Title: Predicting openhole minislam logs from casedhole pulsed neutron capture logs using neural networks
1Predicting open-hole mini-slam logs from
cased-hole pulsed neutron capture logs using
neural networks
- Dong (Cynthia) Xue
- August 10th
- Baker Hughes Houston Technology Center
2Acknowledgements
- Sheng fang, Dan Georgi
- Allen Gilchrist, Pingjun Guo
- Darryl Trcka, Freddy Mendez, M.Milan, S.Mathur
- Tsili Wang, Weidong Li.
- Many thanks to all of those I haven't named
3Outline
- Introduction
- Motivation
- Background
- Data selecting and preprocessing
- Selection of PNC Curves
- Removal of Outliers
- Optimization of neural network
- Improvement of Generalization
- Selection of the Training, Validation and Test
Samples - Selection of the Neurons
- Results
- Data from Thailand
- Data from Colorado
4Motivation
- The result of using coil tubing as production
casing - Unavailable Mini-slam data in open holes
- Neutron porosity
- Resistivity
- Density porosity
- Available pulsed neutron data in cased hole
- Neural network as a prediction tool
- Strengths, assumptions, and limitations of a
neural-network approach - Accuracy and variability of prediction results
5Neural Network
- Biases, a sigmoid layer, and a linear output
layer are capable of approximating any function. - Overcome overfitting by improving generalization
- Regularization.
- Early stopping.
-
- Conjugate gradient, Quasi-Newton, and
Levenberg_Marquardt - LM fastest convergence
- Bayesian regularization (BR) better
generalization
6Prior Work
- R.C.Odom - 2000
- Predicted neutron porosity and density parameters
- Used a Multi-Layer network
- M.Mullen - 2001
- Predicted open hole triple combo data.
- Used a commercial neural network software
package. - J.Quirein - 2003
- A good assessment of neutron porosity and density
porosity. - Unavailable to accurately predict deep
resistivity - Non-linearity of the predictions due to the coals
and depth misalignment
7Pulsed Neutron Capture (PNC)
- Logging measurement from the Reservoir
Performance Monitor (RPM). - Generator, SS, LS, XLS
- Time decay spectra for SS and LS detector
- Inelastic energy
- Capture energy
- Count rates or ratios
8Input PNC curves and targets
- Input logs PNC curves
- 1. GR gamma ray ,
- 2. CSS capture gamma ray count rate
from short space(ss) detector, - 3. CLS capture gamma ray count rate
from long space(ls) detector, - 4. SGFC corrected formation sigma,
- 5. RONE ss to ls capture gamma ray
ratio. RONE CSS/CLS, - 6. RIN ss to ls inelastic gamma ray
ratio. RIN ISS/ILS, - 7. SGB1 dual exponential borehole
sigma from ss. - Output logs the mini-slam data
- ILDI deep induction resistivity,
- NPHIOHI neutron porosity,
- RHOBI density.
9Correlation between logs
- Relate open hole data to case hole data from
deterministic method - A comparison of
- cased hole and open
- hole cross plots
10Correlation of open-hole resisitivity with
cased-hole logs
- Single measurements (GRI -- Best Single Input
Measurement)
2. Double measurements (SGFC)
3.Three measurements7. All the seven
measurements
11The stepwise regression of cased hole inputs for
both training and application wells
12The uplift Effect
- Best single input measurement v.s. All seven
measurements
13Removal of Outliers
- Polynomial approximation a discontinuous
function - Removal of discontinuity
- Outliner removal,
- Additional measurements adoption.
14Selection of data samples
- Data samples Training, Validation and Test
- Guard against using too small a percentage of
samples for training.
15Selection of the neurons
- The number of neurons in hidden layers.
- Too few neurons - Underfitting
- Too many neurons - Overfitting
- BR always has less variation
- Less neutron is generally best.
16Results One Well from Thailand
- The top half well is Training well for
resistivity
ILDI
CLS
SGFC
GR
ILDI
CLS
SGFC
GR
17The bottom half well is application well for
resistivity
ILDI
CLS
SGFC
GR
ILDI
CLS
SGFC
GR
18Optimization of training (1)
19Optimization of training(2)
- First half training
- Second half application
- Every other 100 ft for training the rest for
application.
20Two offset wells
21Conclusions and future work
- Neural network predictions are feasible.
- Problems
- Near-far ratio is sensitive to the presence of
gas. - Difficult to characterize borehole and diffusion
effects. - Existing complicated geological environments.
- Solutions
- Establish a model predictability assessment
system to compensate measurements - Adopt RPM-C data instead of RPM-A.
- Include special environments as inputs .
22 23The training well for neutron porosity
24The application well for neutron porosity
25The training well for density
26The application well for density
27(No Transcript)
28(No Transcript)