852
7 Purlin 8-15-1 0.75433
9 Logsin 8-18-1 0.4786
10 Tangsig 8-20-1 0.39964

Table 4. Sample of data utilized for ROP estimator model [6].


S. No. Depth Bit Drilling rate Bit weight Rotary Tooth Jet ECD Pore
(ft) Number (ft/hr) (1,000 lb/in) speed wear Impact (lb/gal) gradient
(rpm) (lb/gal)
1 2150 2 171 0.82 120 -0.5 0.882 8.93 8.365
2 2155 7 20 0.57 110 -0.125 0.819 9.06 8.365
3 3591 8 160 0.82 120 -0.5 1.29 9.11 8.365
4 5190 10 82 1.63 120 -0.75 1.29 9.11 8.365
5 5872 11 49 2.45 120 -0.875 1.29 9.11 8.365
6 6000 12 43 2.45 120 -0.25 1.29 9.11 8.365
7 6080 16 64 1.63 120 -0.625 1.062 9.49 8.365
8 6322 17 36 2.45 120 -0.875 0.772 9.67 8.365
9 6592 18 27 2.85 120 -1 0.772 9.67 8.365
10 6679 19 14 0.41 120 -0.625 1.338 9.69 8.365
APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

Table 5: Sensitivity analysis of conventional ROP models and ANN based ROP estimator
models.

Model RMSE AMPE R2


BY+MVRA 0.497321 0.150785 0.7133
ANN 0.39964 0.084232 0.86174

Figure 3. Optimum Feedforward Back propagation ANN developed for ROP prediction.

Figure 4. Plot of predicted versus measured Bourgoyne and Young/ANN ROP model value.
APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

Figure 5. Comparison between estimated and actual ROPs.

CONCLUSIONS
A real time ANN estimator model is successfully developed and tested on Kinabalu field data for
the prediction of drilling rate of penetration. The results indicate high degree of estimation
accuracy as compared with conventional Bourgoyne and Young ROP model. The proposed ANN
model can be adjusted according to the changing wellbore conditions and availability of data in
real time operation. The drilling parameters such as weight on bit and bit rotation speed are
adjusted to achieve most effective rate of penetration for drilling operations. Real time analysis
of ROP helps the drilling engineer for proper selection of various drilling and hydraulic
parameters easily. The results clearly show that ANN based ROP model is more efficient for
drilling oil and gas fields and can be used for proper planning and successful execution of real
time drilling operations.
APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

ACKNOWLEDGEMENT
We thank Rajiv Gandhi Institute of Petroleum Technology for providing us financial
support and research infrastructure.

NOMENCLETURE

a1 = formation strength parameter

a2 = exponent of the normal compaction trend

a3 = under compaction exponent

a4 = pressure differential exponent

a5 = bit weight exponent

a6 = rotary speed exponent

a7 = tooth wear exponent

a8 = hydraulic exponent

b = bearing constant
B = fractional bearing wear
d = bit diameters, in
D= well depth, ft

W/ d = weight on bit per inch of bit diameter, 1,000 lb/in

N = rotary speed, rpm


u= a function of bit weight per inch, rotary speed, and rotating time
REFERENCES
[1]. A. Bahari, B. A. Syed, “Trust region approach to find constants of Bourgoyne and young
penetration rate model in khangiran Iranian gas field” in Lat. Ameri. & Carri. Petro. Eng Conf.,
SPE-107520-MS, April, 2007
[2]. M. H. Bahari, A. Bahari, and N.F. Moharrami, “Determining Bourgoyne and Young Model
Coefficients using Genetic algorithm to predict drilling rate Asian network for scientific
information” in J. Appl. Sci.8 (17):3050-3054, 2008.
[3]. A.T. Bourgoyne, and F. S. Young, “A multiple regression approach to optimal drilling and
abnormal pressure detection” in SPE-AIME Sixth Conf. on Drill. and Ro. Mec. SPE 4238, 1973.
[4]. T. Eren, and M. E. Ozbayoglu, “Real-Time drilling rate of penetration performance monitoring”
in SPE Oil and gas In. Conf. and Exh., SPE 129126-MS, Mumbai, India, 2010.
APCEC17 INTERNATIONAL CONFERENCE ON ADVANCES IN PETROLEUM CHEMICAL
& ENERGY CHALLENGES

[5]. J. W. Graham and N.L. Muench, “ Analytical determination of optimum bit weight and rotary
speed combinations” in SPE-AIME 34th Annual Fall Meeting, SPE 1349-G, Dallas Oct. 4-7,
1959
[6]. S. Irawan, A.M.A. Rahman, and S.Q. Tunio, “Optimization of weight on bit during drilling
operation based on rate of penetration model” in Res. J. of App.Sci.Vol4 (12) pp 1690-1695,
2012.
[7]. H. Rahimzadeh, M. Mostofi, A. Hashemi, and Salahshoor, “Comparison of the penetration rate
models using field data for one of the gas field in Persian Gulf area” in SPE International Oil and
Gas Conf. and Exhi., SPE 131253, Beijing, China, 2012.
[8]. J.W. Speer, “A method for determining Optimum drilling Technique” in Drill. and Prod. Prac. ,
API, 1958.

View publication stats

You might also like