| dc.description.abstract | Estimated ultimate recovery (EUR) predictions are important in the petroleum industry. Many researchers have worked on implementing accurate EUR predictions. In this study, we used machine learning techniques to help predict the EUR range. We analyzed 200 Barnett shale wells with less than 170 months production history. We forecasted the production profile for each well using the modified Arps hyperbolic decline model. With the EUR values for 200 wells available, we forecasted the EUR of wells with limited production history by using three machine learning techniques, neural networks (NNet), support vector machine (SVM) and random forest (RF).
The results show that the 200 sorted EUR values predicted with the commercial decline analysis software, ValNav, follows a lognormal distribution as indicated on a log-probability paper plot. The P90, P50 and P10 EUR values were identified and the low P10/P90 value of 2.3 shows a low variance of EUR in this geologic area.
The production data were separated into eight groups and processed before being fed into the 3 machine learning algorithms. A four-fold cross-validation technique was employed to reduce the generalization error of the trained classifiers. The details of these 3 algorithms were also introduced. NNet performed best with highest test accuracy of 0.97 among the three machine learning algorithms employed with wells of 170 months’ production history. In addition, we also tested the EUR prediction performance with 24, 48, 96, and 170 months’ production history. The result shows that when we predict the wells’ EUR with increasing production history, we could achieve more accurate forecasting performance.
The results in this project can be used to help oil and gas companies make financial decisions based on available production data in the same geologic area. Also, this project can also help provide a basis for researchers who are interested in this direction.
Robustness analysis was implemented. The robustness of the algorithm is defined as the total distance of misclassified types to the correct types. Less total distance corresponds to more reliable and more stable performance for each individual algorithm. The NNet gives more robust performance with 100% misclassified samples classified into the types within one type distance to the correct types. RF is least robust. As the production history increases, the robustness of the three algorithms increases. | en |
