Abstract:
The determination of residual oil reserves is a complex task that involves high qualified specialists and resource-intensive calculations. Classical oil reservoir simulation methods cannot provide proper level of reliability due to increasing of complexity caused by oil reservoir depleting. Stagnant oil zones arising from heterogeneity of shear modulus of medium cannot be determinate by oil reservoir simulation due to use of effective values of porosity and absolute and phase permeability. The problem of determination of stagnant zones can be solved by using more flexible methods as a machine learning in general and the neural network in particular. The flexibility of these methods allow to build a model that can consider both physical properties of reservoir and the it operating history. In this paper the model based on Convolutional Neural Network (CNN) was built. The choice of CNN based on the assumption that this type of neural network does analysis in a similar way to a human. CNN is also simplifying the approach to distance depended values because the actual coordinate dependence which may not be the continuous function due to border conditions replaces with zone-based dependence produced by convolution. In this case the predicting value function becomes continuous and satisfies Cybenko theorem. The main idea of this model is transforming neighbor wells data of some point in 2D space into nxn matrix which is similar to grayscale image. The stack of matrices passes to the CNN as different channels of full-color image to find dependencies between image and an actual value which will be predict after successful learning.
