Machine learning models to support reservoir production optimization

Traditionally, numerical simulators are used in combination with an optimization algorithm to determine optimum controls that maximize total oil production or net present value (NPV) over the life of the reservoir. These simulators are complex dynamic models that consider geological information, rock and fluid properties, as well as information about the completion of the wells. This complexity results in a high computational time and pose a challenge for the application of gradient-based optimization algorithms, since calculation of gradients of the objective function with respect to controls may demand several evaluations of the simulation model. This paper proposes the use of a machine learning model, based on artificial neural networks, to represent the non-linear dynamic behavior of the reservoir. The proposed approach was applied to data generated with a synthetic reservoir simulation model showing promising results.