Theoretical Research Of Reservoir Production Optimization Based On Gradient Approximation Methods

Reservoir real-time optimization management as an important technology in smart field,consists of reservoir automatic history matching and production optimization, which can beconsidered as the reservoir optimization problem. The gradient type optimization method withits gradient information calculated by adjoint method seems to be the most efficient algorithm.However, the adjoint code needed to be incorporated into the code of reservoir simulator, andthe computation is extremely complicated, which prohibits its application in real fields. Thesedefaults can be avoided by stochastic gradient method; however the accuracy of stochasticgradient is not accurate, which is only a rough approximation of true gradient. Thus, agradient approximation optimization algorithm based on combining stochastic gradient andfinite difference approximation, named as SGFD method, was proposed in the dissertation,which significantly improved the accuracy of stochastic gradient with much lowercomputational cost than finite difference method. In production optimization, based ondifferent stochastic gradient methods, a group of corresponding SGFD methods wereproposed. As the obtained optimal well operation strategy by stochastic gradient methodswere not very smooth, a smoothness technology was provided to modify these algorithms,and hence a set of new methods were provided. Then these methods were used to solve theproduction optimization mathematical model of the reservoir production system. Theensemble Kalman filter method for automatic history matching was modified by covariancelocalization method based on streamlines. Moreover, based on an ensemble of reservoirmodels, a new parameterization method was proposed to reduce the high dimension of historymatching objective function to a low dimension problem with the dimension related to thenumber of reservoir models. Then SGFD algorithm combined with the parameterizationmethod was applied to solve history matching problem. The results show that the SGFDalgorithm can significantly improve the accuracy of stochastic gradient, and obtain muchbetter optimal well operation strategy as well as higher net present value. The optimized welloperation strategy not only can be easily applied for practical production but also improve theeffect of water-flooding. The modified ensemble Kalman filter method solved the problems offilter divergence and spurious pseudo-correlations, which obtained a good matching result using a small size of ensemble. The large-scale reservoir optimization application proved thevalidity and effect of the parameterization method and SGFD gradient approximationalgorithm, in which a good data matching result and a reasonable geological model wereobtained, and the optimal well operation strategy was provided. Therefore, the proposedapproaches provide a new methodology and technology for the reservoir real-time productionoptimization application.