A method for predicting reservoir stratified injection and production based on LSTM-Attention

Abstract

The present invention relates to the technical field of oil reservoir exploitation, and in particular to a method for predicting oil reservoir stratified injection and production based on LSTM-Attention. The method uses an oil reservoir numerical simulator to simulate the injection and production process of a multi-layer heterogeneous oil reservoir, constructs a production data set, pre-processes the data in the production data set, and then divides the production data set into a training set and a test set; constructs an LSTM-Attention neural network model, which includes an LSTM part, an attention mechanism, and a fully connected layer; applies the trained LSTM-Attention neural network model under the production modes of four injections and four productions and four injections and two productions; uses the trained LSTM-Attention neural network model to predict the oil production, combines the oil field economic benefit analysis, and optimizes the water injection strategy, that is, by adjusting the water injection amount of the water injection well, the oil reservoir production is maximized and the cost is minimized, thereby improving the net present value and economic benefit of the oil field. The present invention can also effectively predict the oil production and pressure of each layer, and the prediction accuracy is as high as more than 95%.

Description

Technical Field

[0001] The present invention relates to the technical field of oil reservoir exploitation, and in particular to a method for predicting oil reservoir stratified injection and production based on LSTM-Attention. Background Art

[0002] With the deepening of reservoir development, water-driven reservoirs generally enter the high water-cut stage. This stage is characterized by increasingly prominent interlayer contradictions, continuous increase in water content, and a significant decrease in oil production. The traditional general water injection strategy for the entire reservoir has been difficult to maintain the effectiveness of water injection in the face of complex geological conditions, and may even aggravate the uneven development between layers. As a key means to improve the recovery rate in the high water-cut period, stratified injection and production technology has received widespread attention. This technology can regulate the water injection and oil production of each layer in a targeted manner by differentially managing different layers. This refined operation helps to maintain the pressure balance of the reservoir, reduce interlayer contradictions, improve oil recovery efficiency, and ultimately achieve effective development of the remaining oil. Reasonable stratified injection and production can not only enhance the mining efficiency, but also has important significance for improving the economic benefits of water-driven reservoirs. However, the implementation of stratified injection and production technology also faces many challenges. For example, how to accurately predict the production changes of each layer, how to optimize the injection and production scheme to adapt to dynamic geological conditions, etc. In this context, it is very necessary to use advanced artificial intelligence technologies such as deep learning to predict the production data of stratified injection and production in oil reservoirs. The deep learning model can learn complex nonlinear relationships from a large amount of historical data and make high-precision predictions of future production indicators. This provides a scientific basis for optimizing stratified injection and production plans, and helps to further improve the efficiency and economic benefits of reservoir development.

[0003] The existing technologies have made contributions to the development of reservoir stratified injection and production technology, but the traditional reservoir numerical simulation method is a qualitative study of specific blocks, which requires the construction of complex geological models and requires high experience of engineers, making it difficult to promote its use. However, some algorithms applied to reservoir stratified injection and production still need to be improved. BP neural network and convolutional neural network are not applicable to the prediction of production data of time series, and Gaussian process regression algorithm takes a long time in terms of calculation cost. Therefore, for complex reservoir geological environment and seepage process, in order to ensure the effectiveness of reservoir stratified injection and production, it is very necessary to propose a neural network model for predicting historical production data of reservoir stratified injection and production. Summary of the invention

[0004] In order to solve the problem that the traditional reservoir numerical simulation method is a qualitative study for a specific block and requires the construction of a complex geological model, which is not easy to promote and use, the present invention provides a reservoir stratified injection and production prediction method based on LSTM-Attention. The main steps of the method are as follows:

[0005] S1: Use a reservoir numerical simulator to simulate the injection and production process of a multi-layer heterogeneous reservoir, build a production data set, and preprocess the data in the production data set, dividing the production data set into a training set and a test set after preprocessing;

[0006] S2: Build an LSTM-Attention neural network model, which includes an LSTM part, an attention mechanism, and a fully connected layer. The LSTM part is responsible for capturing long-term dependencies in time series data, and the attention mechanism is used to dynamically adjust the model's attention to data at different levels. After the model is built, use the training set and test set to train and verify the LSTM-Attention neural network model, and fine-tune the network parameters of the model.

[0007] S3: Apply the trained LSTM-Attention neural network model under the production modes of four injections and four extractions and four injections and two extractions;

[0008] S4: Combined with the economic benefit analysis of the oil field, the water injection strategy is optimized, that is, by adjusting the injection volume of the water injection well, the trained LSTM-Attention neural network model is used to predict the oil production, so as to maximize the reservoir production and minimize the cost, thereby improving the net present value and economic benefits of the oil field and maximizing the oil field economy.

[0009] A storage device stores instructions and data for implementing a reservoir stratified injection and production prediction method based on LSTM-Attention.

[0010] A reservoir stratified injection and production prediction device based on LSTM-Attention comprises: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a reservoir stratified injection and production prediction method based on LSTM-Attention.

[0011] The technical solution provided by the present invention has the following beneficial effects: by coupling the long short-term memory neural network (LSTM) and the attention mechanism, the present invention uses the historical data of the reservoir injection well as the input data of the LSTM-Attention neural network, and the oil production and pressure of the production well as the output data, thereby effectively improving the calculation speed and accuracy of the reservoir production prediction. Under the production modes of four injections and four productions and four injections and two productions, the trained neural network model can effectively predict the oil production and pressure of each layer, with a prediction accuracy of more than 95%. At the same time, it also realizes the calculation of the injection and production schemes of different layers to achieve efficient production of the reservoir. Taking the net present value as the optimization target, by optimizing the injection volume and water injection strategy, the reservoir development efficiency and economic benefits are ultimately improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0013] Figure 1 It is a flowchart of a method for predicting oil reservoir stratified injection and production based on LSTM-Attention in an embodiment of the present invention;

[0014] FIG. 2( a ) is a well pattern diagram of an oil reservoir in a four-injection and four-production production mode when the porosity fields are the same in an embodiment of the present invention;

[0015] FIG. 2( b ) is a well pattern diagram of an oil reservoir under a four-injection and four-production production mode when the permeability fields are the same in an embodiment of the present invention;

[0016] FIG3( a ) is a well pattern diagram of an oil reservoir in a four-injection and four-production production mode when the porosity fields are the same in an embodiment of the present invention;

[0017] FIG3( b ) is a well pattern diagram of an oil reservoir under a four-injection and four-production production mode when the permeability fields are the same in an embodiment of the present invention;

[0018] Figure 4 LSTM-Attention neural network structure diagram in an embodiment of the present invention;

[0019] Figure 5 LSTM unit structure diagram in an embodiment of the present invention;

[0020] Figure 6 is a schematic diagram of the attention mechanism in an embodiment of the present invention;

[0021] Figure 7 is the LSTM-Attention cross-validation loss value in the embodiment of the present invention;

[0022] Figure 8 It is a schematic diagram of the prediction of pressure and oil production of each layer under the four-injection and four-production production mode in an embodiment of the present invention;

[0023] Fig. 9 It is a schematic diagram of the prediction of pressure and oil production of each layer under the four-injection and two-production production mode in an embodiment of the present invention;

[0024] Fig.10 It is a schematic diagram of the operation of the hardware device in the embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0026] Example 1

[0027] Please refer to Figure 1 , Figure 1 It is a flowchart of a method for predicting oil reservoir stratified injection and production based on LSTM-Attention in an embodiment of the present invention, which specifically includes:

[0028] S1: Use a reservoir numerical simulator to simulate the injection and production process of a multi-layer heterogeneous reservoir, build a production data set, and preprocess the data in the production data set, dividing the production data set into a training set and a test set after preprocessing;

[0029] The present invention considers the oil reservoir production modes of four injections and four productions and four injections and two productions respectively, wherein in the four injections and four productions production mode, the injection and the production are respectively located at the same layer, FIG2(a) is a well network structure diagram of the oil reservoir under the four injections and four productions production mode when the porosity field is the same, FIG2(b) is a well network structure diagram of the oil reservoir under the four injections and four productions production mode when the permeability field is the same; in the four injections and two productions production mode, the positions of the production wells are located at the middle two layers, FIG3(a) is a well network structure diagram of the oil reservoir under the four injections and four productions production mode when the porosity field is the same, FIG3(b) is a well network structure diagram of the oil reservoir under the four injections and four productions production mode when the permeability field is the same; wherein the black nodes on the left sides of FIG2(a), FIG2(b), FIG3(a), and FIG3(b) represent the positions of the injection layers, and the red nodes on the right sides represent the positions of the production layers. The process from red to dark blue in Figure 2(a) and Figure 3(a) represents the values ​​of the two-dimensional porosity field. Different colors represent different values ​​of the porosity field. For example, the red in Figure 2(a) represents the porosity field in the range of 0.275-0.295, and the dark blue represents the porosity field in the range of 0.100-0.119. The process from red to dark blue in Figure 2(b) and Figure 3(b) represents the values ​​of the two-dimensional permeability field. Different colors represent different values ​​of the permeability field. For example, the red in Figure 2(b) represents the permeability field in the range of 139-151, and the dark blue represents the permeability field in the range of 30-42. The meanings of the colors in Figure 3(a) and Figure 3(b) are similar.

[0030] S2: Build an LSTM-Attention neural network model, which includes an LSTM part, an attention mechanism, and a fully connected layer. The LSTM part is responsible for capturing long-term dependencies in time series data, and the attention mechanism is used to dynamically adjust the model's attention to different levels of data, thereby improving the accuracy of predictions. After the model is built, use the training set and test set to train and verify the LSTM-Attention neural network model, fine-tune the network parameters of the LSTM-Attention neural network model, and improve the reliability of the LSTM-Attention neural network.

[0031] S3: Apply and verify the trained LSTM-Attention neural network model under the production modes of four injections and four extractions and four injections and two extractions; verify the generalization ability and applicability of the model by comparing the prediction results of the model under different production modes.

[0032] In the production mode of four injections and four productions and four injections and two productions, LSTM-Attention is used to predict the oil production and second pressure of the production wells. The prediction effect of the model under different reservoir development strategies can be comprehensively evaluated, which verifies the strong generalization ability and high applicability of the model. Among them, four injections and four productions are to inject water and produce oil in four different layers at the same time, while four injections and two productions are to inject water in four layers and produce oil in the middle two layers (see Figure 2 (a), Figure 2 (b), Figure 3 (a), Figure 3 (b)). In the production mode of four injections and four productions, the second pressure and oil production of the four layers are predicted respectively (see Figure 8 ), among which, the oil production of layer 2 (layer2) is higher, while the oil production of layer 4 (layer4) is lower. In the early stage of production, the pressure of the four layers increases rapidly to the upper threshold. In this embodiment, the upper threshold is 10MPa. In the late stage of production, the oil production of layer 1 (layer1), layer 2 (layer2) and layer 3 (layer3) shows a decreasing trend. Under the production mode of four injections and two productions, the second pressure and oil production of the two middle layers of the production well are predicted respectively (see Fig. 9 ), where the oil production of layer1 increases slowly, while the oil production of layer2 tends to increase in the early stage of production, but decreases rapidly in the later stage of production and tends to stabilize. The second pressure of layer1 and layer2 is maintained at 10MPa. The predicted second pressure is to detect whether the pressure is lower than the upper threshold of the formation pressure. If the pressure is lower than or equal to the upper threshold of the formation pressure, it means that the production at this time is safe and can continue to produce. If the pressure is higher than the upper threshold of the formation pressure, it means that the production is unsafe at this time and corresponding treatment is required to ensure production safety.

[0033] Combining the well pattern diagrams of Figure 2(a), Figure 2(b), Figure 3(a), and Figure 3(b), it can be seen that layers with higher permeability and porosity tend to have higher oil production. The LSTM-Attention neural network has a high prediction accuracy for oil production and pressure (see Figure 8-Figure 9 ), can reach more than 95%, providing reliable technical guarantee and solid theoretical basis for reservoir production.

[0034] S4: Combined with the economic benefit analysis of the oil field, the water injection strategy is optimized, that is, by adjusting the injection volume of the water injection well, the trained LSTM-Attention neural network model is used to predict the oil production, so as to maximize the reservoir production and minimize the cost, thereby improving the net present value and economic benefits of the oil field and maximizing the oil field economy.

[0035] The specific implementation steps of step S1 are as follows:

[0036] S11: First, a reservoir numerical simulator is used to generate production data of reservoirs in different layers, including data sets of various production methods such as four injections and four productions and four injections and two productions. These data include key characteristic parameters such as reservoir water injection rate, oil production and pressure, and the data are exported from the reservoir numerical simulator into an easy-to-read and write xlsx format. At the same time, a reservoir numerical simulator is used to generate a 200×30 size porosity field and permeability field (see Figure 2(a), Figure 2(b), Figure 3(a), Figure 3(b)), and the generation process of reservoir stratified injection and production for 2 years is simulated under the same porosity field and permeability field. The water injection rate, oil production and pressure of different layers of the reservoir are obtained, and the water injection rate and the first pressure are used as the input of the LSTM-Attention neural network, and the oil production and the second pressure are used as the output of the LSTM-Attention neural network. The structure diagram of the LSTM-Attention neural network model is shown in the figure. Figure 4 As shown, it includes a long short-term memory neural network (LSTM) part, an attention mechanism part, and a fully connected layer (Fully Connected Layer) connected in sequence.

[0037] S12: Preprocess the exported data, including data cleaning, missing value processing, outlier detection, time series segmentation and standardization. First, remove the outliers and use linear interpolation to fill the missing values. Use the sliding window conversion technology to convert the reservoir production data into time series data, that is, through the sliding window segmentation, the pressure is divided into a first pressure representing the past pressure and a second pressure representing the future pressure. The sliding window size is 12, that is, the water injection volume and pressure of the past 10 time segments are used to predict the oil production and pressure of the next 2 time segments, such as Figure 4 As shown, , is the input of the LSTM-Attention neural network, where for The amount of water injected into each layer during the time period, for The pressure of each layer of the injection well during the time period, , is the output of the LSTM-Attention neural network, for Oil production of each layer in each time period, for The pressure of each layer of the production well during the time period.

[0038] The input production data is standardized using the maximum and minimum normalization method to eliminate the impact of different dimensions:

[0039]

[0040] Among them, min is the input data The minimum value of the input data The maximum value of .

[0041] Finally, the input data is divided into training set and test set in a ratio of 8:2.

[0042] The implementation steps of S2 are:

[0043] S21: In order to solve the gradient vanishing and gradient exploding problems when the neural network model processes reservoir production data, the LSTM part is introduced. The LSTM part is a three-layer LSTM network. Each layer of LSTM grid contains four interactive units, namely input gate, forget gate, hidden state and output gate (see Figure 5 ). The LSTM part must determine what new information is stored in the memory cell (LSTM Cell), which includes the input gate and the tanh layer. The input gate determines what needs to be updated in the memory cell. The value of the input gate for:

[0044]

[0045] The layer creates a new candidate value vector as

[0046]

[0047] in, is the weight matrix of the input gate, is the bias term of the input gate, is the hidden state at the previous moment, is the current input, is the sigmoid function and tanh is the hyperbolic tangent function.

[0048] The forget gate needs to determine the forgotten cell state information. The expression of the forget gate is:

[0049]

[0050] in, is the weight matrix of the forget gate, is the bias term of the forget gate, It is the hidden state at the previous moment.

[0051] The memory cells are updated according to the forget gate and the input gate. The updated cell state formula is:

[0052]

[0053] in, is the output of the forget gate, is the cell state at the current moment, is the cell state at the previous moment, is the input gate value, is a candidate value.

[0054] The output gate is obtained by passing the current input and the hidden state of the previous time step through the fully connected layer and the sigmoid function. The output gate formula is:

[0055]

[0056]

[0057] in, is the weight matrix of the output gate, is the bias term of the output gate, is the hidden state at the current moment, is the output gate node.

[0058] In order to solve the differences in predicted oil production between different time periods and different reservoir segments, an attention mechanism is introduced to assign different weights to the predicted data (see Figure 6 ), dynamically adjust the model's attention to the characteristics of different time steps and different layers of the reservoir, so as to better capture the key information in reservoir production. The attention mechanism is essentially an attention mechanism for the elements (Value) in the source (Source). i ) is weighted summed, while query (Query) and key (Key i ) is used to calculate the corresponding Value i The weight coefficient of the attention mechanism is calculated as:

[0059]

[0060]

[0061]

[0062] in, The length of the source. iRepresents the elements in Source, Query represents the query, Key i Indicates the key, represents the norm of Query, Represents Key i The norm of represents the attention score weight, is the similarity between the query and the key value, , is the probability distribution function model of the attention mechanism, i=1,2,..., L x ,j=1,2,..., L x .

[0063] The attention mechanism can select key information from a large number of reservoir production data. The larger the weight coefficient, the more focused it is on the corresponding Value. i Value, in other words, weight represents the importance of information.

[0064] S22: Divide the preprocessed reservoir data into a training set and a test set according to a preset ratio. The preset ratio in this embodiment is 8:2. Use the training set and the test set to train and verify the LSTM-Attention neural network model, respectively. During the model training process, the model performance is optimized by fine-tuning hyperparameters such as batch size and learning rate. Cross-validation is used to compare the loss values ​​of the training set and the test set to verify the reliability of the LSTM-Attention neural network model for production data prediction. The model uses mean square error MSE as the loss function. MSE is a commonly used quantitative indicator to measure the difference between the predicted value and the true value. It is particularly used for the accuracy evaluation of continuous numerical predictions. The specific definition is as follows:

[0065]

[0066] in, is the predicted value of the model, is the true value of the sample, and Represents the number of samples. The learning rate is set to 1e-4, the batch size is set to 32, the training rounds are 100, and the Adam optimizer is used to update the network parameters of LSTM-Attention. The network model with a smaller loss value on the validation set is selected for subsequent experiments (see Figure 7 ), after cross-validation, it was found that when the training rounds reached 50 rounds, the loss functions on the validation set and the test set basically tended to be stable, and this round was selected as the candidate for the optimal model.

[0067] The present invention optimizes the water injection strategy by calculating the net present value of oil reservoir production. In dealing with actual oil reservoir development problems, the net present value method is one of the commonly used analysis methods in the economic evaluation process. In the evaluation of oil reservoir economic benefits, the net present value NPV is usually used as an indicator, that is, the sum of the present values ​​of cash inflows and cash outflows within the service life, and the calculation method is as follows:

[0068]

[0069] Among them, CI(n) is The cash inflow in the first year is the cash inflow from crude oil sales, CO(n) is the cash inflow from the first year. The cash outflow in 2017 is the treatment cost of water injection, i is the discount rate, The year of production, is the crude oil price, is the annual oil production, is the treatment cost per ton of injected water, The annual water injection volume is used as the input of LSTM-Attention to predict the oil production of different layers of the reservoir. By adjusting the water injection volume of the reservoir and calculating the net present value (NPV) based on the predicted oil production, the water injection scheme with the maximum net present value can be obtained, which effectively improves the efficiency and economic benefits of reservoir development.

[0070] Example 2

[0071] A reservoir stratified injection and production prediction device 401 based on LSTM-Attention, such as Fig.10 As shown, it includes: a processor 402 and a storage device 403; the processor 402 loads and executes instructions and data in the storage device 403 to implement the reservoir stratified injection and production prediction method based on LSTM-Attention.

[0072] Example 3

[0073] A storage device, wherein the storage device stores instructions and data for implementing the LSTM-Attention-based reservoir stratified injection and production prediction method.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A reservoir stratified injection and production prediction method based on LSTM-Attention, characterized by: The method includes: S1: Use a reservoir numerical simulator to simulate the injection and production process of a multi-layer heterogeneous reservoir, construct a production data set, and the data in the production data set include water injection volume, oil production volume, and pressure. Preprocess the data in the production data set, and divide the production data set into a training set and a test set after preprocessing; the preprocessing includes sliding window segmentation of the time series and normalization processing. Through sliding window segmentation, the pressure is divided into a first pressure representing the past pressure and a second pressure representing the future pressure; S2: Build an LSTM-Attention neural network model, which includes an LSTM part, an attention mechanism, and a fully connected layer. The LSTM part is responsible for capturing long-term dependencies in time series data, and the attention mechanism is used to dynamically adjust the model's attention to data at different levels. After the model is built, use the training set and test set to train and verify the LSTM-Attention neural network model, and fine-tune the network parameters of the model. S3: Apply the trained LSTM-Attention neural network model to predict the oil production and the second pressure of the production well under the production modes of four injections and four productions and four injections and two productions. Four injections and four productions are to inject water and produce oil at four different layers at the same time, and four injections and two productions are to inject water at four layers and produce oil at the middle two layers. Under the production mode of four injections and four productions, the second pressure and oil production of the four layers are predicted respectively, among which the oil production of layer2 is higher, while the oil production of layer4 is lower. In the early stage of production, the pressure of the four layers increases rapidly to the upper threshold. In the late stage of production, the oil production of layer1, layer2 and layer3 shows a decreasing trend. Under the production mode of four injections and two productions, the second pressure and oil production of the two middle layers of the production well are predicted respectively, among which the oil production of layer1 increases slowly, while the oil production of layer2 has an increasing trend in the early stage of production, but decreases rapidly in the later stage of production and tends to be stable. The second pressures of layers1 and layer2 are both kept at the upper threshold. S4: Combined with the economic benefit analysis of the oil field, the water injection strategy is optimized, that is, by adjusting the injection volume of the water injection well, the trained LSTM-Attention neural network model is used to predict the oil production, so as to maximize the reservoir production and minimize the cost, thereby improving the net present value and economic benefits of the oil field and maximizing the oil field economy.

2. The method for predicting oil reservoir stratified injection and production based on LSTM-Attention as claimed in claim 1, characterized in that: The specific implementation steps of step S1 are: S11: Use a reservoir numerical simulator to generate a porosity field and permeability field, simulate the generation process of reservoir stratified injection and production under the same porosity field and permeability field, obtain the water injection rate, oil production rate and pressure of different layers of the reservoir, use the water injection rate and the first pressure as the input of the LSTM-Attention neural network model, and use the oil production rate and the second pressure as the output of the LSTM-Attention neural network model; S12: Use sliding window transformation technology to convert reservoir production data into time series data, and use maximum and minimum normalization to preprocess the input data: Among them, min is the input data The minimum value of the input data Finally, the production data set is divided into training set and test set according to the preset ratio.

3. The method for predicting oil reservoir stratified injection and production based on LSTM-Attention as claimed in claim 1, characterized in that: The specific implementation steps of step S2 are: S21: In order to solve the gradient vanishing and gradient exploding problems when the LSTM-Attention neural network model processes reservoir production data, the LSTM part is introduced. The LSTM part is a three-layer LSTM network. Each layer of the LSTM grid contains four interactive units, namely the input gate, the forget gate, the hidden state and the output gate; The input gate determines the content of the memory cell that needs to be updated. The value of the input gate for: in, is the weight matrix of the input gate, is the hidden state at the previous moment, is the current input, is the sigmoid function, is the bias term of the input gate; The forget gate needs to determine the forgotten cell state information, and the forget gate output The expression is: in, is the weight matrix of the forget gate, is the bias term of the forget gate; The memory cells are updated according to the forget gate and the input gate. The updated cell state formula is: in, is the cell state at the current moment, is the cell state at the previous moment, is the input gate value, is a candidate value; The output gate is obtained by passing the current input and the hidden state of the previous time step through the fully connected layer and the sigmoid function. The output gate formula is: in, is the weight matrix of the output gate, is the bias term of the output gate, is the hidden state at the current moment, is the output gate node; In order to solve the differences in predictions between different time periods and different reservoir segments, the attention mechanism is introduced to assign different weights to the prediction data; the calculation formula of the attention mechanism is: in, , indicating the length of Source; Value i Represents the elements in Source, Query represents the query, Key i Indicates the key, represents the norm of Query, Represents Key i The norm of represents the attention score weight, is the similarity between the query and the key value, , is the probability distribution function model of the attention mechanism, i=1,2,..., L x ,j=1,2,..., L x ; S22: During training and verification, the LSTM-Attention neural network model uses mean square error MSE as the loss function: in, is the predicted value of the model, is the true value of the sample, represents the sample size; Use the Adam optimizer to update the network parameters of the LSTM-Attention neural network model and fine-tune the LSTM-Attention neural network model.

4. The method for predicting oil reservoir stratified injection and production based on LSTM-Attention as claimed in claim 2, characterized in that: The specific implementation steps of step S4 are: In dealing with actual reservoir development problems, the net present value (NPV) is used as an indicator for evaluating the economic benefits of reservoirs, that is, the sum of the present values ​​of cash inflows and cash outflows within the useful life. The calculation method is as follows: Among them, CI(n) is The cash inflow in the first year is the cash inflow from crude oil sales, CO(n) is the cash inflow from the first year. The cash outflow in 2017 is the treatment cost of water injection, i is the discount rate, is the production year, N is the total number of years of production data, is the crude oil price, is the annual oil production, is the treatment cost per ton of injected water, is the annual water injection volume; The water injection volume and first pressure of the reservoir are used as the input of the LSTM-Attention neural network model to predict the oil production of different layers of the reservoir. By adjusting the water injection volume of the reservoir and calculating the net present value (NPV) based on the predicted oil production, the water injection plan with the highest net present value can be obtained, which effectively improves the efficiency and economic benefits of reservoir development.

5. A storage device, characterized in that: The storage device stores instructions and data for implementing the reservoir stratified injection and production prediction method based on LSTM-Attention as described in any one of claims 1 to 4.

6. A reservoir stratified injection and production prediction device based on LSTM-Attention, characterized by: include: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the reservoir stratified injection and production prediction method based on LSTM-Attention as described in any one of claims 1 to 4.

Images