A Smart Splitting Method for Dynamic Production of Unconventional Reservoir Fracturing Wells

Abstract

This invention relates to the field of unconventional reservoir development technology, and particularly to an intelligent method for dynamic production segmentation of fractured wells in unconventional reservoirs. The method includes: collecting parameters; processing outliers and standardizing the parameters; constructing a deep neural network dynamic production prediction model using particle swarm optimization to optimize hyperparameters; introducing fractured segment or stratigraphic category embeddings as conditional variables based on a conditional domain adversarial network; aligning the joint distribution of labeled and unlabeled samples on features and stratigraphic categories through adversarial training of a feature extractor and a domain discriminator; applying a total production constraint; quantifying the uncertainty of the production segmentation results using the Monte Carlo method; calibrating the uncertainty threshold based on the cross-validation results of labeled samples; adding unlabeled well segmentation results below the threshold as high-confidence pseudo-labeled samples to the training set; and retraining the model until completion. This invention provides a high-precision and highly generalizable technical solution for production segmentation in unconventional reservoirs.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development engineering technology, and in particular to an intelligent splitting method for dynamic production of unconventional reservoir fracturing wells. Background Technology

[0002] Unconventional reservoirs often exhibit "vertical multi-layered" or "thin interbedded" development characteristics, with rapid lateral lithological changes, resulting in extremely strong macroscopic heterogeneity. Such reservoirs require directional or horizontal well drilling combined with multi-stage fracturing technology to form a complex fracture network that effectively connects multiple gas- or oil-bearing thin layers, thereby achieving efficient resource utilization. In practical engineering, a single well often performs multi-stage fracturing and production simultaneously across multiple layers, only obtaining the overall total production, not the individual production of each fracturing segment or layer. This creates a typical "big data, small sample" dilemma: few samples with production labels, and many samples without production labels but rich in parameters. Accurately allocating the total production to each fracturing segment or layer (i.e., dynamic production allocation) is crucial for evaluating fracturing effectiveness, predicting new well productivity, optimizing subsequent fracturing parameters, and adjusting development strategies.

[0003] Currently, there are three main methods for determining the production contribution of each fracturing section or layer: The first is the allocation coefficient method, which allocates production based on static parameters such as sand body thickness, permeability, porosity, and gas saturation. This method is simple to operate, but it completely ignores the influence of fracturing engineering parameters and the dynamic changes in the production regime, and the physical mechanism is incomplete. The second is the production profile testing method, which can provide relatively accurate production data for each section, but the testing cost is high and the cycle is long, making it difficult to apply in large-scale well groups. The third is the numerical simulation method, which relies on limited geological parameter input and a large number of simplified assumptions. When the production of each well varies greatly, the prediction accuracy is not high, and the calculation is large and time-consuming. When facing a large number of multi-stage fracturing wells or multi-layer production wells, the above methods all have problems of insufficient accuracy, poor operability, or insufficient economy, making it difficult to meet the actual needs of large-scale production allocation.

[0004] In recent years, data-driven machine learning methods have provided new technical means for production prediction. Deep neural networks have shown significant advantages in establishing nonlinear mapping relationships between geological and engineering parameters and production. However, in the actual scenario of unconventional reservoir fracturing wells, if a model trained on a small number of labeled samples is directly used to predict the production of a large number of unlabeled samples, the prediction accuracy of the model will drop significantly due to the significant differences in the data distribution of the two types of samples in multiple features such as geological and engineering parameters (i.e., domain bias). Existing transfer learning methods only align the marginal distribution of global features and ignore the intrinsic correlation between features and production. In complex field data scenarios, this can easily lead to negative transfer, i.e., incorrectly mixing the feature distributions of different types of samples, which reduces the prediction accuracy. In addition, production splitting itself has inherent uncertainty. How to quantify and utilize this uncertainty to screen reliable results and guide model optimization remains an urgent problem to be solved. Therefore, there is an urgent need to develop an intelligent production splitting method that can overcome domain bias and effectively utilize uncertain information. Summary of the Invention

[0005] To address the shortcomings of existing technologies that cannot directly and accurately obtain the individual production of each fracturing section or layer, the purpose of this invention is to provide an intelligent method for dynamically dividing the production of fracturing wells in unconventional reservoirs, comprising the following steps: Step 1: Collect geological parameters, fracturing operation parameters, well control parameters and dynamic production data of tagged and untagged samples in the study area. Use the interquartile range double truncation tail reduction method to process outliers, and use the Z-score standardization method to jointly standardize the parameters of tagged and untagged samples. Step 2: Using the geological parameters, fracturing operation parameters, and well control parameters of labeled samples as inputs, and the daily gas production as output, with the objective of maximizing the sum of the coefficients of determination of the cross-validated training and test sets, a deep neural network dynamic production prediction model is constructed using the particle swarm optimization algorithm to optimize hyperparameters. Step 3: Based on the production prediction model, introduce the fracturing segment or layer category embedding as a conditional variable based on the conditional domain adversarial network. Through adversarial training of the feature extractor and the domain discriminator, align the joint distribution of labeled and unlabeled samples on features and layer categories. At the same time, apply the total production constraint to ensure that the sum of the predicted production of each fracturing segment or layer is equal to the actual total production, thereby realizing knowledge transfer and production splitting from labeled samples to unlabeled samples. Step 4: Use the Monte Carlo method to quantify the uncertainty of the production splitting results, calibrate the uncertainty threshold based on the cross-validation results of labeled samples, add the splitting results of unlabeled wells below the threshold as high-confidence pseudo-labeled samples to the training set, retrain the model and split the remaining unlabeled wells in a new round, iterate repeatedly until all unlabeled wells are split or the maximum number of iterations is reached.

[0006] Furthermore, the tagged samples mentioned in step 1 are single-layer production wells with production labels, while the untagged samples are multi-stage fracturing and multi-layer production wells with only total production data.

[0007] Furthermore, the double-truncation shortening method for the interquartile range described in step 1 includes: (1) In the formula, x abn,i These are the parameter values ​​after outlier processing; x i These are the original parameter values; Q 1 is the 25th percentile; Q 3 represents the 75th percentile; IQR represents the interquartile range. k This is the robustness coefficient.

[0008] Furthermore, the velocity and position update equations for the particles in the particle swarm algorithm described in step 2 are as follows: (7) (8) In the formula, v ij Indicates the current speed, where i = 1, 2, ..., M s , M s The number of particles; j = 1, 2, ..., M d , M d For the search space dimension; t Indicates the inertia factor; c 1. c 2 represents the learning factor; ϑ 1. ϑ 2 represents a random number in the range [0, 1]. O ij Indicates the first i The optimal position of each particle; O j Indicates the optimal position of the global particles; X ij Indicates the first i The particle in the first j The current position of the dimension; t This indicates the current iteration number.

[0009] Furthermore, the conditional domain adversarial network described in step 3 includes a shared feature extractor, a yield predictor, and a domain discriminator.

[0010] Furthermore, in step 3, the loss function of the domain discriminator is to minimize the binary cross-entropy loss of the domain classification: (11) In the formula, Losses due to domain confrontation; N t This represents the number of samples in the training batch. For each field, the source field is 0 and the destination field is 1. D Domain discriminator ； ⊗ represents a multilinear mapping, indicating the tensor product between the feature representation and the formation embedding; or For layer category; x s The input features are from the source domain (labeled samples); x t The input features are for the target domain (unlabeled samples); or s The hierarchical category label for the source domain; or t The hierarchical category label for the target domain.

[0011] Furthermore, step 4 also includes constructing an adaptive loss function, which dynamically adjusts the contribution of uncertain samples by combining weighted mean square error, uncertainty penalty, and regularization term: (20) In the formula, Losses due to uncertainty; N s The number of samples; α w , α u , α r These are the weighting coefficients for the weighted mean square error, the uncertainty penalty term, and the uncertainty regularization term, respectively. s rel，i For the first i The relative uncertainty of each sample; For the first i Predicted yield for each sample; y i For the first i The actual output of each sample; u To avoid the smallest constant that is divided by zero.

[0012] Compared with the prior art, the advantages of the present invention are: (1) The intelligent splitting algorithm provided by the present invention uses the particle swarm optimization algorithm to automatically search for the optimal hyperparameter combination of the deep neural network and builds a high-precision yield prediction model based on labeled samples, providing a reliable foundation for subsequent domain migration.

[0013] (2) The intelligent splitting algorithm provided by the present invention introduces conditional domain adversarial network into the production splitting problem. By introducing the fracturing segment or layer category as a conditional variable, the joint distribution of features and production is aligned, and accurate knowledge transfer from a small number of labeled samples to a large number of unlabeled samples is achieved, overcoming the defect of negative transfer caused by only aligning the edge distribution.

[0014] (3) The intelligent splitting algorithm provided by the present invention quantifies the uncertainty of the splitting results through Monte Carlo random sampling and designs an iterative self-optimization mechanism to dynamically expand the training set with high-confidence splitting results as pseudo-label samples, forming a virtuous cycle of data augmentation and model optimization.

[0015] (4) The intelligent splitting algorithm provided by the present invention breaks through the bottleneck of the single physical assumption of the traditional method and the lack of transferability of the existing data-driven method, and provides a high-precision and highly generalizable technical solution for unconventional reservoir production splitting.

[0016] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow provided by the present invention; Figure 2 This is a diagram showing the relative uncertainty distribution of a single-layer well in a 5-fold cross-validation embodiment of the present invention. Figure 3 This is a graph of the loss function for the first output splitting iteration in an embodiment of the present invention; Figure 4 This is a graph showing the variation of the number of samples in each layer and the number of multi-layer wells with high uncertainty as a function of the number of iterations in an embodiment of the present invention. Figure 5 This is a box plot showing the distribution of prediction uncertainties for each layer in an iteratively split multi-layer well in an embodiment of the present invention. Detailed Implementation

[0018] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0019] To address the limitations of existing technologies in directly and accurately obtaining the individual production of each fracturing segment or layer on a large scale, such as... Figure 1As shown, this invention provides an intelligent splitting method for dynamic production of unconventional reservoir fracturing wells, comprising the following steps: Step 1: Collect geological parameters, fracturing operation parameters, well control parameters and dynamic production data of tagged and untagged samples in the study area. Use the interquartile range double truncation tail reduction method to process outliers, and use the Z-score standardization method to jointly standardize the parameters of tagged and untagged samples. Step 2: Using the geological parameters, fracturing operation parameters, and well control parameters of labeled samples as inputs, and the daily gas production as output, with the objective of maximizing the sum of the coefficients of determination of the cross-validated training and test sets, a deep neural network dynamic production prediction model is constructed using the particle swarm optimization algorithm to optimize hyperparameters. Step 3: Based on the production prediction model, introduce the fracturing segment or layer category embedding as a conditional variable based on the conditional domain adversarial network. Through adversarial training of the feature extractor and the domain discriminator, align the joint distribution of labeled and unlabeled samples on features and layer categories. At the same time, apply the total production constraint to ensure that the sum of the predicted production of each fracturing segment or layer is equal to the actual total production, thereby realizing knowledge transfer and production splitting from labeled samples to unlabeled samples. Step 4: Use the Monte Carlo method to quantify the uncertainty of the production splitting results, calibrate the uncertainty threshold based on the cross-validation results of labeled samples, add the splitting results of unlabeled wells below the threshold as high-confidence pseudo-labeled samples to the training set, retrain the model and split the remaining unlabeled wells in a new round, iterate repeatedly until all unlabeled wells are split or the maximum number of iterations is reached.

[0020] In step 1, geological parameters (including but not limited to sonic transit time, density, porosity, water saturation, permeability, initial formation pressure, etc.), engineering parameters (including but not limited to sand addition intensity, fluid intensity, displacement, etc.), and well control parameters (including but not limited to wellhead oil pressure, wellhead casing pressure, bottomhole flowing pressure, etc.) and corresponding daily gas production data are first collected from tagged samples (single-stage production wells with production labels) and untagged samples (wells with multi-stage fracturing and multi-stage production but only total production). When a single-stage production well is a multi-stage fracturing well, a tagged sample set is constructed based on the average geological parameters and fracturing operation parameters of each fracturing stage.

[0021] Then, outliers are handled using the interquartile range double truncation shrinking method: (1) In the formula, x abn,i These are the parameter values ​​after outlier processing; x i These are the original parameter values; Q 1 is the 25th percentile; Q3 represents the 75th percentile; IQR represents the interquartile range. k The robustness coefficient is set to 1.5.

[0022] By merging the feature data of labeled and unlabeled samples and then performing Z-score normalization, the domain discriminator of the conditional domain adversarial network can focus on true domain differences rather than scale variations. (2) In the formula, x norm This represents the standardized parameter value; m Represents the characteristic mean; s It represents the standard deviation.

[0023] In step 2, the deep neural network consists of an input layer, hidden layers, and an output layer, connected by neurons, possessing hierarchical learning capabilities. The steps for building the deep neural network model mainly include weight initialization, forward propagation, loss calculation, back propagation, and weight and bias updates. The deep neural network minimizes the loss function to make the predicted output as close as possible to or equal to the actual output. N s For a data system with a sample size, a mean squared error (MSE) loss function can be constructed: (3) In the formula, This represents the loss function used for modeling labeled samples. N s For the sample size, y i Indicates the first i The actual output of each sample; For the first i Predicted yield for each sample.

[0024] Among them, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination R are used. 2 As an evaluation metric for machine learning models: (4) (5) (6) In the formula, This represents the average of actual output.

[0025] The appropriate selection of hyperparameters for deep neural network models has a significant impact on the model's prediction accuracy and generalization ability. In this invention, the particle swarm optimization algorithm is first used to optimize the hyperparameters of the deep neural network model with the goal of maximizing the sum of the coefficients of determination between the training and test sets. Then, a production capacity prediction model based on labeled samples is constructed based on the deep neural network model. The particle swarm optimization algorithm is a cooperative random search algorithm that finds the optimal solution by changing the position and velocity of particles in space to find the optimal position of a single particle and the globally optimal position of the particle swarm. For models with... M s individual particles M d The velocity and position update equations for the particle in the 3D search space are: (7) (8) In the formula Indicates the current speed, where i = 1, 2, ..., M s , M s The number of particles; j = 1, 2, ..., M d , M d For the search space dimension; t Indicates the inertia factor; c 1. c 2 represents the learning factor; ϑ 1. ϑ 2 represents a random number in the range [0, 1]. Indicates the first i The optimal position of each particle; Indicates the optimal position of the global particles; X ij Indicates the first i The particle in the first j The current position of the dimension; t This indicates the current iteration number.

[0026] In step 3, the Conditional Domain Adversarial Network (CDL) addresses the domain offset problem between the source domain (labeled samples) and the target domain (unlabeled samples) by introducing conditional variables and a joint distribution alignment mechanism. It adjusts the domain discriminator based on the discriminative information predicted by the classifier, thereby accurately aligning the joint distribution of features and strata. The core principle of the CDL involves two competing neural network components: a feature extractor that learns to generate domain-invariant representations, and a domain discriminator that attempts to distinguish whether an input sample comes from the source or target domain. Through this minimax optimization process, the feature extractor is trained to "fool" the discriminator by generating cross-domain indistinguishable features, thereby achieving implicit distribution alignment without explicit distribution matching constraints.

[0027] Conditional domain adversarial networks consist of three core components: (1) Shared Feature Extractor F ( x The input features are mapped to a shared feature space through continuous nonlinear transformations in the feature extractor. To capture the conditional relationship between features and hierarchical categories, a [structure / mechanism] is constructed. ψ ( or Embedded layer categories. Specifically, first, the discrete layer category labels are... or ∈ {1, 2, ..., a} is represented as a standard basis vector e η ∈ ℝ a Then, it is projected into a continuous dense vector space through a learnable linear transformation: (9) In the formula, ψ ( or ) represents the layer category embedding vector; or For layer category labels; For embedding weight matrix; For the real number field; a The number of layer categories; d e For the embedding dimension; The deviation vector is used. This transformation maps discrete stratigraphic categories to a continuous vector space that preserves the similarity between stratigraphic layers, thus enabling subsequent domain alignment to be conditioned on stratigraphic-specific discriminative information.

[0028] (2) Yield predictor: a deep neural network model based on pre-trained particle swarm optimization. g (·), and added residual predictors for specific layers. h (·) to enhance adaptability: (10) In the formula, For the first or Predicted yield of the layer; F ( x () is a shared feature extractor for input features x The feature mapping output; g (·) represents the output of a deep neural network model; h (·) denotes a specific layer residual predictor that enhances adaptability by learning the differences between the source and target domains.

[0029] (3) Domain discriminator D A binary classifier that receives conditional feature representations and outputs the probability that a sample belongs to the source or target domain. D ( F ( x ) ⊗ ψ ( or )) → [0, 1]; The multilinear mapping ⊗ represents the tensor product between the feature representation and the formation embedding, by conditionalizing the discriminator into a joint representation. F ( x ) ⊗ ψ ( or Conditional domain adversarial networks align the joint distribution of features and hierarchical categories within a domain, rather than independent marginal distributions; feature distributions are considered aligned when the discriminator outputs of the two domains are close to 0.5, corresponding to a domain adversarial loss of approximately 0.693.

[0030] The loss function of the domain discriminator is to minimize the binary cross-entropy loss of the domain classification: (11) In the formula, It is a domain confrontation loss; N t It is the number of samples in the training batch; It is a domain label (source domain is 0, target domain is 1); x s The input features are from the source domain (labeled samples); x t The input features are for the target domain (unlabeled samples); or s The hierarchical category label for the source domain; or t The hierarchical category label for the target domain.

[0031] A gradient inversion layer is introduced to achieve domain-invariant feature learning. Domain class transformation is performed during forward propagation, and the sign of the gradient is inverted during backpropagation. (12) In the formula: ld It is a domain adaptive strength parameter that is dynamically adjusted through an adaptive strategy. This is to prevent adversarial training from disrupting the stability of early feature learning. l d Dynamic scheduling using a sigmoid-based progressive boosting strategy (13) In the formula, l max It is the upper limit of the adaptability. c It is a scaling factor that controls the steepness of the sigmoid transformation. β The training progress ratio is represented by the current number of training steps divided by the total number of training steps; in the initial training phase, l d ≈0, allowing the feature extractor and the predictor to build meaningful representations in the source domain without prematurely being subjected to adversarial interference; as training progresses, l d near l max This gradually strengthens domain alignment constraints and promotes effective cross-domain feature transfer.

[0032] Meanwhile, to ensure that the sum of the predicted production of all stratigraphic categories / fractured sections matches the actual total production, a multi-stage fractured well total production prediction loss is defined: (14) In the formula, Predicting production losses for multi-stage fracturing wells; This represents the sum of the predicted production from all fractured stages in a multi-stage fracturing well. Y This represents the actual total production from a multi-stage fractured well. By alternately optimizing the predictor and domain classifier, the conditional domain adversarial network achieves alignment of conditional feature distributions. This allows the model to effectively adapt to domain shifts between multi-stage and single-stage wells while maintaining prediction accuracy.

[0033] In step 4, the task of Monte Carlo uncertainty quantification is to introduce randomness into the model prediction process, constructing a prediction distribution through multiple independent samplings, and thus quantifying the uncertainty; given input features x and layer category or The model maintains the activation of the Dropout layer during the inference phase, performing operations on the same input. k The second forward propagation yields... k Independent prediction samples The randomness of each sampling comes from the random deactivation of neurons in the hidden layer (deactivation probability is...). p ), that is, the first k This prediction can be expressed as: (15) In the formula, x The input feature vector consists of the geological parameters, construction parameters, and well control parameters corresponding to that layer. k Number of Monte Carlo samples; i k For the first k The effective parameter set for each sampling; This is the prediction function for the conditional domain adversarial network model.

[0034] Based on the above sampling results, the predicted mean is calculated as the final point prediction: (16) At the same time, the absolute uncertainty is quantified by the standard deviation of the sampling results: (17) In the formula, s The standard deviation is denoted as .

[0035] To eliminate the influence of dimensions and enable uncertainty comparisons across stratigraphic and well levels, relative uncertainty is further defined: (18) In the formula, s rel Due to relative uncertainty, u To avoid the smallest constant when dividing by zero, we take 10. -6 .

[0036] Relative uncertainty s rel Uncertainty threshold used to guide the dynamic adjustment of production allocation weights. T Defined as the critical value that distinguishes between reliable and uncertain predictions; determined using cross-validation based on initial labeled samples. T The value; when s rel ≤ T When the prediction is considered reliable, the weights are directly set to the predicted mean; when... s rel > T At that time, the weight of unreliable predictions is reduced through exponential decay, and the final capacity allocation ratio for each layer is as follows: (19) In the formula, α l For the first l The production allocation ratio of each layer; m The number of layers contained in the well; w l For the first l The proportion of the total production of composite layers.

[0037] An adaptive loss function is constructed, which dynamically adjusts the contribution of uncertain samples by combining weighted mean square error, uncertainty penalty, and regularization term: (20) In the formula, Losses due to uncertainty; s rel，i For the first i The relative uncertainty of each sample; α w , α u , α r These are the weighting coefficients for the weighted mean square error, the uncertainty penalty term, and the uncertainty regularization term, respectively, which are taken as 1.0, 0.5, and 0.1 in this paper; for a single-layer well, The mean square error in is For multi-layer wells, however, it is Therefore, the total loss for the training task, including both labeled and unlabeled samples, can be expressed as: (twenty one) In the formula, Total loss; α d Assign domain adversarial loss weights; when the main task loss is significantly reduced... α d Moderately increase to enhance domain alignment; conversely, when losses increase, α d This reduces the alignment strength and prevents negative migration; the entire adjustment process remains smooth and gradual to ensure training stability.

[0038] After each round of uncertainty quantification, the overall relative uncertainty of each unlabeled well is calculated (defined as the average uncertainty at each time point during the dynamic production period); if this value is less than or equal to a threshold... T If the well is identified as a high-confidence sample, the splitting results of each fractured section or layer are used as pseudo-labels, which are combined with the corresponding feature and layer category labels to form new labeled samples added to the training set. Then, the model is retrained using the expanded training set to perform a new round of splitting and screening of the remaining unlabeled wells. This process is iterated until no new high-confidence samples are screened out or the maximum number of iterations is reached.

[0039] Example 1: Taking the tight sandstone gas reservoir of Sulige Gas Field as an example, the field application is carried out. It has three main producing layers from bottom to top: Benxi Formation, Shanxi Formation and Shihezi Formation. The main development mode of Sulige Gas Field is multi-layer directional well co-production, that is, a single well simultaneously develops multiple layers of the same or different types.

[0040] A total of 142 single-layer wells were collected, including 47 from the Benxi Formation, 50 from the Shanxi Formation, and 45 from the Shihezi Formation. 360 multi-layer wells were also collected, corresponding to 240 construction sequences in the Benxi Formation, 345 in the Shihezi Formation, and 700 in the Shanxi Formation. These included 6 geological parameters, 3 fracturing parameters, and 3 well control parameters. Production was segmented based on a 450-day dynamic production target. The basic parameters and ranges after outlier screening are shown in Table 1. The data in the table indicate that the parameter ranges of multi-layer wells are mostly larger than those of single-layer wells, with significant domain differences between the two. Therefore, it is necessary to use domain countermeasures to align skewed parameters.

[0041] Table 1. Application Well Basic Parameters and Range Deep neural network model training is carried out based on single-layer samples. Since the number of single-layer samples is moderate, in order to balance the PSO optimization efficiency and computational cost, and to prevent model overfitting, the number of hidden layers in the deep neural network model is set to 3. The hyperparameter optimization range is shown in Table 2.

[0042] Table 2. Optimization Range of Hyperparameters for Deep Neural Network Models To determine the uncertainty threshold for reliable samples after production splitting. T Five-fold cross-validation was performed on the initial single-layer sample, and the Monte Carlo uncertainty assessment was conducted through 50 random forward passes. The relative uncertainty was calculated using equation (18). s rel The average of 5-fold cross-validation s rel Distribution and cumulative average relative error, such as Figure 2 As shown, 82.30% of the validation samples... s rel ≤0.15, the cumulative average relative prediction error within this range is 0.2539; when the interval is expanded... s rel When the cumulative average relative prediction error is ≤0.20, it rises to 0.3023, an increase of 19.1%, indicating that... s rel Samples with a value greater than 0.15 will significantly increase the relative error. Therefore, selecting... T =0.15 is used as a threshold to represent the statistical calibration boundary between reliable and unreliable predictions.

[0043] Production splitting was performed on multi-layered wells. All wells in the Benxi and Shihezi Formations were split after four iterations, while all wells in the Shanxi Formation were split after six iterations. The 5-fold cross-validation results for each split are shown in Table 3. For the initial single-stratum sample, the cross-validation R for all three strata is...2 All values ​​exceeded 0.887, confirming that the model can capture the complex nonlinear relationship between geological engineering parameters and yield; with the gradual introduction of pseudo-labeled samples, the R-value decreased due to the increased geological heterogeneity of the augmented dataset. 2 The value decreased slightly but remained above 0.86, stabilizing after iteration 3. This indicates that the iterative framework maintained robust generalization without overfitting.

[0044] Table 3. Results of Iterative Splitting Average 5-Fold Cross-Validation The loss function curve for the first iteration of splitting is as follows: Figure 3 As shown, all loss functions converge to a stable state after 100 iterations. The conditional domain adversarial loss stabilizes at around 0.693, indicating that the domain classifier can no longer effectively distinguish the source of features. This state reflects that the adversarial process between the feature extractor and the domain classifier has reached equilibrium, where the feature extractor has successfully learned the domain-invariant representation, aligning the feature distribution between single-formation wells and multi-formation wells, thereby achieving domain invariance in the feature space.

[0045] Figure 4 The number of single-layer samples and high-uncertainty multi-layer wells was analyzed as a function of iterations. By the end of the third iteration, approximately 60% of the multi-layer wells had been successfully split. This indicates that after incorporating low-uncertainty multi-layer wells as pseudo-labels into the single-layer sample pool, effective knowledge transfer was achieved as iterations progressed, confirming the applicability of the model for production allocation in multi-layer wells.

[0046] Box plot of uncertainty prediction results for each layer of all multi-layer wells in the production splitting iteration process is shown below. Figure 5 As shown, from Figure 5 As can be seen from the results, since each iteration is based on the newly supplemented sample to remodel the wells with high uncertainty layers, the results of the 3rd to 5th iterations show that the range of relative uncertainty does not decrease with the number of iterations; however, as the iterations proceed, the average and median values ​​of relative uncertainty show a consistent downward trend in all three formations.

[0047] The mean relative uncertainty of the Benxi group and the Shihezi group dropped to around 0.11 after the second iteration, while the mean relative uncertainty of the Shanxi group, which had a larger sample size, also dropped to around 0.12 after the third iteration. Finally, the mean relative uncertainty of the Benxi group was 0.107, that of the Shanxi group was 0.105, and that of the Shihezi group was 0.101.

[0048] Therefore, the method proposed in this invention can realize the intelligent splitting of dynamic production of unconventional reservoir fracturing wells, laying the foundation for evaluating fracturing effect and optimizing fracturing parameters.

[0049] In summary, this invention provides an uncertainty-guided conditional domain adversarial adaptive iterative yield splitting method. First, outlier processing and standardization are performed on the collected raw data to ensure that labeled and unlabeled samples are analyzed at the same feature scale. Second, a particle swarm optimization algorithm is used to automatically search for the optimal hyperparameter combination of a deep neural network, constructing a high-precision yield prediction model based on labeled samples, providing a reliable foundation for subsequent domain transfer. Furthermore, a conditional domain adversarial network is introduced into the yield splitting problem. By introducing fracturing segment or layer category as a conditional variable, the joint distribution of features and yield is aligned, achieving accurate knowledge transfer from a small number of labeled samples to a large number of unlabeled samples, overcoming the defect of negative transfer caused by merely aligning marginal distributions. Finally, the uncertainty of the splitting results is quantified through Monte Carlo random sampling, and an iterative self-optimization mechanism is designed to dynamically expand the training set using high-confidence splitting results as pseudo-labeled samples, forming a virtuous cycle of data augmentation and model optimization. This invention overcomes the bottlenecks of traditional methods' single physical assumptions and the lack of transfer capabilities in existing data-driven methods, providing a high-precision and highly generalizable technical solution for unconventional reservoir yield splitting.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligently splitting dynamic production data in unconventional reservoir fracturing wells, characterized in that: Includes the following steps: Step 1: Collect geological parameters, fracturing operation parameters, well control parameters and dynamic production data of tagged and untagged samples in the study area. Use the interquartile range double truncation tail reduction method to process outliers, and use the Z-score standardization method to jointly standardize the parameters of tagged and untagged samples. Step 2: Using the geological parameters, fracturing operation parameters, and well control parameters of labeled samples as inputs, and the daily gas production as output, with the objective of maximizing the sum of the coefficients of determination of the cross-validated training and test sets, a deep neural network dynamic production prediction model is constructed using the particle swarm optimization algorithm to optimize hyperparameters. Step 3: Based on the production prediction model, introduce the fracturing segment or layer category embedding as a conditional variable based on the conditional domain adversarial network. Through adversarial training of the feature extractor and the domain discriminator, align the joint distribution of labeled and unlabeled samples on features and layer categories. At the same time, apply the total production constraint to ensure that the sum of the predicted production of each fracturing segment or layer is equal to the actual total production, thereby realizing knowledge transfer and production splitting from labeled samples to unlabeled samples. Step 4: Use the Monte Carlo method to quantify the uncertainty of the production splitting results, calibrate the uncertainty threshold based on the cross-validation results of labeled samples, add the splitting results of unlabeled wells below the threshold as high-confidence pseudo-labeled samples to the training set, retrain the model and split the remaining unlabeled wells in a new round, iterate repeatedly until all unlabeled wells are split or the maximum number of iterations is reached.

2. The dynamic output intelligent splitting method according to claim 1, characterized in that, The tagged samples mentioned in step 1 are single-layer production wells with production labels, while the untagged samples are multi-stage fracturing and multi-layer production wells with only total production data.

3. The dynamic output intelligent splitting method according to claim 1, characterized in that, The double-truncation tail reduction method for interquartile range described in step 1 includes: （1） In the formula, x abn,i These are the parameter values ​​after outlier processing; x i These are the original parameter values; Q 1 is the 25th percentile; Q 3 represents the 75th percentile; IQR represents the interquartile range. κ This is the robustness coefficient.

4. The dynamic output intelligent splitting method according to claim 1, characterized in that, The velocity and position update equations for the particles in the particle swarm algorithm described in step 2 are as follows: （7） （8） In the formula, v ij Indicates the current speed, where i = 1, 2, ..., M s , M s The number of particles; j = 1, 2, ..., M d , M d For the search space dimension; τ Indicates the inertia factor; c 1. c 2 represents the learning factor; ϑ 1. ϑ 2 represents a random number in the range [0, 1]. O ij Indicates the first i The optimal position of each particle; O j Indicates the optimal position of the global particles; X ij Indicates the first i The particle in the first j The current position of the dimension; t This indicates the current iteration number.

5. The dynamic output intelligent splitting method according to claim 1, characterized in that, The conditional domain adversarial network described in step 3 includes a shared feature extractor, a yield predictor, and a domain discriminator.

6. The dynamic output intelligent splitting method according to claim 1, characterized in that, In step 3, the loss function of the domain discriminator is to minimize the binary cross-entropy loss of the domain classification: （11） In the formula, Losses due to domain confrontation; N t This represents the number of samples in the training batch. For the domain label, the source domain is 0 and the destination domain is 1; D Domain discriminator ； ⊗ represents a multilinear mapping, indicating the tensor product between the feature representation and the formation embedding; η For layer category; x s These are the input features of the source domain; x t The input features are for the target domain; η s The hierarchical category label for the source domain; η t The hierarchical category label for the target domain.

7. The dynamic output intelligent splitting method according to claim 1, characterized in that, Step 4 also includes constructing an adaptive loss function, which dynamically adjusts the contribution of uncertain samples by combining weighted mean square error, uncertainty penalty, and regularization term: （20） In the formula, Losses due to uncertainty; N s The number of samples; α w , α u , α r These are the weighting coefficients for the weighted mean square error, the uncertainty penalty term, and the uncertainty regularization term, respectively. σ rel，i For the first i The relative uncertainty of each sample; For the first i Predicted yield for each sample; y i For the first i The actual output of each sample; υ To avoid dividing by zero, the smallest constant.

Images