Study on intelligent simulation and prediction algorithm of oil-water phase change during polymer flooding development of reservoir

In this paper, a deep learning method integrating physical constraints is proposed to simulate and predict the phase change of oil and water during polymer flooding development. This method combines Long Short-Term Memory (LSTM) and convolutional neural network (CNN) to construct an intelligent algorithm framework to realize the spatio-temporal dynamic simulation of oil-water phase state. Through the Encoder-Decoder architecture, the data-driven model is combined with the reservoir physical laws, the space-time features are extracted by ConvLSTM module, and the multi-scale feature fusion is realized by U-Net structure. In addition, the mass conservation equation is embedded in the loss function as a physical constraint, and a dynamic weight multi-objective optimization algorithm is introduced to balance data fitting, physical consistency and gradient smoothness. In the modeling of polymer adsorption effect, Langmuir isothermal adsorption equation is used as an auxiliary constraint, and a fully connected network is introduced to modify it. Taking BZ25-1 block of offshore heavy oil reservoir as an example, the intelligent algorithm is used to simulate and predict the phase change of oil and water. The results show that the intelligent predicted value is highly consistent with the actual water cut, and the relative error of predicted oil recovery is only 4.8%, and the calculation time is significantly less than that of traditional numerical simulation methods. The intelligent algorithm can effectively capture reservoir heterogeneity and local fingering effect, which provides strong support for optimizing polymer injection scheme, evaluating oil displacement efficiency and formulating subsequent development strategy.