Accelerated Production Forecasting with Deep Learning
Published at Jan 1, 2021
Summary
This project delivered a deep learning model for forecasting well production in the Kuparuk reservoir, using well construction data, reservoir parameters, operational data, and offset injection rates and pressures. The model generated forecasts in approximately 10 minutes, compared to 6 hours required by traditional reservoir simulators—a 36x reduction in time to insights.
Project Highlights
- Trained a deep learning model to forecast production using comprehensive well and reservoir data
- Achieved forecast accuracy on par with established reservoir simulation models
- Leveraged distributed computing on HPC via Spark to run tens of thousands of sensitivity cases for rapid scenario analysis
- Published results to AWS Athena tables for easy, scalable access and analysis by reservoir engineers
- Incorporated model outputs into pattern health reviews and drillsite reviews, supporting operational changes and decision making
- Developed a user-friendly interface for engineers to run custom cases and analyze impacts in minutes
- Enabled rapid estimation of operational impacts, such as pump downtime, supporting field management and optimization
- Provided SHAP value explainability for forecasts, making results more intuitive and transparent for users
- Expanded access to forecasting tools from a small group of simulator experts to the entire engineering team, enabling broader scenario analysis and collaboration
Technical Innovation
- Applied deep learning to reservoir production forecasting, replacing time-intensive simulation workflows
- Integrated distributed computing and cloud data storage for scalable, high-throughput analysis
- Automated scenario generation and result retrieval for large-scale field analysis
Impact
The accelerated forecasting workflow transformed reservoir management in Kuparuk, enabling engineers to analyze the impact of operational changes in minutes rather than hours. The deep learning model’s accuracy and speed supported more agile decision making, improved field optimization, and enhanced collaboration between reservoir and production engineers.