Professional Documents
Culture Documents
181-Article Text-292-1-10-20200406
181-Article Text-292-1-10-20200406
181-Article Text-292-1-10-20200406
PETROVIETNAM JOURNAL
Volume 6/2019, p. 48 -
57 ISSN-0866-854X
Summary
Oil production forecast is a major challenge in the oil and gas industry. Simulation model and prediction results play an important role in field
operation and management. Currently, production forecast problems are resolved mainly by using pure traditional prediction methods. Generally,
production forecast by dynamic simulations does not provide reliable results in case where a lot of uncertain parameters remains when the dynamic
model is constructed.
In fact, in Vietnam, the dynamic models of fractured reservoirs give unreliable results and differ with actual performance. It is a challenge to
build and design reasonable production plans for fractured granite reservoirs in Vietnam. In order to replace the disadvantages of simulation model
by different methods, a growing trend of research in the world is predictive tools by using machine learning algorithms.
The paper introduces the applicability of machine learning through the artificial neural network to predict oil production for basement formation
- Bach Ho field. The research results show that Artificial Neural Network (ANN) model has improved the ability to predict production with high
accuracy.
Key words: Artificial Neural Network, machine learning, oil production, reservoir management, Bach Ho field.
1. Introduction
Date of receipt: 20/2/2019. Date of review and editing: February 20 - April 1, 2019.
PETROVIETNAM
One of the most popular machine learning methods - the output layer back, through the hidden layers, to the
ANN is employed for this purpose. In computer science, input layer. This process is repeated until the connection
ANN is formed of computer architecture, inspired by bio weights produce an output which is with a tolerance of the
logical neural networks (the central nervous systems of desired output [2].
animals, particularly, the brain) and used to estimate or
The selection of an optimum architecture of a model
approximate functions that can depend on a large number
is a difficult task requiring a procedure of trial and error [5].
of inputs and are generally unknown. ANN is gener ally
Thus, several networks with various numbers of hid den
presented as systems of connected “neurons” which can
units, training algorithms, and activation functions are
compute values from inputs and are capable of machine
attempted and the generalization error is estimated for
learning or pattern recognition, thanks to their adaptive
each. The network with the minimum estimated gen
nature. Figures 1 and 2 show the basic biological neuron eralisation error is chosen.
structure and representation of artificial neuron.
3. Production data of basement formation of Bach Ho
2. Neural networks
field
70000 500
Prediction
Training data interval Interval 450
60000
400
50000
350
300
40000
NOP,Reservoir
pressure
(at)
Beginning of 250
water ooding
30000
Production
Injection
day),
day)
(m3/
(ton/
rate
rate
200
150
20000
100
10000
50
0 0
0 50 100 150 200 250 300 350 400
Time (Months)
2018, the total reservoir achieved a cumulative oil mum error. The testing set used 47 data months (16.5%)
production of 180 million tons, accounting for 86% of to fine-tune the network model. It is not applied for training
Vietsovpet ro's oil production, with an average oil rate of ing and validation process, only used to identify optimal
6,000 tons per day, and an average water cut of 60% [6]. network architecture, to select a suitable network model
and assess their performance.
4. Network architecture
- ANN network architecture
Neuron architecture is composed of five inputs and
The best results were obtained from ANN sub model
three outputs. The inputs are the average field oil
sisting of 2 hidden layers and 50 neurons for each one.
production rate (FOPR) at time t, the average field liquid
The node in the hidden and output layers is activated
production rate (FLPR) at time t, the average reservoir
through Sigmoid function and trained by the Backpropagation
pressure (FPR) at time t, the average water injection rate
Neural Network algorithm (BPNN).
(FWIR) at time t+1 and the number of production wells (NP)
at time t+1. The outputs are the average field oil production 4.2. Long-term production prediction
rate (FOPR) at time t+1, the average field liquid production
- Data pre-processing
rate (FLPR) at time t+1, the average reservoir pressure
(FPR) at time t+1. The selection of an optimum neural The first set used 236 data months (from May 1993 to
network architecture can be achieved using a trial and er December 2012) to build a network model. The second set
ror approach. Figure 3 shows the oil production rate from used 60 data months (from January 2013 to Decem ber
September 1988 to May 2018. 2017) to predict oil production rate, liquid production rate,
and reservoir pressure. The training set used 160 data
4.1. Short-term production prediction months (67%) to calculate gradient and update the net
work weights and biases. The validation set used 38 data
- Data pre-processing
months (16.5%) to evaluate the quality of the training
Normally, an accurate network model can be achieved process. Training can be stopped when the performance
without adequate data. Therefore, before training model, of the model on the validation dataset provides a mini
production data have to guarantee high reliability to avoid mum error. The testing set used 38 data months (16.5%)
peculiar answers from trained network model. However, to fine-tune the network model. It is not applied for training
depending on the problem, there may be special features ing and validation process, only used to identify optimal
from the data that are able to test its quality. One way to network architecture, to select a suitable network model
check the quality is to view the graphical representations and assess their performance.
of the data in question, in the hope of selecting a reason
- ANN network architecture
able subset while eliminating portions. As presented in
Figure 3, the oil field production rate is time dependent The best results were obtained from ANN sub model
and was split into two sets. The first set (from May 1993 to sisting of 1 hidden layer and 60 neurons for each one. The
December 2016) used 284 data months to build the network node in the hidden and output layers is activated through
model. The second set (from January 2017 to May 2018) Sigmoid function and trained by the Backpropagation
used 15 data months to predict the average oil production Neural Network algorithm (BPNN)
rate, liquid production rate, and reservoir pressure.
5. Assessing and comparing the production prediction
results of ANN model and those of the dynamic simu
To avoid overfitting or underfitting results and im prove lation model
the generalization of the network model, the first set was
5.1. Evaluating short-term production prediction results
subdivided randomly into three parts: training, validation, from the ANN model
and testing. The training set used 190 data months (67%)
to compute the gradient and update the network weights The statistic method is used to assess the accuracy of
and biases. The validation set used 47 data months (16.5%) the ANN model in the training, validation, and testing
to evaluate the quality of the training process. Training can process (Table 1) through the average absolute error (AE)
be stopped when the performance of the model on the and average relative error (ARE) of three parameters: oil
validation dataset provides a mini production rate , liquid production rate, and reservoir pressure:
PETROVIETNAM
- Training sets:
Data_Training
st + AE: 526 tons/day, 637 tons/day, 6at;
40000 350
35000 300 + ARE: 3.11%, 3.13%, 2.47%;
30000
250
25000 - Validation set:
200 Reservoir
pressure
(at)
20000
150
+ AE: 998 tons/day, 1112 tons/day, 6.67at;
Production
day)
(ton/
rate
15000
100
10000
5000 50 + ARE: 5.51%, 5.26%, 2.76%;
Figure 4. Performance of short-term prediction training set. The errors are in the allowable limit. The results
of training, validation and testing processes are
described in Figures 4, 5 and 6.
Data_Validation
1stDataset To study the robustness and accuracy of the
40000 350
network approach, with respect to predicting oil
35000 300
reservoir production, the second dataset was used to
30000
250
predict the reservoir oil production. The predict ed
25000
200
Production
day)
(ton/
rate
20000
Reservoir
pressure
(at)
M_Data_Testing FLPR M_Data_Testing FPR + AE: 1215 tons/day, 1261 tons/day, 7.69at;
PETROVIETNAM
4000 25000
200
3000 20000 Reservoir
pressure
(at)
Production
day)
(ton/
rate
150
15000
2000 100
10000
1000 50
5000
0 284 286 288 290 292 294 296 298 300 302 00 50 100 150 200 0 250
Time (Months) Time (Months)
Data_Predicted FOPR H_Data_Prediction FOPR
H_Data_Validation FOPR H_Data_Validation FLPR
Liquid
day)
(ton/
rate
9000 35000 300
8000 30000 250
7000
25000
6000 200
20000
5000 Production
day)
(ton/
rate
150
Reservoir
pressure
(at)
4000 15000
280 285 290 295 300 305 100
10000
Time (Months) 5000 50
Data_Predicted FLPR H_Data_Prediction FLPR
00 50 100 150 200 0 250
Time (Months)
150
Reservoir
pressure
(at)
The errors are in the allowable limit. The results of
100
training, validation and testing processes are described in
50
Figures 8, 9, and 10.
0 284 286 288 290 292 294 296 298 300 302 To study the robustness and accuracy of the network
Time (Months)
Data_Predicted FPR H_Data_Prediction FPR approach, the second dataset was used to predict the res
ervoir oil production. The predicted reservoir oil rate val
Figure 7. Prediction of average oil production rate, liquid production rate and reservoir
pressure (from January 2017 to April 2018).
ues agree with the historical values indicating the training
network can serve as a practical robust reservoir utility
tion management tool (Figure 11). The network provides
Data_Training
1stDataset reservoir oil rates with an average AE of 698 tons/day and
40000 350
35000 300 average ARE of 12.61%, as illustrated in Table 2.
30000 250
25000 200 5.3. Comparing the production prediction results of
20000
Reservoir
pressure
(at)
150
15000 ANN model and dynamic simulation model results in
Production
day)
(ton/
rate
100
10000
5000
50 the short term and in the long term
00 50 100 150 200 0 250
Time (Months)
- Comparing the results of short-term production
H_Data_Train FOPR H_Data_Train FLPR M_Data_Train FLPR prediction and those of long-term production prediction
M_Data_Train FOPR H_Data_train FPR M_Data_Train FPR
From Figures 12 and 13, it is obvious that short term
Figure 8. Performance of long-term prediction training set. oil production prediction of ANN model (284 data
PETROVIETNAM
Data_Prediction FOPR
Data_Prediction FOPR nd
2 Dataset FOPR nd
Data_Prediction Data_Prediction FOPR
9000 2 Dataset nd
2 Dataset 2nd Dataset
9000 9000
8000
9000
8000 8000
7000
8000
7000 7000
6000
7000
6000 6000
day)
(ton/
rate
Oil
5000
6000
5000 5000
4000
day)
(ton/
rate
Oil
day)5000
(ton/
rate
Oil
day)
(ton/
rate
Oil
4000 4000
3000
4000
3000 3000
2000
3000
2000
1000
2000 2000
1000
0 1000
1000 246 256 266 246 256 266 276 286 296 306 286 296 306 286
0 276 266 296 306
236 0 236 236 246 256 Time (Months) 276
0 236 246 256 266 276 286 296 306
Time (Months) Time (Months)
Data_Predicted FOPR Time (Months)H_Data_Prediction FOPR
Data_Predicted FOPR H_Data_Prediction FOPR Data_Predicted FOPR H_Data_Prediction FOPR
Data_Predicted FOPR H_Data_Prediction FOPR
Liquid
day)
(ton/
rate
10000
10000 Liquid
Liquid
day)
(ton/
day)
(ton/
rate
rate
6000
10000
8000
8000 5000
8000
6000 4000
6000
6000
4000 3000
4000 246 256 266 276 286 256 266 276 286 296 2000
236 4000 246 296
236 236 246 Time
256 266 (Months)
276 286 296 1000
Time (Months)
Data_Predicted FLPR Time (Months)H_Data_Prediction FLPR
Data_Predicted FLPR H_Data_Prediction FLPR 0 230 240 250 260 270 280 290 300
Data_Predicted FLPR H_Data_Prediction FLPR
Time (Months)
FOPR_Actual FOPR_Simulation FOPR_ANN
Data_FPR
300 Data_FPR Figure 14. The results of oil production rate prediction.
300 Data_FPR
300
250
250
250 Data_FLPR
200
200 16000
200
150 15000
Reservoir
pressure
(at)
150 14000
150 13000
100
Reservoir
Reservoir
pressure
pressure
(at)
(at)
100 12000
100 11000
50 10000
50
9000
0 Liquid
day)
(ton/
rate
8000
50 0 230 260
240270 280250
290 300 0 230 240 250 260 270 280 290 300 Time 7000
(Months) 230 240 250 260 270 280 290 300 Time (Months) )
6000
5000
FPR _Actual
Time (Months)FPR _ANN 4000
FPR _Actual FPR _ANN FPR 3000
_Actual FPR _ANN Figure 11. 2000
Prediction of average oil production rate, liquid production rate and reservoir pressure
1000
(from January 2013 to December 2017).
0 230 240 250 260 270 280 290 300
Time (Months)
Data_Prediction FOPR
FLPR _Actual FLPR _Simulation FLPR _ANN
2nd Dataset
7000
Figure 15. The results of liquid production rate prediction.
6000
5000 months) gives better results than long-term oil
4000
day)
(ton/
rate
Oil
Data _FPR
400
350
300
250
Reservoir
pressure
(at)
200
150
100
50
0
230 240 250 260 270 280 290 300
Time (Months)
FPR_Actual FPR_Simulation FPR_ANN
high accurate prediction, corresponding to production Observation: From the experimental results, the per
input data. The ANN method forecasts input data without form of the two models was evaluated, showing:
being based on the subjective experiment of professors.
- Dynamic simulation model:
After the training process, the ANN model will actively
determine weights for each input parameter and their The model is controlled by a constant value of liquid
relationship. yes, the results of the ANN model are more production rate (LPR) in the period from January 2013
trustworthy than the traditional prediction method. to December 2017. As a matter of fact, LPR did not
remain stable due to production operation (well shut-in,
- In the training set, the network will regulate input
facility maintenance, weather conditions…). The
parameters to satisfy mean squared error value, ANN's
qualification assump tion of production depends on
convergence ability depends on original arguments, so
experiment, simulator's subjective and field development
many sensitivity scenarios must be run to choose the
plan. On the other hand, the simulation oil production
best original arguments. On the other hand, the training
rate has high deviation compared with history at starting
in complicated networks becomes more difficult than
forecasting date, the model does not obtain reliability
shallow and narrow networks, in which case the
and neither does it capture geological complications,
optimisation is more likely to converge to some useless
rock property distribution, fracture network, and
local optima. Ideally, we would like to design a model of
hydrodynamic connec tivity of granite basement. Until
reasonable complexity but powerful representation for now, there is not a granite basement simulation method
the data we feed into it. Furthermore, to avoid overfitting
that is accurate, reliable and widely recognized.
the model, the size of the training data has also to be
considered in the design. Therefore, taking all these - ANN model:
concerns into account and after several trials on the
The parameters: oil production rate, liquid production
validation dataset, one layer of the hidden layer with the rate, and reservoir pressure are very close to the actual
proper number of neurons fits the best. data, the trend of results and actual production match
- Figures 14, 15 and 16 show comparison between closer than the dynamic model. however, the confine
the results of long-term production prediction of the ANN ment of the ANN model only applies to predict short term.
model and those of the dynamic simulation model.
PETROVIETNAM
6. Conclusion and recommendation using machine learning. SPE Argentina Exploration and
Production of Unconventional Resources Symposium, Buenos
This research work aims to present a new approach to
Aires, Argentina. 1 - 3 June, 2016.
predict oil production rate based on the historical production
data. The results of methodology show prediction problem 2. Brian D. Ripley. Pattern recognition and neural networks.
generalization ability on the ANN model, become an effective Cambridge University Press. 1996.
implement to resolve variable problems in operation and 3. John R. Koza, Forrest H. Bennett, David Andre, Martin A.
management field production techniques. Keane. Automated design of both the topology and sizing of
The ANN model has many features: data learning possibility, analog electrical circuits using genetic programming. Springer.
adaptation, a decision with deficiency or noise data, which are a 1996: p. 151 - 170.
significant advantage compared with numeri cal simulation.
4. Jerome H. Friedman. Data mining and statistics: What's
the connection. Computing Science and Statistics.
- ANN application will be more effective when the first stages 1997.
such as training samples, extracting characteristics and pre-
5. Jeff Heaton. Introduction to neural networks for java (2nd
processing are well done;
edition). Heaton Research. 2008.
- ANN model postulates more time to train and
6. Tran Van Hoi, Nguyen Van Duc, Pham Xuan Son.
adjust network argument;
Oil exploration and development in the foundation rock of Bach
- As for the future work, other particularly different algorithms Ho mine: Documentation, facts and lessons learned. Vietsovpetro.
and input data effected to production prediction such as well 2018: pages 7 - 20.
bottom hole pressure, choke size, and gas lift rate will be
7. Yunan Li, Yifu Han. Decline curve analysis for production
integrated.
forecasting based on machine learning.
Acknowledgement SPE Symposium: Production Enhancement and Cost Optimization,
Kuala Lumpur, Malaysia. 7 - 8 November, 2017.
This work is supported by the Ministry of Industry and Trade
of Vietnam, project No. CNKK.001/19.
8. A. Mirzaei Paiaman, S. Salavati. The application of artificial
References
neural networks for the prediction of oil production flow rate.
Energy Sources, Part A: Recovery, utilization, and environmental
1. Q.Cao, R.Banerjee, S.Gupta, J.Li, W.Zhou, B.Jeyachandra.
effects. 2012; 34(19): p. 1834 - 1843.
Data driven production forecasting