Sequential Neural Simulation: A New Approach for Stochastic Reservoir Modelling

摘要

This paper proposes a new and practical neural network for stochastic reservoir modelling. The new approach known as sequential neural simulation (SNS) is developed based on the use of an ensemble of supervised neural networks within the conventional sequential simulation framework. Compared to the sequential Gaussian simulation (SGS), SNS uses a data-driven conditional distribution and can handle multivariate inputs and non-linearity. It does not require any (cross) variogram models and does not use a parametric function for simulation. The formulation of the algorithm however does not guarantee the reproduction of the desired statistics. The methodology is demonstrated in a fluvial-deltaic reservoir in China. Prorosity is simulated using different number of inputs followed by oil in place estimation. The study shows the simplicikty of the methodology for the incorporation of multiple soft information.