Sie sind auf Seite 1von 5

Permeability Prediction in Carbonate Reservoirs using Specific Area, Porosity and Water

Saturation

M.Sitouah* , M.Al-Hamoud and Y.Bougerira, Sclumberger Ltd, O. Abdullatif, King Fahd University of Petroleum
and Minerals

Abstract multi disciplinary and integrating research,


involving experts from engineering,
This paper presents a comparative study of the mathematics and computer science. ANN is one
capabilities of Extreme Learning Machines of the most used algorithms in the prediction of
(ELM), Decision Trees (DT) and Artificial reservoir properties. Ideally a good predictor
Neural Networks (ANN), in the prediction of algorithm should produce high prediction
permeability from specific surface area, porosity accuracy with minimal computational
and water saturation. ANN has been applied in complexity. The motivation in the present paper
the prediction of various oil and gas properties is to show the performance of the extreme
but with limitations such as computational learning machines (ELM) and decision tress (
instability due to its lack of global optima. ELM DT) against the application of back propagation
and DT are recent advances in Artificial neural network (BPNNs) .
Intelligence with improved architectures and A total of 66 points have been used in this
better performance. research work. Among them, 20 points were
The techniques were optimized and applied to used for testing and validating the algorithms,
the same carbonate reservoir field dataset . where 46 points were used to train and learn the
Following the popular convention and to ensure different paradigms.
fairness, a stratified sampling approach was used
to randomly extract 70% of the dataset for Theory and Methodology
training while the remaining 30% was used for
testing. ELM is a unified framework of broad type of
The results showed that ELM performed best generalized single-hidden layer feed forward
with the highest correlation coefficient, lowest networks, first developed by Huang et al. Unlike
root mean square error and shortest execution traditional popular learning methods, ELM
time. This agrees perfectly with the literature requires less human interventions and can run
that ELM has a more compact architecture thousands times faster than those conventional
optimized for faster execution than the original methods. ELM automatically determines all the
ANN. DT was also found to be a promising network parameters analytically, which avoids
technique for reservoir modeling. trivial human intervention and makes it efficient
The results showed that ELM performed best in online and real-time applications.
with the highest correlation coefficient, lowest ELM was proposed to solve the problem caused
root mean square error and shortest execution by gradient-based algorithms. In the other hand
time. This agrees perfectly with the literature Decision Tree Learning, as a data mining and
that ELM has a more compact architecture machine learning technique uses a decision tree
optimized for faster execution than the original as a predictive model which maps observations
ANN. DT was also found to be a promising about a problem to conclusions about the
technique for reservoir modeling. problem's target value. More descriptive names
for the tree models are classification trees or
Introduction regression trees. In these tree structures, leaves
represent classifications and branches represent
Recently there has been a lot of interest in the conjunctions of features that lead to those
area of artificial intelligence research. This is a classifications. Decision tree (DT) learning is a

© 2014 SEG DOI http://dx.doi.org/10.1190/segam2014-0021.1


SEG Denver 2014 Annual Meeting Page 2709
common method used in data mining with the Tree (CART) analysis and Boosting Trees can
goal of creating a model that predicts the value be used for both regression-type and
of a targetvariable based on several input classification-type problems.
variables. Each interior node corresponds to one
of the input variables; there are edges to children Some of the algorithms that are used for
for each of the possible values of that input constructing decision trees include: Gini
variable. Each leaf represents A value of target impurity and Information gain. Decision tree are
variable given the values of the input variables simple to understand and interpret. It requires
represented by the path from the root to the leaf. little data preparation. It is able to handle both
numerical and categorical data. It uses a white
box model such that if a given situation is
Extreme Learning Machine (ELM) observable in a model the explanation for the
condition is easily explained by boolean logic. It
ELM is a unified framework of broad type of is possible to validate a model using statistical
generalized single-hidden layer feedforward tests making it possible to account for the
networks. Unlike traditional popular learning reliability of the model. It is robust and performs
methods, ELM requires less human well with large data in a short time.
interventions and can run thousands times faster
than those conventional methods. ELM However, the problem of learning an optimal
automatically determines all the network decision tree is known to be NP-complete.
parameters analytically, which avoids trivial Consequently, practical decision-tree learning
human intervention and makes it efficient in algorithms are based on heuristic algorithms
online and real-time applications. The various such as the greedy algorithm where locally
versions available for ELM are: the basic optimal decisions are made at each node. Such
Extreme learning machine, Incremental extreme algorithms cannot guarantee to return the
learning machine and online sequential extreme globally optimal decision tree similar to ANN.
learning machine. It was proposed to solve the Also, decision-tree learners create over-complex
problem caused by gradient-based algorithms. trees that do not generalize the data well referred
ELM randomly chooses the input weights and to as overfitting requiring the use of additional
analytically determines the output weights of mechanisms such as pruning to avoid this
SLFN. ELM has much better generalization problem, thus increasing the complexity of
performance with much faster learning speed. implementation. Similar to the traditional ANN
However, ELM does not consider the network during its inception, decision trees have some
structure features and the involved problem concepts that are hard to learn because they
properties and its generalization performance cannot be expressed them easily, such as XOR,
with much faster learning speed. However, ELM parity or multiplexer problems. In such cases,
does not consider the network structure features the decision tree becomes prohibitively large.
and the involved problem properties and its
generalization performance is also limited.

Classification Tree analysis is when the X H


1
predicted outcome is the class to which the data 1

belongs while Regression Tree analysis is when


K
the predicted outcome can be considered a real
number (e.g. the price of a house, or a patient’s X H
length of stay in a hospital). Tree-based 2 2

techniques such as CHi-squared Automatic P


Interaction Detector (CHAID) and Random O
Forest are used for classification tasks while X H
others such as Classification And Regression 3 3

© 2014 SEG DOI http://dx.doi.org/10.1190/segam2014-0021.1


SEG Denver 2014 Annual Meeting Page 2710
estimate and model other properties which
usually derived from empirical equations.
Training for Permeability Prediction using Decition Trees
3.5
CC = 0.96
3

2.5

Actual
1.5

0.5

-0.5
-0.5 0 0.5 1 1.5 2 2.5 3 3.5
Predicted

Training for Permeability Prediction using Extreme Learning Machine


3.5
CC = 0.99
3

2.5

Actual 1.5

Results and Discussion 1

Different neural network algorithms proposed in 0.5

this study were able to predict permeability from 0

porosity, specific surface area and irreducible -0.5


-0.5 0 0.5 1 1.5 2 2.5 3 3.5
water saturation. The agreement between Predicted
measured data and the network predictions Testing for Permeability Prediction using Extreme Learning Machine
4
demonstrate a successful implementation and CC = 0.96
validation of the algorithm’s ability to map the 3.5

complex nonlinear relationship between the 3

inputs and the output. 2.5

2
Actual

The results show the high performance of ELM 1.5

and DT compared to BPANN’s, with correlation 1

coefficient of 99.3874 % in training process and 0.5


96.0688 % in the testing process for ELM. For 0
DT algorithm, the correlation coefficient in -0.5
about 96.4580 % in the training phase and -1
94.0747 % in the testing phase. In term of root 0 0.5 1 1.5
Predicted
2 2.5 3 3.5

mean square error and computing time , ELM


shows the highest performance again with
RMSE = 0.10, CT ( computing time) = 0.046 in
the training and RMSE = 0.36, CT = 0.005 in
the testing process. For DT , the training process
gave RMSE = 0.24, CT= 0.10 and RMSE =
0.44, CT = 0.01.

The performance for these two neural networks


to predict reservoir properties can be used to

© 2014 SEG DOI http://dx.doi.org/10.1190/segam2014-0021.1


SEG Denver 2014 Annual Meeting Page 2711
Training for Permeability Prediction using Decition Trees
3.5
Actual
3 Predicted

2.5
Permeability (log)

2
Comparison of DT, ELM and
1.5
ANN
1

0.5
1
0
0.9
-0.5
0 5 10 15 20 25 30 35 40 45 50
Training Data 0.8
Testing for Permeability Prediction using Decition Trees
3.5 0.7
Actual
3 Predicted
0.6

CC
2.5
0.5
Permeability (log)

2
0.4
1.5
0.3
1
0.2
0.5

0.1
0

-0.5
0
0 2 4 6 8 10
Testing Data
12 14 16 18 20
8 neurons
10 neurons
12 neurons
14 neurons
DT ELM ANN
Testing for Permeability Prediction using Decition Trees
3.5
CC = 0.94
3

2.5

2
Actual

1.5

0.5

-0.5
-0.5 0 0.5 1 1.5 2 2.5 3 3.5
Predicted

© 2014 SEG DOI http://dx.doi.org/10.1190/segam2014-0021.1


SEG Denver 2014 Annual Meeting Page 2712
http://dx.doi.org/10.1190/segam2014-0021.1

EDITED REFERENCES
Note: This reference list is a copy-edited version of the reference list submitted by the author. Reference lists for the 2014
SEG Technical Program Expanded Abstracts have been copy edited so that references provided with the online metadata for
each paper will achieve a high degree of linking to cited sources that appear on the Web.

REFERENCES
Al-Bazzaz, W. H., Y. W. Al-Mehanna, and A. Gupta, 2007, Permeability modeling using neural-networks
approach for complex Mauddud-Burgan carbonate reservoir: Proceedings of the Middle East Oil and
Gas Show and Conference, Society of Petroleum Engineers, SPE-105337-MS.
Ali, K., 1994, Neural networks: A new tool for the petroleum industry?: Proceedings of the European
Petroleum Computer Conference, Society of Petroleum Engineers, SPE-27561-MS,
http://dx.doi.org/10.2118/27561-MS.
Aminzadeh, F., and F. Brouwer, 2006, Integrating neural networks and fuzzy logic for improved reservoir
property prediction and prospect ranking: 76th Annual International Meeting, SEG, Expanded
Abstracts, 1752–1756.
Asquith, G., and C. Gibson, 2004, Basic well log analysis for geologists, 2nd ed.: AAPG.

© 2014 SEG DOI http://dx.doi.org/10.1190/segam2014-0021.1


SEG Denver 2014 Annual Meeting Page 2713

Das könnte Ihnen auch gefallen