CN-121071408-B - Method and system for predicting yield of reservoir based on dual-attention mechanism and storage medium

CN121071408BCN 121071408 BCN121071408 BCN 121071408BCN-121071408-B

Abstract

The invention discloses a method, a system and a storage medium for predicting the yield of a centralized reservoir based on a dual-attention mechanism, relates to the field of oilfield development, and aims to solve the problems that the existing method is difficult to capture the bidirectional dependency relationship and local characteristics of long time sequence data at the same time and has insufficient effect on characteristic weight distribution and time sequence information focusing. The method is technically characterized by comprising the steps of S1, collecting field development data of a polymer flooding oil reservoir and screening optimization features, S2, constructing a CNN-BiLSTM-DAM fusion model, wherein the function implementation process of the model is that local features of input data are extracted based on a convolution layer, dynamic weights are distributed for the local features of each time step through an attention mechanism layer, forward and backward dependency relations of weighted time sequence feature data are captured through a bidirectional circulating structure of BiLSTM layers, weight adjustment is conducted on the time sequence features output by BiLSTM layers through the attention mechanism layer, global time dependency is captured through BiLSTM layers, and S3, the optimized features are input into the trained CNN-BiLSTM-DAM fusion model to predict the yield of the polymer flooding oil reservoir.

Inventors

GONG YAOHUA
GENG YANKAI
LIU BO
Qu Lanyanlin
HOU XIANGRUI
ZHAO LING
XING YUTONG
SUN XIANDA

Assignees

东北石油大学三亚海洋油气研究院
东北石油大学

Dates

Publication Date: 20260508
Application Date: 20251106

Claims (9)

1. A method for predicting yield of a polymer reservoir based on a dual-attention mechanism is characterized by comprising the following steps: s1, acquiring field development data of a polymer flooding oil reservoir, cleaning the data, and screening optimization features through correlation analysis to obtain an optimization feature set; s2, constructing a CNN-BiLSTM-DAM fusion model, wherein the CNN-BiLSTM-DAM fusion model comprises a multivariable time sequence feature input layer, a convolution layer, a first attention mechanism layer, a first layer BiLSTM layer, a second attention mechanism layer, a second layer BiLSTM layer, a Dropout layer and an output layer, wherein the convolution layer is used for extracting local features of input data, the first attention mechanism layer is used for distributing dynamic weights to the local features of each time step, the first layer BiLSTM layer is used for capturing forward and backward dependency relations of weighted time sequence data through a bidirectional circulation structure, the second attention mechanism layer is used for carrying out weight adjustment on the time sequence features output by the first layer BiLSTM, focusing time sequence information with obvious influence on a prediction result, and the second layer BiLSTM layer is used for capturing global time dependency; S3, inputting the optimized feature set into the trained CNN-BiLSTM-DAM fusion model, and predicting the yield of the reservoir.
2. The method for predicting yield of a reservoir based on dual attention mechanisms of claim 1, wherein said convolution layer uses one-dimensional convolution (Conv 1D), filter number filters=64, convolution kernel size kernel_size=1, activation function activation= 'relu', filling pattern padding= 'same'.
3. The method for predicting yield of a reservoir based on a dual-attention mechanism of claim 2, wherein the number of hidden units of BiLSTM layers BiLSTM _ UNITS =64.
4. The method for predicting yield of a reservoir based on a dual-attention mechanism of claim 3, wherein the output layer is a full connection layer, the output dimension is set to be 1, and the activation function is linear for outputting a predicted value of yield of the reservoir.
5. The method for predicting yield of a reservoir based on a dual-attention mechanism of claim 4, wherein in S1, optimization features are screened by correlation analysis, specifically, pearson correlation analysis is adopted to screen optimization features, and a pearson correlation coefficient calculation formula is as follows: Wherein, the And Respectively takes values for two variables of the ith observed value, And The average of the variables.
6. The method for predicting yield of a reservoir based on a dual-focus mechanism as set forth in claim 5, wherein said washing of the data in S1 comprises filling missing values in the data and interpolating outliers in the data.
7. The method for predicting yield of a reservoir based on a dual-focus mechanism as set forth in claim 6, wherein a Root Mean Square Error (RMSE), a root mean square percentage error (MAPE) and a determination coefficient are adopted in the training process of the CNN-BiLSTM-DAM fusion model ) As model evaluation indexes, the calculation formulas are respectively as follows: Wherein, the As a result of the fact that the value, In order to be able to predict the value, For a true yield of the product, In order to predict the yield of the product, N is the number of samples, which is the average of the true yield.
8. A dual-attention mechanism-based reservoir yield prediction system is characterized by comprising a program module corresponding to the steps of the method of any one of claims 1-7, and executing the steps in the dual-attention mechanism-based reservoir yield prediction method during running.
9. A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program configured to implement the steps in the dual-attention mechanism based reservoir yield prediction method of any one of claims 1 to 7 when invoked by a processor.

Description

Method and system for predicting yield of reservoir based on dual-attention mechanism and storage medium Technical Field The invention relates to the technical field of oilfield development, in particular to a method and a system for predicting yield of a reservoir based on a dual-attention mechanism and a storage medium. Background The polymer-flooding reservoir development comprises different stages, wherein only water injection exploitation is carried out in the early exploitation stage, no polymer injection data are generated, and when the polymer injection quantity reaches a certain degree and the oil field exploitation index is not obviously improved, the block can stop the polymer injection to enter the later stage based on the professional knowledge of the oil field, the polymer is not injected in the later stage, and the polymer injection residual pressure is relied on or the water injection exploitation is continued, so that the polymer injection characteristic is not generated. In the existing prediction method for the yield of the polymer-driven reservoir, the change of the yield of the crude oil is explored partially depending on the injection parameters, but the method fails due to lack of injection data in the non-injection stage. In the traditional time sequence prediction method, single variable time sequence prediction (such as ARIMA and SARIMA models) only focuses on the time sequence relation of a single variable, the multivariable characteristics such as the number of open wells, the liquid injection quantity and the like cannot be fused, a single cyclic neural network (RNN) has the gradient vanishing problem, a long-short-term memory network (LSTM) can process long time sequence data, but is difficult to capture two-way dependency relations at the same time, a Convolutional Neural Network (CNN) can extract local characteristics but ignores long-term dependency of the time sequence data, and a single attention mechanism has limited effects on characteristic weight distribution and time sequence information focusing, so that the precision of characteristic extraction and time sequence processing cannot be optimized at the same time. Therefore, for the yield prediction of the non-injection phase of the reservoir, a prediction method capable of fusing multivariable features, capturing long-time-sequence bi-directional dependence, dynamically optimizing features and time sequence weights is needed. Disclosure of Invention The invention aims to solve the technical problems that: The existing time-series oil reservoir yield method fuses multivariable characteristics, is difficult to capture the bidirectional dependency relationship and local characteristics of long time-series data at the same time, and has insufficient effects of characteristic weight distribution and time-series information focusing, so that the yield trend of the stage cannot be accurately predicted. The invention adopts the technical scheme for solving the technical problems: the invention provides a method for predicting yield of a reservoir based on a dual-attention mechanism, which comprises the following steps: s1, acquiring field development data of a polymer flooding oil reservoir, cleaning the data, and screening optimization features through correlation analysis to obtain an optimization feature set; S2, constructing a CNN-BiLSTM-DAM fusion model, wherein the function implementation process of the model is that local features of input data are extracted based on a convolution layer, dynamic weights are distributed to the local features of each time step through an attention mechanism layer, forward and backward dependency relations of weighted time sequence feature data are captured through a BiLSTM-layer bidirectional circulating structure, weight adjustment is carried out on the time sequence features output by a BiLSTM layer through the attention mechanism layer, and then global time dependency is captured through a BiLSTM layer; S3, inputting the optimized feature set into the trained CNN-BiLSTM-DAM fusion model, and predicting the yield of the reservoir. Further, the CNN-BiLSTM-DAM fusion model comprises a multi-variable time sequence feature input layer, a convolution layer, a first attention mechanism layer, a first layer BiLSTM layer, a second attention mechanism layer, a second layer BiLSTM layer, a Dropout layer and an output layer, wherein the convolution layer is used for extracting local features of input data, the first attention mechanism layer is used for distributing dynamic weights to the local features of each time step, the first layer BiLSTM layer is used for capturing forward and backward dependency of time sequence data through a bidirectional circulation structure, the second attention mechanism layer is used for carrying out weight adjustment on the time sequence features output by the first layer BiLSTM, focusing is used for influencing remarkable time sequence information on a prediction