Beruflich Dokumente
Kultur Dokumente
Cite this article as: PETROL. EXPLOR. DEVELOP., 2016, 43(4): 611615. RESEARCH PAPER
Abstract: The porosity and permeability distribution in four layers of the Cretaceous Ilam Formation was simulated using optimized
artificial intelligent algorithms based on conventional logging data of 50 wells in Mansuri oil field in Iran. First, the neutron porosity,
interval transit time and density wireline logs in five key wells with core data were used as input parameters to calculate porosity and
permeability of the reservoirs using backpropagation artificial neural network (BP neural network) and Support Vector Regression
methods, and based on the correlation between the calculated results and the core tested results, BP neural network method was taken to
do the physical property calculation. Then, the porosity and permeability distribution of the four layers were modeled using kriging
geostatistical algorithms. The results show that Layers 2.1 and 2.2 are high in porosity, Layers 1, 2.1 and 2.2 are high in permeability,
while Layer 3 is not reservoir; and the porosity and permeability are higher in the north and lower in the south on the whole.
Key words: reservoir; physical property modeling; BP neural network; Support Vector Regression; Mansuri oil field; Zagros region
Fig. 1. Position and stratigraphic column of Mansuri Oilfield (modified from references[47]).
over the whole Mansuri Oilfield, the smooth effect of the Table 1. Correlation coefficient between conventional log pa-
Kriging algorithm has been obviously reduced, and the rameters and porosity and permeability
modeling results are reliable. Finally, the attribute model Conventional logs W/porosity W/permeability
obtained from modeling was used to analyze the reservoir Sonic differential time 0.73 0.60
property of four zones in the Ilam Formation of Mansuri
Density wireline log 0.69 0.58
Oilfield.
Neutron porosity 0.65 0.63
2.1. Neural network modeling Gamma Ray 0.54 0.45
Firstly, the core porosity and permeability data of four wells Photoelectric 0.50 0.48
were used to train the artificial neural network, and the Conductivity temperature 0.43 0.39
physical property data of cores taken from the fifth well were
Table 2. Statistics of data used in artificial neural network
used to test the network, so as to guarantee the accuracy. At
Number of Number of
the time of training the network, the conventional log Data set
porosity data permeability data
parameters having the best correlativity to porosity and
Training set 712 356
permeability were screened out. The correlation coefficient of
Validation set 152 76
some conventional log parameters and physical property data
Testing set 152 76
is listed as in Table 1, showing that the neutron porosity, sonic
All data 1 016 508
differential time and density log value have the highest
correlativity to porosity and permeability, therefore, the placed in training set to ensure the trained network to be
neutron porosity, sonic differential time and density log are reliable.
selected as the input parameters in the paper. All the data used
2.2. Support vector regression modeling
in artificial nueral network training and the statistics of them
are listed as in Table 2, showing that about 70% input data are The input data set of support vector regression modeling is
612
ALI Dashti et al. / Petroleum Exploration and Development, 2016, 43(4): 611615
the same as that of BP neural network method, including network. Because the permeability has more influential
neutron porosity, sonic differential time and density log. A factors than the porosity, the permeability prediction accuracy
forward selection and backward elimination mechanism was is not as good as porosity prediction accuracy. The correlation
run to select these logs as the input data set. In the forward coefficient in Table 1 also shows that the correlativity between
selection process, the model was designed without any conventional log parameters and permeability is not as good
variable in the start point; in the next steps, all of variables as that between conventional log parameters and porosity. As
were input in the model and their effects were monitored by a shown in Fig. 2c and 2d, the correlation coefficient between
selected model comparision standard; iterative computation permeability precited by two methods and permeability
was ceaselessly conducted and did not stop until the input and analyzed from cores is 0.825 and 0.814 respectively. Ob-
output data of the model have the highest correlativity. The viously, when BP neural network and support vector regres-
function of the model comparision standard is to recognize the sion methods are used to model the porosity and permeability
best correlation coefficient between input variables and respectively, the correlativity of BP neural network method is
outputs. In the backward elimination step, all variables were superior to that of support vector regression method, possibly
input at the start point and would be eliminated one by one in due to the repetitive network training process adopted in the
the next stages. The model comparision standard would record neural network method. As listed in Table 1 and shown in Fig.
every variable elimination effect on accuracy and stop the 2, the correlativity between the porosity and permeability
process at the most reliable state. figured out by BP neural network method and that measured
from cores is obviously better than the correlativity between
2.3. Method correlation and optimization
log parameters and core porosity and permeability. Therefore,
The correlation between the porosity and permeability the BP neural network method is adopted in the paper to
predicted from the BP neural network and support vector calculate the porosity and permeability of target zones in 50
regression methods and anzlyzed from the cores are shown as wells and build attribute model.
in Fig. 2. As shown in Fig. 2a and 2b, the correlation
2.4. Physical property distribution modeling
coefficient between porosity predicted by two methods and
porosity analyzed from cores is 0.912 and 0.899 respectively, The planar size of geologic model of the Mansuri Oilfield
ensuring the the porosity prediction accuracy of the trained is 40 km 6 km, and the planar grid size is 100 m100 m; the
Fig. 2. Crossplot of physical property values analyzed from cores and predicted by BP neural network and support vector regression
methods (Rcorrelation coefficient).
613
ALI Dashti et al. / Petroleum Exploration and Development, 2016, 43(4): 611615
geologic model is subdivided into 283 layers vertically, with zone 2.1 has the highest porosity value, and it can be regarded
vertical grid size of 1 m; as a result, the model is composed of as the main reservoir of Ilam Formation; the porosity
6 792 000 grids. The Kriging algorithm was used to model the distribution exhibits an increase trend from the south to the
physical property parameters of four zones (zone 1, zone 2.1, north. Zone 2.2 has the secondary high porosity value in Ilam
zone 2.2 and zone 3) in Ilam Formation, with the modeled Formation; zone 1 has lower porosity value, almost exbithing
porosity and permeability distribution of the four target zones a nonporous rock; whereas zone 3 has the lowest porosity
as shown in Figs. 3 and 4 respectively. As shown in Fig. 3, value, and it is classified into nonreservoir.
3
Fig. 4. Permeability distribution in Ilam zones based on modeling (Kpermeability, 10 m2).
614
ALI Dashti et al. / Petroleum Exploration and Development, 2016, 43(4): 611615
Based on the calculation results of the artificial neural Zagros Fold-Thrust belt, Iran. Marine and Petroleum Geology,
network method, the Kriging geostatistical algorithm is used 2004, 21(7): 829843.
to model the porosity and permeability in four zones of Ilam [8] JAMES G A, WYND J G. Stratigraphic nomenclature of Ira-
Formation of the Mansuri Oilfield, showing that zones 2.1 and nian oil consortium agreement area. AAPG Bulletin, 1965,
2.2 are high porosity zones, zones 1, 2.1 and 2.2 are 49(12): 21822245.
permeable zones, the physical property of zone 1 is [9] MOTIEI H. Geology of Iran, the stratigraphy of Zagros. Te-
complicated due to the development of local fractures, and hran: Geological Survey of Iran Publication, 1993.
zone 3 is a nonreservoir, whose porosity and permeability [10] ADABI M H, ASADI-MEHMANDOSTI E. Microfacies and
value are all relatively low. The reservoir porosity and geochemistry of the Ilam Formation in the Tang-E Rashid
permeability distribution exhibits a characteristic of high in area, Izeh, SW Iran. Journal of Asian Earth Sciences, 2008,
the north and low in the south on the whole. 33(3/4): 267277.
615