Loading...
Predicting oil production from stochastic AVA inversion attributes
Powers, Hayden A.
Powers, Hayden A.
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2019
Date Submitted
Keywords
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
2019-12-11
Abstract
Interpreting seismic information for future drilling locations is vital to the success of many oil and gas companies. These interpretations are primarily made on time or depth volumes or from derived attributes. However, inversions of these seismic data sets can solve for Earth properties of the reservoir and surrounding formations. This dissertation analyzes results from stochastic amplitude versus angle (AVA) inversions to train machine learning algorithms with the end goal of predicting cumulative oil production. We utilize two AVA inversions, the first from the SEAM Life of Field Model (SEAM) and the second from field data acquired over an offshore reservoir in West Africa (WAF). The two main algorithms we use are a Naive Bayesian Classifier (NBC) and Multidimensional Scaling (MDS), to make production predictions in both fields. MDS is an unsupervised method often used to understand how similar information is in a low-dimensional space. The NBC is a supervised technique we employ to predict high or low oil producing reservoir locations. Previously known wells in the reservoir determine the decision to classify locations or wells as high or low. The wells are split into high and low groups based upon their cumulative oil productions or total injected water. Sensitivity testing on the accuracy of the classifier is a major project focus, and requires evaluating many inputs to the NBC. We capitalize on the bulk accuracy from cross validation to evaluate the experiment parameters. The execution of the cross validation is exhaustive and based upon the number of omitted wells. In general, SEAM has a lower bulk accuracy than WAF, but has lower variance associated with changing the boundaries between high and low producing wells. We validate the boundaries by investigating the lower-dimensional space to understand specific misclassifications. Lastly, we incorporate the RMS error from the SEAM AVA inversion to aid in the quality control of the full reservoir classification results. The final reservoir classifications are accurate and prove the merit of using the NBC as a classification algorithm.
Associated Publications
Rights
Copyright of the original work is retained by the author.