The Importance of Integrating Subsurface Disciplines with Machine Learning when Predicting and Optimizing Well Performance – Case Study from the Spirit River Formation
- John Hirschmiller (GLJ Petroleum Consultants Ltd.) | Anton Biryukov (Verdazo Analytics) | Bertrand Groulx (Verdazo Analytics) | Brian Emmerson (Verdazo Analytics) | Scott Quinell (GLJ Petroleum Consultants Ltd.)
- Document ID
- Society of Petroleum Engineers
- SPE Annual Technical Conference and Exhibition, 30 September - 2 October, Calgary, Alberta, Canada
- Publication Date
- Document Type
- Conference Paper
- 2019. Society of Petroleum Engineers
- Spirit River, Data Analytics, Machine Learning, optimization, Unconventional
- 7 in the last 30 days
- 245 since 2007
- Show more detail
- View rights & permissions
|SPE Member Price:||USD 9.50|
|SPE Non-Member Price:||USD 28.00|
This machine learning study incorporates geoscience and engineering data to characterize which geological, reservoir and completion data contribute most significantly to well production performance. A better understanding of the key factors that predict well performance is essential in assessing the commercial viability of exploration and development, in the optimization of capital spending to increase rates of return, and in reserve and resource evaluations.
Machine learning models provide an objective, analytical means to interpret large, complex datasets. Generally, such models demand large databases of consistently evaluated data. As geological data is interpretive, often varying from one geologist to another, or from one pool to another, it can be difficult to incorporate geological data into regional machine learning models. Consequently, efforts to use machine learning in the oil and gas industry to predict well performance are often focused exclusively on engineering completion technology. However, this case study has utilized a regional geological Spirit River database with consistent petrophysical evaluation methodology across the entire play. This geological database is complemented with public completion and fracture data and production data to build predictive models using inputs from all subsurface disciplines.
Redundancies in the data were identified and removed. Features explaining a significant proportion of the variance in production were also removed if their effect was captured by more fundamental, correlated features that were more straightforward to interpret. The dataset was distilled to 13 key features providing predictions with a similar precision to those obtained using the full-featured dataset.
The thirteen features in this case study are a combination of geological, reservoir and completion data, underlining that an approach integrating both geoscience and engineering data is vital to predicting and optimizing well performance accurately for future wells.
|File Size||1 MB||Number of Pages||20|
Breiman, L., 2001, Random Forests: in Machine Learning, 45: 5. https://doi.org/10.1023/A:1010933404324
Chen, T. and Guestrin, C., 2016, XGBoost: A Scalable Tree Boosting System: in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 785–794 http://doi.acm.org/10.1145/2939672.2939785
Molnar, C., 2017plus later updates, online book, Interpretable Machine Learning: https://christophm.github.io/interpretable-ml-book