A progressive prediction method for inter-well fracture channeling based on a time-series prototype network

Abstract

The application provides a fracturing interwell channeling gradual prediction method based on a time sequence prototype network, relates to the field of oil and gas field development and engineering, and specifically comprises the following steps: based on the pressure response curve form and numerical characteristics of the observation well in the historical construction, combining the reservoir fluid percolation theory and the rock mechanics characteristics, the interwell channeling is divided into multiple types. Based on the fracturing channeling type classification standard, the historical fracturing time sequence data is systematically labeled and pretreated to form a labeled data set, and the data is pretreated. A time sequence prototype network model suitable for the time sequence data of the fracturing variable length is designed, including a feature encoder and a prototype matching layer. The trained time sequence prototype network model is used for gradual prediction and risk early warning of the channeling type. The technical scheme of the application overcomes the problem in the prior art that the fracturing channeling prediction method cannot quantitatively express the uncertainty of the prediction, is easy to cause false positives or false negatives in the early construction, and misleads the site decision.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development and engineering, and specifically to a progressive prediction method for inter-well hydraulic crossflow based on a time-series prototype network. Background Technology

[0002] The effective development of unconventional oil and gas resources relies heavily on large-scale horizontal well hydraulic fracturing technology, which unlocks reservoir productivity by creating complex fracture networks. However, in the factory-style operation mode of densely distributed wells on a platform, due to the small well spacing and complex underground stress field, the high-pressure fluid and expanding fractures generated during fracturing operations on a single well can easily interfere with adjacent production wells or temporarily shut-down wells. This phenomenon is known as "fracturing crosstalk" or "fracturing crosstalk." Crosstalk can lead to serious consequences such as abnormal pressure increases in adjacent wells, damage to production tubing, and even wellbore integrity failure, resulting not only in significant safety risks and economic losses but also severely restricting the full utilization of reserves and the improvement of development efficiency. Therefore, achieving early and accurate identification and type diagnosis of crosstalk risks is a key prerequisite for proactive prevention and control, ensuring the safety and efficiency of fracturing operations.

[0003] Currently, on-site monitoring and assessment of pressure channeling primarily rely on engineers' manual, experience-based analysis of the morphology of pressure monitoring curves from adjacent wells. This method is highly dependent on individual expert experience, lacks quantitative and standardized judgment criteria, and different engineers may arrive at different conclusions regarding the same pressure response pattern, leading to highly subjective and unrepeatable results. More importantly, manual analysis is typically lagging, only identifying pressure channeling when signs are relatively obvious, making it difficult to provide a valuable early warning window for real-time dynamic adjustments to construction parameters.

[0004] With the development of digital and intelligent technologies, machine learning methods have brought new possibilities for the automatic identification of fracturing crosstalk. Some studies have attempted to use fracturing operation data and adjacent well response data as features, employing traditional machine learning classification models (such as support vector machines and random forests) to determine the type of crosstalk. However, these methods have significant limitations. First, they typically rely on static statistical features (such as average pressure and maximum pressure rise) extracted from complete historical fracturing data, and cannot effectively handle variable-length time-series data that continues to grow over time, thus failing to achieve "progressive" prediction during the fracturing process. Second, these models are mostly "black box" models, and their prediction results lack interpretability. Engineers find it difficult to understand the basis for the model's judgments, thereby reducing their trust in and willingness to adopt the model's results. Most importantly, existing methods cannot effectively handle the inherent uncertainty in early predictions. In the early stages of fracturing operations, the amount of data is limited, and the pressure response pattern has not yet fully emerged, making it difficult for any model to make a high-confidence, accurate judgment. Existing methods typically provide a definitive classification result directly, failing to quantify this predictive uncertainty, which can easily lead to false alarms or missed alarms in the early stages of fracturing, misleading on-site decision-making.

[0005] Therefore, there is a need for a progressive prediction method for inter-well hydraulic crosstalk based on a time-series prototype network that can achieve a smooth transition from early warning to late confirmation and improve accuracy. Summary of the Invention

[0006] The main objective of this invention is to provide a progressive prediction method for inter-well crosstalk based on a time-series prototype network, in order to solve the problem that existing crosstalk prediction methods cannot quantitatively express the uncertainty of the prediction, which can easily lead to false alarms or omissions in the early stages of construction and mislead on-site decision-making.

[0007] To achieve the above objectives, this invention provides a progressive prediction method for inter-well hydraulic crosstalk based on a time-series prototype network, specifically including the following steps:

[0008] S1, based on the pressure response curve morphology and numerical characteristics of observation wells in historical construction, combined with reservoir fluid seepage theory and rock mechanics properties, divides inter-well pressure channeling into multiple types and constructs a classification standard for fracturing pressure channeling types.

[0009] S2, based on the classification criteria for fracturing and cracking types in step S1, systematically annotate and preprocess the historical fracturing construction time series data to form the labeled dataset required for training the time series prototype network model, and preprocess the data.

[0010] S3, a time-series prototype network model adapted to the time-series data of hydraulic fracturing length is designed, including: a feature encoder and a prototype matching layer. The feature encoder realizes fixed-dimensional embedding of time-series data, the prototype matching layer establishes the association between features and hydraulic fracturing type, and the model training is completed by optimizing the composite loss function.

[0011] S4 utilizes the trained temporal prototype network model for progressive prediction and risk warning of compression types.

[0012] Furthermore, step S1 specifically includes the following steps:

[0013] S1.1, inter-well pressure channeling types include: weak transmission or no effective communication type, pore or micro-fracture transmission type, pressure fracture progressive communication type, pressure fracture direct communication type, and stress-activated natural fracture type.

[0014] S1.2, formulate three-dimensional quantitative judgment rules based on pressure numerical characteristics, curve morphology characteristics and derivative characteristics to distinguish different inter-well pressure crossflow types.

[0015] S1.3 Statistical analysis is performed on the pressure response data of all observation wells in the historical dataset. A threshold search range is set based on the reservoir characteristics of the target block. Multiple candidate threshold combinations are then generated using a grid search method based on this search range. The optimization objective is to achieve the highest weighted Kappa coefficient between the automatic classification results and the manual judgment results from domain experts under different threshold combinations. The threshold combinations are iteratively optimized using the Particle Swarm Optimization (PSO) algorithm to finally determine the optimal threshold suitable for the target block. .

[0016] Furthermore, step S1.2 specifically includes:

[0017] When the maximum pressure increase of the observation well Below the first threshold When judged as weak transmission or no effective communication, that is... ;in, , For the observation well pressure, This refers to the start time of construction. For time steps, This is the function for finding the maximum value.

[0018] when And the average rate of pressure increase Below the second threshold It was determined to be of the pore or microcrack propagation type, i.e. and ;in, , The time it takes for the pressure to reach its peak.

[0019] When the pressure curve shows an "S" shape or a step-like growth, it is determined to be a progressively communicating type of pressure fracture. The quantitative standard is that the first derivative sequence of the pressure curve has a broad peak.

[0020] when and It was determined to be a direct communication type of pressure fracture; among which , , and These are the third and fourth thresholds, respectively.

[0021] When the pressure curve suddenly accelerates upward after a steady or declining phase, it is determined to be a stress-activated natural crack type. The quantitative standard is that the second derivative of the pressure curve shows a positive pulse in the middle and late stages of construction.

[0022] Furthermore, step S2 specifically includes the following steps:

[0023] S2.1, Collect the discharge volume of the construction well. Construction well pressure and observation well pressure The time-series data is aligned by time step.

[0024] S2.2, perform data preprocessing, including: using linear interpolation to handle missing data, ensuring the continuity of time series data, and... , and Z-score standardization is performed separately. The formula for Z-score standardization is as follows:

[0025] ;

[0026] in, For the data mean, The standard deviation of the data. and These are the data before and after standardization, respectively.

[0027] S2.3, based on the classification criteria for fracturing and hydraulic channeling types, each preprocessed time series data sequence is... Label with compression type The label uses one-hot encoding format.

[0028] S2.4 divides the labeled dataset into training, validation and test sets.

[0029] Furthermore, step S3 specifically includes the following steps:

[0030] S3.1, using a one-dimensional convolutional 1D-CNN to capture local correlation features within a short time window in fracturing time series data and filter high-frequency noise, the 1D-CNN calculation formula is as follows:

[0031] ;

[0032] in, This represents a one-dimensional convolution operation. For convolution kernel weights, This is fracturing time series data. For bias terms, For the extracted local features, This is the activation function.

[0033] S3.2, will Input the data into a Long Short-Term Memory (LSTM) network to capture long-term dependencies in fracturing time-series data:

[0034] ;

[0035] in, For LSTM Input of time steps, For LSTM The hidden state of the time step. It is the sigmoid activation function. It represents the Hadamah accumulation. Input gates Forgotten Gate Cell state With output gate The weight matrix, Input gates respectively Forgotten Gate Cell state With output gate The bias term is given by tanh, which is the hyperbolic tangent function.

[0036] S3.3 utilizes a progress-aware attention mechanism to dynamically extract key temporal features for judging pressure crosstalk based on the current construction progress.

[0037] S3.4, the LSTM hidden states are weighted and summed based on the attention weights to obtain the final fixed-dimensional embedding representation. :

[0038] ;

[0039] in, For attention weights, For LSTM The hidden state of the time step. This represents the length of the original time-series data.

[0040] S3.5, establish the association between the embedded representation and the compression type in the prototype matching layer.

[0041] Furthermore, step S3.3 specifically includes the following steps:

[0042] S3.3.1, Calculate the construction schedule The calculation formula is:

[0043] ;

[0044] in, This represents the actual length of the current time series data. The total data length is estimated based on the construction plan.

[0045] S3.3.2, will Discretize into A progress box, get the progress box .

[0046] S3.3.3, via a fully connected network Encoded as query vector The key features used to characterize the current progress are calculated using the following formula:

[0047] ;

[0048] in, As a schedule weight, This is a bias term.

[0049] S3.3.4, Calculate the query vector and The similarity is used to obtain the attention weight. The calculation formula is:

[0050] ;

[0051] in, Here is the attention weight matrix. For query vector transpose, For LSTM The hidden state of the time step.

[0052] Furthermore, step S3.5 specifically includes the following steps:

[0053] S3.5.1, Types of inter-well pressure channeling Under each progress bin, define a set of trainable prototype vectors. A category center initialization strategy is adopted, which initializes categories in the training set that belong to the category center. And the progress falls into the progress box samples Construction data Take the average as initial value The calculation formula is:

[0054] ;

[0055] in, For training focus belongs to And in the progress box The number of samples, This is the set of indices for the corresponding samples.

[0056] S3.5.2, Quantization and similarity The calculation formula is:

[0057] ;

[0058] in, express L2 norm, express L2 norm, The range of values ​​is , Indicates an exact match. Indicates no association. Indicates a complete reversal.

[0059] S3.5.3, the similarity vector is converted into a probability distribution of the compression type using the Softmax function. The calculation formula is:

[0060] ;

[0061] in, For temperature parameters, It is a natural exponential function. For progress box Down The class's compressed prototype vector.

[0062] Furthermore, step S3.5 also includes the following steps:

[0063] S3.5.4 uses a composite loss function to train the model, optimizes network parameters using gradient descent, and employs cross-entropy classification loss. The probability distribution used to constrain the model output to match the true labels of the samples, ensuring classification accuracy, is calculated using the following formula:

[0064] ;

[0065] ;

[0066] in, For the sample The true label, For batch size, For the schedule weight function, For hyperparameters, For the sample Construction progress; Represents each type of inter-well pressure channeling The corresponding serial number.

[0067] S3.5.5, Prototype Clustering Loss The Euclidean distance used to constrain the embedding representations of similar samples and their corresponding prototype vectors is calculated using the following formula:

[0068] ;

[0069] in, For the sample Construction data, For progress box Lower sample The prototype vector of the type;

[0070] Total loss function The calculation formula is:

[0071] ;

[0072] in, This is a hyperparameter.

[0073] Furthermore, step S4 specifically includes the following steps:

[0074] S4.1, For the ongoing fracturing operation to be predicted, from the start time to the current time... Real-time collection of well discharge volume Construction well pressure Observation well pressure The collected data is processed in real time to form the current time-series data sequence. And calculate the current construction progress. and the progress box corresponding to the current construction progress. .

[0075] S4.2, the current time series data sequence and Generate real-time embedding representations from a pre-trained temporal prototype network model. , call Calculate the following 5 types of prototype vectors. The cosine similarity with each prototype vector is used to obtain the probability distribution of each compression type at the current time through the Softmax function. The compression type corresponding to the highest probability is selected as the current prediction result. The calculation formula is:

[0076] ;

[0077] in, This is the maximum value index function.

[0078] S4.3, Calculate the overall confidence level of the current prediction. :

[0079] ;

[0080] ;

[0081] in, It is the highest probability value currently predicted. Based on The confidence calibration function, This is the sensitivity coefficient.

[0082] The present invention has the following beneficial effects:

[0083] This invention proposes a progressive prediction method for inter-well crosstalk based on a time-series prototype network. This method addresses the subjectivity and inconsistency inherent in traditional methods relying on human experience by establishing a quantified crosstalk classification standard. The designed time-series prototype network model effectively handles variable-length construction time-series data and, through learning the "prototype" patterns of five typical crosstalk types under different construction progresses, achieves probabilistic prediction of crosstalk types, significantly improving the accuracy of early identification. The built-in progress-aware confidence assessment mechanism quantifies the uncertainty of prediction results, providing reliability indicators for engineering decisions and effectively avoiding false alarms and missed alarms in early predictions. This method realizes a shift from "post-analysis" to "in-process early warning," providing key technical support for the real-time dynamic adjustment of fracturing construction parameters and possessing significant engineering application value for ensuring the safe and efficient development of unconventional oil and gas resources. Attached Figure Description

[0084] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0085] Figure 1 A flowchart of a progressive prediction method for inter-well hydraulic crossflow based on a time-series prototype network according to the present invention is shown.

[0086] Figure 2 This is a typical Type I stress curve, characterized by weak transmission or no effective communication.

[0087] Figure 3 It is a typical pore or microcrack transmission type (Type II) pressure curve.

[0088] Figure 4 It is a typical pressure curve diagram of progressive communication type (Type III) of hydraulic fracturing.

[0089] Figure 5 This is a typical pressure curve diagram of a direct-connection type (Type IV) hydraulic fracture.

[0090] Figure 6 It is a typical stress-activated natural crack type (V-type) pressure curve.

[0091] Figure 7 This is a training loss graph for a temporal prototype network model.

[0092] Figure 8 It is a complete pressure curve diagram of adjacent observation wells.

[0093] Figure 9 It is a dynamic evolution diagram of the predicted probability and confidence level of pressure channeling type over construction time. Detailed Implementation

[0094] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0095] like Figure 1 The method for progressive prediction of inter-well crossflow based on time-series prototype networks, as shown, includes the following steps:

[0096] S1, based on the pressure response curve morphology and numerical characteristics of observation wells in historical construction, combined with reservoir fluid seepage theory and rock mechanics properties, divides inter-well pressure channeling into multiple types and constructs a classification standard for fracturing pressure channeling types.

[0097] S2, based on the classification criteria for fracturing and cracking types in step S1, systematically annotate and preprocess the historical fracturing construction time series data to form the labeled dataset required for training the time series prototype network model, and preprocess the data.

[0098] S3, design a time-series prototype network model adapted to the time-series data of pressure fracturing length, including: a feature encoder and a prototype matching layer. The feature encoder realizes fixed-dimensional embedding of time-series data, the prototype matching layer establishes the association between features and pressure fracturing type, and the composite loss function is used to optimize the model training to ensure that the model has the ability to extract time-series features and match pressure fracturing type.

[0099] S4 utilizes the trained temporal prototype network model for progressive prediction and risk warning of compression types.

[0100] This example uses the fracturing operation of a shale gas extraction platform as an example. The platform has 10 horizontal wells with a spacing of 300 meters between them. In this case, fracturing was carried out on one of the wells (Well S), and the pressure response of the adjacent observation well (Well O) was monitored. The fracturing time was expected to be 2 hours and 55 minutes.

[0101] Specifically, based on the pressure response curve morphology and numerical characteristics of observation wells during historical construction, and combined with reservoir fluid seepage theory and rock mechanics properties, inter-well pressure channeling is classified into five typical types. Each type corresponds to different reservoir communication mechanisms and risk levels. Step S1 specifically includes the following steps:

[0102] S1.1, the inter-well pressure channeling types include: weak transmission or no effective communication type (Type I), pore or microfracture transmission type (Type II), pressure fracture progressive communication type (Type III), pressure fracture direct communication type (Type IV), and stress activated natural fracture type (Type V).

[0103] Weak Transmission or No Effective Communication (Type I): The fracturing fluid seeps only through the reservoir matrix pores in a very weak manner, without forming effective fractures or connecting channels; the pressure increase in the observation well is minimal, having no substantial impact on production, and the risk level is "low". A typical pressure curve is shown below. Figure 2 As shown.

[0104] Pore ​​or microfracture propagation type (Type II): Fracturing fluid slowly diffuses into the observation well through the reservoir's native pores or microfracture network; the pressure in the observation well shows a gradual upward trend without abrupt changes, and there is no significant risk of tubing damage or production loss. The risk level is "low to medium." A typical pressure curve is shown below. Figure 3 As shown.

[0105] Type III progressive fracturing: The fracturing fracture gradually extends and expands from the construction well, forming a staged, incomplete connection with the observation well; the pressure curve of the observation well shows an "S-shaped" or step-shaped increase, indicating a potential risk of tubing fatigue damage, with a risk level of "medium-high". A typical pressure curve is shown in the figure below. Figure 4 As shown.

[0106] Type IV direct-connection fracturing: The fracturing fracture rapidly penetrates the construction well and the observation well, forming a highly conductive connecting channel; the pressure in the observation well rises sharply, making it prone to wellbore integrity failure (such as casing deformation). The risk level is "high". A typical pressure curve is shown below. Figure 5 As shown.

[0107] Stress-activated natural fracture type (V-type): Fracturing operations trigger reservoir stress field reconstruction, activating the natural fracture network and forming complex flow channels; the pressure in the observation well rises sharply after a stable initial phase, easily leading to permanent production damage, with a risk level of "extremely high". A typical pressure curve is shown below. Figure 6 As shown.

[0108] S1.2, formulate three-dimensional quantitative judgment rules based on pressure numerical characteristics, curve morphology characteristics and derivative characteristics to distinguish different inter-well pressure crossflow types.

[0109] S1.3 Statistical analysis is performed on the pressure response data of all observation wells in the historical dataset. A threshold search range is set based on the reservoir characteristics of the target block. Multiple candidate threshold combinations are then generated using a grid search method based on this search range. The optimization objective is to achieve the highest weighted Kappa coefficient between the automatic classification results and the manual judgment results from domain experts under different threshold combinations. The threshold combinations are iteratively optimized using the Particle Swarm Optimization (PSO) algorithm to finally determine the optimal threshold suitable for the target block. .

[0110] Statistical analysis was performed on all historical fracturing operation data of the shale gas extraction platform (a total of 192 effective pressure response sequences), and a threshold search range was set: ∈[1.6,2.4]MPa ∈[0.32,0.80]MPa / h, ∈[4.6,6.3]MPa, The threshold values ​​are ∈ [0.92, 2.1] MPa / h. Based on this range, 1782 candidate threshold combinations are generated using a grid search method (if the number of combinations is too large and the computational load is too high, a strategy of coarse search followed by fine search can be used to determine the threshold combinations in multiple rounds). The optimization objective is to maximize the weighted Kappa coefficient between the automatic classification results and the manual judgment results of domain experts under different threshold combinations. The threshold combinations are then iteratively optimized using the Particle Swarm Optimization (PSO) algorithm to finally determine the optimal threshold combination. =2.14MPa =0.52MPa / h =5.12MPa =1.14MPa / h, at which point the Kappa coefficient reaches 0.92, showing excellent consistency with the expert's manual judgment result.

[0111] Specifically, for the five types of pressure channeling mentioned above, three-dimensional quantitative judgment rules based on "pressure numerical characteristics + curve morphology characteristics + derivative characteristics" are formulated to ensure accurate differentiation of different types. Step S1.2 is as follows:

[0112] When the maximum pressure increase of the observation well Below the first threshold When judged as weak transmission or no effective communication, that is... ;in, , For the observation well pressure, This refers to the start time of construction. For time steps, This is the function for finding the maximum value.

[0113] when And the average rate of pressure increase Below the second threshold It was determined to be of the pore or microcrack propagation type, i.e. and ;in, , The time it takes for the pressure to reach its peak.

[0114] When the pressure curve exhibits an "S"-shaped or step-shaped increase, it is identified as a progressively communicating type of pressure fracture. The quantification standard is the first derivative of the pressure curve, i.e., the instantaneous rate of increase. The sequence has a distinct broad peak.

[0115] when and It was determined to be a direct communication type of pressure fracture; among which , , and These are the third and fourth thresholds, respectively.

[0116] When the pressure curve suddenly accelerates upward after a steady or declining phase, it is determined to be a stress-activated natural crack type. The quantitative standard is that the second derivative (rising acceleration) of the pressure curve shows a significant positive pulse in the middle and late stages of construction.

[0117] Specifically, step S2 includes the following steps:

[0118] S2.1, Collect the discharge volume of the construction well. Construction well pressure and observation well pressure The time-series data is aligned by time step.

[0119] S2.2, perform data preprocessing, including: using linear interpolation to handle missing data, ensuring the continuity of time series data, and... , and Z-score standardization is performed separately to eliminate the interference of unit differences on model training. The Z-score standardization calculation formula is as follows:

[0120] ;

[0121] in, For the data mean, The standard deviation of the data. and These are the data before and after standardization, respectively.

[0122] S2.3, based on the classification criteria for fracturing and hydraulic channeling types, each preprocessed time series data sequence is... Label with compression type The labels use one-hot encoding format (example: Type III compression corresponds to label [0,0,1,0,0]), which is suitable for model classification tasks.

[0123] S2.4 divides the labeled dataset into a training set, a validation set, and a test set. The training set is used for learning model parameters to ensure that the model fits the compressed type features. The validation set is used for hyperparameter tuning (such as embedding dimension and learning rate) to avoid model overfitting. The test set is used to evaluate the model's generalization performance and verify the model's ability to predict unseen data.

[0124] Collect all historical fracturing operation data for shale gas extraction platforms, including well discharge rates. Construction well pressure and observation well pressure The time-series data, with a time step of 1 second and precisely aligned to timestamps, resulted in a total of 21,921,095 data entries. Based on the aforementioned quantitative classification standards, the historical fracturing operation time-series data was systematically labeled and preprocessed. Linear interpolation was used to fill in 592 missing data points. , , Z-score normalization was performed to eliminate the influence of units. One-hot encoding was then applied to the preprocessed time-series sequences for labeling, resulting in type I labels [1,0,0,0,0], type II labels [0,1,0,0,0], type III labels [0,0,1,0,0], type IV labels [0,0,0,1,0], and type V labels [0,0,0,0,1]. This formed the labeled dataset required for training the temporal prototype network model. ,in It is each preprocessed time-series data sequence. yes The corresponding compression type tag.

[0125] Finally, the labeled dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for learning model parameters to ensure that the model fits the compressed data type features. The validation set was used for hyperparameter tuning (such as embedding dimension and learning rate) to avoid overfitting. The test set was used to evaluate the model's generalization performance and verify the model's ability to predict unseen data.

[0126] Specifically, step S3 includes the following steps:

[0127] S3.1 The feature encoder adopts a hybrid structure of "1D-CNN+LSTM+progress-aware attention mechanism" to fully capture the local patterns and global correlations of temporal data.

[0128] One-dimensional convolutional 1D-CNN (32 kernels, 12 kernel size) is used to capture local correlation features within short time windows in fracturing time series data and filter high-frequency noise. The 1D-CNN calculation formula is as follows:

[0129] ;

[0130] in, This represents a one-dimensional convolution operation. For convolution kernel weights, This is fracturing time series data. For bias terms, For the extracted local features, This is an activation function used to introduce nonlinear features and enhance the model's ability to fit complex local patterns.

[0131] S3.2 utilizes an LSTM (128 hidden layers, 2 layers, dropout rate 0.25) to capture long-term dependencies in fracturing time-series data, addressing the problem that traditional CNNs cannot handle long-term temporal correlations. The LSTM receives local features from the output of the 1D-CNN. As input, its output is the hidden state at each time step. ,in Integrating the first The temporal information before and after a time step can characterize long-term dependencies. LSTM uses the input gate... Forgotten Gate Cell state With output gate To achieve long-term information memory.

[0132] Will Input the data into a Long Short-Term Memory (LSTM) network to capture long-term dependencies in fracturing time-series data:

[0133] ;

[0134] in, For LSTM Input of time step (i.e.) (the tth row) For LSTM The hidden state of the time step. It is the sigmoid activation function (output range [0,1]). This represents the Hadamard product (element-wise multiplication). Input gates Forgotten Gate Cell state With output gate The weight matrix, Input gates respectively Forgotten Gate Cell state With output gate The bias term, It is the hyperbolic tangent function.

[0135] S3.3 utilizes a progress-aware attention mechanism to dynamically extract key temporal features for judging pressure crosstalk based on the current construction progress, thereby improving the relevance of the embedded representation.

[0136] S3.4, the LSTM hidden states are weighted and summed based on the attention weights to obtain the final fixed-dimensional embedding representation (in this embodiment, the embedding representation has a dimension of 128). :

[0137] ;

[0138] in, For attention weights, For LSTM The hidden state of the time step. The length of the original time series data, regardless of How it changes, The dimension is always fixed, in this embodiment The dimension is always fixed at 128, which solves the problem of standardized representation of variable-length time series data.

[0139] S3.5 In the prototype matching layer, the association between the embedded representation and the compression type is established through the process of "prototype vector definition - similarity calculation - probability output" to ensure the interpretability of the prediction results.

[0140] Specifically, step S3.3 includes the following steps:

[0141] S3.3.1, Calculate the construction schedule The calculation formula is:

[0142] ;

[0143] in, This represents the actual length of the current time series data. The total data length is estimated based on the construction plan.

[0144] S3.3.2, will Discretize into One progress bar (default) (i.e., 0~10%, 10~20%, ..., 90~100%), to obtain the progress box. In this embodiment, Discretize into 10 progress bins (0~10%, 10~20%, ..., 90~100%) to obtain the progress bins. .

[0145] S3.3.3, via a fully connected network Encoded as query vector The key features used to characterize the current progress are calculated using the following formula:

[0146] ;

[0147] in, As a schedule weight, This is a bias term.

[0148] S3.3.4, Calculate the query vector and The similarity is used to obtain the attention weight. The calculation formula is:

[0149] ;

[0150] in, Here is the attention weight matrix. For query vector transpose, For LSTM The hidden state of the time step.

[0151] Specifically, step S3.5 includes the following steps:

[0152] S3.5.1, Types of inter-well pressure channeling Under each progress bin, define a set of trainable prototype vectors. . It is a model pair Type compression in A condensation of typical characteristics of the progress stage, its dimensions and... Consistent. A category center initialization strategy is adopted, assigning categories belonging to the training set... And the progress falls into the progress box samples Construction data Take the average as initial value The calculation formula is:

[0153] ;

[0154] in, For training focus belongs to And in the progress box The number of samples, As the index set for the corresponding samples, this embodiment ultimately yields 50 prototype vectors. (k=1~5, b=1~10).

[0155] S3.5.2, Quantifying and embedding current construction data With the current progress box Prototypes of each category similarity The higher the similarity, the more the current data conforms to the typical characteristics of this type of compression crosstalk. The calculation formula is:

[0156] ;

[0157] in, Representation of embedding The L2 norm (Euclidean norm). express The L2 norm (Euclidean norm). The range of values ​​is , Indicates an exact match. Indicates no association. Indicates a complete reversal.

[0158] S3.5.3, the similarity vector is converted into a probability distribution of the compression type using the Softmax function. The calculation formula is:

[0159] ;

[0160] in, This is a temperature parameter used to control the sharpness of the probability distribution, with a value of 0.9. It is a natural exponential function. For progress box Down The class's compressed prototype vector.

[0161] Specifically, step S3.5 also includes the following steps:

[0162] S3.5.4 employs a composite loss function to train the model and optimizes network parameters (including 1D-CNN convolutional kernels, LSTM gating parameters, attention weights, prototype vectors, etc.) using the gradient descent algorithm. The specific design is as follows:

[0163] Total loss function Classification loss based on cross-entropy Compared with prototype clustering loss Weighted composition. Cross-entropy classification loss. The probability distribution used to constrain the model output to match the true labels of the samples, ensuring classification accuracy, is calculated using the following formula:

[0164] ;

[0165] ;

[0166] in, For the sample The true label (one-hot encoding). This represents the batch size, with a value of 256. For the schedule weight function, This is a hyperparameter with a value of 0.6, used to assign higher weights to samples with high construction progress to optimize the model's predictive performance in the early stages of construction. For the sample Construction progress; Represents each type of inter-well pressure channeling The corresponding serial number.

[0167] S3.5.5, Prototype Clustering Loss The Euclidean distance between the embedding representations of similar samples and their corresponding prototype vectors is used to constrain the representativeness of the prototype vectors. The calculation formula is as follows:

[0168] ;

[0169] in, For the sample Construction data, For progress box Lower sample The prototype vector of the type.

[0170] Total loss function The calculation formula is:

[0171] ;

[0172] in, This is a hyperparameter used to balance the two types of loss, and its value is 0.4.

[0173] Specifically, by collecting construction data in real time, using dynamic operation models, and quantifying confidence levels, progressive prediction and risk warning of pressure surge types are achieved, providing precise support for on-site construction decisions. Step S4 specifically includes the following steps:

[0174] S4.1, For the ongoing fracturing operation to be predicted, from the start time (t0=0) to the current time... Real-time collection of well discharge volume Construction well pressure Observation well pressure The collected data is processed in real time (including time alignment and standardization) to form the current time-series data sequence. And calculate the current construction progress. and the progress box corresponding to the current construction progress. .

[0175] S4.2, the current time series data sequence and Generate real-time embedding representations from a pre-trained temporal prototype network model. , call Calculate the following 5 types of prototype vectors. The cosine similarity with each prototype vector is used to obtain the probability distribution of each compression type at the current time through the Softmax function. The compression type corresponding to the highest probability is selected as the current prediction result. The calculation formula is:

[0176] ;

[0177] in, This is the maximum value index function.

[0178] S4.3, Calculate the overall confidence level of the current prediction. :

[0179] ;

[0180] ;

[0181] in, It is the highest probability value currently predicted. Based on The confidence calibration function, The sensitivity coefficient is set to 6.

[0182] As construction progresses, the current time-series data sequence Continuously growing, By continuously increasing the value and repeating step S4, the prediction result of the compression type can be obtained. With confidence level The coordinated dynamic evolution. Complete pressure curves of adjacent observation wells, as shown... Figure 8 As shown, the prediction results for key time nodes are as follows:

[0183] At 15 minutes into construction, the progress is 8.57%, belonging to the first progress box, with a probability distribution of [0.496, 0.181, 0.161, 0.101, 0.061]. The prediction result is Type I, with a probability of 0.496 and a confidence level of 0.233.

[0184] At 30 minutes into construction, the progress is 17.14%, belonging to the second progress box, with a probability distribution of [0.172, 0.193, 0.482, 0.105, 0.048]. The prediction result is Type III, with a probability of 0.482 and a confidence level of 0.347.

[0185] At 45 minutes of construction, the progress is 25.71%, belonging to the 3rd progress box, with a probability distribution of [0.098, 0.123, 0.612, 0.107, 0.060]. The prediction result is Type III, with a probability of 0.612 and a confidence level of 0.526.

[0186] At 60 minutes of construction, the progress is 34.29%, belonging to the 4th progress box, with a probability distribution of [0.075, 0.092, 0.685, 0.088, 0.060]. The prediction result is Type III, with a probability of 0.685 and a confidence level of 0.637.

[0187] At 75 minutes of construction, the progress is 42.86%, belonging to the 5th progress box, with a probability distribution of [0.053, 0.071, 0.763, 0.072, 0.041]. The prediction result is Type III, with a probability of 0.763 and a confidence level of 0.732.

[0188] At 90 minutes of construction, the progress is 51.43%, belonging to the 6th progress box, with a probability distribution of [0.032, 0.048, 0.857, 0.041, 0.022]. The prediction result is Type III, with a probability of 0.857 and a confidence level of 0.840.

[0189] At 105 minutes of construction, the progress is 60.00%, belonging to the 7th progress box, with a probability distribution of [0.021, 0.035, 0.896, 0.031, 0.017]. The prediction result is Type III, with a probability of 0.896 and a confidence level of 0.887.

[0190] At 120 minutes of construction, the progress is 68.57%, belonging to the 7th progress box, with a probability distribution of [0.015, 0.024, 0.928, 0.022, 0.011]. The prediction result is Type III, with a probability of 0.928 and a confidence level of 0.923.

[0191] At 135 minutes of construction, the progress is 77.14%, belonging to the 8th progress box, with a probability distribution of [0.009, 0.016, 0.954, 0.014, 0.007]. The prediction result is Type III, with a probability of 0.954 and a confidence level of 0.952.

[0192] After 150 minutes of construction, the progress is 85.71%, belonging to the 9th progress box, with a probability distribution of [0.005, 0.010, 0.978, 0.006, 0.001]. The prediction result is Type III, with a probability of 0.978 and a confidence level of 0.977.

[0193] At 165 minutes of construction, the progress is 94.29%, belonging to the 10th progress box, with a probability distribution of [0.002, 0.003, 0.995, 0.000, 0.000]. The prediction result is Type III, with a probability of 0.995 and a confidence level of 0.995.

[0194] The dynamic evolution of the predicted probability and confidence level of pressure channeling type over construction time is shown in the figure below. Figure 9 As shown.

[0195] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A time-series prototype network-based progressive prediction method for interwell fracture pressure channeling, characterized in that, Specifically comprising the following steps: S1, based on the pressure response curve shape and numerical characteristics of the observed well in the historical construction, combined with the reservoir fluid percolation theory and rock mechanical properties, the interwell channeling is divided into multiple types, and a fracturing channeling type classification standard is constructed; S2, based on the fracturing channeling type classification standard of step S1, the historical fracturing construction time series data is systematically labeled and preprocessed to form a labeled data set required for training the time series prototype network model, and the data is preprocessed; S3, a time series prototype network model suitable for fracturing variable length time series data is designed, including a feature encoder and a prototype matching layer, the fixed dimension embedding of time series data is realized through the feature encoder, the association between features and channeling types is established through the prototype matching layer, and the model training is completed through the optimization of the composite loss function; S4, the trained time series prototype network model is used for progressive prediction and risk warning of channeling type; Step S1 specifically comprises the following steps: S1.1, the interwell channeling type includes: weak transmission or no effective communication type, pore or microfracture transmission type, fracturing fracture progressive communication type, fracturing fracture direct communication type and stress activated natural fracture type; S1.2, three-dimensional quantitative judgment rules based on pressure numerical characteristics, curve shape characteristics and derivative characteristics are formulated to distinguish different interwell channeling types; S1.3, statistical analysis is performed on all the observed well pressure response data in the historical data set, a threshold search range is set according to the reservoir characteristics of the target block, and a plurality of candidate threshold combinations are generated based on the search range through a grid search method, the highest weighted Kappa coefficient between the automatic classification results under different threshold combinations and the artificial identification results of the field experts is the optimization target, and the threshold combination is iteratively optimized by combining a particle swarm optimization algorithm PSO, and finally the optimal threshold suitable for the target block is determined ; Step S3 specifically comprises the following steps: S3.1, the local correlation features in the short time window of the fracturing time series data are captured through one-dimensional convolution 1D-CNN, and the high-frequency noise is filtered, and the calculation formula of 1D-CNN is as follows: ； wherein, represents a one-dimensional convolution operation, is a convolution kernel weight, is a fracturing timing data, is a bias term, is an extracted local feature, is an activation function; S3.2, to An input long short-term memory network (LSTM) is used to capture the long-time dependence of the fracturing time series data. ； in, For LSTM Input of time steps, For LSTM The hidden state of the time step. It is the sigmoid activation function. It represents the Hadamah accumulation. Input gates Forgotten Gate Cell state With output gate The weight matrix, Input gates respectively Forgotten Gate Cell state With output gate The bias term, tanh is the hyperbolic tangent function; S3.3, the progress perception attention mechanism is used to dynamically extract the time series features critical to channeling judgment according to the current construction progress; S3.4, weighting sum of LSTM hidden states based on attention weights to get the final fixed-dimension embedding representation : ； wherein, is the attention weight, is the hidden state of the LSTM at the time step, is the original time series data length; S3.5, the association between embedded representation and channeling type is established in the prototype matching layer; Step S4 specifically comprises the following steps: S4.1, for the ongoing fracturing operation to be predicted, from the starting time to the current time , real-time acquisition of the construction well discharge , construction well pressure , observation well pressure , and real-time processing of the acquired data to form a current time series data sequence , and calculate the current construction progress and the progress box corresponding to the current construction progress ; S4.2, the current time series data sequence and Input the trained time series prototype network model to generate real-time embedding representation , call 5 prototype vectors under, calculate The cosine similarity of each prototype vector, get the probability distribution of each channeling type at the current time through the Softmax function , select the channeling type corresponding to the maximum probability as the current prediction result , the calculation formula is: ； wherein is a maximum index function; S4.3, compute the overall confidence of the current prediction : ； ； wherein, is the maximum probability value of the current prediction, is a confidence calibration function based on is a sensitivity coefficient.​ 2. The time-series prototype network-based progressive prediction method for interwell fracture pressure channeling of claim 1, wherein, Step S1.2 specifically is: When the maximum pressure rise in the observation well is lower than the first threshold it is determined to be weak transmission or no effective communication type, i.e. ; wherein , is the observation well pressure, is the construction start time, is the time step, is the maximum value function; When and the average rate of pressure rise is lower than the second threshold value , it is determined to be a pore or micro-fracture transmission type, i.e. and ; wherein, , is the time when the pressure reaches the peak value; When the pressure curve shows "S" shape or step growth, it is determined as fracturing fracture progressive communication type, and the quantization standard is that the first derivative sequence of the pressure curve exists wide peak; When and is determined as a direct communication type of fracture; wherein , , and are a third threshold value and a fourth threshold value, respectively. When the pressure curve rises rapidly after the stable segment or the decline segment, it is determined as stress activated natural fracture type, and the quantization standard is that the second derivative of the pressure curve appears positive pulse in the later stage of construction.

3. The time-prototype network based progressive prediction method of interwell fracture pressure communication according to claim 1, wherein, Step S2 specifically comprises the following steps: S2.1, collect time series data of construction well rate , construction well pressure , and observation well pressure , and align by time step; S2.2, data preprocessing, including: using linear interpolation method to deal with data missing, ensuring the continuity of time series data, and 、 and respectively, Z-score standardization, Z-score standardization formula as follows: ； wherein, is the data mean, is the data standard deviation, and are the data before and after normalization, respectively. S2.3, according to the fracturing channeling type classification standard, for each piece of pre-processed time series data sequence Tagging with channeling type , the label is in one-hot encoding format; S2.4, the labeled data set is divided into training set, validation set and test set.

4. The time-prototype network based progressive prediction method of interwell pressure communication during fracturing according to claim 1, characterized in that, Step S3.3 specifically comprises the following steps: S3.3.1, Calculate construction progress The calculation formula is: ； wherein, is the actual length of the current timing data, is the total data length estimated based on the construction plan; S3.3.2, to discretize into progress bins, resulting in progress bins ; S3.3.3, by a fully connected network encoding into a query vector , for characterizing the feature focus of the current progress, the formula is: ； wherein, is a progress weight, is a bias term; S3.3.4, compute query vector with similarity to , the formula is: ； wherein, is an attention weight matrix, is a query vector is the transpose of is the hidden state of the LSTM at time step.

5. The time-prototype network based progressive prediction method of interwell fracture pressure communication of claim 1, wherein, Step S3.5 specifically comprises the following steps: S3.5.1, for each interwell pressure channeling type Under each progress bin, a set of trainable prototype vectors is defined With the category center initialization strategy, the construction data of the samples belonging to the class and whose progress falls into the progress bin are averaged to obtain the initial value of the prototype vector, and the calculation formula is as follows: , for each prototype vector​​ ； wherein, is the number of samples in the training set belonging to class and in the progress bin , is the index set of the corresponding samples; S3.5.2, quantization with the degree of similarity , the calculation formula is: ； wherein, represents the L2 norm of represents the L2 norm of the value range of , represents a perfect match, represents no association, represents a perfect reverse; S3.5.

3. Convert the similarity vector into a probability distribution of the type of the burst by a Softmax function The calculation formula is: ； wherein, is a temperature parameter, is a natural exponential function, is a progress bin Down Class of channeling prototype vectors.

6. The time-prototype network-based progressive prediction method of interwell fracture pressure channeling according to claim 5, wherein, Step S3.5 further comprises the following steps: S3.5.4, the model is trained with a composite loss function, the network parameters are optimized by gradient descent algorithm, and the cross entropy classification loss The probability distribution of the constrained model output is consistent with the sample true label, and the classification accuracy is ensured. The calculation formula is: ； ； in, For the sample The true label, For batch size, For the schedule weight function, For hyperparameters, For the sample Construction progress; Represents each type of inter-well pressure channeling The corresponding serial number; S3.5.5, prototype clustering loss To constrain the Euclidean distance between the embedding representation of a same-class sample and the corresponding prototype vector, the formula is as follows: ； in, For the sample Construction data, For progress box Lower sample The prototype vector of the type; Total loss function The formula for calculating the total loss function is: ； wherein is a hyperparameter.

Images