Hadietal 2019
Hadietal 2019
net/publication/337129737
CITATIONS READS
13 198
3 authors, including:
All content following this page was uploaded by Farqad Hadi on 24 May 2020.
This paper was prepared for presentation at the Abu Dhabi International Petroleum Exhibition & Conference held in Abu Dhabi, UAE, 11-14 November 2019.
This paper was selected for presentation by an SPE 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 Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers 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 SPE copyright.
Abstract
Over the years, the prediction of penetration rate (ROP) has played a key rule for drilling engineers due it
is effect on the optimization of various parameters that related to substantial cost saving. Many researchers
have continually worked to optimize penetration rate. A major issue with most published studies is that
there is no simple model currently available to guarantee the ROP prediction.
The main objective of this study is to further improve ROP prediction using two predictive methods,
multiple regression analysis (MRA) and artificial neural networks (ANNs). A field case in SE Iraq was
conducted to predict the ROP from a large number of parameters. A databases from one well drilled in
carbonate environments were subjected to the predictive methods. Each raw dataset is described by eight
parameters including rate of penetration (ROP), true vertical depth (TVD), weight on bit (WOB), bit
rotational speed (RPM), torque (TQ), flow rate (Q), equivalent circulating density (ECD), standpipe
pressure ( SPP), and unconfined compressive strength (UCS). First, both MRA and ANNs are tested as
predictive methods. The prediction capacity of each model was also verified by using two-based error
metrics: the determination coefficient (R2) and the mean square error (MSE ).
The current results support the evidence that MRA and ANNs are able to be effectively utilize the drilling
data, and thus provide accurate ROP prediction. However, more attention to the multiple regression analysis
is required where it is implemented for ROP prediction. ANNs appear to be more conservative in predicting
ROP than MRA as indicated by a higher value R2 (0.96) and lower value MSE (1.89) of the ANN model.
Considering the input parameters, the obtained results showed that TVD, WOB, RPM, SPP, and ECD had
the greatest effect on estimated ROP-conditions, followed in decreasing by pump flow rate, drilling torque,
and rock strength. Another important point that highlights in this study is that the drilling rate may increase
with depth in carbonate rocks because of their heterogeneity. This study presents new models to estimate
ROP from other parameters which can help the driller to achieve an optimal drilling rate through monitoring
controllable parameters.
Introduction
It is important to maximize the penetration rate while maintaining the drilling operation safety. Rate of
penetration (ROP) is an indicator of drilling performance or it is the biggest variable affecting total rig
2 SPE-197663-MS
hours. According to Li et al. (2015) and Al-AbdulJabbar et al. (2018) , the expenditures of drilling phase
are around 50% of the well budget, and thus an accurate prediction of ROP can be useful in optimizing
drilling parameters to achieve the minimum cost per foot. Prediction of ROP can be also considered as a
consistent index to outline the changes of formation drillability by considering the effects of drilling
parameters, bit types, and bit wear on ROP model (Andrews et al., 2007; Hareland & Nygaard, 2007; Shi
et al., 2016; Elkatatny et al., 2017) . Furthermore, ROP prediction can help in making decisions when a
sudden slowness in the drilling process is happened. Moreover, many researchers have shown the
decreasing of ROP with burial depth in normal pressure sections, and thus any deviation from this trend
occurs when abnormally pressured zones are encountered (Jorden & Shirley, 1966; Hadi et al., 2019).
It is extremely important to compare the results of different computing and modeling techniques to find
ROP predictions. Such these techniques are multiple regression analysis (MRA), artificial neural networks
(ANNs), and simulation (Bourgoyne and Young, 1974; Andrews et al., 2007; Moradi et al., 2010;
Arabjamaloei & Shadizadeh, 2011; Moran et al., 2010; Al-AbdulJabbar et al., 2018) . The researchers have
identified various controllable and uncontrollable factors affecting ROP which can be divided into different
categories: (1) personnel/rig efficiency, (2) rock elastic and mechanical characteristics, (3) bit operating
conditions including bit type, weight on bit (WOB) and rotary speed (RPM), (4) bit hydraulics including
jet velocity and bottom hole cleaning, and (5) the drilling fluid properties including mud density and mud
viscosity (Ahmed et al., 2018) . However, modeling a universal ROP-model is still have some limitations.
First, most of the models in the existing studies have interconnected parameters, and changing one of them
can affect the others. Second, they did not consider all relevant parameters that affecting ROP which will
decrease the accuracy of the proposed models. Another reason is associated with the geological
uncertainties of uncontrollable parameters including rock properties (Shi et al., 2016; Al AbdulJabbar et al.,
2018; Anemangely et al., 2018).
Work performed by several investigators have presented many mathematical equations that relate ROP
to other drilling parameters. Galle and Woods (1963) developed an analytical model to calculate ROP
utilizing weight on bit, rotary speed, and bit type, but it fails to consider the formation rock properties. In
1974, Bourgoyne and Young presented a statistical model for predicting ROP from past drilling parameters
using MRA. The model is able to obtain a relatively better prediction results, and it is being widely used
for a real-time drilling optimization, but the results of this model is highly depending on eight undetermined
constants. The main limitation exposed by this model is the common way to determine these constants,
which is based on the human experience or using the multiple regression which are both not precise enough
(Bahari & Seyed, 2007; Anemangely et al., 2018) . Considering the MRA technique, a few mathematical
models have been found in the literature to predict ROP from other drilling data in the field of interest.
Unlike statistical approaches, the artificial neural networks (ANNs) have recently gained a great
popularity in petroleum engineering for their ability to learn from examples and their better performance in
capturing the system nonlinearity than conventional regression techniques. Jahanbakhshi et al. (2012)
proved the efficiency of ANN in drilling planning and real-time operation using some drilling parameters
in addition to formation type and rock mechanical properties. Moran et al. (2010) and Al-AbdulJabbar et
al. (2018) also showed ANNs can be effectively used to estimate ROP and total drill-time. Arabjamaloei
and Shadizadeh (2011) applied three techniques to estimate ROP. These techniques are artificial neural
networks, multivariate regression, and Bourgoyne-Young methods. A comparison of the results of the three
methods showed the superiority of ANNs than other methods (Anemangely et al., 2018). Elkatatny et al.,
2017 also demonstrated that ANNs have a high prediction capacity of ROP by taking into consideration the
influence of the drilling fluid properties and the drilling mechanical properties on ROP. However, a major
issue with most published field studies is that there is no empirical model currently available to estimate
ROP using ANNs.
SPE-197663-MS 3
New models that relate the drilling rate with the drilling depth, weight on bit, torque, rotary speed, flow
rate, circulating pressure, equivalent circulating density, and rock strength have been developed in this study
using artificial neural networks and multiple regression analysis. The predictive methods were trained and
validated based on offset data taken from one well drilled in Basrah oil fields, Iraq. Each raw dataset is
composed of nine parameters: seven drilling operational parameters, one uncontrollable parameter, and
penetration rate. To test whether an improvement is achieved in this study, two performance indicators
including determination coefficient (R2) and mean square error (MSE) have been used for a comparison
between the presented models.
Data Overview and Methodology
When the data is sampled and understood thoroughly, it can be described quantitatively by a mathematical
model. The real field data in this study were assembled from one well drilled in Basra oil fields, Iraq. The
production section of most Basrah oil fields is mainly composed of carbonate (limestone). Table 1 shows
the input parameters with their statistical evaluations, which contains 1705 drilling data points for each rate
of penetration (ROP), true vertical depth of drilling (TVD), weight on bit (WOB), torque (TQ), bit rotational
speed (RPM), flow rate (Q), standpipe pressure (SPP), equivalent circulating pressure (ECP). With respect
to the unconfined compressive strength (UCS) magnitudes, the equation that presented by Hadi et al. (2017)
has been used in this study to determine this mechanical property for carbonate formations as a function of
porosity.
Before presenting the input-output data to the training phase of ANNs, all the datasets were normalized
between the ranges [-1, 1] using Eq. 1, where Xnorm is the normalized value of input parameter, X is the input
parameter, and Xmin and Xmax are the minimum and maximum values of input parameters, respectively. In
fact, the scaling of data within these ranges [-1, 1] is an important task to be performed in order to meet the
algorithm requirements and to avoid computational problems (Zhang et al., 1998).
(1)
Two predictive methods were conducted in this study to estimate ROP using available drilling data in
the well of interest (Table 1) . These methods are multiple regression analysis (MRA) and artificial neural
networks (ANNs).
Table 1—Statistical evaluation of input parameters used in this study
predicting ROP using eight unknown functions; however, these functions requires many input parameters
which are not easy to be computed and the model did not provide satisfactory accuracy. In addition, only a
limited number of models have been found in the literature to estimate ROP using MRA.
In general, the ANN structure is composed of three components: Weights, biases, and activation function
(Figure 1). The inputs (Xj) for each neuron are multiplied by an adjustable weight factor of the neuron (Wi)
and summed together with the constant bias value (Qi). The resulting ni is the input to the activation function
yi, where i represents the number of neurons and j represents the number of input parameters. Simply, a
neuron performs two tasks: (1) a weighted summation of its input array with bias value which produced ni,
and (2) the application of activation transfer function to this summation to give an output yi. Various types
of the activation function including the sigmoid, linear, and threshold are possible but the tangent sigmoid
function (S-shaped), as nonlinear transfer function, was preferred in this study (Figure 1) . The advantage
of the tangent sigmoid function is to squash the possible output from a neuron to typically [-1, 1] by
differentiating the outputs for both the hidden layer and the output layer (Philip, 2001) . In other words, the
sigmoid function convert inputs to a continuously changing outputs (Han et al., 2019).
Three different processes are commonly applied in the creation of the ANN model: training, validation,
and testing.
SPE-197663-MS 5
Training Phase. A trained neural network can act as an expert in the formation category so that it is
possible to detect important predictive models by relating the input parameters with the desired output. In
this study, the supervised back-propagation neural network (BPNN) training algorithm is used because it
allows iteration and adjustment of the connection weights of the networks so that the overall error between
two outputs (desired and computed) can converge to a threshold minimum, ideally zero (Shahab, 2000;
Moran et al., 2010; Hadi & Nygaard, 2018) . Briefly, the basic steps needed for the BPNN learning
processes: enter the training file into the network; calculate the output; compare the desired and computed
outputs; and readjust the connection weights until the difference between the desired and computed outputs
is minimized.
Validation Phase. A cross-validation process is an important step to ensure that the ANN model can
successfully predict data not seen in the learning data set (Hadi & Nygaard, 2018) . In general, ANN models
have a large number of model parameters and can therefore overfit the training data, especially if the training
data are noisy (Shahin et al., 2004) . In this status, the network tends to memorize insignificant details of
calibration data, and thus diminishing the prediction capacity of the ANN model. Thereby, a cross-
validation process is applied to estimate when it is best to stop the training process in order to ensure that
oscillation did not occur. The raw dataset in Table (1) was divided into two subsets on an arbitrary basis:
training and validation, in the proportions of 70% and 30%, respectively.
Performance Measures
To check how well does the developed models truly represent the set of data, the prediction capacity of
multiple regression analysis and artificial neural networks were determined by using two performance
criteria: the determination coefficient (R2) and the mean square error (MSE). The R2 and MSE-values can
be computed using Eqs. 1 and 2, respectively.
(1)
(2)
Where n is the number of sample in the database; mi and pi are the simulated and observed values at step
i, respectively; and 𝑚
̅ and 𝑝̅are the mean and standard deviation of the observed values, respectively (Hadi
& Nygaard, 2018).
Figure 2—The partial regression leverage plots for ROP prediction using MRA.
SPE-197663-MS 7
Accordingly, the TVD, WOB, RPM, SPP, and ECD appear to be more conservative in predicting ROP
rather than the drilling torque, flow rate, and rock strength. While TVD, WOB, RPM, and SPP have highly
positive effect on ROP, ECD has a negative effect on ROP. Another notable point that can be observed
from Figure 2 is that the ROP is relatively increased with TVD, which is different from what is previously
thought to be, as the ROP should be decreased with increasing depth (i.e., decreasing rock strength). In fact,
this can be attributed to the heterogeneity of carbonate rocks because of their complex depositional
environments and digenetic processes which alter the rock mechanical properties to varying degrees (Hadi
& Nygaard, 2018) . In other words, these processes result in a complex pore structure in which ductile,
brittle, fractured rocks, and vugular pores may all exist within small interval (Hadi et al., 2017) , resulting
a large variance in ROP. This conclusion is agreed with the absence effect of UCS on ROP prediction
(Figure 2) . Seeing how the scattering around each curve in Figure 2 also proved the fact that it is extremely
difficult to evaluate the actual effect of each influential parameter on ROP predictions.
Figure 3 along with Eq. 3 represent the developed model for a prediction of ROP using MRA. The Pvalue,
R2, and MSE of the presented model are less than 0.001, 0.83, and 3.7361, respectively, which refer to the
superiority of MRA in predicting ROP. However, the drawbacks of using MRA is that the distribution of
the predicted output is more narrow than that of the original data set (Balan et al., 1995) . The negative
values are another inherent challenge associated with using MRA in predicting the output model (Figure
3).
(3)
Figure 3—Cross-correlation plot between actual and predicted ROP using multiple regression analysis.
Figure 4—ROP prediction using ANNs: Training (left) and validation (right).
Equation 4 along with Table 2 present the weights and biases of the neurons of the ANN-model, where
n is the normalized value, i is the index for neurons, is the weights between input and hidden layers for
neuron i, is the weights between hidden and output layers for neuron i, b1 is the bias between input and
hidden layers of neural network, and b2 is the bias between hidden and output layers of neural network. It
is important mentioning that the results of this study are different from previous works in that an empirical
1 0.0574 −7.633 −2.447 −0.7302 0.0041 −0.074 −1422 −0.020 −1.709 1996.4 24.512
2 0.0743 −0.6303 −0.7372 0.09136 0.01653 0.0172 −802 0.0768 6.425 805.7
3 0.0279 −1.1444 0.7983 −0.1182 −0.0015 0.0071 337 −0.126 8.350 −514.4
Conclusions
Artificial neural networks and multiple regression analysis were applied in this study to estimate the
penetration rate using 1705 data points in carbonate formations based on seven controllable parameters
(drilling depth, weight on bit, rotary speed, standpipe pressure, drilling torque, flow rate, equivalent
circulating density) and one uncontrollable parameter (unconfined compressive strength). The following
points can be concluded from the presented study:
SPE-197663-MS 9
➢ Both techniques are consistent tools for estimating ROP using other drilling parameters, but the
predicted output of MRA could be narrower than that of the original data set.
➢ The higher determination coefficient (0.96) and the lower mean square error (1.89) of ANNs than
that obtained from MRA indicates the outperformance of ANNs in predicting ROP.
➢ While TVD, WOB, RPM, SPP, and ECD had the greatest effect on estimated ROP, flow rate, drilling
torque, and rock strength appear to be less conservative in predicting ROP.
➢ If the presented models are applied for a future prediction of ROP using input data that fall outside
of the presented ranges in this study, the results of ROP may not be reliable.
➢ As ROP follows quite complex and nonlinear interconnections with other parameters, selecting all
relevant parameters in one predictive model is typically difficult.
➢ Increasing the drilling depth does not always decrease ROP in carbonate environments.
➢ The rock strength tests, as an indicator of formation drillability, can increase the accuracy of the
presented models.
References
Bourgoyne, A. T., and Young, F.S. 1974. A Multiple Regression Approach to Optimal Drilling and Abnormal Pressure
Detection. Society of Petroleum Engineers Journal, 14. https://doi.org/10.2118/4238-PA
Ahmed, O., Adeniran, A., & Samsuri, A. 2018. Rate of Penetration Prediction Utilizing Hydromechanical Specific
Energy. Drilling. https://doi.org/10.5772/intechopen.76903
Al-AbdulJabbar, A., Elkatatny, S., Mahmoud, M., & Abdulraheem, A. 2018. Predicting rate of penetration using artificial
intelligence techniques. Society of Petroleum Engineers - SPE Kingdom of Saudi Arabia Annual Technical
Symposium and Exhibition 2018, SATS 2018, (April), 23–26.
Alkinani, H. et al. 2019. Applications of Artificial Neural Networks in the Petroleum Industry: A Review. SPE Middle
East Oil and Gas Show and Conference, MEOS, Proceedings.
Andrews, R., Hareland, G., Nygaard, R., Calgary, U., & Engler, T. 2007. Methods of Using Logs T o Quantify Drillability.
SPE 106571
Anemangely, M., Ramezanzadeh, A., Tokhmechi, B., Molaghab, A., & Mohammadian, A. 2018. Drilling rate prediction
from petrophysical logs and mud logging data using an optimized multilayer perceptron neural network. Journal of
Geophysics and Engineering, 15(4), 1146–1159. https://doi.org/10.1088/1742-2140/aaac5d
Arabjamaloei, R., & Shadizadeh, S. 2011. Modeling and optimizing rate of penetration using intelligent systems in an
Iranian southern oil field (ahwaz oil field). Petroleum Science and Technology, 29(16), 1637–1648. https://
doi.org/10.1080/10916460902882818
Avcı E. 2018. An Artificial Neural Network Approach for the Prediction of Water-Based Drilling Fluid Rheological
Behaviour. International Advanced Researchers. 02(02), 124–13.
Bahari, A., & Seyed, A. B. 2007. Trust-region approach to find constants of bourgoyne and young penetration rate model
in khangiran iranian gas field. Proceedings of the SPE Latin American and Caribbean Petroleum Engineering
Conference, 2, 823–828.
Balan, B., Mohaghegh, S., & Ameri, S. 1995. State-of-the-art in permeability determination from well log data: part 1-A
comparative study, model development. Eastern Regional Conference and Exhibition, 2(3), 1–10. https://
doi.org/10.2118/30978-MS
Elkatatny, S. M., Tariq, Z., Mahmoud, M. A., & Al-AbdulJabbar, A. 2017. Optimization of rate of penetration using
artificial intelligent techniques. 51st US Rock Mechanics / Geomechanics Symposium 2017, 3, 1620–1627.
Galle, E. M., & Woods, H. B. 1963. Best constant weight and rotary speed for rotary rock bit. Drilling and Production
Practice 1963, 48–73.
Hadi, F., Nygaard, R., and Hilgedick, S. 2017. Probabilistic-Analysis of Deviated Wellbore Stability in Carbonate
Formations of SE Iraq. 51 Th American Rock Mechanics Association.
Hadi, F. A., & Nygaard, R. 2018. Shear wave prediction in carbonate reservoirs: Can artificial neural network outperform
regression analysis? 52nd U.S. Rock Mechanics/Geomechanics Symposium.
10 SPE-197663-MS
Hadi, F., Eckert, A., & Almahdawi, F. 2019. Real-time pore pressure prediction in depleted reservoirs using regression
analysis and artificial neural networks. SPE Middle East Oil and Gas Show and Conference, MEOS, Proceedings,
2019- March.
Han, J., Sun, Y., & Zhang, S. 2019. A data driven approach of ROP prediction and drilling performance estimation.
International Petroleum Technology Conference 2019, IPTC 2019.
Hareland, G., & Nygaard, R. 2007. Calculating unconfined rock strength from drilling data. US Rock Mechanics
Symposium.
Jahanbakhshi, R., Keshavarzi, R., & Jafarnezhad, A. 2012. Real-time prediction of rate of penetration during drilling
operation in oil and gas wells. 46th US Rock Mechanics / Geomechanics Symposium 2012, 4, 2390–2396.
Jorden, J. R., & Shirley, O. J. 1966. Application of Drilling Performance Data to Overpressure Detection. Journal of
Petroleum Technology, 18(1), 1387–1394. https://doi.org/10.2118/1407-PA
Mohaghegh, S. 1995.Neural network:What it can do for petroleum engineers. JPT, Journal of Petroleum
Technology,47(1),42.
Moradi, H., Bahari, M. H., Sistani, M. B. N., & Bahar, A. 2010. Drilling rate prediction using an innovative soft computing
approach. Scientific Research and Essays, 5(13), 1583–1588.
Moran, D., Ibrahim, H., Purwanto, A., & Osmond, J. 2010. Sophisticated ROP prediction technologies based on neural
network delivers accurate drill time results. Society of Petroleum Engineers - IADC/SPE Asia Pacific Drilling
Technology Conference 2010, 100–108.
Philip, N. S. 2001. Studies in Artificial Neural Network Modeling (Vol. 2) . Cochin University of Science and technology.
Shahab, M. 2000. Virtual-Intelligence Applications in Petroleum Engineering: Part 1—Artificial Neural Networks.
Society of Petroleum Engineers. Journal of Petroleum Science and Engineering, Sep 1.
Shahin, M. A., Maier, H. R., & Jaksa, M. B. 2004. Data Division for Developing Neural Networks Applied to Geotechnical
Engineering. Journal of Computing in Civil Engineering, 18(2), 105–114. https://doi.org/10.1061/ (ASCE)0887-
3801(2004)18:2(105)
Shi, X., Liu, G., Gong, X., Zhang, J., Wang, J., & Zhang, H. 2016. An Efficient Approach for Real-Time Prediction of
Rate of Penetration in Offshore Drilling. Mathematical Problems in Engineering, 2016. https://
doi.org/10.1155/2016/3575380
Wai Li, Xionghu Zhao, Yufei Li, Yihui Ji, Hui Peng, Lan Liu, Q. Y. 2015. Laboratory investigations on the effects of
surfactants on rate of penetration in rotary diamond drilling. Journal of Petroleum Science and Engineering, 134, 114–
122.
Zhang, G., Patuw, B.E., and Hu, M. Y. 1998. Forecasting with artificial neural networks. International Journal of
Forecasting, 14, 35–62.