Research mathematician Andrew Conn (pictured) launched the Reservoir Management and Production Optimization project last year to develop algorithms to optimize petroleum production network simulator parameters, using proxy models and structural constraints. The project will also make its open code available to developers through IBM's Open Collaborative Research (OCR) program.
Launching Reservoir Management and Production Optimization
Conn, an advanced analytics and optimization researcher, was asked to join the technical committee of the Norwegian University of Science and Technology's Center for Integrated Operations while working with Norway's Statoil in 2006. He took advantage of the multiple in-person meetings each year to strengthen IBM's relationship with NTNU. This included working on the OCR project with IBM's summer interns from the university. They focused on optimizing the extraction of oil and gas from the subsea. The goal: create models and simulate numerous scenarios to locate and manage petroleum in the subsea more rapidly and efficiently than currently possible.
Improving the search for underwater oil
The OCR team has relied on various optimization techniques in looking for underwater oil supplies. Such techniques typically use either line search or trust region methods.
Using the line search method, researchers determine where they will begin and in what direction they will go and iterate. Using the trust region method, researchers create a model which is compared with the actual function to be optimized. If there is reasonable agreement between the model and the behavior of the actual function, then the researchers will expand the region of applicability of their model to include a wider area of potential exploration. Conversely, if the actual objective does not behave like the model, the researchers downsize the region of exploration. Iterating on this paradigm, the goal is to home in on the solution in a region for which one has created a sufficiently accurate model.
 |
| Diagram of undersea drilling platform |
The team is also working at simulating the entire field that they are trying to optimize. The model for these simulations includes the various components (wells, pipelines, manifolds, separators) that make up an exploration application. Simulations might be done for pressure drops on the pipelines, for how wells behave, and other real-world oil extraction scenarios. Normally for such problems, some derivatives are unavailable – and both discrete and continuous variables are involved – which have significant consequences for the optimization methods that can be used.
Conn and his colleagues at the Center for Integrated Operations compared their optimization approach with NOMAD Black Box optimization software – a generally well-considered package for optimization without derivatives. Where Conn's algorithm required four iterations to determine an appropriate approximate solution, NOMAD's required 351. Where Conn's team needed to render 82 well simulations, NOMAD needed 23,402. Likewise, Conn's team rendered 1,662 pipeline simulations; whereas NOMAD needed 15,602. Furthermore, the resulting IBM-NTNU result had a greater than 10 percent improvement over NOMAD’s approximate optimal solution.
“You would be surprised at the number of people who don’t know that IBM is engaged in this kind of work,” Conn said. “They can’t believe that we are in the business of helping oil companies save money by optimizing their exploration processes.
“Through the OCR, we are getting companies inside and outside the petroleum industry to understand that if they can improve their models, combined with the optimization, by even one percent, they are going to save millions and millions of dollars.”
This is just one of IBM Research's several upstream petroleum projects. And these techniques can be broadly applied to other industries where simulation and optimization with both discrete and continuous variables are required.