Monthly Luncheon With Guest Speaker Hayden Powers of Colorado School of Mines

04/11/2019 @ 11:30 am – 1:00 pm

Map Unavailable

Date(s) - 04/11/2019
11:30 am - 1:00 pm

Wynkoop Brewery

Categories No Categories

Please Join Us April 11, 2019

With Guest Speaker Hayden Powers

located at Wynkoop Brewery Company

1634 18th Street in Denver, CO

Talk starts promptly at Noon.

Application of machine learning to predict oil production from stochastic AVA inversion attributes

H. Powers1, W. Trainor-Guitton1, and G. Michael Hoversten2 Colorado School of Mines1, Chevron2

The focus of our work is to use machine learning, specifically a Naive Bayes Classifier (NBC), Multi-Dimensional Scaling (MDS), and a simple Artificial Neural Network (ANN), to classify and predict oil production from stochastic AVA inversion attributes. The output attributes of the inversion include density, porosity, Vp/Vs, and p-impedance. The inversion generates hundreds to thousands of samples from the converged chain for every model location. These attribute samples are clipped to be within a defined radius of each wellbore and are then used to train and classify the NBC and ANN. The algorithm cross validates by testing all combinations of wells, based on the number omitted. The results of the cross validation are evaluated using MDS to understand why wells are predicted correctly or incorrectly. MDS is valuable because it incorporates an understanding of similarity between 3D structures around the individual wellbores.

This analysis was done on both the SEAM Life of Field Model (SEAM) and a field dataset from offshore West Africa (WAF). SEAM has 11 producing wells, while WAF only has 3. Three data points is insufficient to running a cross validation, so the training set for WAF is extended to include the 3 injecting wells as pseudo-producers. The results for WAF using the NBC are more accurate than SEAM, showing validity in the extension of the dataset. The NBC also struggles near the fault blocks of SEAM, where the covariance in the inversion most likely extends past the fault planes. Lastly, the ANN has yet to be applied to WAF, but the early results show improvement over the NBC for SEAM. The ANN does a binary prediction of high and low producing which is the most comparable to the NBC. However, once there is more confidence in the predictions, this will be changed to a continuous cumulative oil prediction. The final product is determining and using the optimal parameters from the two supervised methods of the entire reservoir to predict high grade drilling locations and give an uncertainty measurement to those predictions.


Bookings are closed for this event.