Applying soft computing methods to improve the computational tractability of a subsurface simulation–optimization problem

Abstract Formal optimization strategies normally evaluate hundreds or even thousands of scenarios in the course of searching for the optimal solution to a given management question. This process is extremely time-consuming when numeric simulators of the subsurface are used to predict the efficacy of a scenario. One solution is to train artificial neural networks (ANNs) to stand in for the simulator during the course of searches directed by some optimization technique such as the genetic algorithm (GA) or simulated annealing (SA). The networks are trained from a representative sample of simulations, which forms a re-useable knowledge base of information for addressing many different management questions. These concepts were applied to a water flood project at BP's Pompano Field. The management problem was to locate the combination of 1–4 injection locations that would maximize Pompano's simple net profit over the next 7 years. Using a standard industry reservoir simulator, a knowledge base of 550 simulations sampling different combinations of 25 potential injection locations was created. The knowledge base was first queried to answer questions concerning optimal scenarios for maximizing simple net profit over 3 and 7 years. The answers indicated that a considerable increase in profits might be achieved by expanding from an approach to injection depending solely on converting existing producers to one involving the drilling of three to four new injectors, despite the increased capital expenses. Improved answers were obtained when the knowledge base was used as a source of examples for training and testing ANNs. ANNs were trained to predict peak injection volumes and volumes of produced oil and gas at 3 and 7 years after the commencement of injection. The rapid estimates of these quantities provided by the ANNs were fed into net profit calculations, which in turn were used by a GA to evaluate the effectiveness of different well-field scenarios. The expanded space of solutions explored by the GA contained new scenarios that exceeded the net profits of the best scenarios found by simply querying the knowledge base. To evaluate the impact of prediction errors on the quality of solutions, the best scenarios obtained in searches where ANNs predicted oil and gas production were compared with the best scenarios found when the reservoir simulator itself generated those predictions during the course of search. Despite the several thousand CPU hours required to complete the simulator-based searches, the resulting best scenarios failed to match the best scenarios uncovered by the ANN-based searches. Lastly, results obtained from ANN-based searches directed by the GA were compared with ANN-based searches employing an SA algorithm. The best scenarios generated by both search techniques were virtually identical.