Multilabel classification for inflow profile monitoring

MACSPro'2019 - Modeling and Analysis of Complex Systems and Processes, Vienna
21 - 23 March 2019

Ivan Vrabie, Dmitry Ignatov, Pavel Spesivtsev, Vladimir Zyuzin, Dmitry Kurgansky, Svyatoslav Elizarov

Conference website http://macspro.club/

Website https://exactpro.com/
Linkedin https://www.linkedin.com/company/exactpro-systems-llc
Instagram https://www.instagram.com/exactpro/
Twitter https://twitter.com/exactpro
Facebook https://www.facebook.com/exactpro/
Youtube Channel https://www.youtube.com/c/exactprosystems

Transcript

1. Schlumberger-Private
   Multilabel Classification for Inflow Profile Monitoring
   Dmitry I. Ignatov, Pavel Spesivtsev, Dmitry Kurgansky, Ivan Vrabie, Svyatoslav Elizarov, Vladimir Zyuzin
   22.03.2019
   

   View Slide

2. Schlumberger-Private
   The inflow zones (sources) problem
   Active inflow source
   Inactive inflow
   source
   Goal:
   • Determine the active
   and inactive inflow
   sources
   Effects:
   • Efficiency decrease
   Applications:
   • Decision making:
   reperforating
   • Understanding
   performance: A small
   amount of active inflow
   sources or small
   productivity of the
   sources.
   BHP
   WHP Q
   

   View Slide

3. Schlumberger-Private
   The machine learning approach
   Target variable: The binary vector of size N of active and
   inactive inflow sources.
   Features: Time series of surface flowrates, BHP and WHP.
   Models: Random Forest, XGBoost, SVM, kNN, CNN and LSTM
   Auxiliary methods:
   • Feature engineering
   • Dimensionality reduction
   • Ensemble of algorithms
   • Cascade algorithms
   

   View Slide

4. Schlumberger-Private
   Data
   Numerical simulator
   Input parameters:
   • Wellbore geometry
   • Distribution of volume fractions of phases
   • Choke size
   • …
   Output variables
   • Surface flowrates
   • Bottomhole pressure
   • Wellhead pressure
   • Vector of active and non-active sources
   5000 numerical simulations
   Train test split: 4:1
   • Wellbore
   geometry
   • Distribution
   of volume
   fractions of
   phases
   • Choke size
   • Qo, Qg, Qw
   • BHP
   • WHP
   • Target vector
   Numerical
   simulator
   

   View Slide

5. Schlumberger-Private
   Feature generation
   Features obtained using TSFresh:
   • Average
   • Standard deviation
   • Median
   • Dispersion
   • Min/Max value
   • Trend
   • Number of min/max values
   • …
   • ~ 1200 features
   Features used for solving the multi-label
   classification problem
   

   View Slide

6. Schlumberger-Private
   Experiment 1
   The approach of independent classifiers for each of the inflow sources.
   Normalization: standard, min-max
   Dimension reduction: PCA, ICA
   Methods: RF, SVM, kNN and XGBoost
   Best results: XGBoost + standard normalization + PCA
   0/1 loss: 0.36
   

   View Slide

7. Schlumberger-Private
   Experiment 2
   Ensemble approach: top 10 algorithms + majority voting. 0/1 loss: 0.31
   Algorithms:
   1. XGBoost + PCA + standard norm.
   2. RF + PCA + standard norm.
   3. XGBoost + PCA + min-max norm.
   4. RF + PCA + min-max norm.
   5. XGBoost + PCA
   6. RF + ICA
   7. SVM + ICA + standard norm.
   8. RF + PCA
   9. SVM + PCA + min-max norm.
   10. kNN + ICA + standard norm.
   Accuracy matrix
   

   View Slide

8. Schlumberger-Private
   Experiment 3
   Cascade classifier approach:
   1. Predict the number of working sources (100% accuracy).
   2. Obtain probabilities of class 1 for each source separately.
   3. Sort probabilities in descending order.
   4. Select the sources with the highest probabilities.
   0/1 loss: 0.44
   

   View Slide

9. Schlumberger-Private
   Experiment 4
   Enlarged feature space approach
   • 300 ICA components over the time series
   • 300 PCA components over the new features
   • Number of active sources
   Method: XGBoost
   0/1 loss: 0.26
   

   View Slide

10. Schlumberger-Private
    Summary and Conclusions
    • The problem of inflow profile monitoring is hard to handle.
    • The XGBoost method over initial data + extracted features achieved a 0/1 loss of 0.26
    • The sources that are closer to the surface are easier to predict.
    • The results are better than a random guess and they show a potential possibility in
    improvement.
    XGBoost, PCA Ensemble of
    10 algorithms
    Cascade
    classifier
    XGBoost +
    PCA + ICA
    0/1 loss 0.36 0.31 0.44 0.26
    

    View Slide