MACHINE LEARNING AIDED PRODUCTION DATA ANALYSIS �FOR �ESTIMATE ULTIMATE RECOVERY FORECASTING

Liuyi Jin

​

Committee

Prof. John Lee

Prof. Duane McVay

Prof. Yoonsuck Choe

Outline

• Introduction
• Data Processing
• Machine Learning Algorithms
• Conclusions

​

2

Introduction

• Unconventional Reservoir: “that cannot be produced at economic flow rates or that does not produce economic volumes of oil and gas without assistance from mass stimulation treatments or special recovery processes and technologies”

----Holditch 2003

3

Introduction

4

Conventional

Reservoir

Decline Curve Analysis

Arps decline model

Unconventional

Reservoir

Modified Arps decline model

Duong’s method, stretched exponential, power’s law

Introduction

• Correctly and reliable estimate ultimate recovery (EUR) is important in providing oil & gas companies with solid decision basis

5

Sharma (Sharma and Lee, 2016) prepared a comprehensive improved workflow for the EUR prediction in unconventional reservoirs

Introduction

• Machine Learning Application in Oil & Gas Company
• Jia (Jia and Zhang 2016) used neural networks to forecast production from the Barnett Shale, and he achieved much more accurate predictions than that with conventional empirical models.

6

Introduction

• Machine Learning Application in Oil & Gas Company

​

• In Anifowose’s paper (Anifowose et al. 2012), artificial neural networks (ANN) and SVM are both employed to predict porosity and permeability of oil and gas reservoirs with carbonate platforms. The results show that SVM performs better than ANN

​

• Aulia (Aulia et al. 2014) states that a subset of the bottom hole pressure (BHP) can be a contributing factor to the oil recovery factor in the field, so Aulia combines Latin Hypercube Monte Carlo (LHMC) and random forest (RF) to identify such subset.

7

Data Process

Preprocessing:

• 200 gas wells
• Barnett Shale
• Predicted using ValNav

​

• 360 months and the abandonment rate, whichever comes first would be assumed to be the end of production

8

Source: DrillingInfo 2017

Data Process

• Forecasting: Modified Hyperbolic Model

9

Robertson 1988

Data Process

10

The switch point was determined at the point where the decline rate (%yr) is equal to 6.5%

Data Process

• EUR Distribution: Lognormal
• Labelling

​

​

​

• Uniform Dimensionality

360 months 🡪 170 months

11

Data Process

• Four-fold cross validation: used in finding the best model parameter
• Four of them will be used for cross-validation, the fifth group will be used as the test data set.
• The overall training accuracy of this algorithm in our problem will be the average of the accuracies that we get from the four repeated training and validation cases

12

Machine Learning Algorithms

• Neural Networks (NNet)
• Support Vector Machine (SVM)
• Random Forest (RF)

13

Machine Learning Algorithms

• Neural Networks (NNet):
• Convolutional Neural Networks (CNN)
• Recurrent Neural Networks (RNN)
• Multi-layer Perceptron (MLP)

14

Simple to implement

High Nonlinearity

​

Machine Learning Algorithms

• Neural Networks (NNet): MLP

15

Normalization needed for logistic activation function

Machine Learning Algorithms

• Neural Networks (NNet): MLP

16

Machine Learning Algorithms

• Neural Networks (NNet): MLP

17

Parameters:

hidden_layer#: 1

hidden_neurons: 163

input neurons: 170

output neurons: 4

activation function: logistic

solver: lbfgs

learning rate = 0.1

momentum = 0.5

Machine Learning Algorithms

• Neural Networks (NNet): MLP

18

Parameters:

hidden_layer#: 1

hidden_neurons: 163

input neurons: 170

output neurons: 4

activation function: logistic

solver: lbfgs

learning rate = 0.1

momentum = 0.5

Machine Learning Algorithms

19

Machine Learning Algorithms

20

Upper bound to lower bound

Solve a convex optimization problem

Machine Learning Algorithms

21

Machine Learning Algorithms

22

SVM Kernel

Sequential Minimal Optimization (SMO)

Machine Learning Algorithms

• Support Vector Machine (SVM)

​

23

Machine Learning Algorithms

• Support Vector Machine (SVM)

​

24

Parameters:

C: 1.0

kernel: linear

Stopping tolerance: 1e-3

the confidence of classifying this well into correct type is 85% when using support vector machine

Machine Learning Algorithms

• Random Forest (RF)

Adopting a principle that a group of “weak learners” can come together to form a “strong learner”

Start with decision trees

25

Machine Learning Algorithms

• Random Forest (RF)
• 2 values to determine which attribute to pick up: Gini index vs. entropy gain

26

Machine Learning Algorithms

• Random Forest (RF)
• we have 87.5% percent confidence that this classification was implemented correctly.

27

Conclusion

• (1) We forecasted production profile for 200 Barnett Shale gas wells using modified hyperbolic model, and found that the EUR values for wells in this data set followed a lognormal distribution, with a variability (P10/P90 ratio) of 2.32, indicating a highly consistent data set with minimal dispersion.

​

• (2) We successfully used three machine learning algorithms, Neural Networks, Support Vector Machine, and Random Forest to forecast EURs for wells with only limited production histories, following training the algorithms with the EUR values information.

​

• (3) The training set accuracies of all three machine learning algorithms were all 100%. After the model training process on training dataset and validation dataset using cross validation technique, we get the test dataset accuracies for each were above 0.85. In particular, neural network algorithm provided the best test accuracy at 0.925, which indicates a 92.5% confidence when classify a well into a EUR range type.

​

28