Oil reservoir economic benefit optimization system based on improved LSTM

Abstract

The invention discloses an oil reservoir economic benefit optimization system based on an improved LSTM (Long Short Term Memory), and relates to the field of oil reservoir development and economic evaluation. Data collection: collecting oil reservoir dynamic data: daily oil production and water content in numerical simulation software CMG; and collecting economic data: international crude oil price, global GDP growth rate, and global crude oil production; the data in the step 2 are subjected to noise reduction preprocessing, abnormal values of an oil reservoir dynamic data curve are removed through statistical analysis, the curve is smoothed, and the accuracy and stability of the model are improved; model training: a prediction model is constructed by adopting an LSTM algorithm added with an attention mechanism, and the model learns association between each parameter and economic benefits based on historical data and predicts future economic benefits; and economic optimization: selecting an economic benefit index net present value NPV, using a response surface method, optimizing oil reservoir injection parameters, searching an optimal combination of different parameter combinations, and maximizing the net present value in the cold production process.

Description

Technical Field

[0001] The present invention relates to the field of oil reservoir development and economic evaluation, and in particular to an oil reservoir economic benefit optimization system based on improved LSTM. Background Art

[0002] In the process of reservoir development, economic benefit evaluation is an important part of decision-making. Traditional economic evaluation methods mainly rely on empirical formulas or simple statistical analysis. However, these methods have obvious limitations when dealing with nonlinear, multi-dimensional and dynamically changing data in reservoir development. Moreover, they can only rely on numerical simulation software to predict the dynamic changes of future reservoir production indicators. It is difficult to accurately predict future economic benefits, leading to decision-making errors and economic losses. If you want to establish a system that can accurately and efficiently evaluate the economic benefits of future reservoir development, you must introduce a more advanced evaluation model that supports complex interactions between multiple variables.

[0003] At present, there are literatures that have given optimization methods for the reservoir economic benefit optimization system. The "Research on Optimization and Economic Evaluation of Fracturing Parameters in Tight Oil Reservoirs" gives an optimal solution for fracturing well parameters based on the particle swarm optimization algorithm. The single well field data is used as the model input, and the algorithm model solves the maximum net present value.

[0004] Patent No. CN2020102928890 provides a large-scale reservoir injection and production optimization system based on a machine learning agent model, which provides an optimization solution to the shortcomings of traditional evolutionary algorithms that require a large number of numerical simulations during optimization design.

[0005] Patent No. CN2023115551256 provides an efficient optimization decision-making method for well location and well control parameters based on a reduced-order model. It jointly optimizes the well location layout and well control parameters during the reservoir production process, takes economic indicators as the optimization target, and establishes a well location and well control parameter optimization model.

[0006] The above documents only perform economic optimization on historical data, do not consider future economic benefits, and fail to make long-term plans. The document "Oil Well Production Prediction Method for Complex Block Reservoirs Based on Improved LSTM" proposes a production prediction method based on an improved LSTM model, which reduces the prediction error. Patent No. CN2022106030166 discloses a three-dimensional underground reservoir dynamic prediction method based on a convolutional Fourier neural network, which realizes high-precision prediction of three-dimensional reservoir models. The above two documents only consider the reservoir production itself, do not involve economic evaluation, and give one-sided suggestions for the decision-making investment of reservoir production managers.

[0007] In view of the problems of insufficient accuracy, limited dynamic adaptability and poor optimization effect of development parameters in economic benefit evaluation during reservoir development, the present invention forms an oil reservoir economic benefit optimization system based on an improved LSTM based on an LSTM algorithm and a response surface method with an attention mechanism, so as to solve the limitations of traditional methods in dynamic data processing and multivariable optimization, make up for the shortcomings of traditional economic evaluation methods in dynamic changes and high-dimensional data processing, promote the scientific nature of investment in reservoir development decisions and the improvement of economic benefits, and promote the development of intelligent reservoir development and economic optimization technology. Summary of the invention

[0008] The purpose of the present invention is to provide a reservoir economic benefit optimization system based on improved LSTM to solve the problems raised in the above background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: a reservoir economic benefit optimization system based on improved LSTM, comprising the following steps:

[0010] Step 1: Experimental design;

[0011] Step 2: Data collection: using numerical simulation software CMG to collect reservoir dynamic data: daily oil production, water content; and economic data: international crude oil prices, global GDP growth rate, and global crude oil production;

[0012] Step 3: Perform noise reduction preprocessing on the data in step 2, remove outliers from the reservoir dynamic data curve through statistical analysis, smooth the curve, and improve the accuracy and stability of the model;

[0013] Step 4: Model training: Use the LSTM algorithm with attention mechanism to build a prediction model. The model learns the relationship between each parameter and economic benefits based on historical data to predict future economic benefits.

[0014] Step 5: Economic optimization, select the economic benefit indicator net present value (NPV), use the response surface method to optimize the reservoir injection parameters, find the optimal combination of different parameter combinations, and maximize the net present value in the cold production process.

[0015] Preferably, in step 1, the central composite design method CCD of the response surface method RSM is used for experimental design, and the central composite design method CCD includes cubic points, central points and axial points.

[0016] Preferably, in step 3, a Savitzky-Golay filter is used to eliminate noise in the data while maintaining the trend of the data.

[0017] Preferably, the Savitzky-Golay filter converts the polynomial fitting process into a weighted average form, that is, a convolution formula is used to achieve smoothing, and the formula is as follows:

[0018]

[0019] Among them, y′ i is the data point after filtering and smoothing; y i+j is the original data point in the window; ω j is the weight coefficient of the filter, ω j The calculation of depends on the window size 2m+1 and the polynomial order n and remains constant for each data point.

[0020] Preferably, in step 4, LSTM uses a forget gate, an input gate, and an output gate to determine which information needs to be retained, forgotten, or output. Through the synergistic effect of these gates, LSTM can flexibly transmit information in different time steps to ensure that the model effectively learns important temporal dependencies in long sequences. The calculation formulas of the forget gate, the input gate, and the output gate are:

[0021] f t =σ(W f h t-1 +W f x t +b f )#(2);

[0022] i t =σ(W i h t-1 +W i x t +b i )#(3);

[0023] o t =σ(W o h t-1 +W o x t +b o )#(4);

[0024] Among them, σ is the sigmoid activation function, the output is between 0 and 1, h t-1 is the hidden state of the previous time step, x t is the input of the current time step, W f , W i , W i are the weight matrices of the forget gate, input gate, and output gate, respectively, and f t 、i t , o t are the outputs of the forget gate, input gate, and output gate, respectively, and bf , b i , b o They are the bias vectors of the forget gate, input gate, and output gate respectively;

[0025] The LSTM calculation process is as follows:

[0026]

[0027] h t =o t tanh(C t )#(7);

[0028] in, is the candidate memory state, W C , b C are the weight matrix and bias vector of the candidate memory state, respectively, C t , C t-1 are the memory states of the current time step and the previous time step, respectively, h t is the hidden state of the current time step, that is, the output of the LSTM unit.

[0029] Preferably, after using the attention mechanism in step 4, the model will consider the importance of different parts of the input sequence when generating the output.

[0030] Preferably, the attention mechanism includes calculating an attention score, calculating an attention weight, and calculating a context vector;

[0031] a. Calculate attention score

[0032] For each time step t, calculate the current hidden state h t With each hidden state h in the input sequence i Similarity score:

[0033] e ti =α(h t ,h i )#(8);

[0034] Among them, α is a similarity function, common choices are dot product or additive similarity function;

[0035] b. Calculate attention weights

[0036] Perform softmax normalization on the attention score to get the attention weight:

[0037]

[0038] c. Calculate the context vector

[0039] The context vector is obtained by weighted summing the hidden states of the input sequence using the attention weights:

[0040]

[0041] Preferably, the NPV calculation formula in step 5 is as follows:

[0042]

[0043] Where: C t is the cash flow in year t; r is the discount rate, usually based on the market interest rate or the cost of capital; n is the life of the project; I0 ​​is the initial investment cost;

[0044] C t =R t -E t =Q t ·P t -E t #(12)

[0045] Where: R t is the total income in year t, Q t is the crude oil production in year t; P t is the crude oil price in year t; E t is the total expenditure in year t.

[0046] Preferably, the response surface method in step 5 usually assumes that the response variable y and the input variables x1, x2, ..., x k The relationship between can be approximated by a binomial model; the commonly used second-order polynomial model formula is:

[0047]

[0048] Among them, β0 is the intercept term, representing the linear effect of the input variable, and β i is the second-order effect coefficient, representing the quadratic effect of each variable, β ij is the interaction effect coefficient, indicating that x i and x j The interaction between them.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) The present invention can effectively capture the time dependency and nonlinear relationship in historical reservoir data by introducing the LSTM algorithm with the attention mechanism. Compared with the traditional method, the present invention has significantly enhanced the ability to dynamically adapt to the fluctuations of oil prices and production costs. The error in predicting the net present value in the next three years is reduced by 10%-25%. The accuracy of economic benefit evaluation can be maintained in the market changes, providing a more reliable economic benefit evaluation for reservoir development.

[0051] (2) The present invention combines the response surface method and optimizes reservoir development parameters through experimental design. Compared with the traditional method, the net present value of the present invention after optimizing the development parameters is increased by 10%-15%, which significantly improves the economic benefits of reservoir development;

[0052] (3) The present invention provides dynamic and intelligent decision support for reservoir development by combining the LSTM model and the response surface method. It can quickly generate optimization plans under different oil prices and production costs, shortening the decision cycle by 30%-50%. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Table 1. Cube point experimental combinations;

[0054] Table 2 Center point experimental combination;

[0055] Table 3 Axis point experimental combinations;

[0056] Table 4 Summary of injection parameter experimental combinations;

[0057] Table 5: Annual net cash flow and annual net present value table of combination 1; Table 6: NPV table of nine experimental combinations;

[0058] Table 7 Regression coefficient table;

[0059] Figure 1 Point and line graphs of daily oil production and water content for nine experimental groups;

[0060] Figure 2 De-noised daily oil production map of combination 1;

[0061] Figure 3 Denoised daily oil production map of combination 2;

[0062] Figure 4 Denoised daily oil production map of combination 3;

[0063] Figure 5 De-noised daily oil production map of combination 4;

[0064] Figure 6 Denoised daily oil production map of combination 5;

[0065] Figure 7 De-noised daily oil production map of combination 6;

[0066] Figure 8 De-noised daily oil production map of combination 7;

[0067] Fig. 9 De-noised daily oil production map of combination 8;

[0068] Fig.10 De-noised daily oil production map of combination 9;

[0069] Fig.11 Oil price denoised image;

[0070] Fig.12 Daily oil production forecast chart for combination 1;

[0071] Fig.13 Daily oil production forecast chart for combination 2;

[0072] Fig.14 Daily oil production forecast chart for combination 3;

[0073] Fig.15 Daily oil production forecast chart for combination 4;

[0074] Fig.16 Daily oil production forecast chart for combination 5;

[0075] Fig.17 Daily oil production forecast chart for combination 6;

[0076] Fig.18 Daily oil production forecast chart for combination 7;

[0077] Fig.19 Daily oil production forecast chart for combination 8;

[0078] Fig. 20 Daily oil production forecast chart for combination 9;

[0079] Fig.21 Oil price forecast chart;

[0080] Fig. 22 Response surface three-dimensional plot;

[0081] Fig.23 Injection rate-injection concentration plot. DETAILED DESCRIPTION

[0082] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0083] Example

[0084] Reservoir A is set as a common heavy oil reservoir, with a burial depth of 800 meters, an effective thickness of 8m, a reservoir pressure of 6MPa, a porosity of 19%, and a permeability of 1000mD. It was put into production in 2013, and the economic benefits deteriorated with water flooding in the first five years. Chemical flooding and cold production were adopted from 2018 to 2023.

[0085] The present invention provides an oil reservoir economic benefit optimization system based on improved LSTM, comprising the following steps:

[0086] Step 1: Experimental design;

[0087] Central composite design (CCD) using response surface methodology can effectively evaluate response surface models;

[0088] Select two injection parameters: injection rate and injection concentration, and design the experiment.

[0089] The variable levels for reservoir A can be set as follows:

[0090] Injection speed (X1): L1 = 100m 3 / d, L2=450m 3 / d, L3=800m 3 / d

[0091] Injection concentration (X2): L1 = 500 mg / L, L2 = 4250 mg / L, L3 = 8000 mg / L

[0092] Response variable: Net present value NPV(Y),

[0093] The numerical simulation software was used to simulate and obtain the daily oil production and water content data of each group as shown in the attached figure. Figure 1 .

[0094] Step 2: Data collection: using numerical simulation software CMG to collect reservoir dynamic data: daily oil production, water content; and economic data: international crude oil prices, global GDP growth rate, and global crude oil production;

[0095] Step 3: Perform noise reduction preprocessing on the data in step 2, remove outliers from the reservoir dynamic data curve through statistical analysis, smooth the curve, and improve the accuracy and stability of the model;

[0096] Step 4: Model training: Use the LSTM algorithm with attention mechanism to build a prediction model. The model learns the relationship between each parameter and economic benefits based on historical data to predict future economic benefits.

[0097] Use the improved LSTM prediction model, input daily oil production data and oil price data, divide the data into training set and test set, and predict daily oil production and oil prices from 2024 to 2027.

[0098] When predicting daily oil production, add water content data, and the prediction results are as shown in the attached Figure 12-20 .

[0099] When forecasting oil prices, we add the global GDP growth rate and global crude oil production data. The forecast results are as follows: Fig.21 .

[0100] Step 5: Economic optimization, select the economic benefit indicator net present value (NPV), use the response surface method to optimize the reservoir injection parameters, find the optimal combination of different parameter combinations, and maximize the net present value in the cold production process.

[0101] In this embodiment, in step 1, the central composite design method CCD of the response surface method RSM is used for experimental design. The central composite design method CCD includes cubic points, central points and axial points.

[0102] a.Cubic points: extreme combinations of each factor, i.e., combinations of low and high levels.

[0103] Table 1 Cube point experimental combination

[0104] Injection speed Injection concentration 800 500 100 500 100 8000 800 8000

[0105] b. Center point: The center point is the combination of the middle levels of each factor.

[0106] Table 2 Center point experimental combination

[0107] Injection speed Injection concentration 450 4250

[0108] c. Axis point: used to increase the estimation of quadratic terms and capture nonlinear effects. The choice of axis point is usually determined by the α value, where α is usually where k is the number of factors.

[0109] In order to ensure that the value of α is closer to the actual range, the value of α is reduced and α=1.05 is taken.

[0110] Table 3 Axis point experimental combinations

[0111] Injection speed Injection concentration 82.5 4250 817.5 4250 450 312.5 450 8187.5

[0112] The above experiments are summarized as follows:

[0113] Table 4 Summary of injection parameter experimental combinations

[0114]

[0115]

[0116] In this embodiment, a Savitzky-Golay filter is used in step 3 to eliminate noise in the data while maintaining the trend of the data.

[0117] In this embodiment, the Savitzky-Golay filter converts the polynomial fitting process into a weighted average form, that is, a convolution formula is used to achieve smoothing, and the formula is as follows:

[0118]

[0119] Among them, y′ i is the data point after filtering and smoothing; y i+j is the original data point in the window; ω j is the weight coefficient of the filter, ω j The calculation of depends on the window size 2m+1 and the polynomial order n and remains constant for each data point.

[0120] The daily oil production data and water content data of reservoir A from 2013 to 2023, as well as the international crude oil price data, global GDP growth rate, and global crude oil production are collected and the data are denoised.

[0121] The daily oil production is denoised and the results are shown in the attached figure. Figure 2-10 ;

[0122] The oil price is denoised, and the results are shown in the attached Fig.11 .

[0123] In this embodiment, in step 4, LSTM uses a forget gate, an input gate, and an output gate to determine which information needs to be retained, forgotten, or output. Through the synergistic effect of these gates, LSTM can flexibly transmit information in different time steps to ensure that the model effectively learns important temporal dependencies in long sequences. The calculation formulas of the forget gate, the input gate, and the output gate are:

[0124] f t =σ(W f h t-1 +W f x t +b f )#(2);

[0125] i t =σ(W i h t-1 +W i x t +b i )#(3);

[0126] o t =σ(W o h t-1 +Wo x t +b o )#(4);

[0127] Among them, σ is the sigmoid activation function, the output is between 0 and 1, h t-1 is the hidden state of the previous time step, x t is the input of the current time step, W f , W i , W i are the weight matrices of the forget gate, input gate, and output gate, respectively, and f t 、i t , o t are the outputs of the forget gate, input gate, and output gate, respectively, and b f , b i , b o They are the bias vectors of the forget gate, input gate, and output gate respectively;

[0128] The LSTM calculation process is as follows:

[0129]

[0130] h t =o t tanh(C t )#(7);

[0131] in, is the candidate memory state, W C , b C are the weight matrix and bias vector of the candidate memory state, respectively, C t , C t-1 are the memory states of the current time step and the previous time step, respectively, h t is the hidden state of the current time step, that is, the output of the LSTM unit.

[0132] In this embodiment, after using the attention mechanism in step 4, the model will consider the importance of different parts of the input sequence when generating output.

[0133] In this embodiment, the attention mechanism includes calculating the attention score, calculating the attention weight, and calculating the context vector;

[0134] a. Calculate attention score

[0135] For each time step t, calculate the current hidden state h t With each hidden state h in the input sequence i Similarity score:

[0136] e ti =α(ht ,h i )#(8);

[0137] Among them, α is a similarity function, common choices are dot product or additive similarity function;

[0138] b. Calculate attention weights

[0139] Perform softmax normalization on the attention score to get the attention weight:

[0140]

[0141] c. Calculate the context vector

[0142] The context vector is obtained by weighted summing the hidden states of the input sequence using the attention weights:

[0143]

[0144] In this embodiment, the NPV calculation formula in step 5 is as follows:

[0145]

[0146] Where: C t is the cash flow in year t; r is the discount rate, usually based on the market interest rate or the cost of capital; n is the life of the project; I0 ​​is the initial investment cost;

[0147] C t =R t -E t =Q t ·P t -E t #(12)

[0148] Where: R t is the total income in year t, Q t is the crude oil production in year t; P t is the crude oil price in year t; E t is the total expenditure in year t.

[0149] The initial investment in the reservoir is I0 = 2.7 million US dollars, r = 10%,

[0150] Table 5 Annual net cash flow and annual net present value table of combination 1

[0151]

[0152]

[0153] Therefore, we can get:

[0154]

[0155] Similarly, the NPV values ​​of all combinations are calculated as follows:

[0156] Table 6 NPV table of nine experimental combinations

[0157] combination Injection speed Injection concentration NPV 1 800 500 25740747.68 2 100 500 11224535.51 3 100 8000 11412302.22 4 800 8000 27429352.54 5 450 4250 27100936.44 6 82.5 4250 9988501.348 7 817.5 4250 27258839.48 8 450 312.5 25435923.97 9 450 8187.5 27176374.33

[0158] In this embodiment, the response surface method in step 5 usually assumes that the response variable y and the input variables x1, x2, ..., x k The relationship between can be approximated by a binomial model; the commonly used second-order polynomial model formula is:

[0159]

[0160] Among them, β0 is the intercept term, representing the linear effect of the input variable, and β i is the second-order effect coefficient, representing the quadratic effect of each variable, β ij is the interaction effect coefficient, indicating that x i and x j The interaction between them.

[0161] Fitting response surface models;

[0162] The response surface model is a quadratic polynomial, as follows:

[0163]

[0164] in:

[0165] Y is the net present value NPV; β0 is the constant term, β1, β2 are the regression coefficients of the first-order effect; β 11 , β 22 is the regression coefficient of the second-order effect, β 12 is the regression coefficient of the interaction effect.

[0166] Substitute the experimental data into the quadratic polynomial and fit it to get the regression coefficient as follows:

[0167] Table 7 Regression coefficient table

[0168]

[0169]

[0170] From the three-dimensional graph of the curve, we can see that the optimal combination is:

[0171] X1=639.22

[0172] X2=6994.32

[0173] At this time, the net present value is: Y = 29313722.39

[0174] From the optimal combination of parameter values, it can be seen that when the injection rate of the chemical flooding stage of the reservoir is set to 639.22m 3 / d, when the injection chemical flooding concentration is set to 6994.32 mg / L, the economic net present value of the reservoir by the end of 2027 is the largest, which is US$29,313,722.39.

[0175] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A reservoir economic benefit optimization system based on improved LSTM, characterized by: The following steps are involved: Step 1: Experimental design; Step 2: Data collection: Collect reservoir dynamic data in numerical simulation software CMG: daily oil production, water content; And collect economic data: international crude oil prices, global GDP growth rate, global crude oil production; Step 3: Perform noise reduction preprocessing on the data in step 2, remove outliers from the reservoir dynamic data curve through statistical analysis, smooth the curve, and improve the accuracy and stability of the model; Step 4: Model training: Use the LSTM algorithm with attention mechanism to build a prediction model. The model learns the relationship between each parameter and economic benefits based on historical data to predict future economic benefits. Step 5: Economic optimization, select the economic benefit indicator net present value (NPV), use the response surface method to optimize the reservoir injection parameters, find the optimal combination of different parameter combinations, and maximize the net present value in the cold production process.

2. The reservoir economic benefit optimization system based on improved LSTM according to claim 1, characterized in that: In step 1, the central composite design method CCD of the response surface method RSM is used for experimental design. The central composite design method CCD includes cubic points, central points and axial points.

3. The reservoir economic benefit optimization system based on improved LSTM according to claim 1, characterized in that: In step 3, a Savitzky-Golay filter is used to remove the noise from the data while maintaining the trend of the data.

4. The reservoir economic benefit optimization system based on improved LSTM according to claim 3 is characterized in that: The Savitzky-Golay filter converts the polynomial fitting process into a weighted average form, that is, a convolution formula is used to achieve smoothing. The formula is as follows: Among them, y′ i is the data point after filtering and smoothing; y i+j is the original data point in the window; ω j is the weight coefficient of the filter, ω j The calculation of depends on the window size 2m+1 and the polynomial order n and remains constant for each data point.

5. The reservoir economic benefit optimization system based on improved LSTM according to claim 1, characterized in that: In step 4, LSTM uses the forget gate, input gate, and output gate to decide which information needs to be retained, forgotten, or output. Through the synergy of these gates, LSTM can flexibly transmit information at different time steps to ensure that the model effectively learns important temporal dependencies in long sequences. The calculation formulas for the forget gate, input gate, and output gate are: f t =σ(W f h t-1 +W f x t +b f )#(2); i t =σ(W i h t-1 +W i x t +b i )#(3); o t =σ(W o h t-1 +W o x t +b o )#(4); Among them, σ is the sigmoid activation function, the output is between 0 and 1, h t-1 is the hidden state of the previous time step, x t is the input of the current time step, W f , W i , W i are the weight matrices of the forget gate, input gate, and output gate, respectively, and f t 、i t , o t are the outputs of the forget gate, input gate, and output gate, respectively, and b f , b i , b o They are the bias vectors of the forget gate, input gate, and output gate respectively; The LSTM calculation process is as follows: h t =o t fishy(C) t )#(7); in, is the candidate memory state, W C , b C are the weight matrix and bias vector of the candidate memory state, respectively, C t , C t-1 are the memory states of the current time step and the previous time step, respectively, h t is the hidden state of the current time step, that is, the output of the LSTM unit.

6. The reservoir economic benefit optimization system based on improved LSTM according to claim 1, characterized in that: After using the attention mechanism in step 4, the model will consider the importance of different parts of the input sequence when generating output.

7. The reservoir economic benefit optimization system based on improved LSTM according to claim 1, characterized in that: The attention mechanism includes calculating the attention score, calculating the attention weight, and calculating the context vector; a. Calculate attention score For each time step t, calculate the current hidden state h t With each hidden state h in the input sequence i Similarity score: in ti =α(h t ,h i )#(8); Among them, α is a similarity function, common choices are dot product or additive similarity function; b. Calculate attention weights Perform softmax normalization on the attention score to get the attention weight: c. Calculate the context vector The context vector is obtained by weighted summing the hidden states of the input sequence using the attention weights:

8. The reservoir economic benefit optimization system based on improved LSTM according to claim 1, characterized in that: The NPV calculation formula in step 5 is as follows: Where: C t is the cash flow in year t; r is the discount rate, usually based on the market interest rate or the cost of capital; n is the life of the project; I0 ​​is the initial investment cost; C t =R t -E t =Q t ·P t -E t #(12) Where: R t is the total income in year t, Q t is the crude oil production in year t; P t is the crude oil price in year t; E t is the total expenditure in year t.

9. The reservoir economic benefit optimization system based on improved LSTM according to claim 1, characterized in that: In step 5, the response surface method usually assumes that the response variable y and the input variables x1, x2, ..., x k The relationship between can be approximated by a binomial model; the commonly used second-order polynomial model formula is: Among them, β0 is the intercept term, representing the linear effect of the input variable, and β i is the second-order effect coefficient, representing the quadratic effect of each variable, β ij is the interaction effect coefficient, which means x i and x j The interaction between them.

Images