The goal was to demonstrate a technique for simulation-optimization using genetic algorithms (GA) and artificial neural networks (ANN) during reservoir simulation to improve storage field development decision making.
Lawrence Livermore National Laboratory (LLNL) – project management and research product
Equitrans – storage field and economic data
Livermore, California 94511
Although reservoir simulation is a well-established tool, reservoir simulation coupled with systematic optimization techniques (simulation-optimization) has not been widely applied. When using a reservoir simulator to analyze a field development decision, as the number of competing engineering, economic, and environmental planning objectives and constraints increases, it becomes difficult to track complex interactions and select a single set of development strategies for examination. Using optimization techniques along with simulation runs, the analysis can be extended to all possible combinations of variables, uncovering and highlighting strategies not otherwise obvious to the analyst. This project applied simulation-optimization to the problem of selecting an optimal strategy for a natural gas storage field development decision.
Coupling optimization tools to a reservoir simulator allows for the simultaneous evaluation of reservoir performance and economic, environmental, and policy considerations. The single biggest obstacle to the application of optimization techniques using a reservoir simulator as the forecasting tool is computational time. Extending the use of a simulator to optimization involves hundreds or thousands of simulation runs and can pose a computational problem for many companies.
One solution to this problem is to train artificial neural networks to predict information that a simulator would normally predict. Neural networks are computer codes that seek to recognize patterns and make predictions in ways similar to the human brain. A heuristic technique such as the genetic algorithm then searches for increasingly better strategies (for example, the most productive infill drilling pattern in a field), using the trained networks to evaluate the effectiveness of each strategy in place of multiple simulator runs. After analysis of the results of the search, the best-performing strategies are evaluated using the simulator to confirm their performance. The steps in this process are: (1) create a knowledge base of representative simulations, (2) train the ANN to predict selected simulator results, (3) generate field development plans using the genetic algorithm and use the ANN to predict the results, and (4) verify the optimal plan with the simulator. The number of simulator runs needed for step 1 is relatively small compared to the number of runs that would be required to solve the problem without the GA/ANN optimization approach.
The results of this research indicate that the tools of optimization can facilitate the determination of variations in simulation results in a systematic manner. By enabling planners to experiment freely, the GA/ANN methodology greatly increases the value of reservoir simulators as decision-making tools.
This project looked at a storage company's task of selecting a strategy for determining the optimal combination of facility and operational changes required to offer peaking service in addition to the baseload service already being offered at a specific storage facility. Initial modeling work was done using a 3-D black oil simulator coupled with a deliverability forecasting model to handle the wellbores, surface pipelines, and facilities. A total of 700 runs were required to support the GA/ANN methodology for a given operation scenario: 500 runs to create the knowledge base and 100 verification runs for each of two objective functions. At approximately three minutes/run, this amounted to 35 CPU hours, which was a relatively small investment of machine time. The 700 runs required by the ANN-assisted GA were a small effort compared to the runs required for the full-model-assisted GA. These latter searches employed identical procedures except that a simulator was called to supply predictions instead of obtaining those predictions from the ANNs. In this case, 6,274 unique calls to the simulators, taking 313.7 hours to complete, were required to generate results that were identical to the ANN-assisted results. The difference in effort is almost an order of magnitude.
The optimized solutions to the planning problem that resulted from this exercise were based on a deterministic, “best guess” view of the field's reservoir properties. However, at least some of the uncertainties associated with these properties need to be taken into account in a thorough analysis. After refining the model used for the initial simulation-optimization exercise, the researchers developed and applied procedures for applying the GA/ANN optimization approach so that three sources of uncertainty could be accommodated: alternative hypotheses regarding the permeabilities in a key region of the field, uncertainty regarding the success of remediating existing wells, and risks associated with siting new wells in relatively unknown regions of the field.
The first two sources of uncertainty involve physical properties (permeabilities and skin factors, respectively) that are embedded in the simulation of the reservoir response and, therefore, required substantial changes to the initial knowledge base of simulations. The third source of uncertainty could be examined simply by making changes to the objective functions driving the optimization. Within a few hours, it was possible to construct a new objective function and run many different analyses varying the weights applied to individual terms, whereas completing a single analysis relying on the full simulator model to generate predictions would have required about six days. In the end, the answer to the management problem was to not attempt to provide peak service with the existing array of development options.
This project is completed.
NETL – James Ammer (firstname.lastname@example.org or 304-285-4383)