The design of production systems of gas fields has been a difficult task because of the nonlinear nature of optimization problem and the complex interactions between each operational parameter. Conventional methods, which usually are stated in precise mathematical forms, cannot include the uncertainties regarding vague or imprecise information in the objective and constraint functions. This article proposes a fuzzy nonlinear programming approach to accommodate these uncertainties and applies to a variety of optimization processes. Specifically, the fuzzy-formulation is combined with a hybrid co-evolutionary genetic algorithm for solving optimum gas production rates of each well to minimize investment cost with given constraints in order to enhance ultimate recovery. The synthetic optimization method can find a globally compromised solution and offer a new alternative with significant improvement over the existing conventional techniques. The reliability of the proposed approach is validated by a synthetic practical example yielding more improved results.

The major impediment in the formal optimization of large petroleumproducing fields is the cost of computing the state of the objects being optimized. Previous studies for prediction of reservoir performance with respect to time have used local inflow performance relationships or material balance models. These approaches, however, ignored flow interactions among wells during the optimization process, often resulting in suboptimal operations. In this study, a new polynomial neural network (PNN) with layer over-passing structure has been developed to replace a relatively time consuming reservoir simulator through robust and systematic search algorithm. The networks are subject to some form of training based on a representative sample of simulations that can be used as a re-useable knowledge base of information for addressing many different management questions. The proposed approach significantly reduces computational effort for optimizing the development scheme within reasonable accuracy and outperforms other neural network models.