Abstract
Reservoir modeling data could be divided into two categories: spatial data (i.e. permeability, effective grids, crack, irregular boundaries) and vector data (fluid properties i.e. relative permeability, density, viscosity). This paper is mainly interested in considering the permeability and relative permeability, which are the representatives of the two data types, to construct a proxy model for forecasting saturation and pressure maps in heterogeneous reservoirs during water flooding. The proxy model is built on the dense encoder-decoder network, to learn the reservoir dynamic states at different time steps. Results indicate that the trained proxy model could predict fluid saturation and pressure fields accurately. This paper presents a calibration method, which is adding a constraint to well-blocks. After calibration, the trained proxy model is utilized to calculate reservoir production. The comparison results illustrate that the proxy model can forecast well rates with relatively high accuracy. Compared with traditional reservoir numerical simulators, the proxy model could predict fluid saturation, pressure and well rates with similar accuracy and less time-cost.
NSTL