Search

CN-121980943-A - Heterogeneous oil reservoir carbon dioxide flooding gas suction profile prediction method

CN121980943ACN 121980943 ACN121980943 ACN 121980943ACN-121980943-A

Abstract

The invention particularly discloses a heterogeneous oil reservoir carbon dioxide flooding air suction profile prediction method, and relates to the technical field of oil reservoir development dynamic prediction. The method comprises the steps of firstly, selecting 21 geological static parameters representing reservoir heterogeneity through Latin hypercube sampling, setting a value range to generate a geological model, then applying a 180-month water-gas alternate injection system, extracting key time step dynamic response data, dividing a training set and a testing set, then constructing a static and dynamic combined characteristic structure, constructing a neural network for outputting gas production rate increment, and finally adopting a rolling prediction mechanism to carry out cyclic iteration to realize 180-month complete dynamic prediction of a gas suction profile. The method solves the problems of insufficient coverage of samples, difficult prediction of longitudinal heterogeneity and long-period prediction of physical distortion in the prior art, improves prediction precision and efficiency, and can adapt to different geological conditions.

Inventors

  • ZHENG ZIHAO
  • LIU XINTONG
  • Shi Sitong
  • Ai Tianci
  • LIU JUNDA
  • ZHOU SHENLONG

Assignees

  • 长江大学

Dates

Publication Date
20260505
Application Date
20260123

Claims (7)

  1. 1. A heterogeneous oil reservoir carbon dioxide flooding air suction profile prediction method is characterized by comprising the following specific steps: Step S1, constructing a parameter space based on Latin hypercube sampling, which comprises the following steps: s11, 21 geological static parameters representing the heterogeneity of a reservoir are selected, wherein the parameters comprise the porosity of 7 layers, the permeability of 7 layers and the effective thickness of 7 layers; Step S12, setting the value range of each variable according to the geological features of the target oil reservoir; s13, carrying out layered sampling on the 21-dimensional parameter space by using a Latin hypercube sampling algorithm to generate different geological models; s2, numerical simulation and data extraction under multi-dimensional geological conditions are carried out, and the method specifically comprises the following steps: step S21, uniformly applying a water-gas alternating injection WAG system for 180 months to the generated geological model; Step S22, running each geological model based on a numerical simulator, and extracting dynamic response data of 30 key time steps; s23, dividing the extracted dynamic response data into a training set and a testing set in proportion; S3, constructing a characteristic structure comprising static input and dynamic input, wherein the static input is 21 geological static parameters, and the dynamic input is a layered gas production rate at a historical moment; s4, constructing an oil reservoir neural network ReservoirLSTM, wherein the oil reservoir neural network comprises an input characteristic layer, a plurality of LSTM units and a fully-connected output layer, the dimension of the output layer is set to 21, and the increment of the gas production rate of the corresponding 21 subdivision layers is set; and S5, predicting the data in the step N+1 by using the known data in the previous step N by adopting a rolling prediction mechanism, and filling the predicted result into the tail end of the sliding window as new historical data after inverse normalization, and performing loop iteration to realize the dynamic prediction of the complete inhalation profile for 180 months.
  2. 2. The method for predicting the carbon dioxide flooding suction profile of the heterogeneous oil reservoir according to claim 1, wherein in the step S12, the value range of each variable is that the porosity is 0.15-0.20, the permeability is 0.1-10 mD, and the thickness is 0.8-6 m.
  3. 3. The method for predicting a carbon dioxide flooding suction profile of a heterogeneous oil reservoir according to claim 1, wherein in step S21, the WAG system for alternate water and gas injection is set to switch water/gas injection medium every two months.
  4. 4. The method for predicting carbon dioxide flooding gas profile of a heterogeneous oil reservoir according to claim 1, wherein in step S22, the dynamic response data is a gas production rate of 21 sub-segments.
  5. 5. The method for predicting the carbon dioxide flooding and inspiration profile of the heterogeneous oil reservoir according to claim 1, wherein in the step S3, 21-dimensional static parameters are formed into vectors, and the vectors are copied and spliced into dynamic vectors of each time step to form space-time coupling characteristic input, wherein the dynamic vectors are gas production rates of each time step, and each time step has 21 layers; the splice formula is as follows: ; Wherein, the The vector is input for the model at time step t, As a dynamic vector at the moment of time step t, Is a static parameter vector.
  6. 6. The method for predicting the carbon dioxide flooding suction profile of the heterogeneous oil reservoir according to claim 1, wherein in step S4, in order to adapt to severe fluctuation under the WAG alternate injection condition, the oil reservoir neural network ReservoirLSTM adopts an incremental prediction strategy, and calculates a final predicted value by calculating the variation of the gas production rate, and the formula is as follows: ; Wherein, the Is the gas production rate at time t+1, Is the gas production rate at the time t, Is the predicted gas production rate increment.
  7. 7. The method of claim 6, wherein the neural network ReservoirLSTM uses an L1 loss function to minimize the difference between the predicted delta and the true delta, and uses an Adam optimizer to iteratively update parameters until the model converges.

Description

Heterogeneous oil reservoir carbon dioxide flooding gas suction profile prediction method Technical Field The invention relates to the technical field of oil reservoir development dynamic prediction, in particular to a heterogeneous oil reservoir carbon dioxide flooding air suction profile prediction method. Background Carbon dioxide flooding (CO 2 -EOR) is used as an effective recovery technology and is widely applied to oil and gas field development. In the design of a CO 2 -EOR development scheme, accurate prediction of the layering gas suction rate (gas suction profile) is a key for optimizing injection and production parameters and improving oil displacement efficiency. At present, the industry mainly relies on numerical simulation software to calculate layered gas production/gas suction rate, and meanwhile, in order to obtain reliable statistical rules, monte Carlo random simulation is usually required. However, the prior art suffers from the following drawbacks and deficiencies in predicting stratified inhalation rates: The sample space is not covered enough, the simulation efficiency is low, the problems of sample aggregation or insufficient coverage easily occur when the traditional simple random sampling is used for processing a high-dimensional parameter space, tens of thousands of times of simulation are needed to achieve the statistical significance, the time consumption is extremely long, and the design efficiency of a development scheme is seriously influenced; The longitudinal heterogeneity prediction is difficult, the existing substitution model is used for predicting the accumulated output of a single well, and is difficult to capture the differential fluid response of multiple layers under different physical property combinations, and the requirement of the fine prediction of the air suction profile cannot be met; The physical distortion of long period sequence prediction is that the absolute value is directly predicted by a conventional LSTM model, and error accumulation is easy to generate in the water-gas alternate injection period which is up to 180 months, so that the predicted result violates mass conservation or non-physical oscillation occurs, and the reliability of the predicted result is reduced. In summary, the existing model input-output structure does not fully consider the multi-layer heterogeneous characteristics, and the traditional deep learning model lacks the constraint on the physical rule of the reservoir fluid flow. Therefore, there is a need for a heterogeneous reservoir carbon dioxide flooding gas profile prediction method that can compromise prediction accuracy, computational efficiency, and physical consistency. Disclosure of Invention The invention aims to provide a heterogeneous oil reservoir carbon dioxide flooding gas suction profile prediction method, which aims to solve the problems of insufficient coverage of a sample space, difficulty in longitudinal heterogeneous prediction and long-period prediction physical distortion in the prior art. In order to achieve the above purpose, the invention provides a heterogeneous oil reservoir carbon dioxide flooding air suction profile prediction method, which comprises the following specific steps: Step S1, constructing a parameter space based on Latin hypercube sampling, which comprises the following steps: s11, 21 geological static parameters representing the heterogeneity of a reservoir are selected, wherein the parameters comprise the porosity of 7 layers, the permeability of 7 layers and the effective thickness of 7 layers; Step S12, setting the value range of each variable according to the geological features of the target oil reservoir; s13, carrying out layered sampling on the 21-dimensional parameter space by using a Latin hypercube sampling algorithm to generate different geologic model schemes; s2, numerical simulation and data extraction under multi-dimensional geological conditions are carried out, and the method specifically comprises the following steps: step S21, uniformly applying a water-gas alternating injection WAG system for 180 months to the generated geological model; Step S22, running each geological model based on a numerical simulator, and extracting dynamic response data of 30 key time steps; s23, dividing the extracted dynamic response data into a training set and a testing set in proportion; S3, constructing a characteristic structure comprising static input and dynamic input, wherein the static input is 21 geological static parameters, and the dynamic input is a layered gas production rate at a historical moment; s4, constructing an oil reservoir neural network ReservoirLSTM, wherein the oil reservoir neural network comprises an input characteristic layer, a plurality of LSTM units and a fully-connected output layer, the dimension of the output layer is set to 21, and the increment of the gas production rate of the corresponding 21 subdivision layers is set; and S5, predicting the data in