Dynamic Optimization Method for Carbon Dioxide Mixed Gas Huff and Puff Parameters Based on Multi-Objective Interaction

Abstract

The present invention belongs to the technical field of huff and puff development of tight reservoirs, and specifically discloses a dynamic optimization method for carbon dioxide mixed gas huff and puff parameters based on multi-objective interaction, which is used to solve the problem of how to supplement formation energy and improve oil recovery in the middle and late stages of current tight reservoir development. It includes: selecting a target block for design and collecting a time series prediction sample set; standardizing the sample data, dividing the data set, constructing a TFT prediction surrogate model using a TFT neural network, and using the TFT prediction surrogate model to quickly predict the recovery rate, storage volume, and NPV; determining the objective function, constructing a multi-objective optimization model, and establishing a dynamic optimization model for the huff and puff effect of CO2 mixed gas based on the NSGA-Ⅲ multi-objective optimization algorithm. The present invention effectively improves production efficiency, reduces risks and uncertainties, and realizes balanced optimization among multiple objectives.

Description

Technical Field

[0001] The present invention belongs to the technical field of huff and puff development of tight oil reservoirs, and particularly relates to a method for dynamically optimizing carbon dioxide mixed gas huff and puff parameters based on multi-objective interaction. Background Art

[0002] Most of the existing tight oil development strategies have certain limitations. Among them, although the initial production of depletion exploitation is relatively high, the production decline rate is generally high, the formation energy will rapidly deplete, resulting in low recovery rate and low economic benefits; due to the poor pore permeability conditions of tight reservoirs, the fluidity of oil and gas is poor, and it is difficult to achieve water injection development; although gas flooding is effective quickly, early gas channeling is likely to occur, leading to low gas utilization rate. Huff and puff with gas injection is a general strategy to solve the serious problem of gas channeling in conventional gas flooding, and it is also an effective method to improve the crude oil recovery rate of tight oil reservoirs.

[0003] By summarizing previous studies, huff and puff with CO 2 mixed gas combination is an efficient development method to improve the crude oil recovery rate of tight oil reservoirs, and the production increase effect is higher than that of single gas huff and puff. However, in the process of tight oil development, the reasonable development technical policy boundary for improving the crude oil recovery rate by huff and puff with CO 2 mixed gas is not clear. Therefore, dynamically optimizing the development factors of huff and puff with CO 2 mixed gas is the key to the adjustment of huff and puff development of tight oil reservoirs. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for dynamically optimizing carbon dioxide mixed gas huff and puff parameters based on multi-objective interaction, which effectively solves the problems of how to supplement formation energy and improve crude oil recovery rate in the middle and late stages of current tight oil reservoir development.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a method for dynamically optimizing carbon dioxide mixed gas huff and puff parameters based on multi-objective interaction, including the following steps: S1. Select a target block for design and collect a time series prediction sample set; S2. Standardize the sample data, divide the data set, construct a TFT prediction surrogate model using the TFT neural network, and use the TFT prediction surrogate model to quickly predict the recovery rate, buried volume and NPV; S3. Determine the objective function, construct a multi-objective optimization model, and establish a CO 2 mixed gas huff and puff effect dynamic optimization model based on the NSGA-Ⅲ multi-objective optimization algorithm to ensure a reasonable and efficient working system at different time steps, and realize the dynamic optimization of carbon dioxide mixed gas huff and puff parameters of tight oil reservoirs. 2

[0006] Furthermore, in step S2, the sample data standardization process is to normalize each feature using the min-max standardization method:

[0007] ;

[0008] wherein, is the value after normalization, is the actual value of the feature factor in the dataset, is the maximum value of the feature factor in the original data, is the minimum value of the feature factor in the original data.

[0009] The original data after normalization is divided into a training set, a validation set, and a test set according to the ratio of 6:2:2, which are used for the training of the TFT prediction proxy model, the hyperparameter tuning of the TFT neural network, and the performance evaluation of the TFT prediction proxy model, respectively.

[0010] Furthermore, in step S2, the components of the TFT neural network include input features, a variable selection network, a static covariate encoder, a sequence-to-sequence layer, a gated residual network, a multi-head attention mechanism, a temporal fusion decoder, and quantile outputs.

[0011] Regarding the input features: The TFT neural network divides the input variables into static variables and dynamic variables. Among them, the static variables are divided into static continuous variables and static discrete variables, and the dynamic variables include dynamic time-varying variables and dynamic time-invariant variables. Dynamic time-varying variables refer to variables whose features change over time and cannot be inferred, including recovery factor, buried volume, and net present value; dynamic time-invariant variables are variables applied to the TFT prediction proxy model based on reasoning, including month and cumulative injection volume.

[0012] Furthermore, in step S2, the functions of the variable selection network include: selecting relevant input variables for each time step, and weakening or discarding irrelevant information; the static covariate encoder integrates static features into the TFT neural network and adjusts the time dynamics by encoding the context vector; the sequence-to-sequence layer uses two groups of LSTM networks, one of which is used to process past input features and encode past historical inputs; the other is used to process future input features; the gated residual network is a selection framework using a gating mechanism, and selectively introduces relevant input features through selectable skip connections and gating matrices.

[0013] The TFT prediction proxy model learns to capture long-term dependencies at different time steps through the multi-head attention mechanism, and simultaneously pays attention to different parts of the input sequence in parallel, increasing the model attention weight for specific features.

[0014] The temporal fusion decoder consists of two layers of gated residual networks and a multi-head attention mechanism layer; the first layer of gated residual network combines the output information of the sequence-to-sequence layer with the information from the static features, and then the multi-head attention mechanism layer performs the standard self-attention mechanism on the input information at all time steps. Finally, the output of the multi-head attention mechanism layer and the remaining connections are passed to another layer of gated residual network for non-linear processing.

[0015] Further, in step S2, the TFT neural network determines the prediction interval of the possible target value range of each prediction layer through quantile prediction, and uses quantile regression to generate the prediction interval, so as to obtain information about prediction certainty; the quantile prediction is generated by the linear transformation of the output of the temporal fusion decoder.

[0016] The TFT prediction surrogate model trains the TFT neural network by jointly minimizing the quantile loss of all quantile outputs:

[0017] ;

[0018] ;

[0019] Among them, is the total loss function, is the quantile loss function, is the actual observed value at the th time step, is the predicted value of the model for the th time step at the quantile , is the estimated value of the model for the sample according to the input features, is the training data containing samples , is the set of predicted data of the output quantiles at different time steps , , is the maximum value of the time step.

[0020] The normalized quantile loss is used for the validation and testing of the TFT prediction surrogate model:

[0021] ;

[0022] Among them, is the prediction time span, is the maximum length of the time series, represents the threshold value of the test sample, represents the normalized quantile loss, Indicates the model at time point For the time point Quantile of Predicted value of

[0023] Furthermore, in step S2, the gated residual network is defined as:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] Wherein, Is the input variable parameter, Is the static variable vector from the static selection network, Is the gated residual network with the width multiplication factor Of Is the linear unit activation function, Is to implement the standard normalization function, Is the gated linear unit, Is the sigmoid activation function, Is the weight matrix of the first transformation layer in the GRN, Is associated with Corresponding bias vector, Is the weight matrix of the second transformation layer in the GRN, Is associated with Corresponding bias vector, Is the weight matrix of the third transformation layer in the GRN, Is associated with Corresponding bias vector, Is the weight matrix of the first linear transformation in the GLU, Is the weight matrix of the second linear transformation in the GLU, Is associated with Corresponding bias vector, Is associated with Corresponding bias vector, Represents the element-wise product, Is the width multiplication factor, Is the output after the first transformation layer, Is the output after being processed by the function Of Is Input variable parameter of

[0029] Further, in step S3, the multi-objective optimization model includes three factors: an objective function, optimization variables, and constraint conditions; the objective function is to maximize the recovery rate, maximize the buried volume, and maximize the NPV; the optimization variables include the huff and puff development method and the production regime of each horizontal well. Among them, the production regime includes the mixed gas ratio, gas injection rate, gas injection time, soaking time, oil production rate, flowback time, and huff and puff cycle. It is set that a fixed amount of gas is injected during the gas injection stage of the huff and puff well, and a fixed amount of production is carried out during the flowback stage; regarding the constraint conditions: the key parameters of each huff and puff well, including the daily gas injection volume and daily oil production volume, do not exceed the maximum working capacity limit of each single well and are not lower than zero.

[0030] Further, there is a correlation between the huff and puff parameters of each well under different huff and puff development methods. Let represent the middle horizontal well, and represent two side wells, and the constraint condition expressions are as follows.

[0031] When huff and puff simultaneously, the huff and puff cycles of the three wells are the same, and only the mixed gas ratio, gas injection rate, oil production rate, and huff and puff cycle are different:

[0032] ;

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] Among them, represents the proportion of the mixed gas injected by the th horizontal well, represents the gas injection rate of the th horizontal well, represents the oil production rate of the th horizontal well, , , respectively represent the gas injection times of , and , , , respectively represent the flowback times of , and The soaking time of , , respectively represent , and the flowback time of represents the huff and puff cycle of the th horizontal well, and respectively represent the minimum and maximum values of the mixed gas ratio, and respectively represent the minimum and maximum values of the gas injection rate, and respectively represent the minimum and maximum values of the oil production rate, and respectively represent the minimum and maximum values of the gas injection time, and respectively represent the minimum and maximum values of the soaking time, and respectively represent the minimum and maximum values of the flowback time, and respectively represent the maximum and minimum values of the huff and puff cycle.

[0040] When huff and puff is carried out asynchronously, the gas injection time, soaking time and flowback time of each time step of the three wells are the same, and only the mixed gas ratio, gas injection rate, oil production rate and huff and puff cycle are different:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] Among them, and respectively represent the minimum and maximum values of the time of each time step. The gas injection time, soaking time and flowback time of each time step of the three wells are the same.

[0049] When huff and puff is carried out alternately in adjacent wells, During depletion development, there is only one variable, the oil production rate; and have the same throughput cycle, and only the mixture gas ratio, gas injection rate, oil production rate, and huff and puff cycles are different:

[0050] , ;

[0051] , ;

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] , ;

[0057] Among them, represents the ratio of the injected mixture gas, represents the ratio of the injected mixture gas, represents the gas injection rate of represents the gas injection rate of

[0058] Furthermore, in step S3, the TFT prediction proxy model established in step S2 is coupled, and the TFT prediction proxy model is used to quickly predict the recovery rate, buried storage volume, and NPV. Secondly, the NSGA-Ⅲ multi-objective optimization algorithm is called to dynamically regulate the huff and puff development method and the huff and puff working system of each well, ensuring that an efficient and reasonable working system is matched at different time steps, and realizing the dynamic optimization of the CO 2 mixed gas flooding-buried storage effect

[0059] Compared with the prior art, the beneficial technical effects of the present invention are: (1) The TFT prediction proxy model established by the present invention trains the TFT neural network by screening the data set, which is beneficial to more accurately fitting and predicting the huff and puff effect; the CO 2 mixed gas huff and puff effect dynamic optimization model established by the present invention can optimize multiple development objectives with conflicting interests (buried storage volume, recovery rate, and NPV), and combined with the TFT prediction proxy model, obtain the buried storage volume, recovery rate, and NPV optimization schemes based on the benchmark scheme; it is beneficial to maximize the buried storage volume, recovery rate, and NPV, and can provide for CO 2It provides guidance for the optimal design of the mixer throughput development plan, which is of positive significance to reservoir development and management.

[0060] (2) The dynamic optimization model of the CO 2 mixed gas huff and puff effect effectively improves production efficiency, reduces risks and uncertainties; and this model comprehensively considers three development objectives of buried volume, recovery rate and NPV, and realizes the balanced optimization among multiple objectives. Description of the Drawings

[0061] Figure 1 It is the relationship curve between the loss value of the TFT neural network and the number of iterations in Example 1.

[0062] Figure 2 It is the comparison chart of the predicted value and the true value of the recovery rate by the trained TFT prediction proxy model in Example 1.

[0063] Figure 3 It is the comparison chart of the predicted value and the true value of the buried volume by the trained TFT prediction proxy model in Example 1.

[0064] Figure 4 It is the comparison chart of the predicted value and the true value of the NPV by the trained TFT prediction proxy model in Example 1.

[0065] Figure 5 It is the comparison chart of the changing trends of the average formation pressure with time under the production capacity plan and the basic plan in Example 1.

[0066] Figure 6 It is the comparison chart of the changing trends of the recovery rate with time under the production capacity plan and the basic plan in Example 1.

[0067] Figure 7 It is the comparison chart of the changing trends of the structural buried volume with time under the storage plan and the basic plan in Example 1.

[0068] Figure 8 It is the comparison chart of the changing trends of the immobile gas buried volume with time under the storage plan and the basic plan in Example 1. Detailed Description of the Invention

[0069] Example 1: Taking the actual data of a tight sandstone reservoir as an example, the target well group is intercepted from the actual heterogeneous geological model of a certain block. By coupling the tight oil reservoir TFT prediction proxy model with the multi-objective optimization algorithm, a proxy-assisted multi-objective optimization process is established. Selecting the maximum recovery rate, the maximum buried volume and the maximum net present value (NPV) as the objective functions, the multi-level parameters and multi-objective dynamic optimization of the CO 2 mixed gas huff and puff are carried out to realize the dynamic regulation of the working systems of each well, and a set of plans considering different development objectives is formed.

[0070] Based on the above data, a dynamic optimization method for carbon dioxide mixed gas huff and puff parameters based on multi-objective interaction is adopted to jointly optimize the well location and injection-production parameters, including the following steps: S1. Select the target block for design and collect the time series prediction sample set.

[0071] (1) Selection of the target block: The selection of the target block mainly combines the remaining oil occurrence characteristics and seepage characteristics of the block, and at the same time considers avoiding the influence of faults and special structures. An actual model with a size of 1800 × 1510 m is intercepted from the geological model, and three horizontal wells are designed and established.

[0072] (2) Construction of the time series prediction sample set: The time series data involved in this embodiment is set as the huff and puff development mode, as well as the mixed gas ratio, gas injection rate, oil production rate, gas injection time, soaking time, and flowback time of each well, and the crude oil recovery rate, CO 2 buried storage volume, and NPV that change with time as output variables. Set to use a set of production parameters every 5 years, and set the time interval to record a set of sample data every 6 months. The setting of the huff and puff development parameters is mainly to constrain the parameter range according to the actual production situation of the oilfield site. Based on the Monte Carlo sampling method, specific variable values are extracted within the set range of each parameter, and the huff and puff development parameters of each well are respectively assigned values. Based on the reservoir numerical simulation model of the target block, 600 different production plans are randomly selected, and the time step is set to 6 months, and a total of 18,000 data samples under different production times and different production systems can be obtained, constituting the time series prediction sample set required for this embodiment.

[0073] S2. Standardize the sample data, divide the data set, construct a TFT prediction surrogate model using the TFT neural network, and use the TFT prediction surrogate model to quickly predict the recovery rate, buried storage volume, and NPV.

[0074] (1) Standardization processing of the sample data: Since the unprocessed sample data have differences in units and orders of magnitude, in order to ensure the efficiency and accuracy of the training process of the TFT prediction surrogate model, it is necessary to standardize the sample data. In this embodiment, the minimum-maximum normalization method is used to normalize each feature:

[0075] ;

[0076] where is the value after normalization processing, is the actual value of the feature factor in the data set, is the maximum value of the feature factor in the original data, is the minimum value of the feature factor in the original data.

[0077] (2)Design of TFT Neural Network: The design of the TFT neural network aims to use canonical components to effectively represent various input types, including static features that do not change over time, past information, and constructed features of known future inputs. The components of the TFT neural network include input features, variable selection network, static covariate encoder, sequence-to-sequence layer, gated residual network, multi-head attention mechanism, temporal fusion decoder, and quantile output.

[0078] I. Input Features: When the TFT prediction proxy model is used for time series problem prediction, the input variables are divided into two major categories: static variables and dynamic variables. Among them, static variables are divided into static continuous variables and static discrete variables, and dynamic variables include dynamic time-varying variables and dynamic time-invariant variables. Among them, dynamic time-varying variables refer to variables whose features change over time and cannot be inferred, such as features like recovery rate, buried volume, and net present value; dynamic time-invariant variables are variables that can be applied to the TFT prediction proxy model through inference, such as features like month and cumulative injection volume. Dynamic time-varying variables provide sensitivity to time series changes and can capture key features that evolve over time in the data. In contrast, dynamic time-invariant variables provide the TFT prediction proxy model with features that remain stable in the time dimension, enabling the TFT prediction proxy model to better understand time-independent factors.

[0079] II. Variable Selection Network (VSN): The main role of the variable selection network is to select relevant input variables for each time step. By strengthening the main control factors crucial for the prediction task, the VSN can effectively improve the prediction performance of the TFT prediction proxy model. At the same time, the VSN also has the ability to weaken or discard irrelevant information, thereby optimizing the representation of input features and making the TFT prediction proxy model more focused on accurate prediction of the target.

[0080] This strategy of introducing the variable selection network helps to improve the expressive ability of the TFT neural network in processing time series data. By automatically learning and selecting input variables, the TFT prediction proxy model can more flexibly adapt to different tasks and data characteristics. The role of the VSN is similar to an intelligent filter, enabling the TFT prediction proxy model to more accurately capture key information when facing complex time series prediction problems and improving the overall prediction performance.

[0081] III. Static Covariate Encoder (SCE): Compared with other time series prediction architectures, the TFT neural network effectively integrates information from static metadata. The introduction of the static covariate encoder enables static features to be incorporated into the entire network structure, regulating time dynamics by encoding context vectors. This integration enables the TFT neural network to have better adaptability, being able to make full use of static information for prediction while considering time dynamics.

[0082] IV. Sequence-to-Sequence Layer: This part uses two sets of LSTM networks. One set is used to process past input features and encode past historical inputs, and the other set is used to process future input features and act as a decoder. By locally processing known input features and observed input features, it provides an appropriate inductive bias for the temporal ordering of the inputs, thereby improving the performance of the next layer based on the temporal attention architecture.

[0083] V. Gated Residual Network (GRN): The gated residual network is a selective framework that utilizes a gating mechanism. Through optional skip connections and gated matrix filtering, it selectively introduces relevant input features. GRN includes a dense layer followed by a gating mechanism that can selectively cancel values to remove unnecessary information. They can be defined as:

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] where is the input variable parameter, is the static variable vector from the static selection network, is the gated residual network with a width multiplication factor , is the linear unit activation function, is to implement the standard normalization function, is the gated linear unit, is the sigmoid activation function, is the weight matrix of the first transformation layer in GRN, is the bias vector corresponding to , is the weight matrix of the second transformation layer in GRN, is the bias vector corresponding to , is the weight matrix of the third transformation layer in GRN, is the bias vector corresponding to , is the weight matrix of the first linear transformation in GLU, is the weight matrix of the second linear transformation in GLU, is the bias vector corresponding to , is the bias vector corresponding to The corresponding bias vector, represents the element-wise product, is the width multiplication factor, is the output after the first transformation layer, is the output after processing by the function is the input variable parameter of is

[0089] VI. Multi-Head Attention Mechanism: Based on the multi-head attention mechanism in the Transformer model, the TFT prediction proxy model also learns to capture long-term dependencies at different time steps through the multi-head attention mechanism. The TFT prediction proxy model improves it by introducing multiple attention heads, enabling the TFT prediction proxy model to concurrently focus on different parts of the input sequence, thereby better capturing the correlation information in the time series. By increasing the attention weights of the TFT prediction proxy model for specific features, the importance of key features for the prediction results is emphasized. This not only improves the modeling ability of the TFT prediction proxy model for long-term dependencies but also enhances the interpretability of the prediction results of the TFT prediction proxy model.

[0090] VII. Temporal Fusion Decoder (TFD): The temporal fusion decoder consists of two layers of gated residual networks and a multi-head attention mechanism layer. In this structure, the first layer of gated residual network combines the output information from the sequence-to-sequence layer with the information from static features, achieving the fusion of past and static features; then, the multi-head attention mechanism layer performs the standard self-attention mechanism on the input information at all time steps, enabling the TFT prediction proxy model to better capture the long-term dependencies within the sequence; finally, the output of the multi-head attention mechanism and the remaining connections are passed to another layer of gated residual network for non-linear processing.

[0091] VIII. Quantile Output: Different from other prediction models that give single-point estimates, the TFT neural network determines the prediction interval of the possible target value range for each prediction layer through quantile prediction. Quantile regression is used to generate the prediction interval, thereby obtaining more information about prediction certainty. Quantile prediction is generated through a linear transformation of the output of the temporal fusion decoder. The TFT prediction proxy model trains the TFT neural network by jointly minimizing the quantile loss of all quantile outputs:

[0092] ;

[0093] ;

[0094] where is the total loss function, is the quantile loss function, is the actual observation value at the th time step, is the prediction value of the model for the th time step at the quantile is the estimated value of the model for the sample based on the input features, is the set containing samples for training data, is the set of prediction data for different output quantiles at different time steps , , is the maximum value of the time step.

[0095] The normalized quantile loss is used for the validation and testing of the TFT prediction surrogate model:

[0096] ;

[0097] where, is the prediction time span, is the maximum length of the time series, represents the threshold value of the test sample, represents the normalized quantile loss, represents the prediction value of the model at time point for the quantile at time point .

[0098] (3)Dataset division: Based on the TFT neural network time series algorithm, the tight oil reservoir CO 2 mixed gas huff and puff time series prediction sample set established in this embodiment is used for learning and training. According to the ratio of 6:2:2, the original data is divided into a training set, a validation set, and a test set, which are used for the training of the TFT prediction surrogate model, the hyperparameter tuning of the TFT neural network, and the performance evaluation of the TFT prediction surrogate model, respectively.

[0099] (4)Hyperparameter tuning: The hyperparameter optimization process of the TFT neural network is a key step. It involves selecting and adjusting the neural network structure and various parameters in its training process to maximize the performance of the TFT prediction surrogate model. Hyperparameter optimization is carried out through random search, and the number of iterations is set to 200 times. The final hyperparameter configuration of the TFT neural network after tuning is shown in Table 1.

[0100] Table 1 TFT neural network hyperparameter configuration

[0101]

[0102] (5) Prediction effect evaluation: The evaluation method uses three evaluation indexes, namely, mean absolute percentage error ( ), mean square error ( ), and mean absolute error ( ), to evaluate and analyze the prediction effect of the TFT prediction surrogate model for tight oil reservoirs. These evaluation indexes are used to measure the degree of dispersion between the actual values calculated by numerical simulation and the predicted values of the TFT prediction surrogate model. Through these indexes, the accuracy of the TFT prediction surrogate model can be quantified, and its fitting degree to the actual observed values in the validation set can be evaluated. Different evaluation indexes have different focuses, and their specific calculation processes are as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] Among them, is the th actual value, is the th predicted value, and is the number of actual values in the validation set.

[0107] When the values of the three evaluation indexes are smaller, it indicates that the error between the predicted value of the TFT prediction surrogate model and the actual observed value calculated by numerical simulation is smaller, which proves that the established TFT prediction surrogate model has stronger performance ability.

[0108] The evaluation and analysis results of the prediction effect of the TFT prediction surrogate model in this embodiment are shown in Table 2. The relationship curve between the loss value of the TFT neural network obtained by training and the number of iterations is as Figure 1 shown, indicating that the TFT neural network obtained by training has good convergence and can predict errors.

[0109] Table 2 Evaluation indexes of the prediction effect of the TFT prediction surrogate model

[0110]

[0111] From the total of 600 production plans in the time series prediction sample set, a test set of 120 production plans is selected, and the sample data at the end time is used to make a detailed comparison of the prediction effect. The comparison between the predicted values and the true values of the recovery rate, buried volume, and NPV can be seen in Figure 2 , Figure 3 and Figure 4 . From Figures 2 to 4It can be seen that during the test of the TFT prediction proxy model, the model can well fit the actual recovery rate, storage volume, and NPV obtained through numerical simulation calculations. Among them, the fitting effect of the recovery rate is the best, followed by NPV and storage volume. The error between the actual value and the predicted value is within the allowable reasonable range, indicating that the TFT prediction proxy model established in this embodiment can effectively predict the oil displacement-storage effect under different production systems, and at the same time proves that the use of the TFT neural network can achieve accurate prediction of the CO 2 mixed gas huff and puff effect.

[0112] Select the production system data of a certain set of solutions in the test set, use the TFT prediction proxy model based on the TFT neural network to predict its recovery rate, storage volume, and NPV, and compare it with the numerical simulation calculation results. The comparison results show that: the predicted results of the recovery rate are very close to the numerical simulation calculation results, indicating that the TFT prediction proxy model has a high ability to restore the actual production change trend; for the two targets of storage volume and NPV, the model also shows a high prediction accuracy, and the error always remains within an acceptable range. In summary, the error between the prediction results of the TFT prediction proxy model and the numerical simulation calculation results is relatively small, indicating that the prediction effect of the TFT prediction proxy model is within the ideal range, with high credibility, and can replace numerical simulation calculations to achieve rapid response for different targets.

[0113] S3. Determine the objective function, construct a multi-objective optimization model, and establish a CO 2 mixed gas huff and puff effect dynamic optimization model based on the NSGA-Ⅲ multi-objective optimization algorithm to ensure a reasonable and efficient working system is matched at different time steps, and realize the dynamic optimization of CO 2 mixed gas huff and puff parameters.

[0114] (1) Problem analysis: In this embodiment, the maximum recovery rate, maximum storage volume, and maximum NPV are taken as the objectives, and multi-objective dynamic optimization is carried out on the development parameter combinations that change every 5 years during the huff and puff development process of the tight oil reservoir. The 15-year production process is divided into three stages: the early stage, the middle stage, and the late stage of production, and three sets of production system solutions that change with time are optimized.

[0115] The three objective functions of recovery rate, storage volume, and NPV respectively reflect the oil displacement effect, storage effect, and overall economic effect during the huff and puff process. The recovery rate is the main index of the oil displacement effect during the huff and puff process, the storage volume is related to the storage effect during the huff and puff process, and the calculation of NPV not only reflects the economic benefits during the huff and puff process, but also reflects to a certain extent the size of the gas utilization rate of CO 2 mixed gas. The calculation formula for the net present value NPV is as follows:

[0116] ;

[0117] Among them, is the net present value, 10 7 Yuan; is the production time, in years, ; is the discount rate; is the cumulative oil production, m 3 ; is the cumulative gas injection volume, m 3 ; is the current crude oil price, Yuan / m 3 ; is the current gas injection cost, Yuan / m 3 .

[0118] (2) Optimization mathematical model establishment: When conducting the optimization design of CO 2 mixed gas huff and puff parameters, due to the contradictory relationship among the recovery factor, NPV and the buried volume, in order to obtain the best development effect, in this embodiment, the highest benefit is achieved through the simultaneous and independent optimization of three objectives. It can be seen from this that the optimization process is a multi-objective and multi-variable engineering problem. The optimization mathematical model established for this problem mainly includes three factors, namely the objective function, the optimization variables and the constraint conditions.

[0119] I. Objective function: The core of the optimization problem of CO 2 mixed gas huff and puff oil displacement - buried storage effect in tight oil reservoirs is the multi-objective, multi-round and dynamic optimization of huff and puff development parameters. During the huff and puff development process, high productivity does not necessarily mean high buried storage effect and economic benefits. Adjusting the proportion of the injected mixed gas and other huff and puff development parameters may have different degrees of influence on productivity, buried storage effect and economic benefits. To ensure the improvement of the CO 2 buried volume while efficiently producing crude oil, in this embodiment, a multi-objective optimization mathematical model of CO 2 mixed gas huff and puff parameters in tight oil reservoirs is established with the maximization of the recovery factor, the maximization of the buried volume and the maximization of NPV as the objective function. The three objectives are interrelated and pose challenges for the selection and adjustment of huff and puff development parameters.

[0120] II. Optimization variables: In the optimization problem of CO 2 mixed gas huff and puff oil displacement - buried storage effect, the optimization variables involved include the huff and puff development method and the production regime of each horizontal well. Among them, the production regime includes 7 parameters, namely the mixed gas proportion, the gas injection rate, the gas injection time, the soaking time, the oil production rate, the flowback time and the huff and puff cycle. The gas injection in the gas injection stage of the huff and puff well is quantitatively injected, and the production in the flowback stage is quantitatively produced.

[0121] III. Constraints: In the problem of optimizing throughput parameters, the constraint range of the working system changes of horizontal wells needs to be considered. Key parameters such as the daily gas injection volume and daily oil production volume of each cyclic steam stimulation well need to follow the actual operation limits of the oilfield site, that is, these parameters shall not exceed the upper limit of the maximum working capacity of each individual well and shall not be lower than zero. At the same time, there is a certain correlation between the throughput parameters of each well under different cyclic steam stimulation development methods. Use to represent the middle horizontal well, and to represent two side wells. The constraint condition expressions are as follows.

[0122] ① When cyclic steam stimulation is synchronous, the cyclic steam stimulation periods of the three wells are the same, and only the mixed gas ratio, gas injection rate, oil production rate, and number of cyclic steam stimulation rounds are different:

[0123] ;

[0124] ;

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] Among them, represents the ratio of the mixed gas injected by the th horizontal well, represents the gas injection rate of the th horizontal well, represents the oil production rate of the th horizontal well, , , respectively represent the gas injection times of , and , , , respectively represent the soaking times of , and , , , respectively represent the flowback times of , and , represents the The number of huff and puff cycles of the horizontal well, and represent the minimum and maximum values of the mixed gas ratio respectively, and represent the minimum and maximum values of the gas injection rate respectively, and represent the minimum and maximum values of the oil production rate respectively, and represent the minimum and maximum values of the gas injection time respectively, and represent the minimum and maximum values of the soaking time respectively, and represent the minimum and maximum values of the flowback time respectively, and represent the maximum and minimum values of the number of huff and puff cycles respectively.

[0131] ② When huffing and puffing asynchronously, the gas injection time, soaking time, and flowback time at each time step of the three wells are the same, and only the mixed gas ratio, gas injection rate, oil production rate, and number of huff and puff cycles are different:

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] ;

[0137] ;

[0138] ;

[0139] Among them, and represent the minimum and maximum values of the time at each time step respectively, and the gas injection time, soaking time, and flowback time at each time step of the three wells are the same.

[0140] ③ When huffing and puffing in an alternating pattern between wells, During depletion development, there is only one variable, the oil production rate; and have the same huff and puff cycle, and only the mixed gas ratio, gas injection rate, oil production rate, and number of huff and puff cycles are different:

[0141] , ;

[0142] , ;

[0143] ;

[0144] ;

[0145] ;

[0146] ;

[0147] , ;

[0148] Among them, represents the proportion of the injected mixed gas, represents the proportion of the injected mixed gas, represents the gas injection rate of represents the gas injection rate of

[0149] (3) Based on the NSGA-Ⅲ multi-objective optimization algorithm, a dynamic optimization model for the CO 2 mixed gas huff and puff effect is established.

[0150] It couples the TFT prediction surrogate model established in step S2, uses the TFT prediction surrogate model to quickly predict the recovery rate, buried storage volume and NPV. Secondly, it calls the NSGA-Ⅲ multi-objective optimization algorithm to dynamically regulate the huff and puff development method and the huff and puff working system of each well, aiming to ensure that an efficient and reasonable working system can be matched at different time steps, and realize the dynamic optimization of the CO 2 mixed gas flooding-buried storage effect in the tight oil reservoir.

[0151] To implement the NSGA-Ⅲ multi-objective optimization algorithm, the huff and puff parameters are optimized according to the established multi-objective optimization mathematical model of the huff and puff parameters in the tight oil reservoir. Generally speaking, the value range of the crossover coefficient is usually between 0.40 and 0.99, and the value range of the mutation coefficient is usually between 0.0001 and 0.1. The selection of the crossover coefficient and the mutation coefficient takes into account their roles in the algorithm performance. In this embodiment, after hyperparameter tuning, this embodiment selects a crossover coefficient of 0.45, a mutation coefficient of 0.05, a population size of 200, and 300 generations of inheritance as the parameters of NSGA-Ⅲ. The reasonable selection of these parameters helps to efficiently find high-quality optimized solutions for the huff and puff parameters.

[0152] For the selected target well group in this embodiment in a tight oil reservoir, when the recovery factor is relatively large, the buried volume is usually large, while the NPV is usually small. Under relatively high economic benefits, it cannot meet good oil displacement benefits and burial effects, which means there is a contradictory relationship among the three objective functions. In this case, the NSGA-Ⅲ optimization algorithm will generate a set of non-dominated solutions.

[0153] The analysis of the optimization results of this embodiment will be based on the initial design plan of synchronous huff and puff parameters of CO 2 mixed gas huff and puff. Three different huff and puff plans that match the benchmark plan in terms of recovery factor, buried volume, and net present value are selected from the Pareto optimal solution set and used as the "production capacity plan", "burial plan", and "economic plan" respectively. By comparing the different objective function values of each optimized plan with the basic plan, the optimal solution with the same buried volume and NPV as the basic plan is selected as the production capacity plan, the optimal solution with the same recovery factor and NPV as the basic plan is selected as the burial plan, and the optimal solution with the same recovery factor and buried volume as the basic plan is selected as the production capacity plan to generate three sets of optimized plans considering different development objectives. The specific result comparison information is shown in Table 3. Reservoir engineers can select the most suitable optimized plan according to the target focus.

[0154] Table 3 Comparison of calculation results between optimized plan and basic plan

[0155]

[0156] I. Analysis of production capacity plan: Under the same buried volume and NPV, the recovery factor of the production capacity plan is about 3.81% higher than that of the basic plan. The working system of the optimized production capacity plan is shown in Table 4. By coordinating the injection-production huff and puff relationship among different horizontal wells at different production stages in the target well group, each huff and puff well can maximize its own development potential.

[0157] Table 4 Working system of production capacity plan

[0158]

[0159] The changing trends of the average formation pressure and recovery factor over time under the production capacity plan and the basic plan are as Figure 5 and Figure 6 shown. In the initial production stage, both the basic plan and the production capacity plan adopt the synchronous huff and puff method, and the average formation pressure gradually decreases. In the middle and late production stages, the production capacity plan is designed to adopt the asynchronous huff and puff method, effectively raising the formation pressure, thus better supplementing the formation energy and then increasing the crude oil recovery factor.

[0160] II. Analysis of the storage plan: At the same recovery rate and NPV, the storage volume of the storage plan is about 10.15% higher than that of the basic plan. The working system of the optimized storage plan is shown in Table 5. By coordinating the injection-production huff and puff relationship between horizontal wells in the target well group at different production stages, the CO 2 storage effect can be maximized under the premise of meeting the constraints.

[0161] Table 5 Working system of the storage plan

[0162]

[0163] The changing trends of the structural storage volume and the irreducible gas storage volume over time under the storage plan and the basic plan are as Figure 7 and Figure 8 shown. The storage plan design adopts the huff and puff method of alternate wells. Since the middle wells are always depleted during the huff and puff process of the edge wells, the edge wells play a displacement role on the middle wells during the huff and puff process, promoting the flow of crude oil and CO 2 , enabling the accumulation of structural storage and relatively concentrating in the formation. The structural storage volume is higher than that under the simultaneous huff and puff of the basic plan. At the same time, through the adjustment of the injection volume at different production stages, the overall storage volume of the storage plan has been greatly improved.

[0164] III. Analysis of the economic plan: At the same recovery rate and storage volume, the NPV of the economic plan is about 0.38 10 7 yuan higher than that of the basic plan. The working system of the optimized economic plan is shown in Table 6. Through the dynamic optimization of different production stages, the working system of the economic plan is determined, that is, to maximize the NPV under the given conditions, so that the injected gas can have a larger swept area, and the production can be increased while ensuring higher economic benefits.

[0165] Table 6 Working system of the economic plan

[0166]

[0167] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for dynamic optimization of carbon dioxide mixed gas throughput parameters based on multi-objective interaction, characterized in that: The following steps are involved: S1, select the target block for design and collect the time series prediction sample set; S2. Standardize the sample data, divide the data set, use the TFT neural network to build a TFT prediction agent model, and use the TFT prediction agent model to quickly predict the recovery factor, buried storage volume and NPV; S3. Determine the objective function, construct a multi-objective optimization model, and establish a dynamic optimization model for the CO2 mixed gas throughput effect based on the NSGA-Ⅲ multi-objective optimization algorithm to ensure that a reasonable and efficient working system is matched at different time steps, and realize the dynamic optimization of the CO2 mixed gas throughput parameters in tight oil reservoirs; In step S2, the components of the TFT neural network include input features, variable selection network, static covariate encoder, sequence-to-sequence layer, gated residual network, multi-head attention mechanism, time fusion decoder and quantile output; Regarding input features: TFT neural network divides input variables into static variables and dynamic variables. Static variables are divided into static continuous variables and static discrete variables, and dynamic variables include dynamic time-varying variables and dynamic time-invariant variables. Dynamic time-varying variables refer to variables whose characteristics change over time and cannot be inferred, including recovery factor, storage volume and net present value; dynamic time-invariant variables are variables applied to the TFT prediction proxy model based on inference, including month and cumulative gas injection volume; In step S3, the multi-objective optimization model includes three factors: objective function, optimization variables and constraints; The objective function is to maximize recovery, maximize storage volume and maximize NPV; The optimization variables include the throughput development mode and the production system of each horizontal well. The production system includes the mixed gas ratio, gas injection rate, gas injection time, well shut-in time, oil production rate, flowback time and throughput rounds. The throughput well is set to have quantitative injection in the gas injection stage and quantitative production in the flowback stage. Regarding the constraints: the key parameters of each throughput well, including daily gas injection volume and daily oil production, shall not exceed the maximum working capacity of each single well, and shall not be less than zero.

2. The method for dynamic optimization of carbon dioxide mixed gas throughput parameters based on multi-objective interaction according to claim 1, characterized in that: In step S2, the sample data standardization process uses the minimum and maximum standardization method to normalize each feature: Among them, X s is the normalized value, X is the actual value of the characteristic factor in the data set, max(X) is the maximum value of the characteristic factor in the original data, and min(X) is the minimum value of the characteristic factor in the original data; The normalized raw data is divided into training set, validation set and test set according to the ratio of 6:2:2, which are used for training the TFT prediction agent model, tuning the hyperparameters of the TFT neural network and performance evaluation of the TFT prediction agent model respectively.

3. The method for dynamic optimization of carbon dioxide mixed gas throughput parameters based on multi-objective interaction according to claim 2, characterized in that: In step S2, the functions of the variable selection network include: selecting relevant input variables for each time step, and weakening or discarding irrelevant information; The static covariate encoder enables the incorporation of static features into the TFT neural network, regulating temporal dynamics by encoding the context vector; The sequence-to-sequence layer uses two groups of LSTM networks, one of which is used to process past input features and encode past historical inputs; the other is used to process future input features; The gated residual network is a selection framework that uses a gating mechanism to selectively introduce relevant input features through optional skip connections and gated matrix filtering; The TFT prediction agent model learns to capture long-term dependencies at different time steps through a multi-head attention mechanism, while focusing on different parts of the input sequence in parallel and increasing the model's attention weight for specific features; The temporal fusion decoder consists of two layers of gated residual networks and a multi-head attention mechanism layer; the first layer of the gated residual network combines the output information of the sequence-to-sequence layer with the information from the static features, and then performs a standard self-attention mechanism on the input information at all time steps through the multi-head attention mechanism layer. Finally, the output of the multi-head attention mechanism layer and the remaining connections are passed to another layer of the gated residual network for nonlinear processing.

4. The method for dynamic optimization of carbon dioxide mixed gas throughput parameters based on multi-objective interaction according to claim 3 is characterized in that: In step S2, the TFT neural network determines the prediction interval of the possible target value range of each prediction layer through quantile prediction, and generates the prediction interval using quantile regression to obtain information about the prediction certainty; the quantile prediction is generated by the linear transformation of the output of the time fusion decoder; The TFT prediction agent model trains the TFT neural network by jointly minimizing the quantile loss of all quantile outputs: Among them, l is the total loss function, L is the quantile loss function, y t is the actual observed value at the t-th time step, is the model's predicted value for the t-th time step at quantile q, is the estimated value of sample y according to the input features, Ω is the training data containing M samples y, Q is the predicted data set of output quantile q at different time steps t, t=1,...,t max , t max is the maximum value of the time step; Normalized quantile loss is used for validation and testing of the TFT prediction proxy model: Among them, τ is the predicted time span, τ max is the maximum length of the time series, represents the domain value of the test sample, q-Risk represents the normalized quantile loss, Represents the predicted value of the model at time point t-τ for the quantile q at time point τ.

5. The method for dynamic optimization of carbon dioxide mixed gas throughput parameters based on multi-objective interaction according to claim 4, characterized in that: In step S2, the gated residual network is defined as: GRN ω (a,c)=LayerNorm(a+GLU ω (η1)); η1=W 1,ω η2+b 1,ω ; η2=ELU(W 2,ω a+W 3,ω c+b 2,ω ); GLU ω (c)=σ(W 4,ω c+b 4,ω )⊙(W 5,ω c+b 5,ω ); Among them, a is the input variable parameter, c is the static variable vector from the static selection network, GRN ω is a gated residual network with a width multiplication factor ω, ELU is a linear unit activation function, LayerNorm is a standard normalization function, GLU is a gated linear unit, σ is a sigmoid activation function, W 1,ω is the weight matrix of the first transformation layer in GRN, b 1,ω For W 1,ω The corresponding bias vector, W 2,ω is the weight matrix of the second transformation layer in GRN, b 2,ω For W 2,ω The corresponding bias vector, W 3,ω is the weight matrix of the third transformation layer in GRN, b 3,ω For W 3,ω The corresponding bias vector, W 4,ω is the weight matrix of the first linear transformation in GLU, W 5,ω is the weight matrix of the second linear transformation in GLU, b 4,ω For W 4,ω The corresponding bias vector, b 5,ω For W 5,ω The corresponding bias vector, ⊙ represents the element product, ω is the width multiplication factor, η1 is the output after the first transformation layer, η2 is the output after processing by the function ELU, and γ is the input variable parameter of GLU.

6. The method for dynamic optimization of carbon dioxide mixed gas throughput parameters based on multi-objective interaction according to claim 5, characterized in that: There is a correlation between the throughput parameters of each well under different throughput development modes. W2 represents the middle horizontal well, W1 and W3 represent the two side wells, and the constraint condition expression is as follows: When the three wells are in synchronous huff and puff, the huff and puff cycles are the same, with only differences in the gas mixture ratio, gas injection rate, oil production rate, and huff and puff rounds: a min ≤a i ≤a max ; q jmin ≤q ji ≤q jmax ; q pmin ≤q pi ≤q pmax ; t jmin ≤t j1 =t j2 =t j3 ≤t jmax ; t smin ≤t s1 =t s2 =t s3 ≤t smax ; t pmin ≤t p1 =t p2 =t p3 ≤t pmax ; n min ≤n i ≤n max ; Among them, a i represents the ratio of mixed gas injected into the i-th horizontal well, q ji represents the gas injection rate of the i-th horizontal well, q pi represents the oil production rate of the i-th horizontal well, t j1 ,t j2 ,t j3 Respectively represent the gas injection time of W1, W2 and W3, t s1 ,t s2 ,t s3 Respectively represent the soaking time of W1, W2 and W3, t p1 ,t p2 ,t p3 Respectively represent the return time of W1, W2 and W3, n i represents the throughput round of the i-th horizontal well, a min and a max Represent the minimum and maximum value of the mixture ratio, q jmin and q jmax Represent the minimum and maximum values ​​of the gas injection rate, q pmin and q pmax Represent the minimum and maximum oil production rates, t jmin and t jmax Represent the minimum and maximum values ​​of the gas injection time, t smin and t smax Represent the minimum and maximum values ​​of the soaking time, t pmin and t pmax Respectively represent the minimum and maximum value of the return time, n min and n max Represent the maximum and minimum values ​​of the throughput round respectively; When asynchronous huff and puff is used, the gas injection time, well shut-in time and flowback time of the three wells are the same at each time step. Only the mixed gas ratio, gas injection rate, oil production rate and huff and puff rounds are different: a min ≤a i ≤a max ; q jmin ≤q ji ≤q jmax ; q pmin ≤q pi ≤q pmax ; t min ≤t j1 =t j2 =t j3 ≤t max ; t min ≤t s1 =t s2 =t s3 ≤t max ; t min ≤t p1 =t p2 =t p3 ≤t max ; n min ≤n i ≤n max ; Among them, t min and t max Represent the minimum and maximum values ​​of each time step, respectively. The gas injection time, well soaking time and flowback time of the three wells are the same in each time step. When the wells are interspaced, W2 is depleted and developed, and only one variable is the oil production rate. The W1 and W3 have the same throughput cycle, and only the mixed gas ratio, gas injection rate, oil production rate and throughput rounds are different: a min ≤a1,a3≤a max ; q jmin ≤q j1 ,q j3 ≤q jmax ; q pmin ≤q pi ≤q pmax ; t jmin ≤t j1 =t j3 ≤t jmax ; t smin ≤t s1 =t s3 ≤t smax ; t pmin ≤t p1 =t p3 ≤t pmax ; n min ≤n1,n3≤n max ; Among them, a1 represents the proportion of W1 injected into the mixed gas, a3 represents the proportion of W3 injected into the mixed gas, q j1 represents the gas injection speed of W1, q j3 Indicates the gas injection rate of W3.

7. The method for dynamic optimization of carbon dioxide mixed gas throughput parameters based on multi-objective interaction according to claim 6, characterized in that: In step S3, the TFT prediction agent model established in step S2 is coupled, and the TFT prediction agent model is used to quickly predict the recovery factor, storage volume and NPV. Secondly, the NSGA-Ⅲ multi-objective optimization algorithm is called to dynamically control the throughput development mode and the throughput working system of each well to ensure that an efficient and reasonable working system is matched at different time steps, so as to achieve dynamic optimization of the CO2 mixed gas oil recovery and storage effect in tight oil reservoirs.

Images