Professional Documents
Culture Documents
Farsi 2021
Farsi 2021
https://doi.org/10.1007/s11053-021-09852-2
Original Paper
Accurate prediction of pore-pressures in the subsurface is paramount for successful planning and
drilling of oil and gas wellbores. It saves cost and time and helps to avoid drilling problems. As it is
expensive and time-consuming to measure pore-pressure directly in wellbores, it is useful to be able to
predict it from various petrophysical input variables on a supervised learning basis calibrated to a
benchmark wellbore. This study developed and compared three-hybrid machine-learning opti-
mization models applied to a diverse suite of 9 petrophysical input variables to predict pore-pressure
across a 273-m-thick, predominately carbonate, reservoir sequence in the giant Marun oil field (Iran)
using 1972 data records. The analysis identified that the multilayer extreme learning machine model
hybridized with a particle swarm optimization (MELM–PSO) applied to seven input variables by
feature selection provided the most accurate pore-pressure predictions for the full dataset
(RMSE = 11.551 psi (1 psi = 6.8947590868 kPa) for well MN#281). The Savitzky–Golay (SG) filter
was applied to pre-process the data, and the properties were filter-ranked using the wrapping method.
The MELM–PSO model outperformed the pore-pressure prediction accuracy achieved by com-
monly used empirical formulas involving sonic or resistivity log data or calculated pore compress-
ibility. To further verify and generalize the applicability of the MELM–PSO model, it was applied to
two other Marun oil field wells (MN#297 and MN#378) achieving RMSE prediction accuracy of
10.031 psi and 10.150 psi, respectively. These results confirmed that the trained model can be reliably
applied to multiple locations across the Marun oil field for predicting pore-pressure.
KEY WORDS: Pore-pressure prediction, Petrophysical well-log data, Hybrid machine-learning opti-
mization models, Feature selection, Empirical model comparisons, Multilayer extreme learning
machine, Particle swarm optimization.
1
Department of Petroleum Engineering, Faculty of Petroleum
and Chemical Engineering, Science and Research Branch,
Islamic Azad University, Tehran, Iran.
2
Young Researchers and Elite Club, Omidiyeh Branch, Islamic
Azad University, Omidiyeh, Iran.
3
Young Researchers and Elite Club, Ahvaz Branch, Islamic Azad
University, Ahvaz, Iran.
4
DWA Energy Limited, Lincoln LN5 9JP, UK.
5
School of Earth Sciences and Engineering, Tomsk Polytechnic
University, Lenin Avenue, Tomsk, Russia.
6
Petroleum Engineering Department, Petroleum Industry
University, Ahvaz, Iran.
7
Faculty of Engineering, Department of Computer Engineering,
Shahid Chamran University, Ahwaz, Iran.
8
To whom correspondence should be addressed; e-mail:
dw@dwasolutions.com
strated that their random forest model outper- Rlog n
formed the other algorithms in terms of goodness- Pp ¼ dv dv Phyd ; ð2Þ
Rn
of-fit, generalizability and prediction accuracy.
In this paper, three-hybrid algorithms, namely where Pp is pore-pressure, dv is total vertical stress,
MLP–PSO, LSSVM–PSO and MELM–PSO, were Phyd is normal or hydrostatic pressure, Rlog is the
developed and compared for predicting pore-pres- observed resistivity log value, Rn is the resistivity log
sure from well-log data. The dataset evaluated in- associated with a normal pore-pressure profile, and,
cludes the available information comprising n is an empirical constant, for which a value of 1.2 is
effective stress (deff), pore compressibility (Cp), commonly used (Yoshida et al. 1996). EatonÕs (1975)
corrected gamma ray (CGR; adjusted for uranium empirical formula for predicting pore-pressure from
concentrations), uncorrected spectral gamma ray sonic well-log transit time values is expressed as
(SGR), potassium (POTA) from the SGR tool, (Yoshida et al. 1996):
thorium (THOR) from the SGR tool, uranium
DT log n
(URAN) from the SGR tool, the photoelectric Pp ¼ dv dv Phyd ; ð3Þ
absorption factor (PEF), bulk formation density DT n
(RHOB), compressional sonic transit time (DT),
neutron porosity (NPHI) and deep resistivity (ILD). where Pp is pore-pressure, dv is total vertical stress,
The best-performing algorithm was identified for Phyd is normal or hydrostatic pressure, DTlog is the
accurate prediction of pore-pressure using optimized sonic log transit time value, DTn is the sonic log
feature selection from the well-log curves. value when pore-pressure is normal, n is an empiri-
cal constant, for which a value of 3.0 is commonly
used.
PORE-PRESSURE MODELS In 1995, Bowers (1995) proposed a model for
determining pore-pressure based on effective stress,
Calculating subsurface pore-pressure using Ea- the main purpose of which was to calculate effective
ton’s method (Eaton 1975) by solving the equations velocity stresses and use them to calculate pore-
of Terzaghi et al. (1996) as a function of overburden pressure. That method considered compaction dise-
pressure and matrix stress is widely used. Equa- quilibrium and unloading due to fluid expansion as
tion (1) expresses the fundamental relationship be- the main mechanisms involved in generating over-
tween overburden pressure and pore-pressure, thus: pressure. Azadpour et al. (2015) predicted pore-
pressure in an Iranian gas field based on three
S ¼ r þ PP ; ð1Þ empirical methods: Eaton (Eq. 2) (Eaton 1975),
Bowers (velocity-based model) (Bowers 1995) and
where S is overburden pressure; r is matrix stress; the compressibility (Atashbari and Tingay 2012)
and PP is pore-pressure. As a benchmark, a ‘‘nor- methods. They found that the Eaton method, with
mal’’ compaction curve is established using Eq. 1 for an n exponent of about 0.5, provided more accurate
fine-grained sediments for which overburden stress predictions than the other methods. Consequently,
and pore-pressure increase with burial depth as they developed a three-dimensional pore-pressure
‘‘normal’’ compaction progresses. Such compaction prediction model for that field applying statistical
curves vs. depth provide simplistic effective stress extrapolations. The estimated pore-pressure values
and/or pore-pressure relationship that tend to vary from that 3D model proved to be in good agreement
among sedimentary basins due to distinctive with pore-pressures measured directly by using a
lithologies and prevailing subsurface stress regimes wireline-based modular dynamic tester. Taking into
(Swarbrick 2001). It is useful to make a comparison account compaction to address the challenge of
between the ‘‘observed’’ and ‘‘normal’’ compaction disequilibrium applies Eqs. (4) to (6) to resolve
profiles to identify over-pressured and under-pres- unloading conditions.
sured zones (Swarbrick 2001). Eaton’s method pro-
vides well-used empirical equations for deriving V ¼ V 0 þ ArB ð4Þ
pore-pressure from basic well-log data. For example,
" ðU1 Þ #B
Eqs. 2 and 3 express EatonÕs formulas for calculating r
pore-pressure from resistivity and sonic well-logs, V ¼ V 0 þ A rmax ð5Þ
rmax
respectively.
M. Farsi et al.
performing solutions (those with the lowest RMSE sures from multiple depths. At each depth to be
values) were identified by ranking and passed on to tested, the borehole interval is isolated and a con-
the next optimization iteration. The GA applied trolled flow of reservoir fluid is induced. The RFT
three modification processes (crossover, combina- data, recorded in situ, provide valuable information
tion and mutation) to create a new subset of solu- about the reservoir condition, such as temperature,
tions, including the best performers from the pressure, permeability. The fluid samples recovered
previous iteration, for evaluation in the next itera- by the RFT can be analyzed subsequently to provide
tion (Wahab et al. 2015). compositional details of the reservoir fluids. How-
ever, deploying such tools to record multiple pres-
sure points in boreholes is time-consuming and
Repeat Formation Tester (RFT) Downhole Direct expensive and such data are not routinely collected
Pressure Measurements across the entire reservoir sequence in every well
drilled.
Wireline or measurement-while-drilling This study used the data recorded by the RFT
(MWD) techniques can provide direct formation- tool to provide a measured pressure vs. depth data
pressure measurements. Using two packers and a series from a single well. The objective was to use
probe, the repeat formation tester (RFT), and sim- hybrid optimization/machine-learning tools to pre-
ilar tools, can sequentially record formation pres- dict that measured formation pore-pressure vs.
M. Farsi et al.
depth data from the available suite of well-logs. The Multilayer Perceptron
value of doing this is that reliable and verified ma-
chine-learning solutions can be applied to predict Artificial neural networks (ANNs) have be-
pore-pressure in adjacent wells (already drilled and come widely used since their introduction in the
yet-to-be drilled) for which RFT data are not 1990s (Ali 1994). Four factors determine the pre-
available. This would avoid the cost and time asso- diction accuracy of ANN models (Maimon and
ciated with taking direct in situ pressure measure- Rokach 2009): (a) feature selection (i.e., what input
ments in the reservoir sections of each well drilled. variables should be included); (b) network archi-
tecture (i.e., number of layers and neurons); (c)
transfer functions between layers; and (d) training
Machine-Learning Algorithms Evaluated algorithm selection. Multilayer perceptron (MLP) is
the most commonly used feed-forward ANN be-
Optimization and machine-learning can help to cause it is easy to set up and adapted for evaluating
resolve issues in many areas of the oil and gas large and complex datasets (Bishop 2006). The
industry, including reservoirs (Ghorbani et al. Levenberg–Marquardt (LM) is the most common
2017a), formation damage (Mohammadian and algorithm applied for training of MLPs because it
Ghorbani 2015), wellbore stability (Darvishpour tends to converge rapidly and reliably to find high-
et al. 2019), rheology and filtration (Mohamadian performing predictions. However, in large, complex,
et al. 2018), production (Ghorbani and Moghadasi nonlinear datasets, the LM algorithm tends to con-
2014; Ghorbani et al. 2014 2017b) and drilling fluid verge too rapidly and it becomes trapped at local
(Mohamadian et al. 2019). Machine-learning algo- minima. Consequently, to improve the performance
rithms are now the mathematical tools of choice to of MLPs with such datasets, it is beneficial to sup-
provide accurate and reliable predictions of depen- plement the LM algorithm with a more effective
dent variables governed by nonlinear and scattered optimization algorithm.
relationships with other influential variables. As In this study, the PSO is used for that purpose.
such they offer valuable solutions in all sectors of the Based on trials and sensitivity analysis, two hidden
oil and gas industry (Choubineh et al. 2017; Ghor- layers were selected as the optimum structure for an
bani et al. 2017c 2018 2019, 2020a, b; Mohamadian MLP to predict pore-pressure, with hidden layer 1
et al. 2021; Rashidi et al. 2020; Farsi et al. 2021; and hidden layer 2 assigned 10 and 5 neurons,
Ranaee et al. 2021). respectively. Similarly, based on sensitivity analysis
In this study, three machine-learning algorithms performed, the transfer functions ‘‘tansig’’ and
were deployed: (1) least squares support vector ‘‘purelin’’ were selected for hidden layers 1 and 2,
machine (LSSVM); (2) extreme learning machine respectively.
(ELM); and (3) multilayer perceptron (MLP) neural
network. These algorithms were each coupled with
the powerful particle swarm optimization (PSO) to Extreme Learning Machine
develop hybrid models that are effective in predict-
ing pore-pressure rapidly and with high accuracy. The ELM algorithm was introduced in 2006 an
alternative feed-forward neural network algorithm
offering a high computational speed (Huang et al.
Genetic Algorithm (GA) 2006). Applications have shown that ELMs can re-
duce a networkÕs training time and improve its
GA is an evolutionary algorithm that simulates overall performance. Whereas MLPs use time-con-
natural selection and solves problems optimally in suming iterative back-propagation algorithms to
an iterative manner as described in Simon (2013). establish the weights and biases applied to their
High-performance solutions are identified in each hidden layers, ELMs, with a single hidden layer,
iteration and are used preferentially to contribute to select weights and biases randomly for that hidden
the modification involved in generating new solu- layer from a uniform distribution. In the ELM, those
tions to consider for the next GA iteration. Con- randomly selected weights and biases are typically
versely, the worse-performing solutions are not adjusted during the network tuning process
progressively excluded based on their poor fitness (Huang et al. 2006; Wang et al. 2014). The output
comparisons. weights of ELM are derived using the Moore–Pen-
Predicting Formation Pore-Pressure from Well-Log Data
rose generalized inverse of the hidden-layer output. was adopted here to predict pore-pressure from
Figure 2 illustrates the structure of a single-hidden- well-log data.
layer ELM (Yeom and Kwak 2017).
More complex variants of ELMs, such as the
two-hidden-layer ELM, the four-hidden-layer ELM PSO Algorithm
and the multiple-hidden-layer ELM (MELM) have
been applied successfully to more complex, large, The PSO algorithm was developed by Kennedy
nonlinear datasets (Liu et al. 2019a 2019b; Xiao and Eberhart in 1995 (Kennedy 1997; Kennedy and
et al. 2017). Tests suggested that an MELM was Eberhart 1995). This algorithm is a stochastic opti-
more effective to predict pore-pressure from input mization technique originating as an analogy to
data from multiple well-logs. More details about animal swarming behaviors observed in the natural
theoretical description and construction process of world (Yang and Papa 2016). PSO algorithm in
MELM can be found in Xiao et al. (2017). combination of other optimization and machine-
learning algorithms has been used widely to solve
many nonlinear problems in petroleum industry
Least Squares Support Vector Machine achieving high degrees of prediction accuracy
(Atashnezhad et al. 2014). Theoretical background
LSSVM was developed in 1999 by Suykens and and implementation details for the PSO algorithm
Vandewalle (1999) as a modified form of the are available elsewhere (Anemangely et al. 2017;
established regression-based support vector machine Kennedy 1997; Mohamadian et al. 2021).
(SVM) machine-learning algorithm (Vapnik 2013).
There are three main differences between the
LSSVM and SVM algorithms. First, LSSVM uses a Hybrid Machine-Learning Optimization Algorithms
least squares error cost function while SVM uses
nonnegative errors for its cost function. Second, Hybrid LSSVM–PSO Model
LSSVM uses equality constraints while SVM uses
inequality constraints (Yuan et al. 2015). Third, A schematic flow diagram (Fig. 3) illustrates
whereas SVM conducts training using quadratic how LSSVM and PSO algorithms were integrated to
programming (QP), LSSVM conducts training using operate as a hybrid LSSVM–PSO machine-learning
linear programing (LP), which reduces the compu- optimization model to predict pore-pressure effec-
tational complexity and speeds up the algorithm tively in the dataset evaluated. An RBF kernel
(Kisi and Parmar 2016). The LSSVM algorithm has function provides the best pore-pressure predictions.
been employed successfully to solve a wide range of For tuning the hyperparameters influencing the
regression and classification machine-learning tasks LSSVM prediction performance, the PSO deter-
(Adankon and Cheriet 2009; Lima et al. 2010) and it mined optimum values for the regularization
Figure 3. Flow diagram for LSSVM–PSO hybrid machine-learning optimization model applied for PP
prediction.
Table 1. Optimized control parameter for developed hybrid LSSVM–PSO used for pore-pressure prediction
PSO LSSVM
Table 2. RMSE values calculated for a range of hidden layers and neurons in the layers of the MELM–PSO sensitivity analysis for pore-
pressure prediction
5 10 15 20 25
PSO MELM
On the other hand, MELM with a small number of neurons in those layers to consider. The established
hidden layers and neurons is more likely not to narrow ranges were then used as constraints for the
achieve optimal prediction accuracy of dependent developed PSO–MELM–PSO model. The PSO
variables. Therefore, a balance must be struck in (Step 2 procedure) calculates the values of weights
network-element selection to achieve an efficient and biases for the limited defined ranges of number
MELM model with optimal performance. In this of layers and neurons in each hidden layer estab-
study, the PSO algorithm was applied in two steps to lished by Step 1.
construct the optimized MELM network (Fig. 4). The of number of MELM hidden layers, based
Step 1 Identify optimum narrow ranges for on step 1 analysis, was allowed to vary from 3 to 9,
numbers of hidden layers and neurons in each hid- and the number of neurons in each hidden layer was
den layer (replacing the traditional method of trial- allowed to vary from 5 to 25. Table 2 displays the
and-error method). results of the Step 1 analysis performed for the
Step 2 Identify the optimum values for the MELM–PSO hybrid model. The highest pore-pres-
weights assigned to each neuron in each hidden sure prediction accuracy (lowest root mean squared
layer and the biases assigned to each hidden layer error, RMSE value) was achieved in the range of 7–9
(replacing the traditional ELM method of randomly hidden layers with 20–25 neurons in each hidden
assigning these values). layer. Consequently, in Step 2 of the PSO–MELM–
The two-step optimization method developed is PSO model developed for prediction of pore-pres-
able to reduce the computational time of this hybrid sure in the studied dataset, the range of the hidden
algorithm. The method involves an ‘‘initial-pass’’ layers was constrained to vary from 7 to 9, and the
tuning optimization procedure. This was performed range of the number of neurons in each hidden layer
with the objective of narrowing down the ranges of was constrained to vary from 20 to 25.
meaningful number of hidden layers and numbers of
Predicting Formation Pore-Pressure from Well-Log Data
DATA COLLECTION
AND CHARACTERIZATION
PSO MLP
electric absorption factor (PEF), bulk formation Higher NPHI and DT in the upper part of the
density (RHOB), compressional sonic transit time formation tended to correspond with lower pore-
(DT), neutron porosity (NPHI) and deep resistivity pressures, whereas lower RHOB values tended to
(ILD). Sensitivity analysis suggested that the com- correspond with relatively high pore-pressures in
positional variables (POTA, THOR and URAN) that part of the formation (Fig. 8). In contrast with
were not related to pore-pressure. The remaining 9 the other variables, mid-range ILD values corre-
variables show complex relationships with pore- sponded to relatively high pore-pressures in the
pressure and depth and were therefore used as in- upper part of the formation. On the other hand, pore
puts for the machine-learning prediction analysis. compressibility (Cp) and photoelectric absorption
Table 5 statistically summarizes the distributions of factor (PEF) values did not show obvious relation-
the 9 input variables and the dependent variable, ships with pore-pressure in the upper part of the
pore-pressure, across the Asmari reservoir forma- formation. No obvious relationships existed between
tion for the three wells studied MN# 281 (1792 data the values of the input variables and pore-pressure
records), MN# 279 (1225 data records) and MN# 378 in the lower part of the Asmari Formation (Figs. 7
(1225 data records). and 8). The distributions displayed in Figures 7 and
8 confirm that none of the 9 input variables would be
individually reliable on its own for predicting pore-
Variable Distributions vs. Depth and Pore-pressure pressure.
Figure 6. Map showing the location of the Marun oil field, onshore southwest Iran.
Having established the best architecture for the dation technique, because it overcame over-fitting
MLP, several methods were evaluated for selecting issues; eightfold validation divided the entire dataset
multiple mutually exclusive testing and training into 8 non-overlapping sections. A single section was
subsets of data records for model evaluation. One then selected as a subset to be evaluated. The
approach considered was to select randomly 30% of remaining 7 sections of the dataset, in each case,
all available data records for the testing subset, and were assigned to the training subset. For each subset
then assigning the remaining 70% of data records to selection, the MLP was evaluated 80 times (10 times
the training subset. This method failed to prevent for each training / testing subset combination). The
over-fitting during the feature selection process, model with the lowest predicted vs. recorded pore-
leading to some features being attributed too much pressure RMSE was then selected for each of the 10
influence in the predictions. A more successful ap- training / testing subset combination. The average of
proach was achieved by applying an eightfold vali- the 10-best RMSE values obtained for the eightfold
M. Farsi et al.
Table 5. Statistical characterization of the variables constituting the well-log dataset for three Marun oil field wells: MN#281, MN#279 and
MN#378. The variable values for all data records are available to download in a supplementary file (see Supplementary Data) (1
psi = 6.8947590868 kPa, 1ft = 0.3048 m)
Wells Units Psi psi1 API API Barns/ g/cm3 ls/ft (%) ohm-m psi
cm3
MN#281 Mean 4501.3 1.30E06 21.8 41.3 4.00 2.55 64.7 13.9 1162.6 4733.7
1792 Data Re- Standard devi- 302.8 1.34E07 20.2 20.9 0.93 0.18 9.1 5.6 4271.4 340.9
cords ation
Variance 91,639.3 1.79E14 408.4 438.0 0.87 0.03 82.4 31.6 18234882.3 116137.9
Minimum 2146.9 1.06E06 1.1 12.2 1.91 1.20 51.1 1.3 0.4 4092.2
Maximum 5080.9 2.07E06 121.4 143.6 6.33 2.87 117.2 46.7 20,000.0 5375.4
MN#297 Mean 4860.4 1.41E06 52.1 36.1 4.77 2.67 72.1 11.5 814.0 4877.0
1225 data re- Standard devi- 245.0 2.61E07 29.3 15.9 1.83 0.12 8.6 7.0 3446.9 268.4
cords ation
Variance 59,952.4 6.80E14 856.0 254.0 3.33 0.02 73.1 48.4 ######### 71,994.6
Minimum 4371.1 1.10E06 5.1 6.0 2.63 2.35 41.9 1.0 0.4 4474.0
Maximum 5435.0 2.25E06 110.2 78.3 8.68 2.98 89.3 34.7 20,000.0 5543.0
MN#378 Mean 4605.3 1.39E06 60.0 37.5 14.93 2.52 60.0 12.0 1198.2 4860.0
1225 data re- Standard devi- 325.6 1.80E06 9.6 13.0 2.83 0.13 9.6 5.8 4409.9 168.1
cords ation
Variance 105,918.8 3.25E12 91.8 168.8 8.02 0.02 91.8 34.1 ######### 28,235.3
Minimum 3779.8 4.02E05 48.3 13.6 4.29 2.12 48.3 0.0 0.5 4498.0
Maximum 5192.4 4.07E05 100.3 89.5 17.48 2.77 100.3 34.4 20,000.0 5428.0
was then taken as representative of the prediction Identifying the Best-performing Algorithm
performance accuracy of the feature selection eval- for Pore-pressure Prediction
uated. The eightfold validation sequence applied is
illustrated in Figure 9. Evaluating the models with normalized data
The main purpose of this method of imple- avoids systematic biases resulting from different
mentation feature selection was to determine the value scales among the input variables. Thus, Eq. 8
optimal combination of the 9 input variables (fea- was used to normalize all the data variables to range
tures) listed in Table 6, i.e., the combination of from 1 to + 1.
features that achieved the minimum pore-pressure
X X min
RMSE values. The results of the feature selection X norm ¼ 2 1: ð8Þ
X max X min
analysis are presented in Table 7 and Figure 10. The
lowest pore-pressure RMSE value achieved was The studied dataset was divided into two
71.79 psi.3 This was associated with a 7-variable groups—the training and testing subsets. The train-
combination, which excluded variables Z6 (RHOB) ing subset comprised 70% of all data records, which
and Z7 (DT). For models with more than 7 features, were selected randomly and distributed evenly
the RMSE was higher than for the 7-variable com- across the entire range of the dependent variableÕs
bination identified. It is clear from Table 7 that distribution. Then, 15% of data records constituted
inclusion of a variable in one round of feature the validation subset and 15% of data records con-
selection did not guarantee its inclusions in a sub- stituted the testing subset, which were held inde-
sequent round. For example, feature Z2 was selected pendently of the training subset. The testing subset
as one of the best variables for a 3-variable combi- was used to verify the prediction accuracy of opti-
nation but not for a 4-variable combination. How- mum model solutions derived during training in
ever, variable Z2 was selected as one of the variables terms of its prediction repeatability. The testing
for the best 6-variable and 7-variable combinations, subset also provided indications of over-fitting, if it
occurred.
3
1 psi = 6.8947590868 kPa.
Predicting Formation Pore-Pressure from Well-Log Data
Figure 7. Contour plots of the input variables deff, Cp, CGR, SGR and PEF vs. depth and contoured for pore-pressure for Marun oil field
well MN#281 (1 psi = 6.8947590868 kPa).
Figure 8. Contour plots of the input variables RHOB, DT, NPHI and ILD vs. depth and contoured for pore-pressure for Marun oil field
well MN#281 (1 psi = 6.8947590868 kPa).
Table 7. Feature selections applying the eightfold validation method with GA–MLP to find the optimum combination for pore-pressure
prediction (1 psi = 6.8947590868 kPa)
1 Z3 207.515
2 Z3, Z5 162.055
3 Z1, Z2, Z3 123.325
4 Z8, Z3, Z5, Z1 111.430
5 Z1, Z5, Z8, Z3, Z9 97.608
6 Z2, Z1, Z5, Z8, Z6, Z3 75.428
7 Z3, Z5, Z2, Z1, Z8, Z9, Z4 (Best) 71.790
8 Z6, Z1, Z5, Z3, Z8, Z2, Z4, Z9 81.876
9 Z7, Z3, Z2, Z5, Z4, Z9, Z8, Z6, Z1 85.095
Table 8. Pore-pressure prediction accuracy achieved by the hybrid machine-learning optimization algorithm and empirical models applied
to the training subsets ( 70% of the data records) for the 7-feature selection for Marun oil field MN#281 (1 psi = 6.8947590868 kPa)
Models Statistical errors APD (%) AAPD (%) SD (psi) MSE (psi) RMSE (psi) R2
Empirical models Pore compressibility model 0.243 6.679 419.6 176,972 420.7 0.0212
Eaton Mosel (resistivity) 1.546 6.099 369.8 137,727 371.1 0.0003
Eaton model (sonic) 1.126 5.956 356.4 128,078 357.9 0.0018
Hybrid ML-optimizer models MLP–PSO 0.013 1.366 71.7 5326 73.0 0.9455
LSSVM–PSO 0.006 0.862 45.4 2136 46.2 0.9772
MELM–PSO 0.006 0.223 11.1 141 11.86 0.9985
Predicting Formation Pore-Pressure from Well-Log Data
Table 9. Pore-pressure prediction accuracy achieved by the hybrid machine-learning optimization algorithm and empirical models applied
to the validation subsets ( 15% of the data records) for the 7-feature selection for Marun oil field MN#281 (1 psi = 6.8947590868 kPa)
Models Statistical errors APD (%) AAPD (%) SD (psi) MSE (psi) RMSE (psi) R2
Empirical models Pore compressibility model 14.985 15.794 1089.8 1,197,539 1094.3 0.3798
Eaton Mosel (resistivity) 17.066 17.066 1153.6 1,337,556 1156.5 0.3539
Eaton model (sonic) 16.785 16.785 1137.9 1,297,441 1139.1 0.3473
Hybrid ML-optimizer models MLP–PSO 0.043 1.381 82.7 6936 83.3 0.9546
LSSVM–PSO 0.023 0.879 52.1 2827 53.2 0.9884
MELM–PSO 0.023 0.206 12.1 165 12.85 0.9929
Table 10. Pore-pressure prediction accuracy achieved by the hybrid machine-learning optimization algorithm and empirical models applied
to the testing subsets ( 15% of the data records) for the 7-feature selection for Marun oil field MN#281 (1 psi = 6.8947590868 kPa)
Models Statistical errors APD (%) AAPD (%) SD (psi) MSE (psi) RMSE (psi) R2
Empirical models Pore compressibility model 3.291 5.131 285.9 82,448 287.1 0.348
Eaton Mosel (resistivity) 8.602 9.035 482.3 235,101 484.9 0.2034
Eaton model (sonic) 7.155 7.706 402.3 163,395 404.2 0.2684
Hybrid ML-optimizer models MLP–PSO 0.249 1.380 76.5 5942 77.1 0.965
LSSVM–PSO 0.003 0.892 48.4 2422 49.2 0.9666
MELM–PSO 0.003 0.225 11.8 157 12.53 0.9833
Table 11. Pore-pressure prediction accuracy achieved by the hybrid machine-learning optimization algorithm and empirical models applied
to the full subsets ( 100% of the data records) for the 7-feature selections for Marun oil field MN#281 (1 psi = 6.8947590868 kPa)
Models Statistical errors APD (%) AAPD (%) SD (psi) MSE (psi) RMSE (psi) R2
Empirical models Pore compressibility model 2.293 7.050 527.8 281,859 530.9 0.0023
Eaton Mosel (resistivity) 3.849 7.020 516.2 269,928 519.5 0.04
Eaton model (sonic) 3.942 7.087 522.8 275,325 524.7 0.0391
Hybrid ML-optimizer models MLP–PSO 0.033 1.245 70.3 5140 71.7 0.9537
LSSVM–PSO 0.001 0.789 44.8 2073 45.5 0.9806
MELM–PSO 0.001 0.201 10.9 133 11.55 0.9987
dataset. Statistical measures of accuracy achieved MN#281 data, when applied to other wells in the
for data records from these two wells are shown in Marun oil field.
Table 12. A comparison of the results of Table 12 Figures 16 and 17 plot the actual vs. predicted
with those of Tables 8, 9, 10 and 11 confirms the high pore-pressure values achieved by the MELM–PSO
pore-pressure prediction accuracy achieved by the model trained with MN#281 data records and ap-
developed MELM–PSO model, trained with plied to all the data records available for wells MN#
M. Farsi et al.
Figure 11. Predicted vs. measured pore-pressure comparisons for three-hybrid machine-learning optimization models and empirical
models applied to the complete MN#281 dataset of 1792 data records with the 7-variable model (1 psi = 6.8947590868 kPa).
297 and MN# 378. The performance accuracy would need to be recalibrated initially with some
achieved by this algorithm confirms its reliability for direct formation-pressure measurements from at
application across the Marun oil field in wells, or least one well in each of the fields to which it is
sections of the reservoir, for which direct formation- applied.
pressure measurements are not available. The Figures 18 and 19 show the performance of the
method can be used in other fields but, of course, MELM–PSO model in providing accurate pore-
Predicting Formation Pore-Pressure from Well-Log Data
Figure 13. Pore-pressure prediction relative errors (%) compared for the training, validation and testing subsets of the empirical
equation and hybrid machine-learning optimization models evaluated for the MN#281 dataset. Note that both training, validation
and testing subsets are spread across the entire depth interval sampled and are displayed sequentially in these plots for illustrative
purposes only. Note the different vertical scales on the right-side plots.
Predicting Formation Pore-Pressure from Well-Log Data
Figure 14. Shales effects in the Asmari reservoir zone of MH#281 impacting the
empirically calculated Eaton sonic log pore-pressure values (1
psi = 6.8947590868 kPa).
Table 12. Pore-pressure prediction accuracy of the MELM–PSO model, trained with MN#281 data, applied to the complete datasets
available for Marun oil field wells MN#297 and MN#378 and treating them as additional independent testing subsets (1
psi = 6.8947590868 kPa)
Wells APD (%) AAPD (%) SD (psi) MSE (psi) RMSE (psi) R2
Huang, G.-B., Zhu, Q.-Y., & Siew, C.-K. (2006). Extreme learning provement additive for drilling fluids. Journal of Polymer
machine: Theory and applications. Neurocomputing, 70(1–3), Research, 26(2), 33.
489–501. Mohamadian, N., Ghorbani, H., Wood, D. A., Mehrad, M., Da-
Hutomo, P. S., Rosid, M. S., & Haidar, M. W. (2019). Pore voodi, S., Rashidi, S., & Shahvand, A. K. (2021). A geome-
pressure prediction using eaton and neural network method chanical approach to casing collapse prediction in oil and gas
in carbonate field ‘‘X’’ based on seismic data. In IOP con- wells aided by machine learning. Journal of Petroleum Sci-
ference series: Materials science and engineering (Vol. 546, ence and Engineering, 196, 107811.
No. 3, p. 032017). IOP Publishing. https://doi.org/10.1088/17 Mustafa, M., Rezaur, R., Rahardjo, H., & Isa, M. (2012). Pre-
57-899X/546/3/032017/meta. diction of pore-water pressure using radial basis function
Jain, A., & Zongker, D. (1997). Feature selection: Evaluation, neural network. Engineering Geology, 135, 40–47.
application, and small sample performance. IEEE Transac- Osborne, M. J., & Swarbrick, R. E. (1997). Mechanisms for gen-
tions on Pattern Analysis and Machine Intelligence, 19(2), erating overpressure in sedimentary basins: A reevaluation.
153–158. AAPG Bulletin, 81(6), 1023–1041.
John, G. H., Kohavi, R., & Pfleger, K. (1994). Irrelevant features Polito, C. P., Green, R. A., & Lee, J. (2008). Pore pressure gen-
and the subset selection problem. In Machine learning pro- eration models for sands and silty soils subjected to cyclic
ceedings 1994 (pp. 121–129). Morgan Kaufmann. https://doi. loading. Journal of Geotechnical and Geoenvironmental
org/10.1016/B978-1-55860-335-6.50023-4. Engineering, 134(10), 1490–1500.
Kennedy, J. (1997). The particle swarm: social adaptation of Ranaee, E., Ghorbani, H., Keshavarzian, S., Ghazaeipour Abar-
knowledge. Paper presented at the proceedings of 1997 IEEE ghoei, P., Riva, M., Inzoli, F., & Guadagnini, A. (2021).
international conference on evolutionary computation Analysis of the performance of a crude-oil desalting system
(ICEC’97). https://doi.org/10.1080/10.1109/ICEC.1997.59232 based on historical data. Fuel. https://doi.org/10.1016/j.fuel.
6. 2020.120046.
Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. Rashidi, S., Mohamadian, N., Ghorbani, H., Wood, D. A., Shah-
Paper presented at the proceedings of ICNN’95-international bazi, K., & Ahmadi Alvar, M. (2020). Shear modulus pre-
conference on neural networks. https://doi.org/10.1080/10.110 diction of embedded pressurize salt layers and pinpointing
9/ICNN.1995.488968. zones at risk of casing collapse in oil and gas wells. Journal of
Keshavarzi, R., & Jahanbakhshi, R. (2013). Real-time prediction Applied Geophysics, 104205.
of pore pressure gradient through an artificial intelligence Rashidi, S., Mehrad, M., Ghorbani, H., Wood, D. A., Mohama-
approach: A case study from one of Middle East oil fields. dian, N., Moghadasi, J., & Davoodi, S. (2021). Determination
European Journal of Environmental and Civil Engineering, of bubble point pressure and oil formation volume factor of
17(8), 675–686. crude oils applying multiple hidden layers extreme learning
Kisi, O., & Parmar, K. S. (2016). Application of least square machine algorithms. Journal of Petroleum Science and
support vector machine and multivariate adaptive regression Engineering. https://doi.org/10.1016/j.petrol.2021.108425.
spline models in long term prediction of river water pollution. Rehm, B., Schubert, J., Haghshenas, A., Paknejad, A. S., &
Journal of Hydrology, 534, 104–112. Hughes, J. (2013). Managed pressure drilling. Elsevier.
Kohavi, R., & John, G. H. (1997). Wrappers for feature subset Rubin, Y., & Hubbard, S. (2005). Hydrogeophysics, water science
selection. Artificial Intelligence, 97(1–2), 273–324. and technology library. Springer.
Lima, C. A., Coelho, A. L., & Eisencraft, M. (2010). Tackling Satter, A., & Iqbal, G. M. (2015). Reservoir engineering: The
EEG signal classification with least squares support vector fundamentals, simulation, and management of conventional
machines: A sensitivity analysis study. Computers in Biology and unconventional recoveries. Gulf Professional Publishing.
and Medicine, 40(8), 705–714. Shi, Y., & Wang, C. Y. (1986). Pore pressure generation in sedi-
Liu, H. (2017). Principles and applications of well-logging. mentary basins: Overloading versus aquathermal. Journal of
Springer. https://doi.org/10.1007/978-3-662-53383-3. Geophysical Research: Solid Earth, 91(B2), 2153–2162.
Liu, J., Liu, X., Liu, C., Le, B. T., & Xiao, D. (2019). Random Simon, D. (2013). Evolutionary optimization algorithms. Wiley.
search enhancement of incremental regularized multiple Suykens, J. A., & Vandewalle, J. (1999). Least squares support
hidden layers ELM. IEEE Access, 7, 36866–36878. vector machine classifiers. Neural Processing Letters, 9(3),
Liu, J., Liu, X., & Le, B. T. (2019b). Rolling force prediction of 293–300.
hot rolling based on GA-MELM. Complexity, 2019. https:// Swarbrick, R. E. (2001). Challenges of porosity-based pore pres-
www.hindawi.com/journals/complexity/2019/3476521/. sure prediction. Paper presented at the 63rd EAGE confer-
Lyons, W. C., & Plisga, G. J. (2011). Standard handbook of pet- ence & exhibition. https://doi.org/10.3997/2214-4609-pdb.15.
roleum and natural gas engineering. Elsevier. O-25.
Maimon, O., & Rokach, L. (2009). Introduction to knowledge Terzaghi, K., Peck, R. B., & Mesri, G. (1996). Soil mechanics in
discovery and data mining. In Data mining and knowledge engineering practice (3rd edn.). John Wiley & Sons.
discovery handbook (pp. 1–15): Springer. https://doi.org/10. Vapnik, V. (2013). The nature of statistical learning theory.
1007/978-0-387-09823-4_1. Springer.
Mohammadian, N., & Ghorbani, H. (2015). An investigation on Wahab, M. N. A., Nefti-Meziani, S., & Atyabi, A. (2015). A
chemical formation damage in Iranian reservoir by focus on comprehensive review of swarm optimization algorithms.
mineralogy role in shale swelling potential in Pabdeh and PLoS ONE, 10(5), 1–36.
Gurpi formations. Advances in Environmental Biology, 9(4), Wang, S.-J., Chen, H.-L., Yan, W.-J., Chen, Y.-H., & Fu, X.
161–166. (2014). Face recognition and micro-expression recognition
Mohamadian, N., Ghorbani, H., Wood, D. A., & Hormozi, H. K. based on discriminant tensor subspace analysis plus extreme
(2018). Rheological and filtration characteristics of drilling learning machine. Neural Processing Letters, 39(1), 25–43.
fluids enhanced by nanoparticles with selected additives: An Xiao, D., Li, B., & Mao, Y. (2017). A multiple hidden layers
experimental study. Advances in Geo-Energy Research, 2(3), extreme learning machine method and its application.
228–236. Mathematical Problems in Engineering, 2017. https://www.h
Mohamadian, N., Ghorbani, H., Wood, D. A., & Khoshmardan, indawi.com/journals/mpe/2017/4670187/.
M. A. (2019). A hybrid nanocomposite of poly (styrene-me- Yang, X.-S., & Papa, J. P. (2016). Bio-inspired computation and
thyl methacrylate-acrylic acid)/clay as a novel rheology-im- applications in image processing. Academic Press.
Predicting Formation Pore-Pressure from Well-Log Data
Yeom, C.-U., & Kwak, K.-C. (2017). Short-term electricity- SPE/IADC Asia Pacific Drilling Technology. https://doi.org/h
load forecasting using a TSK-based extreme learning ttps://doi.org/10.2118/36381-MS.
machine with knowledge representation. Energies, 10(10), Yu, H., Chen, G., & Gu, H. (2020). A machine learning
1613. methodology for multivariate pore-pressure prediction.
Yoshida, C., Ikeda, S., & Eaton, B. A. (1996). An investigative Computers & Geosciences, 143, 104548.
study of recent technologies used for prediction, detection, Yuan, X., Chen, C., Yuan, Y., Huang, Y., & Tan, Q. (2015). Short-
and evaluation of abnormal formation pressure and fracture term wind power prediction based on LSSVM–GSA model.
pressure in North and South America. Paper presented at the Energy Conversion and Management, 101, 393–401.