Predicting Water Cut Dynamics Using Hybrid Model Combining Machine Learning and Proxy Flow Model

In the oil industry, there is a growing tendency to use proxy models with varying degrees of complexity for operational predictive calculations. Additionally, machine learning methods are being actively developed as part of the digitalization and intelligentization of production processes. Using the synthetic model of an oil reservoir development element, this paper proposes an approach to combining a physically significant fluid-flow model with machine learning methods to predict oil production parameters. A distinctive feature of the synthetic model is the presence of extended zone inhomogeneity in the permeability field. Comparing to the original formulation, the proposed method uses a simplified single-phase fluid-flow model, which was matched to historical data by reconstructing reservoir hydroconductivity parameters with the help of a radial basis function neural network. The recurrent neural network has been trained to forecast the watercut in produced fluid. The use of recurrent neural networks enables the identification of the specific non-monotonic behavior of watercut in the produced fluid, caused by non-stationary injection and production well operating mode. The combination of these models enables the prediction of both the volume and phase composition of the produced fluid.