Method for predicting flow conductivity of support fracture based on Radam optimized BP (Back Propagation) neural network
Through the method of optimizing the BP neural network based on RAdam, the problems of high cost, long time and low calculation efficiency of supporting crack diversion capacity prediction in the prior art are solved, and fast and accurate diversion capacity prediction is achieved, with better accuracy and stability.
Patent Information
- Application Number
- CN202510281972.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-11
- Publication Date
- 2025-06-20
AI Technical Summary
In the prior art, the prediction method to support the crack flow diversion capacity has problems such as high cost, long time consuming and low computing efficiency, and it is difficult to predict the flow diversion capacity quickly and effectively.
The method of optimizing BP neural network based on RAdam is adopted, and the experimental data set supporting crack diversion ability is obtained, abnormal data processing and normalization processing is performed, a three-layer BPNN model is constructed, and the L2 regularization, Kaiming initialization and RAdam optimization algorithm are used to improve the stability and generalization ability of the model.
It achieves rapid and accurate prediction of supporting crack flow diversion capabilities, which has better accuracy and stability than traditional BP methods, reducing cost and time-consuming.
Smart Images
Figure CN120180915A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of oil and gas field development engineering, and particularly to a method for predicting the conductivity of propped fractures based on an RAdam-optimized BP neural network. Background Art
[0002] The hydraulic fracturing technology is one of the key technologies for the development of unconventional oil and gas resources, and is widely used in the stimulation and transformation of unconventional reservoirs, which has greatly promoted the economic and effective development of unconventional oil and gas resources. In order to improve the fracturing effect, reduce costs, and increase benefits, it is necessary to design a more efficient fracturing plan. The prediction of the conductivity of propped fractures is the core link of fracture optimization design. When determining the conductivity of propped fractures, there are mainly two types of methods in the prior art: one is the indoor experimental test method, and the other is the method of constructing a physical model for calculation. However, both of these methods have limitations. Among them, the indoor experimental test method has high cost and long time consumption, and the method of constructing a physical model for calculation has problems such as complex conditions, incomplete consideration of influencing factors, and low calculation efficiency. Based on the above research status, there is an urgent need to provide a method that can quickly predict the conductivity of proppants. Summary of the Invention
[0003] The purpose of the present invention is to provide a method for quickly determining the conductivity of propped fractures. To achieve the above purpose, the technical solution adopted by the present invention is as follows:
[0004] A method for predicting the conductivity of propped fractures based on an RAdam-optimized BP neural network, comprising the following technical features:
[0005] S1. Obtain an experimental data set of the conductivity of propped fractures;
[0006] S2. Identify abnormal data in the experimental data set of conductivity obtained in step S1, and perform normalization processing on the data after removing the abnormal data;
[0007] S3. Construct a BPNN model with a three-layer network structure as the basic model for prediction;
[0008] S4. Use a loss function with L2 regularization to prevent model overfitting;
[0009] S5. Use Kaiming initialization to solve the problem of the disappearance of neuron activity of the ReLU activation function in the model;
[0010] S6. Train the BPNN model with the normalized data obtained in step S2, and adjust the training parameters of the model using the RAdam optimization algorithm;
[0011] S7. Evaluate the performance of the trained model, and use the model to predict the conductivity of propped fractures when the performance requirements are met.
[0012] Further, the experimental data set in the step S1 includes proppant particle size, proppant density, closure pressure, proppant concentration, and conductivity.
[0013] Further, 80% of the experimental data set in the step S1 is used as the training data set, and 20% of the experimental data set is used as the test data set, and the training data set is used for model training.
[0014] Further, in the step S2, the isolation forest algorithm is used to eliminate abnormal data.
[0015] Further, in the step S3, the optimized BPNN with a three-layer network structure is set as the input layer, the output layer, and one hidden layer.
[0016] Further, in the initial stage of training in the step S6, SGD momentum optimization is used, and then at a certain moment, it is switched to the improved Adam optimization according to the potential divergence of the variance, and a rectification term is established to make the adaptive momentum slowly and stably fully expressed as a function of the potential variance.
[0017] Further, in the step S6, the mean square error, root mean square error, mean absolute error, and coefficient of determination are used to evaluate the accuracy and stability of the model.
[0018] On the other hand, the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the prediction method for the propped fracture conductivity as described in any one of the above is realized.
[0019] In summary, the present invention proposes a prediction method for propped fracture conductivity, which realizes the rapid prediction of conductivity by using the RAdam-optimized BP neural network, and the prediction method in the present invention has better accuracy and stability compared with the traditional BP method. Description of the Drawings
[0020] Figure 1 It is a box plot of the abnormal processing of the propped fracture conductivity data of two types A and B applied in the present invention before processing;
[0021] Figure 2 It is a box plot of the abnormal processing of the propped fracture conductivity data of two types A and B applied in the present invention after processing;
[0022] Figure 3 It is a schematic diagram of the construction of the optimized BPNN neural network and its parameters in the present invention;
[0023] Figure 4 It is a schematic diagram of the training process of the RAdam-BP model in the present invention;
[0024] Figure 5 The validation loss curve of the RAdam-BP model of the present invention during training and testing on Class A data;
[0025] Figure 6 The validation loss curve of the RAdam-BP model of the present invention during training and testing on Class B data;
[0026] Figure 7 The schematic diagram of the prediction effect of the RAdam-BP model of the present invention on Class A data using the test data set;
[0027] Figure 8 The schematic diagram of the prediction effect of the RAdam-BP model of the present invention on Class B data using the test data set. Detailed implementation manners
[0028] The present invention will be described in detail below with reference to the accompanying drawings.
[0029] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0030] The objective of the present invention is to provide a method for predicting the propped fracture conductivity based on the RAdam-optimized BP neural network. In order to make the above objectives, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners.
[0031] The prediction of propped fracture conductivity is the core link of fracture optimization design. Currently, it is mainly divided into two types of methods: one is the indoor experimental test method, which mainly predicts the propped fracture conductivity by simulating actual mining conditions and evaluates the influence of different factors on the conductivity; the other is the calculation method of constructing a physical model, which describes the fluid flow of the propped fracture and the embedding deformation of the proppant by establishing a mathematical model, and then predicts the conductivity, or uses the numerical simulation method to simulate the complex mining conditions and the influence of various factors on the conductivity. However, both of these methods have limitations. The experimental test method for propped fracture conductivity consumes a large amount of cost, and the physical model has defects such as complex operation and incomplete consideration of influencing factors, resulting in low calculation efficiency.
[0032] In view of the problems existing in the above two methods, the present invention uses an outlier data processing method based on the Isolation Forest algorithm and a BP neural network prediction model using the RAdam optimization algorithm, L2 regularization constraint, and Kaiming initialization method to more conveniently and quickly predict the propped fracture conductivity. RAdam combines the advantages of Adam and Stochastic Gradient Descent (SGD), ensuring both fast convergence and preventing it from easily falling into local optima at the beginning of training, thus improving the stability of model training. Secondly, the L2 regularization constraint can prevent model overfitting and improve the generalization ability of the model. Finally, to address the problem that neuron activity may disappear with the ReLU activation function, the present invention adopts the Kaiming initialization method to solve it.
[0033] The method steps of the present invention are as follows:
[0034] S1. Obtain an experimental data set of propped fracture conductivity, where the experimental data set includes parameters such as proppant particle size, proppant density, closure pressure, proppant concentration, and conductivity.
[0035] When obtaining the experimental data set of propped fracture conductivity for the target block, it can be divided into a training data set and a test data set according to a certain ratio.
[0036] S2. Identify the outlier data in the experimental data set of conductivity obtained in step S1, and perform normalization processing on the data after removing the outlier data.
[0037] Due to reasons such as equipment abnormalities and human errors, data quality problems such as data anomalies or data invalidity will inevitably occur in the experimental data of fracture conductivity. To improve data quality, the Isolation Forest algorithm is used to process the outlier values of the collected sample data, and the abnormal samples are removed.
[0038] The Isolation Forest algorithm is a fast outlier detection method based on Ensemble, with linear time complexity and high accuracy. Its outlier detection process is as follows
[0039] ① For each sample, randomly select a feature and a splitting point for splitting, and repeat the splitting process until the preset depth is reached (for example, the depth of the tree is less than the logarithm of the number of samples) or all features have been tried. The sample finally reaches the leaf node, and the path length is defined as the number of edges from the root node to the leaf node.
[0040] ② For each sample in the training set, calculate its path length in the decision tree. The longer the path length, the more likely the sample is an outlier.
[0041] ③ Divide the path length of the test sample by the average path length of the training set to obtain the anomaly score of the sample.
[0042] ④ Set a threshold. If the anomaly score of the test sample is lower than this threshold, it is determined as an outlier.
[0043] Among them, Figure 1-2 are the box plots before and after the abnormal data processing of the A and B types of support fracture conductivity data applied in the embodiments of the present invention, respectively. It can be seen that most of the abnormal data can be removed by using the Isolation Forest algorithm, thus providing reliable data for subsequent model training.
[0044] Since the orders of magnitude and dimensions of different parameters are different and their value ranges are large, it affects the accuracy and solution speed of BP neural network modeling. Using min-max normalization processing, this method can map the value range of each parameter to 0 and 1 and retain the relationship between the original data. The formula is as follows:
[0045]
[0046] Among them, x is the sample data, x min is the sample minimum value, x max is the sample maximum value, and x' is the normalization result.
[0047] S3. Construct a BPNN model with a three-layer network structure as the basic model for prediction;
[0048] The Back Propagation Neural Network (BPNN) is a multi-layer feedforward neural network based on the error backpropagation algorithm (error Back Propagation, abbreviated as BP).
[0049] Given the training set as D = {(x1, y1), (x2, y2),..., (x m , y m )}, x i ∈R d , y i = R l Taking Figure 3 as an example, it is a multi-layer feedforward network structure with d attribute descriptions, l output neurons, and q hidden layer neurons.
[0050] Among them, the threshold of the j-th neuron in the output layer is represented by θ j , and the threshold of the h-th neuron in the hidden layer is represented by γ h . The connection weight between the i-th neuron in the input layer and the h-th neuron in the hidden layer is vi h, the connection weight between the h neurons in the hidden layer and the j-th neuron in the output layer is w hj .
[0051] The input received by the h-th neuron in the hidden layer is:
[0052]
[0053] The input received by the j-th neuron in the output layer is:
[0054]
[0055] where b h is the h-th neuron in the hidden layer.
[0056] For the training example (x k , y k ), the output of the neural network is:
[0057]
[0058] Then the mean squared error of the network on (x k , y k ), that is, the loss function is:
[0059]
[0060] There are (d + l + 1)q + l parameters to be determined in the network: d×q weights from the input layer to the hidden layer, q×l weights from the hidden layer to the output layer, q thresholds of the hidden layer neurons, and l thresholds of the output layer neurons. BP is an iterative learning algorithm that updates and estimates the parameters using the generalized perceptron learning rule in each round of iteration.
[0061] The formula for the BP algorithm to adjust the weights is as follows:
[0062] Δw hj = ηg j b h (6)
[0063] Δθ j = -ηg j (7)
[0064] Δv ih = ηe h x i (8)
[0065] Δγ h = -ηe h (9)
[0066] where η is the fixed learning rate.
[0067] Select ReLU as the activation function for the output of the hidden layer neurons. The ReLU function is represented by the following formula:
[0068] ReLU(x) = max(0, x) (10)
[0069] S4. Use the loss function with L2 regularization to prevent the model from overfitting and improve the generalization ability of the model;
[0070] In the loss function with L2 regularization, in addition to the original loss term mean squared error, there is an additional L2 norm regularization term, which is usually expressed as the sum of the squares of the weights. This regularization term is combined with the original loss function to form a new loss function.
[0071] The loss function with L2 regularization is usually expressed as:
[0072]
[0073] L(x, y) is the total loss function with the regularization term added; λ is the regularization parameter used to control the importance of the regularization term, and it is a non - negative real number; n is the number of weight parameters in the model; wi is the i - th weight parameter.
[0074] S5. Use Kaiming initialization to solve the problem of neuron activity disappearance that may occur in the ReLU activation function in the model;
[0075] Kaiming initialization follows the Glorot condition, that is, the variance of the data received by each layer during forward propagation is required to be consistent, and the variance of the parameter gradients of each layer during backpropagation is required to be consistent. Kaiming initialization stipulates that the parameters are random variables with a mean of 0, and uniform distribution and Gaussian distribution are used to create random variables.
[0076] S6. Use the data obtained in step S2 to train the BPNN model, and use the RAdam optimization algorithm as the method for adjusting the training parameters of the model;
[0077] The RAdam optimization algorithm optimizes the objective function. That is, in the initial stage of training, SGD momentum optimization is used, and then at a certain moment, it switches to the improved Adam optimization according to the potential divergence of the variance. And a rectification term is established so that the adaptive momentum can be slowly and stably fully expressed as a function of the potential variance, improving the stability of model training.
[0078] Specifically, as Figure 4 shown in the schematic diagram of the model training process of the present invention, the steps for the RAdam optimization algorithm to solve the objective function are as follows:
[0079] ① Input the learning rate α t, decay rates {β1, β2}, t = 0
[0080] ② Initialize the first - moment estimate w of the gradient t (mean) and the second - moment estimate v t (uncentered variance), and calculate the maximum length ρ of the approximate moving average ∞ , the formula is as follows:
[0081] ρ ∞ = 2 / (1 - β2)-1 (12)
[0082] ③ t = t + 1, calculate the gradient g of the objective function t , update w t and v t , the corrected first - moment estimate is and calculate the maximum length ρ of the approximate moving average t , the formula is as follows:
[0083]
[0084] w t = β1w t-1 +(1 - β1)g t (15)
[0085]
[0086] where represents the gradient solution, J t (θ t-1 ) is the objective function, and θ t-1 is the model parameter at t - 1.
[0087] ④ According to ρ t calculate θ t , if ρ t > 4, use the Adam optimizer, correct the second - moment estimate, and construct the rectification term r t ; obtain the corrected second - moment estimate and the model parameter θ t , the formula is as follows:
[0088]
[0089] If ρ t ≤4, use the SGD + Momentum optimizer to obtain the training parameter θ t .
[0090] ⑤ Output the model parameter θ t .
[0091] S7. Use an evaluation method to assess the performance of the model, and use the model to predict the propped fracture conductivity when the performance requirements are met.
[0092] The present invention adopts the following four model evaluation indicators: mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ). These four indicators evaluate the accuracy and stability of the model from different perspectives.
[0093] MSE (mean squared error) measures the squared difference between the predicted value and the actual value, reflecting the volatility of the model prediction. The formula is as follows:
[0094]
[0095] RMSE (root mean squared error) is the square root of MSE, which can more intuitively present the magnitude of the model error. The formula is as follows:
[0096]
[0097] MAE (mean absolute error) focuses on the translation property of the error of the model, that is, the average difference between the model predicted value and the actual value. The formula is as follows:
[0098]
[0099] The above three indicators comprehensively reflect the bias and variance of the model. At the same time, R 2 is used as this indicator to measure the fitting degree of the model. The larger R 2 , the higher the fitting degree of the model to the data, that is, the stronger the explanatory ability of the model for the target variable. The formula is as follows:
[0100]
[0101] In a calculation example, as Figure 5 、 7 shows, under type A data, the loss values of the optimized model in the training and prediction processes, and the final prediction effect are shown. The MSE of the model is 3.09, the RMSE is 1.76, the MAE is 1.37, and R 2 is 0.98, and the average prediction error rate is 2.14%, and the prediction effect is good.
[0102] In a calculation example, as Figure 6 、 8 shows, under type B data, the loss values of the optimized model in the training and prediction processes, and the final prediction effect are shown. The MSE of the model is 177.68, the RMSE is 13.33, the MAE is 10.5, and R 2It is 0.94, the predicted average error rate is 11.72%, and the prediction effect is good.
[0103] For the prediction of the propped fracture conductivity of the present invention, the Isolation Forest algorithm is first used to identify and remove abnormal data. Secondly, the RAdam optimization algorithm is used to optimize the model parameter adjustment method. The RAdam algorithm combines the advantages of Adam and Stochastic Gradient Descent (SGD), which not only ensures the rapid convergence of the model but also makes it not easy to fall into the local optimal solution in the initial stage of training, thus improving the stability of model training.
[0104] To prevent model overfitting and improve the generalization ability of the model, a loss function with L2 regularization is introduced. Aiming at the problem of neuron activity disappearance that may occur in the ReLU activation function, the Kaiming initialization method is adopted to solve it.
[0105] To effectively evaluate the prediction effect of the model, 80% of the total data is randomly selected as the training set, and the remaining 20% is used as the test set. Evaluating on the prediction set is a key step in evaluating the model performance. The present invention uses four evaluation indicators, namely the Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R 2 ) to evaluate the prediction effect of the model. When the evaluation using the above evaluation indicators meets the performance requirements, the model can be used to quickly predict the conductivity of the propped fracture.
[0106] The present invention first collects the experimental data of the target block, preprocesses the data using the Isolation Forest algorithm and the normalization method, divides the preprocessed data set, inputs the training data set into the RAdam-BP model for training, and the test data set is used to evaluate the model performance. When the model performance is good, the model can be applied to predict the conductivity of this block to help the construction personnel optimize the fracturing design.
[0107] In this article, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.
Claims
1. A method for predicting propped fracture conductivity based on RAdam optimized BP neural network, characterized in that: The following technical features are included: S1. Obtaining experimental data set of propped fracture conductivity; S2, identifying abnormal data in the flow conductivity experimental data set obtained in step S1, and normalizing the data after the abnormal data is removed; S3, construct a BPNN model with a three-layer network structure as the basic model for prediction; S4. Use L2 regularized loss function to prevent model overfitting; S5. Use Kaiming initialization to solve the problem of disappearance of neuron activity of ReLU activation function in the model; S6, training the BPNN model using the normalized data obtained in step S2, and adjusting the training parameters of the model using the RAdam optimization algorithm; S7. Evaluate the performance of the trained model and use the model to predict the propped fracture conductivity when the performance requirements are met.
2. The method for predicting propped fracture conductivity according to claim 1, wherein the experimental data set in step S1 comprises proppant particle size, proppant density, closure pressure, proppant concentration, and conductivity.
3. The propped fracture conductivity prediction method according to claim 1, wherein in step S1, 80% of the experimental data set is used as a training data set, and 20% of the experimental data set is used as a test data set, and the training data set is used for model training.
4. The propped fracture conductivity prediction method according to claim 1, wherein in step S2, an isolation forest algorithm is used to eliminate abnormal data.
5. The propped fracture conductivity prediction method according to claim 1, wherein the optimized BPNN of the three-layer network structure in step S3 is set to an input layer, an output layer, and a hidden layer.
6. The propped fracture conductivity prediction method according to claim 1, wherein in step S6, SGD momentum optimization is used in the initial stage of training, and then at a certain moment, the improved Adam optimization is switched according to the potential divergence of the variance, and a rectification term is established so that the adaptive momentum is slowly and steadily fully expressed as a function of the potential variance.
7. The propped fracture conductivity prediction method according to claim 1, wherein in step S6, mean square error, root mean square error, mean absolute error and determination coefficient are used to evaluate the accuracy and stability of the model.
8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for predicting the conductivity of propped fractures as claimed in any one of claims 1 to 7 is implemented.