Horizontal well production dynamic prediction method based on genetic-neural network
Through the genetic-neural network-based method, combined with GA-BP neural network and soft threshold technology, the uncertainty and computational complexity of traditional horizontal well production dynamic prediction methods are solved, and more efficient prediction accuracy and efficiency are achieved. For multi-data samples and multi-dimensional training sets, it is suitable for dense conglomerate reservoirs and unconventional gas reservoirs.
Patent Information
- Application Number
- CN202310641946.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-01
- Publication Date
- 2024-12-03
AI Technical Summary
The traditional dynamic prediction method for horizontal well production has uncertainty, idealized assumptions and computational complexity, making it difficult to accurately predict reservoir characteristics and complex seam geometry, and has limited processing capabilities for early unstable flow stage data and multi-data samples. , resulting in the prediction results deviating from reality and inefficient calculations.
Using a genetic-neural network-based method, by optimizing the GA-BP neural network, using soft threshold technology and random methods to find the optimal solution in multiple global regions, combining geological and fracturing parameters to dynamic prediction of horizontal well production is applicable. For multi-data samples and multi-dimensional training sets, prediction accuracy and efficiency are improved.
Achieve higher prediction accuracy and efficiency, can effectively guide horizontal well production dynamic research, suitable for dense conglomerate reservoirs and unconventional gas reservoirs, replace traditional prediction software, reduce costs, and maintain high correlation in multiple training sessions and low MSE.
Smart Images

Figure CN119067245A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of oil and gas field development, and in particular to a horizontal well production dynamic prediction method based on genetic-neural network. Background Art
[0002] Horizontal wells and volume fracturing are economical and effective development methods. In recent years, with the gradual deepening of unconventional oil and gas exploration and development, the dynamic prediction of horizontal well production and the prediction of EUR effect are not ideal. The traditional capacity prediction method not only has strong uncertainty in the characterization of reservoir characteristics and complex fracture network geometry, but also introduces some idealized assumptions and simplifications in calculation, which may make the capacity prediction results deviate from reality. In addition, when solving the seepage equation, most of them use grid encryption and iterative calculation, which leads to problems such as long modeling cycle and low calculation efficiency. At the same time, the study area is restricted by factors such as reservoir physical properties, fluid properties, and interlayers. The production law of fracturing levels is affected by many factors such as geology, cobalt completion, fracturing construction, and production system. Therefore, too many idealized assumptions and the complexity of actual production factors limit the application and promotion of mechanism-driven capacity prediction methods in the field.
[0003] Traditional production dynamics analysis methods such as Blasingame Fetkovich Arps have many disadvantages:
[0004] 1. Obtain production data for a long enough period of time as analysis support;
[0005] 2. Unable to analyze data from the early unstable flow stage;
[0006] 3. It is impossible to obtain relevant reservoir physical parameters and is only applicable to production data analysis when the bottom hole pressure remains unchanged;
[0007] 4. The reliability of estimating recoverable reserves of oil reservoirs is relatively poor and is very sensitive to early data errors. Summary of the invention
[0008] In order to solve the problems existing in the prior art, the present invention provides a method for dynamic prediction of horizontal well production based on genetic-neural network, which initializes the production sample to obtain the initial threshold and weight of the initial grid, and then selects, crosses, and mutates the influencing factors of horizontal well production such as well spacing, segment cluster combination, permeability, porosity, interlayer, and crude oil viscosity to perform learning training and correct prediction direction, and uses random methods to seek the optimal solution in multiple global regions. The present invention is particularly suitable for large-scale operation processing in multiple data samples and multi-dimensional training sets. While ensuring a high correlation and less MSE, the GA algorithm is used to optimize the initial weights and thresholds of the original neural network. During the operation of the algorithm, multiple trainings are performed to improve the operation accuracy and efficiency.
[0009] The present invention provides a method for predicting the production dynamics of a horizontal well based on a genetic-neural network, comprising the following steps:
[0010] Collect geological parameters and fracturing parameters of each well area, determine the factors affecting horizontal well production, and establish a horizontal well production dynamic prediction network model based on the optimized GA (genetic algorithm)-BP neural network; the optimized GA-BP neural network uses soft threshold technology for model training and prediction;
[0011] The influencing factors of horizontal well production are initialized as samples, and the initial thresholds and weights of the horizontal well production dynamic prediction network model are obtained according to the initialized influencing factors of horizontal well production;
[0012] Genetic algorithm is used to modify the initial threshold and weight of the horizontal well production dynamic prediction network model;
[0013] Perform selection, crossover and mutation operations on the influencing factors of horizontal well production to obtain the processed influencing factors of horizontal well production;
[0014] The processed horizontal well production influencing factors are input into the horizontal well production dynamic prediction network model for training. During the training, a random method is used in multiple global regions to seek the optimal solution to obtain a trained horizontal well production dynamic prediction network model.
[0015] The trained horizontal well production dynamic prediction network model is used to input sample data to predict the horizontal well production dynamics.
[0016] Preferably, the horizontal well production dynamic prediction network model includes a forward propagation process and a backward propagation process. In the forward propagation process, the input pattern is processed layer by layer from the input layer through the hidden unit layer and turned to the output layer. The state of each layer of neurons only affects the state of the next layer of neurons. If the expected output cannot be obtained in the output layer, it turns to backward propagation and returns the error signal in the opposite direction of the forward propagation. By modifying the weights of each neuron, the error signal is minimized.
[0017] Preferably, the geological parameters and fracturing parameters of each well area are analyzed, the fracturing parameters include reservoir porosity, permeability, saturation, fracture length and fracture height, and the geological parameters include rock anisotropy, horizontal well length and well spacing, to determine the factors affecting horizontal well production.
[0018] Preferably, the determined factors affecting horizontal well production include the length of the horizontal section and the well spacing.
[0019] Preferably, a soft thresholding technique is used for training and prediction based on the modified initial threshold.
[0020] Preferably, the horizontal well production dynamic prediction network model adopts an attention mechanism to select effective features that meet preset rules and features that meet preset scales.
[0021] Preferably, after determining the optimized variable parameters and the target fitness function, all real numbers are mapped to the interval (0, 1) to obtain the range of the target function, the upper and lower limits are determined and mapped, and the minimum value of the target function is obtained based on the mapping relationship between the variable parameters and the fitness.
[0022] Preferably, the fitness value is calculated by solving the minimum mean square error (RMSE) of the training set and the test set samples.
[0023] Because the higher the prediction accuracy, the lower the error. After using the genetic algorithm, the smaller the fitness function value, the more accurate the training is, and the better the prediction accuracy of the model is.
[0024] Preferably, the variable parameters are parameters that affect the production of different horizontal sections in each well area, and the objective function is used as the objective fitness function.
[0025] Preferably, the parameters affecting the production of different horizontal sections in each well area include Poisson's ratio, formation pressure, Young's modulus and minimum principal stress value.
[0026] Preferably, after the horizontal well production dynamic prediction network model inputs sample data, it outputs horizontal well production prediction data and optimal solution variable parameters, assigns the optimal solution variable parameters to the weight matrix and threshold vector of the horizontal well production dynamic prediction network model, and obtains the final variable parameters through mutation and crossover.
[0027] Preferably, an improved LSTM network is used as a training layer, and the improved LSTM network only performs prediction on time series data.
[0028] Preferably, multiple training sessions are performed during the training process.
[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0030] (1) The present invention initializes the production samples to obtain the initial thresholds and weights of the initial grid, and then performs operations such as selection, crossover, and mutation on the factors affecting the production of horizontal wells to perform learning and training and the correct prediction direction. A random method is used in multiple global regions to seek the optimal solution. It is particularly suitable for large-scale operation processing in multiple data samples and multi-dimensional training sets. While ensuring a high correlation and a low MSE, the GA algorithm is used to optimize the initial weights and thresholds of the original neural network. During the algorithm operation process, multiple trainings are performed to improve the operation accuracy and efficiency.
[0031] (2) The GABP production performance prediction model based on the variables of the conglomerate reservoir lithology, anisotropy parameters, reservoir parameters, etc., effectively guides the research on the production performance of horizontal wells, thereby optimizing and improving the production effect of horizontal wells. After certain optimization, the present invention is not only applicable to tight conglomerate reservoirs, but can also be extended to unconventional gas reservoirs.
[0032] (3) The present invention fully considers the influence of fracturing parameters and geological parameters, and belongs to the numerical simulation research in the field of conglomerate oil reservoir development, realizes the research and simulation of the unique production laws of conglomerate oil reservoirs, and effectively replaces traditional prediction software to save a lot of costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 A flow chart of a method for predicting horizontal well production dynamics based on a genetic-neural network according to an embodiment of the present invention;
[0034] Figure 2 This is an example diagram of the fitness change curve of the existing genetic neural network;
[0035] Figure 3 This is an example schematic diagram of the traditional BP neural network prediction model. In the figure, w is the weight matrix and b is the bias.
[0036] Figure 4 Schematic diagram of the optimized GA-BP neural network prediction model of the present invention, wherein X1, X2...X L is the data sequence after being processed by the input layer, S1, S2...S L is the input of each LSTM unit in the hidden layer, C0 and S0 are the different state input variables of the first LSTM neuron in the first layer LSTM, C1,1 and S1,1 are the different state output variables of the first LSTM neuron in the first layer LSTM, and so on, C2,L-1 and S2,L-1 are the different state output variables of the L-1th LSTM neuron in the second layer LSTM, p1, p2, ... p L is the input parameter sequence of the output layer.
[0037] Figure 5 A comparison chart of the annual oil production prediction result of a well area with a well spacing of 300 meters obtained by using the method of the present invention and the actual value;
[0038] Figure 6 This is a comparison chart of the annual oil production prediction result of another well area with a well spacing of 300 meters obtained by using the method of the present invention and the actual value. DETAILED DESCRIPTION
[0039] The specific implementation modes of the present invention are described in detail below.
[0040] Reference Figure 1 The present invention provides a method for predicting the production dynamics of a horizontal well based on a genetic-neural network, comprising the following steps:
[0041] Collect geological parameters and fracturing parameters of each well area, determine the factors affecting horizontal well production, and establish a horizontal well production dynamic prediction network model based on the optimized GA-BP neural network; the optimized GA-BP neural network uses soft threshold technology for model training and prediction;
[0042] The influencing factors of horizontal well production are initialized as samples, and the initial thresholds and weights of the horizontal well production dynamic prediction network model are obtained according to the initialized influencing factors of horizontal well production;
[0043] Genetic algorithms are used to modify the initial thresholds and weights of the horizontal well production dynamic prediction network model; the initial weight threshold is optimized, which greatly reduces the difference in training time, convergence, and generalization error.
[0044] Perform selection, crossover and mutation operations on the influencing factors of horizontal well production to obtain the processed influencing factors of horizontal well production;
[0045] The processed horizontal well production influencing factors are input into the horizontal well production dynamic prediction network model for training. During the training, a random method is used in multiple global regions to seek the optimal solution to obtain a trained horizontal well production dynamic prediction network model.
[0046] The trained horizontal well production dynamic prediction network model is used to input sample data to predict the horizontal well production dynamics.
[0047] Usually, prediction ability is also called generalization ability or promotion ability, and training ability is also called approximation ability or learning ability. Generally speaking, when the training ability is poor, the prediction ability is also poor, and to a certain extent, as the training ability improves, the prediction ability will be improved. However, this trend is not fixed, and it has a limit. When this limit is reached, as the training ability improves, the prediction ability will decrease, which is the so-called "overfitting" phenomenon. The optimized GABP neural network model can solve this problem, retaining the training speed under a large number of sample training, reducing the absolute value of the prediction error of the training data, that is, the individual fitness value. The smaller the individual fitness value, the better the individual is, and the accuracy is improved.
[0048] According to a specific embodiment of the present invention, the horizontal well production dynamic prediction network model includes a forward propagation process and a backward propagation process. In the forward propagation process, the input pattern is processed layer by layer from the input layer through the hidden unit layer and turned to the output layer. The state of each layer of neurons only affects the state of the next layer of neurons. If the expected output cannot be obtained in the output layer, it turns to backward propagation and returns the error signal in the opposite direction of the forward propagation. By modifying the weights of each neuron, the error signal is minimized.
[0049] According to a specific embodiment of the present invention, the geological parameters and fracturing parameters of each well area are analyzed, the fracturing parameters include reservoir porosity, permeability, saturation, fracture length and fracture height, and the geological parameters include rock anisotropy, horizontal well length and well spacing, to determine the factors affecting horizontal well production.
[0050] The reservoir parameters of each study area, such as reservoir porosity, permeability, saturation, fracture length, fracture height, rock anisotropy, etc., are collected and brought into the model as influencing factor parameters to find the main controlling factors affecting the dynamics of oil and gas production, implement the main controlling factors of production capacity in each study area, analyze the reservoir lithology and physical properties, and the properties of reservoir fluids to effectively and accurately evaluate the production characteristics of horizontal wells.
[0051] Reservoir porosity, permeability, saturation, fracture length, fracture height, rock anisotropy, horizontal section length and well spacing are input as training samples.
[0052] According to a specific embodiment of the present invention, the determined factors affecting horizontal well production include the length of the horizontal section and the well spacing.
[0053] According to a specific embodiment of the present invention, soft thresholding technology is used for training and prediction based on the modified initial threshold. Robust optimization fully considers uncertainty in the modeling process and describes variables in the form of a set. Compared with stochastic programming and fuzzy programming, robust optimization does not require a distribution model of uncertain parameters and a fuzzy membership function of uncertain parameters. A soft thresholding technology is introduced to improve the robustness of the model under different reservoir parameters, a deep residual shrinkage network is established, and a set of oil and gas production prediction technologies with strong robustness and overcoming network degradation are created for different parameters under different conditions.
[0054] According to a specific implementation of the present invention, the horizontal well production dynamic prediction network model uses an attention mechanism to select effective features that meet preset rules and features that meet preset scales. The attention mechanism is further introduced into the horizontal well production dynamic prediction network model so that the model can complete the task effectively and efficiently, and the original feature set is screened using a stepwise regression analysis method to obtain a high-quality feature subset.
[0055] According to a specific embodiment of the present invention, after determining the optimized variable parameters and the target fitness function, all real numbers are mapped to the interval (0, 1) to obtain the range of the target function, the upper and lower limits are determined and mapped, and the minimum value of the target function is obtained based on the mapping relationship between the variable parameters and the fitness. Because it is not uniformly distributed, it needs to be converted according to the specific situation.
[0056] According to a specific embodiment of the present invention, the variable parameters are parameters that affect the production of different horizontal sections of each well area, and the objective function is used as the objective fitness function.
[0057] According to a specific embodiment of the present invention, the parameters affecting the production of different horizontal sections in each well area include Poisson's ratio, formation pressure, Young's modulus and minimum principal stress value.
[0058] According to a specific implementation scheme of the present invention, after the horizontal well production dynamic prediction network model inputs sample data, it outputs the horizontal well production prediction data and the optimal solution variable parameters, assigns the optimal solution variable parameters to the weight matrix and threshold vector of the horizontal well production dynamic prediction network model, and obtains the final variable parameters through mutation and crossover.
[0059] According to a specific implementation of the present invention, an improved LSTM network is used as a training layer, and the improved LSTM network only performs prediction on time series data.
[0060] Genetic neural networks typically require much more data than traditional machine learning algorithms, at least thousands or even millions of labeled samples. In general, neural networks are more computationally expensive than traditional algorithms. The most advanced deep learning algorithms, which achieve successful training of really deep neural networks, can take weeks to train completely from scratch. Most traditional machine learning algorithms take less than a few minutes to a few hours or days. The computing power required for a neural network depends largely on the size of the data, but also on the depth and complexity of the network.
[0061] The data in the training set are operated in the order of selection, crossover, and mutation. The purpose of selection is to select excellent individuals from the current population so that they have the opportunity to be used as original data to predict future data. The genetic algorithm embodies this idea through the selection process. The principle of selection is the probability selection of individuals with strong adaptability to predict the next data or multiple data. The crossover operation is the most important genetic operation in the genetic algorithm. A new generation of data can be obtained through the crossover operation. The new individual combines the characteristics of its original data. The crossover embodies the idea of information exchange. Mutation first randomly selects an individual in the population, and randomly changes the value of a string in the string structure data with a certain probability for the selected individual. Like the biological world, the probability of mutation in the genetic algorithm is very low, and the value is usually very small. It is suitable for large-scale operation and processing in multiple data samples and multi-dimensional training sets.
[0062] According to a specific embodiment of the present invention, multiple trainings are performed during the training process.
[0063] Example 1
[0064] According to a specific embodiment of the present invention, the horizontal well production dynamic prediction method based on genetic-neural network of the present invention is described in detail below.
[0065] The present invention provides a method for predicting the production dynamics of a horizontal well based on a genetic-neural network, comprising the following steps:
[0066] Collect geological parameters and fracturing parameters of each well area, determine the factors affecting horizontal well production, and establish a horizontal well production dynamic prediction network model based on the optimized GA (genetic algorithm)-BP neural network; the optimized GA-BP neural network uses soft threshold technology for model training and prediction;
[0067] The influencing factors of horizontal well production are initialized as samples, and the initial thresholds and weights of the horizontal well production dynamic prediction network model are obtained according to the initialized influencing factors of horizontal well production;
[0068] Genetic algorithm is used to modify the initial threshold and weight of the horizontal well production dynamic prediction network model;
[0069] Perform selection, crossover and mutation operations on the influencing factors of horizontal well production to obtain the processed influencing factors of horizontal well production;
[0070] The processed horizontal well production influencing factors are input into the horizontal well production dynamic prediction network model for training. During the training, a random method is used in multiple global regions to seek the optimal solution to obtain a trained horizontal well production dynamic prediction network model.
[0071] The trained horizontal well production dynamic prediction network model is used to input sample data to predict the horizontal well production dynamics.
[0072] Example 2
[0073] According to a specific implementation scheme of the present invention, the horizontal well production dynamic prediction method based on genetic-neural network of the present invention is described in detail below by taking the samples of 100-200 days of single wells in different well spacings of 5 production blocks in a certain place as an example.
[0074] The present invention provides a method for predicting the production dynamics of a horizontal well based on a genetic-neural network, comprising the following steps:
[0075] The geological parameters and fracturing parameters of single well samples of 100-200 days at different well spacings in 5 production blocks in a certain area were collected to determine the factors affecting horizontal well production. A horizontal well production dynamic prediction network model was established based on the optimized GA (genetic algorithm)-BP neural network; the optimized GA-BP neural network used soft threshold technology for model training and prediction;
[0076] The influencing factors of horizontal well production are initialized as samples, and the initial thresholds and weights of the horizontal well production dynamic prediction network model are obtained according to the initialized influencing factors of horizontal well production;
[0077] Genetic algorithm is used to modify the initial threshold and weight of the horizontal well production dynamic prediction network model;
[0078] Perform selection, crossover and mutation operations on the influencing factors of horizontal well production to obtain the processed influencing factors of horizontal well production;
[0079] The fitness changes, initial thresholds and weights are obtained by screening, and then selection, crossover, mutation and other operations are performed for learning and training and the correct prediction direction. Random methods are used in multiple global regions to seek the optimal solution.
[0080] The processed horizontal well production influencing factors are input into the horizontal well production dynamic prediction network model for training. During the training, a random method is used in multiple global regions to seek the optimal solution to obtain a trained horizontal well production dynamic prediction network model.
[0081] The trained horizontal well production dynamic prediction network model is used to input sample data to predict the horizontal well production dynamics.
[0082] Among them, the horizontal well production dynamic prediction network model includes forward propagation and backward propagation. In the forward propagation process, the input pattern is processed layer by layer from the input layer to the hidden unit layer and then transferred to the output layer. The state of each layer of neurons only affects the state of the next layer of neurons. If the expected output cannot be obtained in the output layer, it will enter backward propagation and return the error signal in the opposite direction of forward propagation. By modifying the weights of each neuron, the error signal is minimized.
[0083] Among them, the geological parameters and fracturing parameters of each well area are analyzed. The fracturing parameters include reservoir porosity, permeability, saturation, fracture length and fracture height. The geological parameters include rock anisotropy, horizontal well length and well spacing to determine the factors affecting horizontal well production.
[0084] Among them, the factors affecting horizontal well production determined include horizontal section length and well spacing.
[0085] Among them, soft thresholding technology is used for training and prediction based on the corrected initial threshold.
[0086] Among them, the horizontal well production dynamic prediction network model adopts the attention mechanism to select effective features that meet the preset rules and features that meet the preset scale.
[0087] Among them, after determining the optimized variable parameters and the target fitness function, all real numbers are mapped to the interval (0, 1) to obtain the range of the target function, the upper and lower limits are determined and mapped, and the minimum value of the target function is obtained based on the mapping relationship between the variable parameters and the fitness.
[0088] The fitness value is calculated by solving the minimum mean square error (RMSE) of the training set and the test set samples.
[0089] Because the higher the prediction accuracy, the lower the error. After using the genetic algorithm, the smaller the fitness function value, the more accurate the training is, and the better the prediction accuracy of the model is.
[0090] Among them, the variable parameters are the parameters that affect the production of different horizontal sections in each well area, and the objective function is used as the target fitness function.
[0091] Among them, the parameters that affect the production of different horizontal sections in each well area include Poisson's ratio, formation pressure, Young's modulus and minimum principal stress value.
[0092] Among them, after the horizontal well production dynamic prediction network model inputs sample data, it outputs the horizontal well production prediction data and the optimal solution variable parameters, assigns the optimal solution variable parameters to the weight matrix and threshold vector of the horizontal well production dynamic prediction network model, and obtains the final variable parameters through mutation and crossover.
[0093] Among them, the improved LSTM network is used as the training layer, and the improved LSTM network only predicts time series data.
[0094] During the training process, multiple trainings are performed.
[0095] The prediction results show that:
[0096] ① The relationship between the monthly decline rate and production time corresponding to 60 days, 120 days and 180 days calculated by actual production data has a good fit with the power law exponential model, indicating that the decline rate in the decline stage of gas well production shows a power law decline;
[0097] ② On the basis of using the power law exponential model to establish a model for the decline rate, by optimizing the differential calculation method, the constant decline rate is changed to a variable decline rate, and a new simplified decline curve model is proposed;
[0098] ③ Compared with the existing ARPS harmonic decline model, HSIEH decline model, PLE decline model, and DUONG decline model, the new calculation model shows a better fitting effect, but the relative error is in the range of 2% to 4%. The error is too large to guarantee the accuracy of the results.
[0099] The Adam adaptive learning rate optimizer is applied in the GA-BP model and MLSTM (improved LSTM, which only predicts time series data), and the prediction accuracy is compared with the genetic neural network (GA-BP) and the traditional BP. The training results show that the genetic neural network (GA-BP) of the present invention has good prediction accuracy. At the same time, the influence of the hyperparameters of the network model on the prediction accuracy is experimentally studied, and the correction and optimization of the randomly set initial weight threshold greatly reduces the difference caused by the training time, convergence, and generalization error, and pursues the actual and predicted model error of 1%. With the increase of network depth, the prediction accuracy also increases. The proposed method can be applied to the actual production predicted based on the dynamic prediction research of horizontal well production in the Mahu area.
[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.
Claims
1. A method for predicting horizontal well production dynamics based on genetic-neural network. It is characterized in that The steps include: Collect geological parameters and fracturing parameters of each well area, determine the factors affecting horizontal well production, and establish a horizontal well production dynamic prediction network model based on the optimized GA-BP neural network; the optimized GA-BP neural network uses soft threshold technology for model training and prediction; The influencing factors of horizontal well production are initialized as samples, and the initial thresholds and weights of the horizontal well production dynamic prediction network model are obtained according to the initialized influencing factors of horizontal well production; Genetic algorithm is used to modify the initial threshold and weight of the horizontal well production dynamic prediction network model; Perform selection, crossover and mutation operations on the influencing factors of horizontal well production to obtain the processed influencing factors of horizontal well production; The processed horizontal well production influencing factors are input into the horizontal well production dynamic prediction network model for training. During the training, a random method is used in multiple global regions to seek the optimal solution to obtain a trained horizontal well production dynamic prediction network model. The trained horizontal well production dynamic prediction network model is used to input sample data to predict the horizontal well production dynamics.
2. The method for predicting horizontal well production dynamics based on genetic-neural network according to claim 1, It is characterized in that The horizontal well production dynamic prediction network model includes forward propagation and backward propagation. In the forward propagation process, the input pattern is processed layer by layer from the input layer to the hidden unit layer and then transferred to the output layer. The state of each layer of neurons only affects the state of the next layer of neurons. If the expected output cannot be obtained in the output layer, the process switches to back propagation, returning the error signal in the opposite direction of the forward propagation, and minimizing the error signal by modifying the weights of each neuron.
3. The method for predicting horizontal well production dynamics based on genetic-neural network according to claim 1, It is characterized in that Analyze the geological parameters and fracturing parameters of each well area. Fracturing parameters include reservoir porosity, permeability, saturation, fracture length and fracture height. Geological parameters include rock anisotropy, horizontal well length and well spacing to determine the factors affecting horizontal well production.
4. The method for predicting horizontal well production dynamics based on genetic-neural network according to claim 3, It is characterized in that The factors affecting horizontal well production determined include horizontal section length and well spacing.
5. The method for predicting horizontal well production dynamics based on genetic-neural network according to claim 1, It is characterized in that Soft thresholding technique is used for training and prediction based on the corrected initial threshold.
6. The method for predicting horizontal well production dynamics based on genetic-neural network according to claim 1, It is characterized in that The horizontal well production dynamic prediction network model adopts the attention mechanism to select effective features that meet the preset rules and features that meet the preset scale.
7. The method for predicting horizontal well production performance based on genetic-neural network according to claim 1, It is characterized in that After determining the optimized variable parameters and the target fitness function, all real numbers are mapped to the interval (0, 1) to obtain the range of the target function, the upper and lower limits are determined and mapped, and the minimum value of the target function is obtained based on the mapping relationship between the variable parameters and the fitness.
8. The method for predicting horizontal well production dynamics based on genetic-neural network according to claim 1, It is characterized in that The variable parameters are the parameters that affect the production of different horizontal sections in each well area, and the objective function is used as the target fitness function.
9. The method for predicting horizontal well production performance based on genetic-neural network according to claim 8, It is characterized in that The parameters that affect the production of different horizontal sections in each well area include Poisson's ratio, formation pressure, Young's modulus and minimum principal stress value.
10. The method for predicting horizontal well production performance based on genetic-neural network according to claim 9, It is characterized in that After the horizontal well production dynamic prediction network model inputs sample data, it outputs the horizontal well production prediction data and the optimal solution variable parameters, assigns the optimal solution variable parameters to the weight matrix and threshold vector of the horizontal well production dynamic prediction network model, and obtains the final variable parameters through mutation and crossover.
11. The method for predicting horizontal well production performance based on genetic-neural network according to claim 1, It is characterized in that The improved LSTM network is used as the training layer, and the improved LSTM network only predicts time series data.
12. The method for predicting horizontal well production performance based on genetic-neural network according to claim 1, It is characterized in that During the training process, multiple training sessions are performed.
Citation Information
Cited By
Horizontal well fracturing segmentation method based on mechanical specific energy dynamic correction and dynamic planning
CN122190714A
