Study of Spatial Feature Extraction Methods for Surrogate Models of Numerical Reservoir Simulation

Numerical reservoir simulation is an important technology in reservoir production development, but the computational consumption of numerical simulation is a key factor affecting reservoir history matching, production prediction, and optimization. By constructing a computationally fast machine learning model to learn the mapping relationship between reservoir model parameters and production data, a maximum alternative to the numerical simulation process can be achieved to improve the efficiency of reservoir management and decision making. The current surrogate models of reservoir numerical simulation for large spatial variables, including permeability and porosity fields, often extract spatial features by convolutional neural networks and later use recurrent neural networks to learn the time-series relationships of production data. In this work, we study the method using convolutional neural networks to extract spatial parameters of reservoir models and propose a new module to convert the temporal and spatial features of surrogate models. By converting the spatial features extracted by convolution and adapting the input features and dimensions of the recurrent neural network, maximum extraction of spatial feature parameters is achieved. The proposed method was verified on a 3D reservoir model, and the results indicate that the method can enhance the accuracy of the surrogate model.