Shale gas refracturing potential prediction method and system based on multi-modal data fusion

Abstract

The invention relates to the technical field of shale gas storage development and yield increase, in particular to a shale gas refracturing potential prediction method and system based on multi-modal data fusion. A data set is established according to the obtained shale gas well data, the data set comprises a multi-dimensional sequence data set and a single-valued data set, the multi-dimensional sequence data set comprises logging curve data and fracturing construction curve data, and the single-valued data set comprises geological data, reservoir physical property data, initial fracturing parameter data, production dynamic data and rock mechanics data; marking the refracturing potential of the well section through the production dynamic data; performing data cleaning and standardization processing, and dividing a data set into a training set, a test set and a verification set; and constructing a shale gas refracturing potential prediction model which meets a preset requirement and takes the multi-dimensional sequence data set and the single-valued data set as input and the refracturing effect type as output. According to the method, efficient prediction of the overall refracturing potential of the shale gas well can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of shale gas storage development and production increase, and in particular to a shale gas re-fracturing potential prediction method and system based on multi-modal data fusion. Background Art

[0002] The rapid and high-quality development of the shale gas industry is of great significance to the country's energy structure adjustment and energy conservation and emission reduction. Improving the final recovery rate of shale gas is a key research direction of China's oil and gas development.

[0003] As shale gas well development enters the middle and late stages, the conductivity of the first fracturing cracks and the permeability of nearby formations gradually decrease due to factors such as scaling, particle migration and fracture closure, and the shale gas production capacity is greatly reduced, making it difficult to produce oil and gas in some areas. The new cracks generated by repeated fracturing extend in a different direction from the previous artificial cracks, causing the cracks to redirect, thereby connecting the unused gas layers of the old cracks to a larger extent, which can relieve reservoir damage and increase the ultimate recovery rate of the reservoir, becoming a necessary choice for increasing the production of old shale gas wells. In addition, shale gas wells have the characteristics of high initial production and rapid production decline, while the cost of drilling new horizontal wells is relatively high. Therefore, repeated reservoir fracturing is of great significance to achieving shale gas production increase and stability.

[0004] Key technologies for improving oil recovery, such as repeated fracturing, are highly risky and uncertain, with problems such as low construction success rate, poor production increase effect, and short effective period of production increase. The main reasons for these problems include the lack of a systematic method for determining the remaining resource production of gas wells and the reformability of reservoirs, and the difficulty in accurately grasping the location and range of the repeated fracturing "sweet spot" section. Therefore, predicting the overall repeated fracturing potential of shale gas wellbore, optimizing and accurately locating the repeated fracturing "sweet spot" section are key issues that need to be solved in repeated fracturing.

[0005] At present, the prediction of refracturing potential and the selection of refracturing sections are mainly based on manual evaluation combined with microseismic monitoring technology, temperature sensor monitoring, proppant tracking technology, etc., or complex and large data analysis, which has problems such as long time consumption, high economic cost, and poor portability. The application of artificial intelligence prediction methods in the oil and gas field is mostly for resource production prediction, and only single-value data such as geological parameters and construction parameters are used, without considering the changes in each section of the wellbore. There is a lack of an intelligent refracturing pressure potential prediction method that can select and accurately locate the "sweet spot" section of refracturing. Summary of the invention

[0006] The purpose of the present invention is to provide a shale gas refracturing potential prediction method and system based on multimodal data fusion, so as to solve the problems that there is no systematic way to determine the remaining resource production of gas wells and the reformability of reservoirs in the existing refracturing process, and the location and range of the refracturing "sweet spot" section are difficult to accurately grasp, so as to realize intelligent and digital evaluation of the potential of refracturing.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] In a first aspect, the present invention provides a method for predicting shale gas re-fracturing potential based on multimodal data fusion, which comprises:

[0009] A data set is established according to the acquired shale gas well data, wherein the data set includes a multidimensional sequence data set and a single-value data set, wherein the multidimensional sequence data set includes well logging curve data and fracturing operation curve data, and the single-value data set includes geological data, reservoir physical property data, initial fracturing parameter data, production dynamic data, and rock mechanics data; and the repeated fracturing potential of the well section is marked by the production dynamic data;

[0010] Perform data cleaning and standardization, and then divide the data set into training set, test set and validation set;

[0011] Constructing a shale gas re-fracturing potential prediction model with a multidimensional sequence data set and a single-value data set as input and re-fracturing potential as output, wherein the shale gas re-fracturing potential prediction model includes a multi-layer perceptron model for processing the single-value data set and a gated recurrent unit model based on an attention mechanism for processing the multidimensional sequence data set;

[0012] The loss function is set, and the training set is used to iteratively train and optimize the shale gas re-fracturing potential prediction model. The test set and the validation set are used to test and validate the shale gas re-fracturing potential prediction model, respectively, to obtain a shale gas re-fracturing potential prediction model that meets the preset requirements;

[0013] The refracturing potential of a target shale gas well is predicted using the shale gas refracturing potential prediction model that meets preset requirements.

[0014] Furthermore, the logging curve data includes lithology curve, resistivity curve, porosity curve, organic matter content curve, permeability, saturation, pore fluid, and gas content;

[0015] The fracturing construction curve data includes a pressure curve, a displacement curve and a sand ratio curve;

[0016] and / or, the geological data include mineral composition, rock structure, brittleness, fault distribution and fold morphology;

[0017] And / or, the reservoir physical property data includes gas saturation;

[0018] And / or, the initial fracturing parameter data includes the type and amount of fracturing fluid, the type and amount of proppant;

[0019] And / or, the production dynamic data includes shale gas well production data and pressure gradient data;

[0020] And / or, the rock mechanics data include elastic modulus, Poisson's ratio and ground stress data.

[0021] Furthermore, the parameter settings of the shale gas repeated fracturing potential prediction model include: defining the data set as a tensor with a shape of (N, Q, M), N is the number of samples in the data set, Q is the maximum number of time steps among all variables, M is the number of variables for each time step, the number of hidden layers of the multilayer perceptron model is 2, the number of neurons of the gated recurrent unit model is 64, and the ReLU activation function is used.

[0022] Furthermore, the loss function is set, and the training set is used to iteratively train and optimize the shale gas repeated fracturing potential prediction model, specifically including: setting the loss function to categorical_crossentropy; using the K-fold cross validation method as the hyperparameter optimization algorithm, and selecting Adam as the optimizer; setting the learning rate of the shale gas repeated fracturing potential prediction model to 0.01, and iterating the shale gas repeated fracturing potential prediction model until the accuracy reaches the preset threshold.

[0023] Furthermore, the shale gas re-fracturing potential prediction model was tested and verified by accuracy and Kappa coefficient.

[0024] Furthermore, the output results of the multi-layer perceptron model and the output results of the gated recurrent unit model based on the attention mechanism are concatenated and merged through the concatenation layer, and then enter the activation function layer for classification.

[0025] In the second aspect, the present invention provides a shale gas re-fracturing potential prediction system based on multimodal data fusion, which includes a data acquisition module, a data processing module and a model building module; the data acquisition module is used to establish a data set according to the acquired shale gas well data, the data set includes a multidimensional sequence data set and a single-value data set, the multidimensional sequence data set includes logging curve data, the single-value data set includes geological data, reservoir physical property data, initial fracturing parameter data, production dynamic data and rock mechanics data; the re-fracturing potential of the well section is marked by the production dynamic data; the data processing module is used to perform data cleaning and standardization processing, and divide the data set into a training set, a test set and a validation set; the model building module is used to construct A shale gas re-fracturing potential prediction model is established with a multidimensional sequence data set and a single-value data set as input and a re-fracturing effect type as output, wherein the shale gas re-fracturing potential prediction model includes a multi-layer perceptron model for processing the single-value data set and a gated recurrent unit model based on an attention mechanism for processing the multidimensional sequence data set; a loss function is set, and the shale gas re-fracturing potential prediction model is iteratively trained and optimized using a training set, and the shale gas re-fracturing potential prediction model is tested and verified using a test set and a validation set, respectively, to obtain a shale gas re-fracturing potential prediction model that meets preset requirements; and the shale gas re-fracturing potential prediction model that meets preset requirements is used to predict the re-fracturing potential of a target shale gas well.

[0026] The present invention has the following unexpected beneficial effects:

[0027] The shale gas re-fracturing potential prediction model of the present invention uses a multidimensional sequence data set and a single-value data set as input and the re-fracturing effect type as output, and predicts the re-fracturing potential through a deep learning network model, which can greatly improve the prediction efficiency and reduce the time and economic cost required in the manual prediction process. Different from the traditional prediction method, which is greatly affected by the differences in geological and engineering characteristics and cannot be migrated, this prediction model has good portability and good adaptability to mines under different conditions. In addition, deep learning is used to achieve the optimization and positioning of the re-fracturing "sweet spot" section, thereby realizing the prediction of the re-fracturing potential of the shale gas well as a whole, providing support for realizing the intelligentization of the entire re-fracturing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A schematic flow chart of the method for predicting shale gas re-fracturing potential based on multimodal data fusion according to the present invention is shown.

[0029] Figure 2 A schematic diagram of the algorithm flow of the shale gas re-fracturing potential prediction model of the present invention is shown.

[0030] Figure 3 A schematic diagram of the structure of the shale gas refracturing potential prediction system based on multimodal data fusion according to the present invention is shown. DETAILED DESCRIPTION

[0031] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0032] In one embodiment, see Figure 1 As shown, the present invention provides a shale gas re-fracturing potential prediction method based on multimodal data fusion, which includes:

[0033] S1, establishing a data set according to the acquired shale gas well data, wherein the data set includes a multidimensional sequence data set and a single-value data set, wherein the multidimensional sequence data set includes well logging curve data, and the single-value data set includes geological data, reservoir physical property data, initial fracturing parameter data, production performance data, and rock mechanics data; marking the re-fracturing potential of the well section by using the production performance data;

[0034] S2, perform data cleaning and standardization, and then divide the data set into training set, test set and validation set;

[0035] S3, constructing a shale gas re-fracturing potential prediction model with a multidimensional sequence data set and a single-value data set as input and re-fracturing potential as output, wherein the shale gas re-fracturing potential prediction model includes a multi-layer perceptron model for processing the single-value data set and a gated recurrent unit model based on an attention mechanism for processing the multidimensional sequence data set;

[0036] S4, setting a loss function, using a training set to iteratively train and optimize the shale gas re-fracturing potential prediction model, using a test set and a validation set to test and validate the shale gas re-fracturing potential prediction model respectively, to obtain a shale gas re-fracturing potential prediction model that meets preset requirements; using the shale gas re-fracturing potential prediction model that meets preset requirements to predict the re-fracturing potential of a target shale gas well.

[0037] The shale gas re-fracturing potential prediction model of the present invention uses a multidimensional sequence data set and a single-value data set as input and the re-fracturing effect type as output, and predicts the re-fracturing potential through a deep learning network model, which can greatly improve the prediction efficiency and reduce the time and economic cost required in the manual prediction process. Different from the traditional prediction method, which is greatly affected by the differences in geological and engineering characteristics and cannot be migrated, this prediction model has good portability and good adaptability to mines under different conditions. In addition, deep learning is used to achieve the optimization and positioning of the re-fracturing "sweet spot" section, thereby realizing the prediction of the re-fracturing potential of the shale gas well as a whole, providing support for realizing the intelligentization of the entire re-fracturing process.

[0038] In a preferred embodiment, the logging curve data includes lithology curve, resistivity curve, porosity curve, organic matter content curve, permeability, saturation, pore fluid, and gas content.

[0039] The fracturing construction curve data includes a pressure curve, a displacement curve and a sand ratio curve.

[0040] The geological data include mineral composition, rock structure, brittleness, fault distribution and fold morphology.

[0041] The reservoir physical property data includes gas saturation.

[0042] The initial fracturing parameter data include the type and amount of fracturing fluid, and the type and amount of proppant.

[0043] The production dynamic data includes shale gas well production data and pressure gradient data.

[0044] The rock mechanics data include elastic modulus, Poisson's ratio and ground stress data.

[0045] Exemplarily, the method for establishing the data set specifically includes the following steps:

[0046] S1.1, Data acquisition: Acquire multi-source data sets, including well logging data, fracturing operation data, geological data, reservoir physical property data, initial fracturing parameter data, production dynamic data and rock mechanics data.

[0047] S1.2, Data integration: Unify data formats to ensure data integrity and consistency.

[0048] S1.3, Data classification: Classify the data sets according to their usage into multidimensional sequence data sets and single-value data sets.

[0049] S1.4, data marking: according to the production and economic effect of old wells in gas fields after repeated fracturing, shale gas wells are segmented and marked into two types: qualified repeated fracturing effect and unqualified repeated fracturing effect.

[0050] Furthermore, the S2 specifically includes the following steps:

[0051] S2.1, Data cleaning: Check all collected data and remove abnormal data points caused by sensor failure or human error.

[0052] S2.2, data standardization processing: eliminate false information variations caused by dimensional differences, so as to avoid errors in analysis and calculation due to different indicator dimensions and maintain the original valuable information to the greatest extent.

[0053] In this embodiment, the standardization process adopts the Z-score calculation method, and the calculation formula is: Among them, z is the characteristic data after standardization, X is the original data, μ is the mean value of the data, and σ is the standard deviation of the data.

[0054] Data enhancement can be performed by rotating, flipping, adding a small amount of noise, etc. to increase the diversity of training data.

[0055] Step 2.2, data set division: divide the standardized data set into training set, test set and validation set with a ratio of 7:2:1.

[0056] Furthermore, the S3 specifically includes the following steps:

[0057] S3.1, Model architecture: The collected logging curve data and fracturing operation curve data set are taken as multidimensional sequence data sets, and the sequence data set can be regarded as a set of observation data with the same interval. Therefore, the sequence classification problem of each sublayer in the prediction of the fracturing potential of normal pressure shale gas is equivalent to a multidimensional sequence classification problem. Assume that there is a sequence data T = {T1, T2, …, T n}, the length of each sequence is m, that is, there are m observations, that is, T i = <t i1 ,t i2 ,…,t in >, the corresponding category is y i ,y i ∈{1,2,...,C}, where C represents the number of categories of the data set. The single-valued data of each small layer in the shale gas re-evaluation fracturing potential prediction problem is regarded as a classification problem. The model needs to classify both sequence problems and single-valued data. Therefore, the present invention constructs a shale gas re-fracturing potential prediction model including a multi-layer perceptron model and a gated recurrent unit based on an attention mechanism.

[0058] The length of a fracturing well is generally more than a kilometer, and the average length of each section is 80 to 100 meters. In order to distinguish and sort the sections of the fracturing well, it is necessary to solve the problem of long sequence machine learning. The present invention builds a model based on a recurrent neural network, a type of neural network that is specifically used to process sequential sequence data. For a given input sequence X = {X1, X2, ..., X t}, the neural network iteratively calculates the hidden vector h = {h1,h2,…,h t} and the output sequence y = {y1,y2,…,y t}.

[0059] h t =H(W xh x t +W hh h t-1 +b h );

[0060] y t =W hy h t +b y ;

[0061] Where W xh is the weight matrix from input x to hidden layer h, W hh is the weight matrix from the previous hidden layer to the next hidden layer, b is the bias of the corresponding layer, and H is the activation function of the hidden layer.

[0062] Based on the small amount of data in the field of shale gas extraction, the gated recurrent unit can process and predict sequence data better and faster, and is more suitable for solving practical problems of gas wells. The long short-term memory neural network has been improved on the basis of the traditional recurrent neural network, realizing the memory function of the neural network for long-term input information, which is suitable for processing and predicting long sequence data. The gated recurrent unit is an improvement on the long short-term memory neural network, which simplifies the network parameters of the long short-term memory neural network, while ensuring the long-term memory function of the long short-term memory neural network, which can effectively reduce overfitting.

[0063] In addition, in order to better improve the computational efficiency, generalization ability and performance of the deep learning model, the present invention adds an attention mechanism to the traditional gated recurrent unit network. The purpose of adding the attention mechanism is to solve the problem of short-term memory forgetting in the traditional recurrent neural network. The method is based on the context vector V on the output sequence y. The attention mechanism acts on each step of the time series decoding process, which can force the neural network to review the hidden vector corresponding to each instantaneous quantity in the original sequence again, and leave important and useful information, and use this information to enhance the currently used hidden vector.

[0064] In one scenario example, the attention score of a segment is first calculated, that is, the similarity between the decoder state at the previous moment and the encoder output at each moment is calculated.

[0065] e t,i =F(h i ,s t-1 ), where e t,i is the i-th attention score of the decoder to the encoder output at time t, h i is the i-th output of the encoder, s t-1 is the output state of the encoder at time t-1, and F is the transformation function when calculating the attention score.

[0066] Then the attention score is transformed by Softmax to obtain the probability distribution of the attention score. The calculation formula is β t =softmax(e t ), where e t is the attention score vector of the decoder to the encoder at time t, β t for e t Probability distribution of attention scores.

[0067] Next, the attention vector a at time t is calculated using the probability distribution and the states of all encoders t ,

[0068] Finally, the attention vector is combined with the current input of the decoder to form new input information, and then the decoding operation is performed, as shown in the following formula.

[0069] s t =f([c t ,a t ],s t-1 );

[0070] In the formula, c t is the input of the decoder at time t, and f is the transformation function when calculating the decoded input.

[0071] The present invention adds an attention mechanism to the gated recurrent unit model, so that the sequence modeling problem of shale gas wellbore segmentation can break the distance constraint and use a smaller time interval to improve the recognition accuracy of the model without considering the negative impact of the sequence length.

[0072] The evaluation of the refracturing potential of each section can be regarded as a problem of classifying single-value data that affects the refracturing potential, and a multi-layer perceptron model is used. The calculation of each neuron in the perceptron includes three steps:

[0073] The first step is to superimpose one or more weighted input values ​​to obtain a cumulative value.

[0074] In the second step, the activation function is used to perform nonlinear transformation on the accumulated value.

[0075] The third step is to pass the transformed value to the neurons in the next layer.

[0076] In addition to the input and output layers, the multi-layer perceptron model can also have multiple hidden layers in the middle, and all layers are fully connected.

[0077] In the present invention, it is assumed that the current multilayer perceptron has L layers, the first layer is the input layer, the Lth layer is the output layer, and the input data of the first input layer is {x1, x2, ..., x n}, the output data of the Lth layer is {y1,y2,…,y n}, and the rest are hidden layers. Assume that there are n l neurons, the output data of the lth hidden layer is {h1 (l) ,h2 (l) ,…,h n (l)}The calculation formula is as follows.

[0078]

[0079] In the formula, f() is the activation function, o i (l) ——The i-th neuron in the l-th layer, w ij (l) is the weight from the jth neuron in the l-1th layer to the ith neuron in the lth layer, b i (l) ——The bias of the ith neuron in the lth layer.

[0080] The multi-layer perceptron model will learn each neuron in each layer according to the data set, and use the back propagation algorithm to make the learned neural network model better fit the data set. Assume that the pre-processed fracturing potential prediction data is {(x1,y1),(x2,y2),…,(x n ,y n )}, the calculation formula of the back propagation function is as follows.

[0081] In the formula, y k is the true value of the training sample; Output values ​​for the model.

[0082] The weights W and biases B in the multilayer perceptron also need to be updated repeatedly, and the calculation formula is as follows. The iterative update of network weights and biases can enable the multilayer perceptron to complete some nonlinear and complex tasks.

[0083]

[0084]

[0085] Where α is the learning rate, and its value range is (0,1).

[0086] S3.2, parameter setting: Since the problem to be solved by the present invention is the modeling problem of multivariate sequence plus single-valued data, it is necessary to apply the univariate GRU to the multivariate problem. The data set is defined as a tensor with a shape of (N, Q, M), where N is the number of samples in the data set, Q is the maximum number of time steps among all variables, and M is the number of variables for each time step. The gated recurrent unit model accepts a single input per time step, that is, M=1. This model is a multi-classification problem and accepts M inputs per time step.

[0087] Preferably, the attention gated recurrent unit is combined with the multi-layer perceptron, and the two branches will produce different results, which are concatenated and merged through the Concat layer and finally enter the Softmax layer, i.e., the activation function layer, for classification.

[0088] Furthermore, the S4 specifically includes the following steps:

[0089] S4.1. Determination of loss function: Since the present invention studies classification problems, the loss function is categorical_crossentropy.

[0090] S4.2, Optimization algorithm selection: Different hyperparameter combinations need to be verified based on experience. This paper selects 8 parameters that have the greatest impact on performance for optimization based on experience. In order to eliminate the potential overfitting problem caused by the division of training and test sets, the K-fold cross-validation (K=10) method is used to verify the hyperparameters, and the optimizer is selected as Adam.

[0091] S4.3, model iteration: Set the learning rate of the model to 0.01, and continuously iterate the model until the model prediction results meet expectations.

[0092] S4.5. Model evaluation and verification: evaluate and verify the prediction model through accuracy and Kappa coefficient. Model evaluation and verification is an important means to test the reliability and robustness of the model, and is a key factor in model performance.

[0093] Using accuracy to evaluate the model: Accuracy is the most commonly used indicator for evaluating model performance. It refers to the ratio of the number of samples correctly classified for a given data to the total number. The present invention uses accuracy as one of the evaluation indicators of the model.

[0094] The calculation formula of accuracy acc is In the formula, total is the number of all samples, TP j is the number of true positives, TN jis the number of true negative examples.

[0095] Kappa coefficient evaluation: As a statistical method, it is used to evaluate consistency and the accuracy of multi-classification models. In classification problems, consistency refers to whether the model classification results are consistent with the actual classification results. The Kappa coefficient is calculated based on the confusion matrix, and its value is usually between 0 and 1.

[0096]

[0097]

[0098] Where P O is the prediction accuracy, P e For accidental consistency, a ii is the i-th element on the diagonal, N is the total number of samples involved in the classification, a i+ is the number of samples predicted to be of the i-th class, a +i is the number of samples that actually belong to the i-th class.

[0099] The prediction model is evaluated by the above two evaluation indicators, namely accuracy and Kappa coefficient. If the evaluation result is qualified, the construction of the shale gas repeated fracturing potential prediction model is completed; if the evaluation result is unqualified, the accuracy of the database is tested and the parameters and thresholds of the model are adjusted, and the model is re-established before the model evaluation is carried out.

[0100] The number of neurons in the gated recurrent unit model is 64.

[0101] See also Figure 2 As shown, the trained shale gas re-fracturing potential prediction model includes a multi-layer perceptron model for processing single-value data sets, and a gated recurrent unit model based on an attention mechanism for processing multi-dimensional sequence data sets.

[0102] During the training process, a portion of neurons are randomly "dropped" (set to 0 their output) through Dropout regularization. For example, in a hidden layer with 100 neurons, if the dropout rate is set to 0.5, then at each training iteration, the output of about 50 neurons will be set to 0. The main purpose is to prevent the neural network from overfitting. Overfitting refers to a situation where the model performs very well on the training data, but performs poorly on new, unseen data (test data). By randomly dropping neurons, dropout prevents the gated recurrent unit model from over-relying on certain specific neurons, thereby prompting the neural network to learn more robust and generalizable features.

[0103] The number of hidden layers of the multi-layer perceptron model is 2, and the ReLU activation function is adopted.

[0104] The output results of the multi-layer perceptron model and the output results of the gated recurrent unit model based on the attention mechanism are concatenated and merged through the concatenation layer. Then they enter the activation function layer for classification, that is, the concatenated and merged numbers are normalized by the normalized exponential function to obtain the classification results.

[0105] In one embodiment, see Figure 3 As shown, a shale gas re-fracturing potential prediction system based on multimodal data fusion is provided. The prediction system 10 includes a data acquisition module 11, a data processing module 11 and a model building module 12.

[0106] The data acquisition module 11 is used to establish a data set based on the acquired shale gas well data, and the data set includes a multidimensional sequence data set and a single-value data set. The multidimensional sequence data set includes logging curve data, and the single-value data set includes geological data, reservoir physical property data, initial fracturing parameter data, production dynamic data and rock mechanics data; the repeated fracturing potential of the well section is marked by the production dynamic data.

[0107] The data processing module 12 is used to perform data cleaning and standardization, and divide the data set into a training set, a test set and a validation set.

[0108] The model building module 13 is used to build a shale gas re-fracturing potential prediction model with a multidimensional sequence data set and a single-value data set as input and a re-fracturing effect type as output, wherein the shale gas re-fracturing potential prediction model includes a multi-layer perceptron model for processing the single-value data set and a gated recurrent unit model based on an attention mechanism for processing the multidimensional sequence data set; a loss function is set, and the shale gas re-fracturing potential prediction model is iteratively trained and optimized using a training set, and the shale gas re-fracturing potential prediction model is tested and verified using a test set and a validation set, respectively, to obtain a shale gas re-fracturing potential prediction model that meets preset requirements; and the shale gas re-fracturing potential prediction model that meets preset requirements is used to predict the re-fracturing potential of a target shale gas well.

[0109] In order to enable those skilled in the art to more clearly understand the scheme of the present application, the following examples are listed for illustration.

[0110] 1. Obtain basic data of each section of shale gas re-fractured wells.

[0111] The basic parameters of the target shale gas well are collected, see Table 1 and Table 2 for details. The obtained data set includes single-value data and multi-dimensional sequence data, which lays the foundation for the next step of data preprocessing.

[0112] A comparison table of the difference between the expected shale gas production during repeated fracturing of each section of the target well and the actual production after fracturing is compiled. When the ratio of the expected production to the actual production is greater than 90% and the overall return on investment of the project is above 60%, the repeated fracturing effect of the shale gas well is marked as qualified and recorded as 1, otherwise it is marked as unqualified and recorded as 0.

[0113] Table 1 Parameters of each section of shale gas well in specific embodiment

[0114]

[0115] Table 2 Evaluation of historical effects of repeated fracturing in various sections of a specific example shale gas well

[0116]

[0117] 2. Data preprocessing: The collected refracturing potential data of 2147 fracturing sections of 396 shale gas refracturing wells were screened, and samples with missing values ​​were deleted, leaving 1985 valid data. The basic parameters and refracturing potential data were standardized, and the characteristic data were converted.

[0118] 3. Construct a shale gas re-fracturing potential prediction model with multidimensional sequence data sets and single-value data sets as input and re-fracturing potential as output. This model takes into account both single-value data and continuously changing multidimensional sequence data, and can more comprehensively reflect the characteristic changes in the gas production process.

[0119] 4. Model training and optimization: Based on the preprocessed standard data set for production dynamic prediction, a machine learning model is trained and generated. For the 1,985 samples in the training data set, the number of training iterations is set to 300. Each time, 100 samples are extracted from the training data set for training until the entire training data set is traversed.

[0120] During the training process, categorical_crossentropy was used as the loss function and the hyperparameter optimization algorithm was used for iterative training and model optimization of the deep learning model. The hyperparameter optimization results are shown in Table 3, which made the prediction effect of the model more outstanding and stable. Finally, a recurrent neural network model with high computational efficiency and accuracy was obtained.

[0121] Table 3 Hyperparameter optimization results

[0122]

[0123] 5. Model evaluation.

[0124] For the validation data set, the trained shale gas re-fracturing potential prediction model is used to predict the re-fracturing potential corresponding to each input data in the validation data set, and the accuracy and Kappa coefficient are used as evaluation indicators to compare the prediction results with the re-fracturing potential data in the validation data set to analyze and evaluate the accuracy of the machine learning model. If the prediction effect of the network model does not meet expectations, the network structure and hyperparameters (the number of hidden layers of the multi-layer perceptron, the number of neurons in the gated recurrent unit, etc.) are adjusted and the network model is retrained.

[0125] For the test set, the verified model is used for testing, and the model is continuously adjusted until it passes the test, and finally a trained shale gas re-fracturing potential prediction model is obtained.

[0126] Through the prediction of this embodiment, the accurate prediction of the production potential of shale gas wells and the precise positioning of the repeated fracturing "sweet spot" section are achieved. The system is tested on the test set, and the accuracy of the repeated fracturing "sweet spot" section positioning reaches 91.2%, which proves the effectiveness and practicality of this method in the actual prediction of repeated fracturing potential and the positioning of repeated fracturing "sweet spot" section.

[0127] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or change made by a person skilled in the art based on the present invention is within the protection scope of the present invention.

Claims

1. A method for predicting shale gas re-fracturing potential based on multimodal data fusion, characterized in that: include: A data set is established according to the acquired shale gas well data, wherein the data set includes a multidimensional sequence data set and a single-value data set, wherein the multidimensional sequence data set includes well logging curve data and fracturing operation curve data, and the single-value data set includes geological data, reservoir physical property data, initial fracturing parameter data, production dynamic data, and rock mechanics data; and the repeated fracturing potential of the well section is marked by the production dynamic data; Perform data cleaning and standardization, and then divide the data set into training set, test set and validation set; Constructing a shale gas re-fracturing potential prediction model with a multidimensional sequence data set and a single-value data set as input and re-fracturing potential as output, wherein the shale gas re-fracturing potential prediction model includes a multi-layer perceptron model for processing the single-value data set and a gated recurrent unit model based on an attention mechanism for processing the multidimensional sequence data set; The loss function is set, and the training set is used to iteratively train and optimize the shale gas re-fracturing potential prediction model. The test set and the validation set are used to test and validate the shale gas re-fracturing potential prediction model, respectively, to obtain a shale gas re-fracturing potential prediction model that meets the preset requirements; The refracturing potential of a target shale gas well is predicted using the shale gas refracturing potential prediction model that meets preset requirements.

2. The method for predicting shale gas re-fracturing potential based on multimodal data fusion according to claim 1, characterized in that: The logging curve data include lithology curve, resistivity curve, porosity curve, organic matter content curve, permeability, saturation, pore fluid, and gas content; The fracturing construction curve data includes a pressure curve, a displacement curve and a sand ratio curve; and / or, the geological data include mineral composition, rock structure, brittleness, fault distribution and fold morphology; And / or, the reservoir physical property data includes gas saturation; And / or, the initial fracturing parameter data includes the type and amount of fracturing fluid, the type and amount of proppant; And / or, the production dynamic data includes shale gas well production data and pressure gradient data; And / or, the rock mechanics data include elastic modulus, Poisson's ratio and ground stress data.

3. The method for predicting shale gas re-fracturing potential based on multimodal data fusion according to claim 1, characterized in that: The parameter settings of the shale gas repeated fracturing potential prediction model include: defining the data set as a tensor with a shape of (N, Q, M), N is the number of samples in the data set, Q is the maximum number of time steps among all variables, M is the number of variables for each time step, the number of hidden layers of the multilayer perceptron model is 2, the number of neurons of the gated recurrent unit model is 64, and the ReLU activation function is used.

4. The method for predicting shale gas re-fracturing potential based on multimodal data fusion according to claim 1, characterized in that: The loss function is set, and the training set is used to iteratively train and optimize the shale gas re-fracturing potential prediction model, including: Set the loss function to categorical_crossentropy; Use K-fold cross validation method as the hyperparameter optimization algorithm, and select Adam as the optimizer; The learning rate of the shale gas repeated fracturing potential prediction model is set to 0.01, and the shale gas repeated fracturing potential prediction model is iterated until the accuracy reaches the preset threshold.

5. The method for predicting shale gas re-fracturing potential based on multimodal data fusion according to claim 1, characterized in that: The shale gas re-fracturing potential prediction model was tested and verified by accuracy and Kappa coefficient.

6. The method for predicting shale gas re-fracturing potential based on multimodal data fusion according to claim 1, characterized in that: The output results of the multi-layer perceptron model and the output results of the gated recurrent unit model based on the attention mechanism are concatenated and merged through the concatenation layer, and then enter the activation function layer for classification.

7. A shale gas re-fracturing potential prediction system based on multimodal data fusion, characterized in that: include: A data acquisition module is used to establish a data set based on the acquired shale gas well data, wherein the data set includes a multidimensional sequence data set, a single-value data set, and a repeated fracturing effect type data set, wherein the multidimensional sequence data set includes well logging curve data, and the single-value data set includes geological data, reservoir physical property data, initial fracturing parameter data, production dynamic data, and rock mechanics data; The data processing module is used to clean and standardize the data and divide the data set into training set, test set and validation set; A model building module is used to build a shale gas re-fracturing potential prediction model with a multidimensional sequence data set and a single-value data set as input and a re-fracturing effect type as output, wherein the shale gas re-fracturing potential prediction model includes a multi-layer perceptron model for processing the single-value data set and a gated recurrent unit model based on an attention mechanism for processing the multidimensional sequence data set; a loss function is set, a training set is used to iteratively train and optimize the shale gas re-fracturing potential prediction model, and a test set and a validation set are used to test and validate the shale gas re-fracturing potential prediction model respectively, so as to obtain a shale gas re-fracturing potential prediction model that meets preset requirements; and the shale gas re-fracturing potential prediction model that meets preset requirements is used to predict the re-fracturing potential of a target shale gas well.

Images