Shale oil three-dimensional well pattern well site multi-objective optimization method

Abstract

The present application belongs to the technical field of shale oil horizontal well three-dimensional development, and specifically discloses a shale oil three-dimensional well pattern well location multi-objective optimization method, which is used to solve the problem of how to realize the multi-objective collaborative optimization of recovery efficiency, economic benefit and interwell interference degree. The method comprises the following steps: (1) constructing a multi-field coupled dynamic drainage volume characterization model; (2) based on the dynamic drainage volume characterization model, establishing an interwell interference degree quantitative evaluation system; (3) combining a CNN-LSTM deep learning prediction model to realize multi-source data fusion and high-precision prediction of dynamic drainage volume, recovery efficiency, net present value and interwell interference degree; (4) based on the NSGA-III multi-objective optimization algorithm, implementing multi-objective optimization, and comprehensively balancing the recovery efficiency, net present value and interwell interference degree under the premise of meeting the engineering and economic constraints. The present application provides reliable technical support for the decision and deployment of reservoir engineers under complex development conditions.

Description

Technical Field

[0001] This invention belongs to the field of shale oil horizontal well three-dimensional development technology, and particularly relates to a multi-objective optimization method for shale oil three-dimensional well network well locations. Background Technology

[0002] As increasing reserves and production in unconventional reservoirs under deep and complex formation conditions becomes increasingly urgent, the relationship between formation pressure, fracture network, and inter-well disturbance degree is becoming more and more complex. If the characteristics of inter-well disturbance cannot be accurately assessed and the dynamic seepage law cannot be mastered, it is easy to lead to insufficient reservoir utilization or excessively high density of ineffective wells, resulting in problems such as accelerated production decline and poor economic benefits.

[0003] Currently, well network optimization typically relies on numerical simulation methods, which involve establishing geological models and simulating and predicting pressure and seepage fields. While this approach has shown some effectiveness in conventional oilfields, it falls short in addressing the coupling analysis of microseismic fracture monitoring data, heterogeneous reservoirs, and real-time production data in shale reservoirs. Some methods for manually assessing the degree of inter-well interference (such as pressure field reconstruction or fluid saturation tracking) can be time-consuming and labor-intensive, and are insufficient for rapid, large-scale well network optimization.

[0004] In recent years, the rapid development of artificial intelligence has introduced new ideas for reservoir engineering. Combining AI methods with numerical simulation or fracture mechanics models allows for dynamic prediction and evaluation of different well locations and construction parameters in a short time, providing a new technical approach for optimizing multi-well-location, multi-objective well networks. However, balancing reservoir recovery, economic benefits, and inter-well interference in high-dimensional, nonlinear, and multi-constraint optimization problems, while also taking into account geological uncertainties, remains a key challenge that needs to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-objective optimization method for well locations in shale oil three-dimensional well networks, which effectively solves the key technical problems in the development of shale oil reservoir three-dimensional well networks: how to establish an accurate inter-well interference evaluation method under the conditions of strong reservoir heterogeneity and complex inter-well mutual influence, and on this basis, achieve multi-objective synergistic optimization of recovery rate, economic benefits and inter-well interference degree.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A multi-objective optimization method for shale oil three-dimensional well network well locations includes the following steps: S1, constructing a multi-field coupled dynamic oil drainage volume characterization model based on reservoir pressure field evolution, fluid seepage characteristics and fracture network response, to realize the quantitative and dynamic evaluation of inter-well interference in shale oil reservoirs.

[0008] S2. Based on the multi-field coupled dynamic oil drainage volume characterization model, a quantitative evaluation system for inter-well interference is established. The quantitative evaluation system for inter-well interference measures the spatial overlap and pressure field difference of the oil drainage volume of each well, providing a quantitative basis for well network layout and interference control.

[0009] S3. Combining the CNN-LSTM deep learning prediction model, multi-source data fusion and high-precision prediction are achieved for dynamic oil drainage volume, recovery rate, net present value and inter-well disturbance.

[0010] S4. Based on the NSGA-III multi-objective optimization algorithm, multi-objective optimization is implemented to comprehensively balance the recovery rate, net present value and inter-well interference under the premise of satisfying engineering and economic constraints. Through the organic combination of CNN-LSTM deep learning prediction model and multi-objective optimization, the optimal well spacing and layer spacing layout that takes into account the recovery rate, net present value and inter-well interference is finally obtained.

[0011] Furthermore, in step S1, based on the multiphase flow theory, a dynamic oil drainage volume calculation equation is established:

[0012] V leak ′(t)=φ(p)·[S o,l,m,n (t+1)-S o,l,m,n (t)]·V eff (p)·△p eff (t);

[0013] Among them, V leak φ'(t) represents the dynamic oil drainage volume at time t, φ(p) represents the pressure sensitivity function, p represents the formation pressure, and S o,l,m,n V represents the oil phase saturation of the corresponding grid block in the (l,m,n) spatial coordinates. eff (p) represents the effective modified volume as a function of pressure, Δp eff (t) represents the production pressure difference at time t.

[0014] The numerical simulation process of the multi-field coupled dynamic oil drainage volume characterization model is as follows: First, rock physical parameters, including but not limited to porosity, permeability, and elastic modulus, are collected and combined with stress field distribution to build a three-dimensional geomechanical model; second, the fracture network is corrected using microseismic monitoring data and fracturing pump injection data to generate fracture geometry that closely resembles the actual field conditions; then, the dynamic distribution of pressure and saturation and fracture modification volume of different grids are obtained through numerical simulation; finally, the dynamic distribution of porosity, permeability, formation pressure, saturation, and fracture modification volume of different grids are substituted into the dynamic oil drainage volume calculation equation to calculate the spatiotemporal distribution characteristics and evolution law of dynamic oil drainage volume at each time point.

[0015] Furthermore, in step S2, the quantitative evaluation system for inter-well interference includes the overlap of oil drainage volume, the pressure field difference coefficient, and the time weighting function.

[0016] The formula for calculating the overlap of the drain volume is:

[0017] V overlap (t)=V A (t)∩V B (t);

[0018]

[0019] Among them, V overlap V represents the overlapping region of dynamic oil drain volume. A (t) and V B (t) represents the dynamic drainage volume of well A and well B at time t, respectively, and Roverlap represents the ratio of the drainage volume overlap.

[0020] The formula for calculating the pressure field difference coefficient is:

[0021]

[0022] Among them, D pressure This represents the pressure field difference coefficient, where N is the total number of pressure field mesh elements, and p A (k,t) and p B (k,t) represent the pressure values ​​of the k-th grid cell in wells A and B at time t, respectively.

[0023] The formula for calculating the time-weighted function is:

[0024] w(t) = e -αt ;

[0025] Where w represents the time-weighted function and α is the attenuation coefficient.

[0026] Comprehensive evaluation index I of inter-well interference total The calculation formula is:

[0027]

[0028] Where T is the total time step, and β and γ are both weighting coefficients.

[0029] When 0≤I total When the value is less than 0.3, the inter-well interference level is weak; when 0.3 ≤ I total When I < 0.6, the inter-well interference level is medium; when I ≤ 0.6 total When the value is ≤1, the inter-well interference level is strong interference.

[0030] Furthermore, in step S3, the architecture of the CNN-LSTM deep learning prediction model includes a CNN module and an LSTM module.

[0031] The CNN module utilizes a convolutional neural network to process geological parameters through multi-layer convolution and pooling operations, extracting the spatial distribution characteristics of the parameter field. The geological parameters include porosity, permeability, oil saturation, formation pressure, and fracture modification volume.

[0032] The LSTM module uses a long short-term memory network to capture the time dependence of production dynamics for time series data, including historical production and pressure changes. Through a two-layer LSTM structure, it predicts the evolution trend of dynamic drainage volume and inter-well disturbance over time.

[0033] The CNN-LSTM deep learning prediction model deeply integrates the spatial features extracted by the CNN module, the temporal features captured by the LSTM module, and tabular data, making full use of multi-source information to achieve collaborative prediction, and finally predicts the dynamic oil drainage volume, recovery rate, net present value, and inter-well disturbance degree.

[0034] Further, in step S4, a three-dimensional well network optimization model based on the NSGA-III multi-objective optimization algorithm is constructed. The construction method is as follows: S41, determine the objective function: maximize the recovery rate max f1(x), maximize the net present value max f2(x), and minimize the inter-well disturbance degree min f3(x) as objectives.

[0035] S42. Determine decision variables: Using horizontal well spacing and vertical well spacing as decision variables, optimize the variable set x:

[0036] x = {(X1,Y1),(X2,Y2),…,(X...} n ,Y n ),d h ,Δh}};

[0037] Among them, (X) i ,Y i Let d be the coordinates of the i-th well, i = 1, 2, ..., n; h Δh represents the horizontal spacing between adjacent wells; Δh represents the vertical layer spacing between adjacent wells.

[0038] S43. Determine the constraints:

[0039] Due to reservoir and geological constraints, well locations should not be deployed in areas with porosity below 4% or inaccessible areas.

[0040]

[0041] in, To be the minimum permissible horizontal well spacing, For the maximum permissible horizontal well spacing, To be the minimum allowable horizontal well spacing, This represents the maximum permissible horizontal well spacing.

[0042] Well locations and formations should avoid interlayers and fracture zones:

[0043] d h,min ≤d h ≤d h,max ,Δh min ≤Δh≤Δh max ;

[0044] Where, d h,min d is the minimum allowable horizontal well spacing. h,max Δh is the maximum permissible horizontal well spacing. min Δh is the minimum allowable horizontal well spacing. max This represents the maximum permissible horizontal well spacing.

[0045] Economic constraints: Total investment costs must not exceed the budget ceiling.

[0046]

[0047] Among them, Cost(X) i ,Y i ) is at coordinate position (X i ,Y i Total cost of deploying wells at location ) Budget max This represents the maximum allowable investment budget for the project.

[0048] Furthermore, in step S4, the design and solution method of the NSGA-III multi-objective optimization algorithm includes:

[0049] S51. Population initialization: The Latin hypercube sampling method is used to randomly generate initial schemes for well spacing and layer spacing that meet the constraints, covering the design space.

[0050] S52. Fitness Assessment and Reference Point Generation: The CNN-LSTM deep learning prediction model is called to predict the recovery rate, net present value and inter-well interference of each well pattern layout scheme. The fitness of each individual is calculated based on the objective function value, and the individuals are ranked in combination with the effects of well spacing and layer spacing adjustment.

[0051] S53. Solution set evolution and reference point distribution: Generate reference points in the target space of "recovery rate - net present value - inter-well disturbance" to guide the distribution of Pareto front solutions and ensure a reasonable trade-off between well spacing and layer spacing.

[0052] S53. Crossover Mutation and Iterative Update: Using simulated binary crossover and multi-point mutation operations, well spacing and layer spacing are iteratively adjusted to optimize the target values.

[0053] S54. Convergence Judgment: Set the maximum number of iterations or the convergence condition of the Pareto front solution, and output a set of optimal well network layouts that balance the trade-offs between recovery rate, net present value and inter-well disturbance.

[0054] Furthermore, after solving the NSGA-III multi-objective optimization algorithm, a set of Pareto front solutions is obtained. Each solution corresponds to a set of well network deployment and engineering parameter combinations. The methods for analyzing and comparing the optimization results include:

[0055] S61. The optimization results are presented in the form of Pareto fronts in the three-dimensional target space. Each solution in the Pareto front represents a well network optimization scheme. The spatial distribution of the schemes reflects the trade-off between different objectives in multi-objective optimization.

[0056] S62. Select several representative solutions from the Pareto optimal solution set for analysis. Each solution corresponds to a different well pattern layout scheme. The development decision-maker selects the most suitable scheme from the complete Pareto optimal solution set based on actual needs and preferences.

[0057] S63. Sensitivity and Uncertainty Analysis: For the most suitable scheme selected in the end, the impact on recovery rate, net present value and inter-well disturbance is observed by changing the well spacing, layer spacing, number of wells and horizontal section length.

[0058] S64. Field Experiment and Model Iteration: Based on the comparison between monitoring results and predicted values, continuously revise the CNN-LSTM deep learning prediction model and the inter-well interference evaluation parameters to form a closed loop of "numerical simulation-deep learning-multi-objective optimization-field feedback".

[0059] Compared with the prior art, the beneficial technical effects of the present invention are:

[0060] (1) This invention integrates dynamic drainage volume theory with the quantification of inter-well interference: Traditional well pattern design is difficult to comprehensively measure the reservoir utilization level and inter-well pressure interference. This invention, through a dynamic drainage volume characterization model and a comprehensive evaluation system for interference, tightly couples the dynamic seepage characteristics of shale reservoirs with the inter-well competition relationship, which greatly improves the accuracy of interference identification and avoids the economic losses caused by ineffective well density layout.

[0061] (2) This invention introduces a CNN-LSTM deep learning prediction model to improve prediction efficiency and accuracy: it transforms three-dimensional geological attributes, microseismic data and historical production data into trainable samples, integrates spatial feature extraction of the CNN module and temporal feature learning of the LSTM module, and greatly shortens the evaluation time for multi-well layout; compared with pure numerical simulation, it has better consistency and generalization ability.

[0062] (3) The multi-objective optimization framework of the present invention is perfect, taking into account recovery rate, NPV and inter-well interference: The present invention uses the NSGA-III multi-objective optimization algorithm to perform collaborative optimization of the three objectives. It not only focuses on the final reservoir production (recovery rate) and economic benefits (net present value), but also effectively controls the inter-well interference, ensuring the balance between engineering feasibility and economic benefits in the development plan. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the dynamic oil drain volume of the block in Example 1 during the initial stage of production and the current simulation results.

[0064] Figure 2 This is a comparison chart of the predicted values ​​and actual values ​​of the CNN-LSTM deep learning prediction model on different targets in Example 1.

[0065] Figure 3 This is a schematic diagram of the Pareto front, representing the optimization results of Example 1.

[0066] Figure 4 This is a comparison chart of the recovery rates of the basic scheme and the optimized scheme in Example 1.

[0067] Figure 5 This is a comparison chart of the net present value of the basic scheme and the optimized scheme in Example 1.

[0068] Figure 6 This is a comparison diagram of the inter-well interference between the basic scheme and the optimized scheme in Example 1. Detailed Implementation

[0069] Example 1: A block in an oilfield was selected as the research object for this example. This block is a typical shale oil reservoir with the characteristics of "low porosity, low permeability, and rich in organic matter": the average porosity is about 6.9%, and the matrix permeability is 0.433 × 10⁻⁶. -3 μm 2 The original formation pressure was 58.2 MPa, the reservoir depth was 3500–3750 m, and natural fractures were relatively well-developed in the area. The block adopted a three-dimensional development model, achieving efficient utilization through multi-layer system well deployment; the actual well deployment included 5 horizontal production wells, with well trajectories deployed along the dominant stress direction, and horizontal sections with a length of 2000–3000 m. The production wells adopted a constant pressure depletion production method.

[0070] The multi-objective optimization method for shale oil three-dimensional well network locations provided in this embodiment includes the following steps: S1, constructing a multi-field coupled dynamic oil drainage volume characterization model based on reservoir pressure field evolution, fluid seepage characteristics and fracture network response, to realize the quantitative and dynamic evaluation of inter-well interference in shale oil reservoirs.

[0071] (1) Definition and physical meaning of dynamic drain volume (DDV)

[0072] During shale oil reservoir development, the formation pressure and fluid distribution within the reservoir undergo complex evolution due to geological conditions and development dynamics. DDV (Dynamic Drainage Volume) aims to quantify the effective drainage area at different production stages of the reservoir. DDV accurately characterizes the effective drainage area within the reservoir using a multi-field coupled dynamic drainage volume characterization model that integrates the reservoir pressure field, seepage characteristics, and fracture network response, dynamically reflecting the spatiotemporal variations of the effective drainage area. The multi-field coupled dynamic drainage volume characterization model integrates the following core dynamic characteristics: ① Dynamic evolution characteristics of pressure propagation: revealing the pressure diffusion law of the reservoir under development dynamics; ② Seepage law of oil-water two-phase fluids: quantifying the spatial distribution of fluid saturation changes; ③ Dynamic response of the fracture network: dynamically capturing the fracture geometry and its interaction with the pressure field and seepage field.

[0073] By constructing a multi-field coupled dynamic oil drainage volume characterization model, the spatial distribution characteristics of reservoir development benefits can be displayed intuitively and quantitatively, providing a scientific basis for optimizing inter-well parameters and development strategies.

[0074] (2) Achieve high-precision calculation of dynamic oil drainage volume through multi-field coupled dynamic oil drainage volume characterization model: ① Three-dimensional geological model coupling: Construct a three-dimensional reservoir model based on actual geological data, including geological parameters such as porosity, permeability, and distribution of natural fractures; ② Geomechanics and seepage coupling simulation: Use numerical simulation software to simulate the stress field and fracture network of the reservoir after hydraulic fracturing, and calculate the pressure field and seepage field distribution at different time steps.

[0075] Based on multiphase flow theory, an equation for calculating dynamic oil drainage volume is established:

[0076] V leak ′(t)=φ(p)·[S o,l,m,n (t+1)-S o,l,m,n (t)]·V eff (p)·△p eff (t);

[0077] Among them, V leak φ'(t) represents the dynamic oil drainage volume at time t, φ(p) represents the pressure sensitivity function, p represents the formation pressure, and S o,l,m,n V represents the oil phase saturation of the corresponding grid block in the (l,m,n) spatial coordinates.eff (p) represents the effective modified volume as a function of pressure, Δp eff (t) represents the production pressure difference at time t.

[0078] The actual dynamic oil drain volume calculation equation can be modified according to specific numerical simulation scenarios, taking into account factors such as porosity, oil saturation, and crack extension scale.

[0079] The numerical simulation process of the multi-field coupled dynamic oil drainage volume characterization model is as follows: First, collect the rock physical parameters of the block (porosity, permeability, elastic modulus, Poisson's ratio, fracture distribution, etc.) and build a three-dimensional geomechanical model in combination with the stress field distribution; second, use microseismic monitoring data and fracturing pump injection data to correct the fracture network and generate a fracture geometry that is closer to the actual situation on site; then, use numerical simulation to obtain the dynamic distribution of pressure and saturation and the fracture modification volume of different grids; finally, substitute the porosity, permeability, formation pressure, dynamic distribution of saturation and fracture modification volume of different grids into the dynamic oil drainage volume calculation equation to calculate the spatiotemporal distribution characteristics and evolution law of dynamic oil drainage volume at each time point.

[0080] The validation of the multi-field coupled dynamic drainage volume characterization model includes the following steps: verifying the initial fracture geometry and stimulation volume using microseismic monitoring data; and achieving accurate characterization of inter-well disturbance characteristics based on historical fitting of production dynamic data. The validation results show that the multi-field coupled dynamic drainage volume characterization model has a high descriptive ability for the dynamic response characteristics of shale oil reservoirs, with a fitting accuracy exceeding 85%, providing a reliable basis for well pattern optimization.

[0081] The specific steps are as follows: First, the rock mechanical parameters (elastic modulus, Poisson's ratio), porosity, permeability, and fracture distribution of the block are input into the geomechanics and seepage coupling model to obtain the pressure field distribution and fracture geometry at each time step. Second, the fracture network is corrected using microseismic monitoring data and fracturing pump injection data to generate fracture geometry that more closely reflects the actual situation on site. Finally, after completing the coupled iterative calculation at each time step, the effective mobilized volume corresponding to each well in the block is output according to the dynamic oil drainage volume calculation equation, such as... Figure 1 As shown. By comparing the microseismic monitoring results with historical production data, it was confirmed that DDV has high simulation accuracy in the early stage of fracturing and the initial stage after commissioning (error controlled within 10% to 15%).

[0082] S2. Based on the multi-field coupled dynamic oil drainage volume characterization model, a quantitative evaluation system for inter-well interference is established. The quantitative evaluation system for inter-well interference measures the spatial overlap and pressure field difference of the oil drainage volume of each well, providing a quantitative basis for well network layout and interference control.

[0083] In the three-dimensional development of shale oil reservoirs, a reasonable assessment of inter-well interference is a crucial prerequisite for optimizing well network layout. Therefore, this embodiment proposes a quantitative evaluation system for inter-well interference based on a multi-field coupled dynamic drainage volume characterization model to scientifically measure the degree of interference among multiple wells.

[0084] (1) The degree of spatial overlap of the drain volume is an important indicator for measuring the intensity of inter-well interference. The example uses the overlap of the drain volume to quantify the dynamic interference between wells.

[0085] The formula for calculating the overlap of the drain volume is:

[0086] V overlap (t)=V A (t)∩V B (t);

[0087]

[0088] Among them, V overlap V represents the overlapping region of dynamic oil drain volume. A (t) and V B (t) represents the dynamic oil discharge volume of well A and well B at time t, respectively, R overlap This indicates the overlap ratio of the drain volume. The value of the overlap ratio ranges from 0 to 1, with a larger value indicating more significant inter-well interference.

[0089] (2) To further reflect the degree of dynamic interference between wells, this embodiment introduces a pressure field difference coefficient. The larger the pressure field difference coefficient, the more significant the difference in pressure fields between wells, and the greater the intensity of the interference between wells. The formula for calculating the pressure field difference coefficient is:

[0090]

[0091] Among them, D pressure This represents the pressure field difference coefficient, where N is the total number of pressure field mesh elements, and p A (k,t) and p B (k,t) represent the pressure values ​​of the k-th grid cell in wells A and B at time t, respectively.

[0092] (3) Time-weighted overlap calculation: In order to comprehensively reflect the inter-well interference characteristics at different production stages, this embodiment designs a time-weighted function w(t) to dynamically adjust the weights of interference evaluation. The calculation formula is as follows:

[0093] w(t) = e -αt ;

[0094] Where w represents the time-weighted function and α is the attenuation coefficient.

[0095] By combining the overlap of the drain volume and the pressure field difference coefficient, a comprehensive evaluation index I for inter-well disturbance is defined. total The calculation formula is as follows:

[0096]

[0097] Where T is the total time step; β and γ are weighting coefficients used to balance the contributions of the oil leakage volume overlap and the pressure field difference coefficient to the disturbance degree.

[0098] (4) Inter-well interference evaluation index system: Based on the above calculations, this embodiment establishes an inter-well interference evaluation system that includes the overlap of drain volume, pressure field difference coefficient, and time weighting function. By combining these three indicators, the inter-well interference can be divided into three levels, as shown in Table 1.

[0099] Table 1 Classification Standards for Inter-well Interference

[0100]

[0101] In this embodiment, five horizontal wells within the block are selected as the target well group. The DDV distribution of each well at the same time is superimposed and the pressure field difference coefficient is calculated to obtain the comprehensive evaluation index I of the inter-well interference degree. total In the first six months of the simulation period, due to fracture interference and fracturing stimulation, some well groups I... total The interference level reached 0.55 to 0.60, indicating a medium to strong level. With the adjustment of the production system and the reduction of pressure, the interference gradually decreased to about 0.35 after one year of operation, indicating that the interference would be alleviated in time, which is consistent with the on-site observation.

[0102] S3. Combining the CNN-LSTM deep learning prediction model, multi-source data fusion and high-precision prediction are achieved for dynamic oil drainage volume, recovery rate, net present value (NPV), and inter-well disturbance.

[0103] To address the problem of low computational efficiency in well network layout optimization under complex geological conditions, this embodiment proposes a CNN-LSTM deep learning prediction model based on convolutional neural networks (CNN) and long short-term memory networks (LSTM). This model integrates multidimensional geological parameters and production dynamic data, and can quickly improve the collaborative prediction capability of dynamic drainage volume, recovery rate, net present value, and inter-well interference through efficient multidimensional feature extraction and sequence learning.

[0104] (1) Architecture design of CNN-LSTM deep learning prediction model.

[0105] I. CNN module: Spatial feature extraction.

[0106] Convolutional neural networks are used to process geological parameters (such as porosity, permeability, and fracture distribution). Through multi-layer convolution and pooling operations, the spatial distribution characteristics of the parameter field are extracted.

[0107] By converting geological parameter fields (such as permeability fields and oil saturation fields) into RGB three-channel images, the dynamic changes of reservoir heterogeneity and fracture networks are captured, thereby improving the accuracy of reservoir modeling.

[0108] II. LSTM module: Temporal feature learning.

[0109] For historical production and pressure changes in time series data, LSTM is used to capture the time dependence of production dynamics. A two-layer LSTM structure is used to predict the evolution trend of dynamic drainage volume and inter-well disturbance over time.

[0110] III. Multi-source data fusion: The spatial features extracted by the CNN module, the temporal features captured by the LSTM module, and tabular data (such as well location, horizontal section length, etc.) are deeply fused to fully utilize multi-source information and achieve collaborative prediction. The final predictions include dynamic drainage volume, recovery rate, net present value, and inter-well interference. The multi-source data fusion design enables the model to adapt to different geological conditions and engineering parameters, generating targeted prediction results.

[0111] (2) Model Training: In this embodiment, a total of 400 reservoir numerical simulation schemes were designed and generated. Among them, 320 schemes were used to train the basic framework of the CNN-LSTM deep learning prediction model, 40 schemes were used to determine the hyperparameters of the hybrid neural network, and the remaining 40 schemes were used to test the generalization ability of the CNN-LSTM deep learning prediction model. The production time cycle of each simulation scheme is 3 years, and the dynamic adjustment cycle of the single-well working system is set to 2 months, for a total of 18 time steps, so as to fully reflect the dynamic response characteristics of the reservoir.

[0112] The training data includes the following main inputs: ① Geological and fracture information: permeability, porosity, fracture network distribution; ② Production dynamics data: production rate, water cut, bottom hole flowing pressure; ③ DDV time series: dynamic oil seepage volume output at each time step.

[0113] The CNN module is responsible for extracting spatial features (such as formations, fracture meshes, etc.), the LSTM module is responsible for capturing time series dependencies, and finally outputs the comprehensive predicted values ​​of the target parameters (dynamic drainage volume, net present value, recovery rate, and inter-well disturbance).

[0114] The training set was created using the Python programming language and by calling the INTERSECT numerical simulator from Schlumberger. Specifically, the `os.system()` function was first used to execute a pre-defined sequence of command-line instructions to batch process automatically generated `.afi` files. Based on this, sequential Gaussian simulation (SGSIM) was used to generate random geological samples to ensure data coverage and reasonableness.

[0115] (3) Model Validation: To evaluate the accuracy of different deep learning models in predicting reservoir dynamic seepage volume, this embodiment selects the coefficient of determination (R²). 2 The performance of the CNN model, LSTM model, and CNN-LSTM deep learning prediction model was compared and tested using three metrics: mean absolute error (MAE) and root mean square error (RMSE). The results are shown in Table 2.

[0116] Table 2 Performance evaluation of different reservoir prediction models based on deep learning

[0117] CNN model 0.881 0.0391 0.0455 LSTM model 0.853 0.0458 0.0491 CNN-LSTM deep learning prediction model 0.913 0.0248 0.0297

[0118] Test results show that the CNN-LSTM deep learning prediction model performs well in R... 2 The CNN-LSTM deep learning prediction model in this embodiment significantly outperforms individual CNN or LSTM models in terms of MAE and RMSE, indicating that it can more accurately capture the spatiotemporal evolution of dynamic oil seepage volume in reservoirs and achieve high-precision prediction.

[0119] (4) Test set results: The last 40 reservoir numerical simulation schemes were used for testing. The results showed that the average relative error of production prediction was less than 10%; the R-value of inter-well interference prediction was... 2 The accuracy can reach 0.92; the recovery rate and net present value predictions also have high accuracy, such as... Figure 2 As shown.

[0120] S4. Based on the NSGA-III multi-objective optimization algorithm, multi-objective optimization is implemented to comprehensively balance the recovery rate, net present value and inter-well interference under the premise of satisfying engineering and economic constraints. Through the organic combination of CNN-LSTM deep learning prediction model and multi-objective optimization, the optimal well spacing and layer spacing layout that takes into account the recovery rate, net present value and inter-well interference is finally obtained.

[0121] Supported by the CNN-LSTM deep learning prediction model, this embodiment proposes a three-dimensional well network optimization model based on the NSGA-III multi-objective optimization algorithm. This three-dimensional well network optimization model uses well spacing and layer spacing as key optimization parameters, comprehensively considering recovery rate, net present value, and inter-well interference, to achieve three-dimensional optimization of the well network layout in both horizontal and vertical dimensions.

[0122] (1) Construct a three-dimensional well network optimization model based on the NSGA-III multi-objective optimization algorithm. The construction method is as follows:

[0123] I. Determine the objective function.

[0124] Objective 1: Maximize recovery rate (max f1(x): Improve reservoir utilization and optimize development results; Objective 2: Maximize net present value (max f2(x): Optimize economic returns by comprehensively considering production, oil price, investment cost, and discount rate; Objective 3: Minimize inter-well interference (min f3(x): Reduce interference between horizontal wells and vertical inter-layer wells and improve the rationality of well network deployment.

[0125] II. Determine the decision variables.

[0126] Using horizontal well spacing (which affects the degree of inter-well interference and the reservoir utilization range) and vertical well spacing (which reflects the development level and economic rationality of wells in different formations) as decision variables, the set of variables x is optimized:

[0127] x = {(X1,Y1),(X2,Y2),…,(X...} n ,Y n ),d h ,Δh};

[0128] Among them, (X) i ,Y i Let d be the coordinates of the i-th well, i = 1, 2, ..., n; h Δh represents the horizontal spacing between adjacent wells; Δh represents the vertical layer spacing between adjacent wells.

[0129] S43. Determine the constraints:

[0130] Due to reservoir and geological constraints, well locations should not be deployed in areas with porosity below 4% or inaccessible areas.

[0131]

[0132] in, To be the minimum permissible horizontal well spacing, For the maximum permissible horizontal well spacing, To be the minimum allowable horizontal well spacing, This represents the maximum permissible horizontal well spacing.

[0133] Well locations and formations should avoid interlayers and fracture zones:

[0134] d h,min ≤d h ≤d h,max ,Δh min ≤Δh≤Δh max;

[0135] Where, d h,min d is the minimum allowable horizontal well spacing. h,max Δh is the maximum permissible horizontal well spacing. min Δh is the minimum allowable horizontal well spacing. max This represents the maximum permissible horizontal well spacing.

[0136] Economic constraints: Total investment costs must not exceed the budget ceiling.

[0137]

[0138] Among them, Cost(X) i ,Y i ) is at coordinate position (X i ,Y i Total cost of deploying wells at location ) Budget max This represents the maximum allowable investment budget for the project.

[0139] In this embodiment, when developing a multi-well network in three dimensions, we focus on three objectives: recovery rate, net present value, and inter-well interference. The decision variables include well location coordinates, horizontal section length, and injection-production method.

[0140] (2) The design and solution methods of the NSGA-III multi-objective optimization algorithm include:

[0141] I. Population initialization: The Latin hypercube sampling method is used to randomly generate initial schemes for well spacing and layer spacing that meet the constraints, covering the design space.

[0142] II. Fitness Assessment and Reference Point Generation: The CNN-LSTM deep learning prediction model is called to predict the recovery rate, net present value and inter-well interference of each well pattern layout scheme. The fitness of each individual is calculated based on the objective function value, and the individuals are ranked in combination with the effects of well spacing and layer spacing adjustment.

[0143] III. Solution set evolution and reference point distribution: Reference points are generated in the target space of "recovery rate - net present value - inter-well disturbance" to guide the distribution of Pareto front solutions and ensure a reasonable trade-off between well spacing and layer spacing.

[0144] IV. Crossover Mutation and Iterative Update: Using simulated binary crossover and multi-point mutation operations, well spacing and layer spacing are iteratively adjusted to optimize the target values.

[0145] V. Convergence Criteria: Set the maximum number of iterations or the convergence criteria for the Pareto front solution, and output a set of optimal well network layouts that balance the trade-offs between recovery rate, net present value and inter-well disturbance.

[0146] In this embodiment, the crossover variation coefficient was ultimately chosen to be 0.45, the variation coefficient was 0.05, the population size was 200, and the Pareto front solution was output after 300 iterations. Figure 3 As shown.

[0147] (3) After solving the NSGA-III multi-objective optimization algorithm, a set of Pareto front solutions are obtained, and each solution corresponds to a set of well network deployment and engineering parameter combinations.

[0148] Methods for in-depth analysis and comparative evaluation of optimization results include: I. Presenting the optimization results as Pareto fronts in a three-dimensional target space composed of recovery rate, net present value, and inter-well disturbance. Each solution in the Pareto front represents a well network optimization scheme, and the spatial distribution of these schemes reflects the trade-offs between different objectives in multi-objective optimization.

[0149] II. Select several representative solutions from the Pareto optimal solution set for analysis to understand the characteristics of different well pattern layout schemes. Then, development decision-makers can select the most suitable scheme from the complete Pareto optimal solution set based on actual needs and preferences.

[0150] III. Sensitivity and Uncertainty Analysis: For the most suitable scheme selected in the end, the impact on recovery rate, net present value and inter-well disturbance was observed by changing the well spacing, layer spacing, number of wells and horizontal section length.

[0151] IV. Field Experiment and Model Iteration: Based on the comparison between monitoring results and predicted values, the CNN-LSTM deep learning prediction model and the inter-well interference evaluation parameters are continuously corrected to form a closed loop of "numerical simulation-deep learning-multi-objective optimization-field feedback".

[0152] In the Pareto optimal solution set, three typical optimization schemes were selected for analysis: the first is the high-productivity scheme, which increases the recovery rate from 22% in the basic scheme to 28%, an increase of 6%; the second is the high-efficiency scheme, which increases the net present value from 15 × 10⁻⁶. 7 Yuan increased to 20×10 7 Yuan, an increase of 5×10 7 The third option is a low-inter-well interference scheme, which reduces the inter-well interference level from 0.5 to 0.2, a reduction of 0.3. For example... Figure 4 , Figure 5 and Figure 6 As shown, by optimizing the injection-production relationship and well network parameters, the high-production scheme achieves better reservoir utilization, the high-efficiency scheme achieves significant improvement in economics, and the low-inter-well interference scheme ensures development stability by adjusting the injection-production ratio.

[0153] The application of this embodiment demonstrates that it can effectively solve the problem of three-dimensional well network optimization under inter-well interference conditions. Its optimization scheme exhibits excellent performance in terms of improved recovery rate, increased net present value, and reduced inter-well interference, fully proving the high reliability and robustness of the three-dimensional well network optimization model based on the NSGA-III multi-objective optimization algorithm. This provides reliable technical support for reservoir engineers' decision-making and deployment under complex development conditions.

[0154] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A multi-objective optimization method for well locations in a three-dimensional shale oil well network, characterized in that, Includes the following steps: S1. Construct a multi-field coupled dynamic oil drainage volume characterization model based on reservoir pressure field evolution, fluid seepage characteristics and fracture network response to achieve quantitative and dynamic evaluation of inter-well interference in shale oil reservoirs; S2. Based on the multi-field coupled dynamic oil drainage volume characterization model, a quantitative evaluation system for inter-well interference is established. The quantitative evaluation system for inter-well interference is used to measure the spatial overlap and pressure field difference of the oil drainage volume of each well, providing a quantitative basis for well network layout and interference control. S3. Combining CNN-LSTM deep learning prediction models, multi-source data fusion and high-precision prediction are achieved for dynamic oil drainage volume, recovery rate, net present value and inter-well disturbance. S4. Based on the NSGA-III multi-objective optimization algorithm, multi-objective optimization is implemented to comprehensively balance the recovery rate, net present value and inter-well interference under the premise of satisfying engineering and economic constraints. Through the organic combination of CNN-LSTM deep learning prediction model and multi-objective optimization, the optimal well spacing and layer spacing layout that takes into account the recovery rate, net present value and inter-well interference is finally obtained. In step S2, the quantitative evaluation system for inter-well interference includes the overlap of drainage volume, pressure field difference coefficient, and time weighting function; The formula for calculating the overlap of the drain volume is: ； ； in, The overlapping region represents the dynamic oil drain volume. and These represent the times of well A and well B, respectively. The dynamic oil drain volume, Indicates the percentage of overlap in the oil drain volume; The formula for calculating the pressure field difference coefficient is: ； in, This represents the pressure field difference coefficient. This represents the total number of grid cells in the pressure field. and At time respectively, well A and well B are at time... The Pressure value of each grid cell; The formula for calculating the time-weighted function is: ； in, This represents the time-weighted function. The attenuation coefficient; Comprehensive evaluation index of inter-well interference The calculation formula is: ； in, The total time step, and All are weighting coefficients; when At that time, the inter-well interference level was weak; when At that time, the inter-well interference level was medium; when At that time, the inter-well interference level was strong. In step S3, the architecture of the CNN-LSTM deep learning prediction model includes a CNN module and an LSTM module; The CNN module utilizes a convolutional neural network to process geological parameters through multi-layer convolution and pooling operations, extracting the spatial distribution characteristics of the parameter field. These geological parameters include porosity, permeability, oil saturation, formation pressure, and fracture modification volume. The LSTM module uses a long short-term memory network to capture the time dependence of production dynamics for time series data, including historical production and pressure changes. Through a two-layer LSTM structure, it predicts the evolution trend of dynamic drainage volume and inter-well disturbance over time. The CNN-LSTM deep learning prediction model deeply integrates the spatial features extracted by the CNN module, the temporal features captured by the LSTM module, and tabular data, making full use of multi-source information to achieve collaborative prediction, and finally predicts the dynamic oil drainage volume, recovery rate, net present value, and inter-well disturbance degree. In step S4, a three-dimensional well network optimization model based on the NSGA-III multi-objective optimization algorithm is constructed. The construction method is as follows: S41. Determine the objective function: to maximize the recovery rate. Maximize net present value and minimize inter-well disturbance For the goal; S42. Determine decision variables: Using horizontal well spacing and vertical well spacing as decision variables, optimize the variable set. : ； in, For the first Well location coordinates, ; This refers to the horizontal spacing between adjacent wells. This refers to the vertical spacing between adjacent wells. S43. Determine the constraints: Due to reservoir and geological constraints, well locations should not be deployed in areas with porosity below 4% or inaccessible areas. ， ， ； in, To be the minimum permissible horizontal well spacing, For the maximum permissible horizontal well spacing, To be the minimum allowable horizontal well spacing, This is the maximum permissible horizontal well spacing; Well locations and formations should avoid interlayers and fracture zones: ， ； in, To be the minimum permissible horizontal well spacing, For the maximum permissible horizontal well spacing, To be the minimum allowable horizontal well spacing, This is the maximum permissible horizontal well spacing; Economic constraints: Total investment costs must not exceed the budget ceiling. ； in, For coordinate position The total cost of deploying wells at one location This represents the maximum allowable investment budget for the project.

2. The multi-objective optimization method for shale oil three-dimensional well network locations according to claim 1, characterized in that, In step S1, based on the multiphase flow theory, a dynamic oil drainage volume calculation equation is established: ； in, express The dynamic oil drain volume at any given moment. Represents the pressure sensitivity function. Indicates formation pressure. For the corresponding grid block in Oil phase saturation in spatial coordinates This indicates the effective modified volume as a function of pressure. for Production pressure differential at any given moment; The numerical simulation process of the multi-field coupled dynamic oil drainage volume characterization model is as follows: First, rock physical parameters, including porosity, permeability, and elastic modulus, are collected and combined with stress field distribution to build a three-dimensional geomechanical model; second, the fracture network is corrected using microseismic monitoring data and fracturing pump injection data to generate fracture geometry that closely resembles the actual field conditions; then, the dynamic distribution of pressure and saturation and fracture modification volume of different grids are obtained through numerical simulation; finally, the dynamic distribution of porosity, permeability, formation pressure, saturation, and fracture modification volume of different grids are substituted into the dynamic oil drainage volume calculation equation to calculate the spatiotemporal distribution characteristics and evolution law of dynamic oil drainage volume at each time point.

3. The multi-objective optimization method for shale oil three-dimensional well network well locations according to claim 2, characterized in that, In step S4, the design and solution methods of the NSGA-III multi-objective optimization algorithm include: S51. Population initialization: The Latin hypercube sampling method is used to randomly generate initial schemes for well spacing and layer spacing that meet the constraints, covering the design space. S52. Fitness Assessment and Reference Point Generation: The CNN-LSTM deep learning prediction model is called to predict the recovery rate, net present value and inter-well interference of each well pattern layout scheme. The fitness of each individual is calculated based on the objective function value, and the individuals are ranked in combination with the effects of well spacing and layer spacing adjustment. S53. Solution set evolution and reference point distribution: Generate reference points in the target space of "recovery rate - net present value - inter-well disturbance" to guide the distribution of Pareto front solutions and ensure a reasonable trade-off between well spacing and layer spacing. S53, Crossover Mutation and Iterative Update: Using simulated binary crossover and multi-point mutation operations, well spacing and layer spacing are iteratively adjusted to optimize target values; S54. Convergence Judgment: Set the maximum number of iterations or the convergence condition of the Pareto front solution, and output a set of optimal well network layouts that balance the trade-offs between recovery rate, net present value and inter-well disturbance.

4. The multi-objective optimization method for shale oil three-dimensional well network well locations according to claim 3, characterized in that, After solving the NSGA-III multi-objective optimization algorithm, a set of Pareto front solutions is obtained. Each solution corresponds to a set of well network deployment and engineering parameter combinations. The methods for analyzing and comparing the optimization results include: S61. The optimization results are presented in the form of Pareto fronts in the three-dimensional target space. Each solution in the Pareto front represents a well network optimization scheme. The spatial distribution of the schemes reflects the trade-off between different objectives in multi-objective optimization. S62. Select several representative solutions from the Pareto optimal solution set for analysis. Each solution corresponds to a different well network layout scheme. The development decision-maker selects the most suitable scheme from the complete Pareto optimal solution set based on actual needs and preferences. S63. Sensitivity and Uncertainty Analysis: For the most suitable scheme selected in the end, the impact on recovery rate, net present value and inter-well interference is observed by changing the well spacing, layer spacing, number of wells and horizontal section length. S64. Field Experiment and Model Iteration: Based on the comparison between monitoring results and predicted values, continuously revise the CNN-LSTM deep learning prediction model and the inter-well interference evaluation parameters to form a closed loop of "numerical simulation-deep learning-multi-objective optimization-field feedback".

Images