Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

OTC-28506-MS

Application of Neural Network in Formation Failure Model to Predict Sand


Production

T. Ketmalee and P. Bandyopadhyay, PTTEP

Copyright 2018, Offshore Technology Conference

This paper was prepared for presentation at the Offshore Technology Conference Asia held in Kuala Lumpur, Malaysia, 20-23 March 2018.

This paper was selected for presentation by an OTC 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 Offshore Technology Conference and are subject to correction by the author(s). The material does not necessarily reflect any
position of the Offshore Technology Conference, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Offshore Technology Conference 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 OTC copyright.

Abstract
Sand management is one of the key component of Bongkot production processes. Current sand production
prediction is based on a model which requires sonic and density logs for all the wells. However, a
combination of complex well architecture and focus on reducing well cost resulted in many wells not having
acquired these important logs. This project has implemented new technique of "Artificial Neural Network"
to solve this problem. Using this method, synthetic logs are generated to obtain the values of missing sonic
and density data. These data are then used in the existing sand models to predict sand production potential.
This project was evaluated with three field cases. The sand failure predictions based on synthetic rock
properties matched with actual sand production. Therefore, the sand prediction workflow has been updated
to include log synthetic if acroustic or density log are missing.

Introduction
Rise of Artificial Intelligent
Application of artificial intelligent (AI) have been growing since 2007.[1] There are technology
breakthroughs in many industries such as helthcare, financial, and transportation. In E&P business, AI
impletmention is rapidly increasing.

Business Dilemma
Sand production management is essential to optimize reservoir production. Bongkot field uses sand
prediction model which relies on sonic transit time and bulk density. However, there are a number of wells
where either or both of these logs are not available due to either cost saving, unfavorable well path, or other
operational issues. Consequently, sand models are absent in these wells. As end of 2016, approximately
52% of the wells has sand prediction model, this statistics is depicted in Figure 1. Without comprehensive
information, optimized reservoir management is difficult to achieve. This may lead to erosion problem,
production loss, and, most importantly, safety issue. Example of surface erosion is depicted in Figure 2.
Although, there are various methods that could estimate acoustic velocity or density, but they fail to meet
2 OTC-28506-MS

required accuracy level. This project attempts to solve this challenge by using artificial neural network
(ANN) to predict sonic slowness or density as inputs for sand failure prediction.

Figure 1—Total number of well and number of well with sand failure model in Bongkot asset.

Figure 2—Surface erosional problem due to sand production.

Artificial Neural Network: From Biological to Digital


Artificial neural network is a computer algorithm that imitates human brain. Consider basic element of the
brain called neuron or nerve call, each neuron processes signal detected at synaptic buds then fire electro-
chemical signal to other neurons. The massive interconnected neurons network collectively forms brain.[2]
Similarly, ANN is a network of distributed interconnection of processing units. The system may consist of a
few to millions of artificial neurons. This structure provides ability to handle complex nonlinear problems.
Furthermore, the network is fault tolerance and can solve problem without a priori model.[3],[4]

Application of Neural Network in Formation Failure Model


The Framework
This new approach is to create model based on artificial neural network. The model is trained to predict
sonic travel time or density by "learning" the responding patterns from other relevant logs such as gramma
ray, resistivity, and neutron-porosity. The work flow is composed of eight steps as outlined in Figure 3.
OTC-28506-MS 3

Figure 3—Artificial neural network process

1. Data Exploring
All relevant log data is determine the relationship with sonic slowness and bulk density. Pattern of
data as well as bad data and outliners are identified.
2. Data Cleaning
Quality of data are critical to achieve accurate and performative model. Bad data and outliners
observed from step 1. are removed.
3. Sampling and Splitting
The dataset is sampled to be used for study. Then, the sampled data is spited into training and data
sets.
4. Transforming
Both training and testing sets are transform in order to improve accuracy and training speed. This
process may consists of data transformation and standardization.
5. Training
ANN network learns corresponding pattern among independent and dependent variables. The
model adjusts its prediction parameter until desirable accuracy is achieved.
6. Testing
The trained network is tested again unseen data to assure generalization of the model.
7. Evaluating
The predictive performance comparison among ANN and other synthetic methodologies, namely,
Gardner equation, depth synthetic, porosity synthetic, and seismic inversion are performed.
8. Predicting
In case that ANN outperforms other approaches, the synthetic values from ANN is fed into sand
model to evaluate sand riskiness of particular well.

Cutting Edge Performance


The neuronet was evaluated with three field cases as follow.
1. Well A
This well was drilled without sonic data acquisition. In March 2014, the well was perforated in multiple
reservoirs. Later in February 2015, flowline was found eroded because of sand production. The initial
4 OTC-28506-MS

investigation by correlation and analogy was not conclusive. Thus, machine learning technique was trialed
in this well.
First of all, ANN network was trained with 800 data points from nearby wells. The training yielded
very accurate sonic transit time as illustrated in Figure 4, all responding from system was within 10%
accuracy envelope. Next, the network passed validation with 200 unseen data points as shown in Figure 4.
Comparison of predicting vs actual sonic transit time of training set for well A.
The prediction methodology was also compared with other synthetic methodologies such as Gardner
equation, depth synthetic, porosity synthetic, and seismic inversion. The evaluation illustrated superior
accuracy above other techniques as demonstrated in Figure 6.

Figure 4—Comparison of predicting vs actual sonic transit time of training set for well A.

Figure 5—Model validating of testing set for well A.

Figure 6—Predicting comparison among ANN and other methodologies of testing set for well A.
OTC-28506-MS 5

Verified system was use to predict missing sonic log data in well A. The synthetic sonic was fed into
sand failure model. The sand model identified only one problematic reservoir. Finally, the diagnosis was
compared to the downhole wireline sand scanner ran in October 2016. The auxiliary model faultlessly match
with field investigation as summarized in Figure 7.

Figure 7—Sand model with ANN input accurately identified sand producing reservoir in well A.

2. Well B
Both sonic and density logs are absent in this well. The added perforation was done in July 2000, soon
after, sand production observed from the well. Therefore, two ANN systems were trained to examine this
case; one for sonic log and another for bulk density. Similarly, the cognitive system managed to predict
sonic and density value with a reasonable accuracy as illustrated in Figure 8 - Figure 11. The new algorithm
outperformed conventional methodologies as explained by Figure 12 - Figure 13.

Figure 8—Comparison of predicting vs actual sonic transit time of training set for well B.
6 OTC-28506-MS

Figure 9—Comparison of predicting vs actual density of training set for well B.

Figure 10—Model validating in sonic transit time of testing set for well B.

Figure 11—Model validating in density of testing set for well B.


OTC-28506-MS 7

Figure 12—Predicting comparison among ANN and other methodologies in sonic transit time of testing set for well B.

Figure 13—Predicting comparison among ANN and other methodologies in density time of testing set for well B.

The predictor cautioned two reservoirs that could fail in well B which was inline with field observation,
the result is summarized in Figure 14.

Figure 14—Sand model with ANN inputs accurately identified sand producing reservoir in well B.

3. Well C
8 OTC-28506-MS

Last but not least, well C is located in area which rock mechanic and stress regime are significantly
different than first two wells. Sonic log was not ran in this well. Numerous reservoirs were initially
perforated in June 2012, sand product was detected later in March 2016. According to the workflow, the
model was trained and tested with information from neighboring wells. The intelligent network provide
powerful predictive performance and surpass other methodologies as compared in Figure 15 - Figure 17.

Figure 15—Comparison of predicting vs actual sonic transit time of training set for well C.

Figure 16—Model validating of testing set for well C.

Figure 17—Predicting comparison among ANN and other methodologies of testing set for well C.

In October 2016, wireline scan detector revealed that only top reservoir failed and produced sand. The
field investigation agreed with warning from model as displayed in Figure 18. This case supports that this
system is applicable in wide range of environment.
OTC-28506-MS 9

Figure 18—Sand model with ANN input accurately identified sand producing reservoir in well C.

Workflow updated
These three pilots demonstrated precise predictively of the system. Therefore, ANN has been adopted into
sand risk assessement process since June 2016. Previous and new workflows are compared in Figure 19.
Depicted by a red block in former procedure, the sand production riskiness was predicted based on analogy
if sonic or density log absent. The results was usually subjective or inconclusive. The new method provides
more objective results by replacing analogy with log systhetic by ANN.

Figure 19—Sand risk evaluation workflow was updated with log synthetic process.

Tangible Benefit
This technique has the potential to save about $2.5 million in logging cost saving. On top of that, risk of
wireline sticking is reduced by omitting lengthly sonic tool.
10 OTC-28506-MS

Into The New Territory


This exploratory project proves that AI application in E&P business is prosperous. Numerous advance
initiatives can be built on this establishment. Immediate opportunities that can directly utilize this system
are, to name but a few, seismic calibration, bad log fixing, various log synthetics, permeability prediction.

Conclusion
This paper depicts the potential of rising technology of machine learning in sand management. The artificial
intelligent can predict acoustic velocity and bulk density more accurate than existing methodologies.
Furthermore, sand model that uses these inputs can pinpoint sand production sources. Ultimately, this
revolution advances reservoir management practice, reduces cost, and curtails risk.

Acknowledgement
Authers appreciate permission from PTTEP, Total E&P Thailand, and Shell Integrated Gas Thailand to
allow publication of this paper. Finally, Authers would like to express gratitude many people who support
this study.

References
1. Cristianini, N. 2016. The road to artificial intelligence: A case of data over theory. Retrieved
from http://www.easybib.com/reference/guide/apa/website
2. Freudenrich, C. and Boyd, R. n. d. How Your Brain Works. Retrieved from http://
science.howstuffworks.com/life/inside-the-mind/human-brain/brain.htm
3. Ali, J. K. 1994. Neural Networks: A New Tool for the Petroleum Industry? Presented at the
European Petroleum Computer Conference in Aberdeen, U. K., 15-17 March 1994. SPE 27561.
4. Kanj, M. Y. 1999. Realistic Sanding Predictions: A Neural Approach. Presented at the 1999 SPE
Annual Technical Conference and Exhibition, Houston, Texas, USA, 3-6 Octerber 1999. SPE
56631.

You might also like