Proxy modeling pf reservoir hydrodynamics with graph neural networks

Abstract

The primary goal of hydrodynamic reservoir modeling is to predict the production wells performance, or more precisely, the dependence of this performance on the choice of the reservoir development plan. The achievement of this goal is hampered by the lack of accurate information about the properties of the reservoir. These properties have to be inferred from indirect data, first of all from the historical indicators of already functioning wells. This information is used to perform the adaptation of the numerical reservoir model or proxy models, which are less informative but have the advantage of speed and flexibility. The article proposes a reservoir proxy modeling method based on the use of a specific artificial neural network (ANN). The novel graph convolutional architecture of the ANN takes in the graph data describing the reservoir. The edges and vertices of the graph contain a spatial description of the reservoir along with the history of the well performance. Such architecture makes it possible to train the neural network for a whole class of situations instead of only one case. In accordance with the principles of the Physics-Informed Neural Networks (PINN), the task of the ANN is to derive a kind of formulation of a physical law guiding the system, rather than just a correlation between time series. The advantages of this approach over most ANN-based proxy models used today are, firstly, speed: adjustment to historical data and forecast output are made in seconds even for hundreds of wells; secondly, a certain degree of physical meaningfulness.