An adaptive migration well selection method and related equipment
By constructing a well selection model that includes a Transformer encoder and a domain discriminator, and using adversarial training techniques to align the feature spaces of the source and target domains, the problem of insufficient model generalization ability in well selection during repeated fracturing in cross-block oilfields is solved, thus improving well selection accuracy and efficiency.
Patent Information
- Application Number
- CN202611599880.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-15
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies have insufficient model generalization ability in cross-block oilfield repeated fracturing well selection, and cannot effectively adapt to the differences in characteristic distribution between different blocks, resulting in insufficient prediction accuracy and low well selection efficiency.
An adaptive migration well selection method is adopted. By constructing a well selection model that includes a Transformer encoder, decoder and a domain discriminator, the model is trained under supervision using source domain data and adversarial training is performed between the Transformer encoder and the domain discriminator until the source domain and target domain data are aligned in the feature space, thus achieving smooth cross-domain feature transfer.
The accuracy of repeated fracturing well selection was improved under low-label sample conditions, the well selection error rate was reduced, the input-output ratio of fracturing stimulation was guaranteed, and intelligent well selection across reservoir blocks was realized.
Smart Images

Figure CN123118294A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of oilfield development technology, specifically relating to an adaptive well migration selection method and related equipment. Background Technology
[0002] As old oilfield development enters its mid-to-late stages, single-well productivity often shows a continuous decline. Implementing repeated fracturing operations has become a key means to maintain stable production in old areas and improve overall recovery rates. In existing repeated fracturing well selection operations, early methods primarily relied on the manual experience of professionals, using comprehensive analysis of geological conditions, well history records, and other characteristics to formulate production enhancement measures. In recent years, with the widespread application of deep learning technology in reservoir enhancement, neural network architectures, represented by Transformer, have begun to be widely used in intelligent well selection models due to their excellent performance in time-series modeling and cross-task feature extraction. Currently, mainstream intelligent well selection schemes typically collect a large number of samples with historical production sequences and core construction parameters, train them using models such as Transformer, and attempt to establish a mapping relationship between fracturing parameters and production response. This assists engineers in selecting the target wells with the greatest production potential from a vast number of oil wells, aiming to improve well selection efficiency.
[0003] However, current well selection technologies have significant limitations and pressing technical problems in practical cross-block applications. On the one hand, traditional screening logic based on human experience is relatively simple, with limited adaptability to heterogeneous features such as geology and well history, resulting in low effectiveness and a complete lack of knowledge transfer capabilities for cross-block linkage. On the other hand, when existing Transformer models are directly applied to cross-block well selection tasks, they often face serious inconsistencies in domain distribution. Due to significant differences in formation characteristics and process parameter distributions among different oilfield blocks, models trained on existing labeled data in the source block (source domain) are difficult to directly generalize to new blocks (target domains) with different data distributions. The current problem is that, given the lack of labeled data in the target domain, existing technologies cannot effectively align the feature distribution spaces between different blocks, leading to technical bottlenecks such as insufficient prediction accuracy and limited generalization ability when the model is transferred and applied across blocks, making it difficult to truly achieve intelligent well selection for cross-block linkage under unsupervised conditions. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide an adaptive migration well selection method and related equipment to address the shortcomings of the prior art, and to solve the technical problem that a high-precision repeated fracturing well selection model cannot be constructed in low-development oilfield blocks due to insufficient historical fracturing data.
[0005] The objective of this invention is achieved through the following technical solutions: In a first aspect, the present invention provides an adaptive migration well selection method, comprising: Acquire source domain data and target domain data. The source domain data includes historical production sequence data and fracturing operation parameter data of source domain oil wells. The target domain data includes historical production sequence data and fracturing operation parameter data of target domain oil wells. Construct a well selection model, which includes a Transformer encoder, a decoder, and a neighborhood discriminator; The source domain data and the target domain data are respectively input into the well selection model. The well selection model is trained under supervision using the source domain data. Adversarial training is then conducted between the Transformer encoder and the neighborhood discriminator using the source domain data and the target domain data until the source domain data and the target domain data are aligned in the feature space, thus obtaining the trained well selection model. The actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested are obtained, and the actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested are input into the well selection model that has been trained. The repeated fracturing recommendation score of the oil wells in the target area to be tested is output, and the target operating well is determined based on the repeated fracturing recommendation score. Adversarial training is performed between the Transformer encoder and the neighborhood discriminator using the source domain data and the target domain data until the source domain data and the target domain data are aligned in the feature space, including: Construct a domain discriminator loss function for the domain discriminator, wherein the optimization objective of the domain discriminator loss function is to maximize the accuracy of the domain discriminator in identifying the domain category to which the input features belong; Construct an encoder back optimization loss function, the optimization objective of which is to minimize the discrimination accuracy of the domain discriminator in determining the domain category to which the target domain input features belong; Obtain the supervised classification loss for the source domain data, and then combine the supervised classification loss with the encoder's back optimization loss function in a weighted manner to construct a joint optimization objective function; The source domain data and the target domain data are input into the Transformer encoder to generate source domain mapping features and target domain mapping features respectively. The source domain mapping features and the target domain mapping features are then input into the domain discriminator for domain category discrimination. The network parameters of the Transformer encoder are updated based on the joint optimization objective function, and the network parameters of the neighborhood discriminator are updated based on the neighborhood discriminator loss function; The above parameter update steps are executed alternately, so that the domain discriminator and the Transformer encoder form an adversarial game until the source domain mapping features and the target domain mapping features are aligned in the feature space.
[0006] As a further improvement of the present invention, before inputting the source domain data and the target domain data into the well selection model respectively, the method further includes processing the fracturing operation parameter data to construct an input tensor, specifically including: Extract the actual fluid injection volume, actual sand addition volume, and actual sand addition intensity from the fracturing construction parameter data; The actual liquid inflow, actual sand addition, and actual sand addition intensity are broadcast along the time dimension to generate a parameter matrix of the same length as the historical production sequence data. The historical production sequence data and the parameter matrix are concatenated along the feature dimension to generate an input tensor.
[0007] As a further improvement of the present invention, the step of acquiring source domain data and target domain data further includes an outlier cleaning step based on the 3σ principle: Calculate the mean and variance of the source domain data and the target domain data; Determine whether the data points corresponding to the source domain data and the target domain data meet the first condition. If the first condition is met, the corresponding data points are determined to be outliers and are cleaned. If the first condition is not met, the corresponding data points are filtered. The first condition is:
[0008] In the formula, For data indexing, For indexing source domain data, An index for the target domain data; These are the data points corresponding to the source domain data and the target domain data. The mean of the source domain data or the target domain data. The variance of the source domain data or the target domain data.
[0009] As a further improvement of the present invention, the loss function of the domain discriminator is:
[0010] In the formula, The output of the loss function for the neighborhood discriminator. The number of data points in the source domain data. The number of data points in the target domain data. For data points in the target domain, For data points in the source domain, For Transformer encoders, ( ) represents the output of the neighborhood discriminator; The encoder inverse optimization loss function is:
[0011] In the formula, The loss function is optimized inversely for the encoder; The joint optimization objective function is:
[0012] In the formula, To jointly optimize the objective function, To monitor the classification of losses, The weight hyperparameters are used to inversely optimize the loss function for the encoder.
[0013] As a further improvement of the present invention, during the process of alternately executing the above parameter update steps, a feature alignment verification and early stopping step are also included: Monitor the discrimination accuracy of the domain discriminator; If the discrimination accuracy of the domain discriminator drops to a preset random level, it is determined that the source domain mapping features and the target domain mapping features are aligned in the feature space. If the supervised classification loss is detected to be continuously increasing, an early stop mechanism is triggered to terminate adversarial training.
[0014] As a further improvement of the present invention, the step of determining the target operating well based on the repeated fracturing recommendation score specifically includes: The recommended scores for repeated fracturing of the oil wells in the target area to be tested are sorted from high to low; The top K ranked oil wells in the target domain to be tested are selected as the target operating wells, where K is a positive integer; While outputting the target operating well, the confidence level of the trained well selection model and the ranking of key feature weights of the actual fracturing parameters are also output simultaneously.
[0015] As a further improvement of the present invention, after determining the target operating well based on the repeated fracturing recommendation score, the method further includes a step of updating the trained well selection model, specifically including: When the target domain accumulates a preset number of new oil well fracturing production feedback data, a periodic fine-tuning mechanism is triggered to freeze the parameters of the front network layer of the Transformer encoder and use the production feedback data to perform offline fine-tuning and updates of the decoder and the domain discriminator. When fracturing response data from a new single oil well is received, an online incremental learning mechanism is triggered to perform incremental updates based on the fracturing response data to adjust the weight parameters of the domain discriminator.
[0016] Secondly, the present invention provides an adaptive migration well selection system, comprising: The data acquisition module is used to acquire source domain data and target domain data. The source domain data includes historical production sequence data and fracturing operation parameter data of source domain oil wells; the target domain data includes historical production sequence data and fracturing operation parameter data of target domain oil wells. The model building module is used to build a well selection model, which includes a Transformer encoder, a decoder, and a neighborhood discriminator. The model training module is used to input the source domain data and the target domain data into the well selection model respectively, perform supervised training on the well selection model using the source domain data, and perform adversarial training between the Transformer encoder and the neighborhood discriminator using the source domain data and the target domain data until the source domain data and the target domain data are aligned in the feature space, thereby obtaining the trained well selection model. The well selection recommendation module is used to obtain the actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested, and input the actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested into the trained well selection model, output the repeated fracturing recommendation score of the oil wells in the target area to be tested, and determine the target operating well based on the repeated fracturing recommendation score.
[0017] Thirdly, the present invention provides a computer device, comprising: a processor and a computer-readable storage medium; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program that, when executed by the processor, implements the adaptive migration well selection method as described above.
[0018] Fourthly, the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed as described above in the adaptive migration well selection method.
[0019] The beneficial effects of this invention are as follows: This invention provides an adaptive migration well selection method. By acquiring historical production sequence data and fracturing operation parameter data of source domain oil wells and historical production sequence data and fracturing operation parameter data of target domain oil wells, it provides full-dimensional basic data support for subsequent model training of two types of oil wells, avoiding the training sample bias problem caused by incomplete single-domain data coverage in existing technologies. By constructing a well selection model including a Transformer encoder, decoder, and domain discriminator, it accurately extracts production evolution features based on the Transformer's ability to capture long-distance dependencies of time-series production sequences. The decoder completes the mapping output from features to fracturing recommendation scores, and the domain discriminator achieves dynamic recognition of cross-domain feature differences without the need to introduce additional complex feature adaptation modules. By inputting source domain data and target domain data into the well selection method, the method achieves the desired results. The well model utilizes source domain data for supervised training of the well selection model. It can perform basic model fitting based on known source domain well fracturing effect labels, ensuring the model's basic accuracy in recognizing fracturing patterns within a single domain. Through adversarial training between the Transformer encoder and the neighborhood discriminator using source and target domain data until the source and target domain data are aligned in the feature space, it eliminates the distribution offset between source and target domain well data without relying on a large number of labeled target domain samples, achieving smooth cross-domain feature transfer. By acquiring the actual fracturing parameters and historical production sequences of the target domain wells and inputting them into the trained well selection model, it outputs a repeated fracturing recommendation score for the target domain wells. Based on this score, it determines the target operating wells, directly outputting quantified and sortable recommendation results, avoiding the subjective bias of manual well selection in existing technologies. This invention collectively improves the accuracy of repeated fracturing well selection across reservoir blocks under low-label sample conditions, effectively reducing the well selection error rate for repeated fracturing operations and ensuring the return on investment of fracturing stimulation. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the 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.
[0021] Figure 1 This is a schematic diagram of the adaptive migration well selection method in an embodiment of the present invention; Figure 2 This is a feature distribution diagram (principal component visualization) before and after domain adaptation in an embodiment of the present invention. The left side shows the feature distribution before domain adaptation, where the source and target domains show significant differences; the right side shows the feature distribution after domain adaptation, where the feature distributions of the two domains become significantly closer, achieving feature alignment. Figure 3 This is a performance comparison chart of the model before and after transfer learning in an embodiment of the present invention. Yellow bars indicate the absence of a domain adaptation method, while orange bars indicate the use of the method of the present invention (domain-adaptive Transformer). In all metrics, the model after domain adaptation significantly outperforms the traditional method, improving generalization ability.
[0022] Figure 4 This is a schematic diagram of the training loss curves of the source domain and the target domain in an embodiment of the present invention.
[0023] Figure 5 This is an internal structural diagram of a computer device in an embodiment of the present invention. Detailed Implementation
[0024] To make the objectives and technical solutions of this invention clearer and easier to understand, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.
[0025] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. The described embodiments are only some embodiments of the present invention, and not all embodiments.
[0026] Example 1 In the later stages of development of old oilfields, the effectiveness of relying on manual experience for repeated fracturing well selection is relatively low. Existing Transformer-based intelligent well selection models, when directly migrating across blocks, suffer from domain shifts due to differences in formation and process parameter distribution across different blocks. Furthermore, when the target domain lacks labeled samples, the model's prediction accuracy and generalization ability are insufficient, making unsupervised intelligent well selection across blocks impossible. The technical problem this independent claim aims to solve is to complete the knowledge transfer from the source domain to the target domain, outputting a repeated fracturing recommendation score for the well to be tested, and selecting the target operating well, even when the target domain lacks labeled samples of fracturing effects.
[0027] This application presents an adaptive migration well selection method, the execution process of which is as follows: Figure 1 As shown, the steps include the following.
[0028] S1: Acquire source domain data and target domain data. Source domain data includes historical production sequence data and fracturing operation parameter data of source domain oil wells; target domain data includes historical production sequence data and fracturing operation parameter data of target domain oil wells.
[0029] The source domain is the oilfield block that has accumulated a sufficient number of well samples with fracturing effect labels, and the target domain is the new block where repeated fracturing is to be carried out and the labeled samples are scarce; the historical production sequence records the oil production and fluid production data of the wells before fracturing over time; the fracturing operation parameters characterize the scale of fracturing operation of a single well.
[0030] Before inputting the source domain data and target domain data into the well selection model, fracturing operation parameter processing is performed to construct the input tensor: Extract the actual fluid injection volume, actual proppant injection volume, and actual proppant injection intensity from the fracturing operation parameter data. The actual fluid injection volume represents the total volume of fluid injected during the fracturing operation; the actual proppant injection volume represents the amount of proppant used; and the actual proppant injection intensity represents the scale of proppant injection per unit thickness of the produced layer. These three parameters directly affect the fracture conductivity after fracturing.
[0031] The actual injection volume, actual sand injection volume, and actual sand injection intensity are broadcast along the time dimension to generate a parameter matrix of the same length as the historical production sequence data. The broadcast process replicates and expands a fixed set of fracturing scalar parameters, with the time dimension length being exactly equal to the time step T of the production sequence.
[0032] Historical production sequence data and the parameter matrix are concatenated along the feature dimension to generate the input tensor; the expression is: By broadcasting the three key fracturing parameters in the time dimension and then concatenating them with the production sequence features to construct the input tensor, the limitation of scalar construction parameters being unable to be input into the temporal neural network along with the time-series production is avoided. This achieves the fusion input of the production time-series evolution features and the physical parameters of fracturing construction, thereby improving the physical interpretability of the model's input features.
[0033] Furthermore, in the steps of acquiring source domain data and target domain data, outlier cleaning is performed based on the 3σ principle: Calculate the mean and variance of the corresponding dimensions of the source domain data and the target domain data; Determine whether the data points corresponding to the source domain data and the target domain data meet the first condition:
[0034] In the formula, For data indexing, For indexing source domain data, An index for the target domain data; These are the data points corresponding to the source domain data and the target domain data. The mean of the source domain data or the target domain data. The variance of the source domain data or the target domain data.
[0035] If a data point meets the first condition, it is determined to be an outlier and is cleaned; data points that do not meet the first condition are retained and proceed to the next step.
[0036] S2: Construct a well selection model, which includes a Transformer encoder, decoder, and neighborhood discriminator.
[0037] The Transformer encoder extracts features from the time-series input, consisting of well production and fracturing parameters, and outputs high-dimensional latent features of the well. The decoder receives the encoder's output features, performs feature mapping, and outputs a recommendation score for repeated fracturing. The domain discriminator, as an adversarial training component, is specifically used to determine whether the input latent features originate from the source or target domain. During model initialization, the Transformer network's internal hyperparameters can be configured. As one implementation, the embedding dimension is set to 32, the encoder layer to 3, the decoder layer to 2, the attention heads to 4, the feedforward layer dimension to 64, the Dropout value to 0.1, and the activation function to be GELU.
[0038] S3: Input the source domain data and target domain data into the well selection model respectively, use the source domain data to perform supervised training on the well selection model, and use the source domain data and target domain data to perform adversarial training between the Transformer encoder and the domain discriminator until the source domain data and target domain data are aligned in the feature space, and the trained well selection model is obtained.
[0039] Supervised training is performed using labeled samples from the source domain to learn fracturing parameters. The intrinsic mapping relationship between output responses is established; simultaneously, unlabeled target domain samples are fed in for adversarial training. The Transformer encoder and the domain discriminator form a game: the domain discriminator tries to distinguish which domain a feature comes from, while the encoder tries to generate features that the discriminator cannot distinguish. When the latent features after mapping the two classes of data tend to be consistent in the feature space, domain feature alignment is achieved. Figure 2 The left image shows the feature distribution before and after domain adaptation. Before adaptation, the feature distributions of the source and target domains are separated. The right image shows the feature distributions of the source and target domains mixed after adversarial training alignment is completed. Figure 4 The training loss curves for the source and target domains show that the two types of losses decrease synchronously with the increase of training rounds, indicating that the model continues to converge.
[0040] Adversarial training is performed between the Transformer encoder and the neighborhood discriminator using source and target domain data until the feature space is aligned. The steps are as follows.
[0041] The domain discriminator loss function is constructed, and its optimization objective is to maximize the discriminator's accuracy in classifying the domain to which the input features belong. The discriminator aims to distinguish as accurately as possible whether a feature originates from the source or target domain. The encoder's backpropagation loss function is then constructed, and its optimization objective is to minimize the discriminator's accuracy in classifying the domain to which the input features belong. In other words, the encoder aims to make the output features as confusing as possible to the discriminator.
[0042] The supervised classification loss for the source domain data is obtained, and then weighted and combined with the encoder's back-optimization loss function to construct a joint optimization objective function. The supervised classification loss ensures the accuracy of the source domain well selection prediction task; the encoder's back-optimization loss function is used to complete the domain feature alignment, and the weight coefficients adjust the trade-off between the two losses.
[0043] The source domain data and target domain data are input into the Transformer encoder to generate source domain mapping features and target domain mapping features, respectively; the source domain mapping features and target domain mapping features are then input into the domain discriminator to perform domain category discrimination.
[0044] The network parameters of the Transformer encoder are updated based on the joint optimization objective function; the network parameters of the domain discriminator are updated based on the domain discriminator loss function.
[0045] The parameter update steps are executed alternately, with the domain discriminator and Transformer encoder forming an adversarial game that iterates until the source domain mapping features and target domain mapping features are aligned in the feature space. The two sets of network parameters are updated iteratively using discriminator loss, encoder inverse loss, and joint loss to form an adversarial game. This avoids the limitation of supervised training alone in eliminating cross-domain distribution shifts, achieving latent feature alignment with unlabeled target domain samples and completing cross-block knowledge transfer.
[0046] The domain discriminant loss function is:
[0047] In the formula, The output of the loss function for the neighborhood discriminator. The number of data points in the source domain data. The number of data points in the target domain data. For data points in the target domain, For data points in the source domain, For Transformer encoders, ( ) represents the output of the neighborhood discriminator; The encoder back-optimization loss function is:
[0048] In the formula, The loss function is optimized inversely for the encoder; The joint optimization objective function is:
[0049] In the formula, To jointly optimize the objective function, To monitor the classification of losses, The weight hyperparameters are used to inversely optimize the loss function for the encoder.
[0050] During the alternating parameter update step, feature alignment verification and early stopping steps are performed: the discrimination accuracy of the domain discriminator is monitored; if the discrimination accuracy of the domain discriminator drops to a preset random level, it is determined that the feature space alignment of the source domain mapping features and the target domain mapping features is complete, and the adversarial iteration can be exited; if the supervised classification loss is detected to continue to rise, the early stopping mechanism is triggered to terminate the adversarial training.
[0051] S4: Obtain the actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested, and input the actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested into the trained well selection model, output the repeated fracturing recommendation score of the oil wells in the target area to be tested, and determine the target operating well based on the repeated fracturing recommendation score.
[0052] After training is complete, the model network parameters are frozen, and historical production time series data of the well to be tested, along with parameters of the planned or historical fracturing operations, are fed into the model; the model outputs 0. The recommended score for repeated fracturing is determined by interval or normalization. The higher the score, the higher the potential for increased production from repeated fracturing of the oil well. Candidate operating wells are selected based on the score.
[0053] The source domain has a sufficient number of labeled samples, and supervised training allows the encoder to... The decoder learns the mapping from well time-series features to production potential scores. During adversarial training, the domain discriminator continuously learns to identify differences in features between the source and target domains. To confuse the discriminator, the Transformer encoder actively maps source and target domain samples to the same feature distribution space, eliminating domain offset caused by cross-block operations. In the inference phase, the encoder extracts features from the target domain wells after alignment, and the decoder outputs a sortable recommendation score for repeated fracturing, completing cross-block well selection.
[0054] Target wells are determined based on repeated fracturing recommendation scores, and the following operations are performed: The repeated fracturing recommendation scores of the wells in the target area are sorted from highest to lowest; the top K wells in the target area are selected as target wells, where K is a positive integer. In engineering, K can be set according to the oilfield operation plan, for example, K can be 5 or 8; Simultaneously, the confidence level of the well selection model after training is output, as well as the ranking of the key feature weights of the actual fracturing parameters. The model confidence level represents the credibility of the inference results; the ranking of key feature weights tells engineers which fracturing parameters have the greatest impact on the well's production potential, assisting in engineering decision-making.
[0055] The system extracts the top K wells by sorting them in descending order of scores, and outputs confidence levels and key feature weight rankings. This avoids the shortcomings of simply outputting a list of candidate wells without interpretable evidence, which is not conducive to engineers' review. It achieves quantitative screening of target operating wells and provides a basis for well selection reasoning, which is adapted to the usage habits of oilfield engineering operations.
[0056] Furthermore, after identifying the target wells, the well selection model is updated: when the target domain accumulates a preset number of new well fracturing production feedback data, a periodic fine-tuning mechanism is triggered; the parameters of the front network layers of the Transformer encoder are frozen, and the decoder and domain discriminator are fine-tuned and updated offline using the production feedback data. As an engineering example, the preset cumulative number can be set to 5 wells; the front of the encoder retains the learned cross-domain general features, updating only the output and discrimination parts to avoid forgetting the learned knowledge. In this embodiment, the front network layers are several network layers of the Transformer encoder near the input side.
[0057] Upon receiving fracturing response data from a new single oil well, an online incremental learning mechanism is triggered. This data is input into the well selection model, freezing all network parameters of the Transformer encoder and decoder. An incremental loss function is constructed based on the fracturing response data, and incremental updates are performed only on the weight parameters of the neighborhood discriminator, quickly detecting minute shifts in neighborhood distribution. Periodic offline fine-tuning is suitable for updating the output and discriminator network after accumulating a batch of new samples, without destroying the cross-domain general features already learned by the encoder. Online incremental learning, targeting feedback from a single well, quickly updates the discriminator, maintaining awareness of neighborhood distribution shifts, without requiring complete retraining of the entire network.
[0058] Example 2 An adaptive migration well selection system includes a data acquisition module, a model building module, a model training module, and a well selection recommendation module.
[0059] The data acquisition module is used to acquire source domain data and target domain data. Source domain data includes historical production sequence data and fracturing operation parameter data of source domain oil wells; target domain data includes historical production sequence data and fracturing operation parameter data of target domain oil wells. This module interfaces with the oilfield production database and fracturing operation record table to complete the reading and output of raw time series and operation parameters.
[0060] The model building module is used to build well selection models, which include Transformer encoders, decoders, and domain discriminators, and completes network layer instantiation and hyperparameter initialization.
[0061] The model training module is used to input source domain data and target domain data into the well selection model, respectively. It uses the source domain data to perform supervised training on the well selection model, and uses the source domain data and target domain data to perform adversarial training between the Transformer encoder and the neighborhood discriminator until the source domain data and target domain data are aligned in the feature space, thus obtaining the trained well selection model. It implements all training logic such as loss calculation, alternating parameter update, alignment verification, and early stopping.
[0062] The well selection recommendation module is used to obtain the actual fracturing parameters and historical production sequences of the oil wells in the target area to be tested, and input the actual fracturing parameters and historical production sequences of the oil wells in the target area to be tested into the trained well selection model, output the repeated fracturing recommendation score of the oil wells in the target area to be tested, and determine the target operating wells based on the repeated fracturing recommendation score, and output the candidate wells, confidence scores, and feature weight rankings.
[0063] The data flow between the various modules corresponds to all the execution steps of the method embodiment 1; the system can be implemented by software programs and deployed on a server platform.
[0064] Specific limitations regarding the adaptive migration well selection system can be found in the limitations of the adaptive migration well selection method described above, and the corresponding technical effects are equivalent, so they will not be repeated here. Each module in the aforementioned adaptive migration well selection system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.
[0065] Example 3 This embodiment uses actual well data from multiple blocks in an oilfield to implement a complete adaptive migration well selection method. The source and target domains are divided as follows: Block A, with a large number of samples, complete fracturing operation records, and fracturing production enhancement labels, is designated as the source domain; Block B, where repeated fracturing is to be performed and samples lacking fracturing effect labels, is designated as the target domain. Source domain samples are used for supervised training, while target domain samples are used for unsupervised adaptive training and model testing. The source domain A contains 25 wells that have undergone repeated fracturing; the target domain B contains 15 candidate wells, and no fracturing effect labels are found in the target domain. An exemplary basic dataset is shown in Table 1. Table 1 Basic Dataset of Oil Wells
[0066] In Table 1, the actual fluid injection volume represents the scale of the liquid volume during fracturing operations; the actual proppant addition represents the amount of proppant used, which directly affects the fracture extension morphology and fracture conductivity; and the actual proppant addition intensity measures the proppant density per unit length of the producing layer, which is related to the strength of the flow channel after fracturing is completed. These three parameters together characterize the scale of fracturing operations in a single well.
[0067] For an oil well, its historical production sequence is defined as:
[0068] In the formula, This refers to the daily liquid or oil production on day t. This represents the total time step length of the production sequence.
[0069] The fracturing operation parameter vector is defined as follows:
[0070] In the formula, These represent the actual fluid injection volume, actual sand addition volume, and sand addition intensity of this round of fracturing, respectively.
[0071] 1. Collect the monthly production sequence of each well in the source and target domains for the 12 time steps prior to fracturing; collect three parameters for each well fracturing operation: actual fluid injection rate, actual sand injection rate, and actual sand injection intensity. Perform 3σ anomaly cleaning to remove statistically significant outliers.
[0072] 2. Broadcast the three fracturing scalar parameters in the time dimension to generate a parameter matrix with a length equal to 12 time steps of the production sequence. Concatenate the production sequence and the parameter matrix in the feature dimension to obtain the input tensor for each well.
[0073] To achieve the fusion of time-series production and static construction multimodal features, the construction parameter vector is... Perform broadcast processing to generate a parameter matrix of the same length as the output sequence. :
[0074] The historical production rate and fracturing parameters are concatenated into the final input tensor:
[0075] In the formula, For the final input tensor, This indicates a concatenation operation along the feature dimension. The final input tensor will be... The input is fed into the embedding layer and position encoding module to obtain the input embedding matrix E(X), where the embedding dimension d=32.
[0076] 3. Select a Transformer encoder model with 3 layers, 2 decoder layers, 4 attention heads, 32 embedding dimensions, 64 feedforward dimensions, 0.1 dropout, and activate GELU; configure a neighborhood discriminator network; set the weight hyperparameter λ of the encoder's back-optimized loss function to 0.2.
[0077] 4. Supervised training is conducted using labeled samples from the source domain; samples from both the source and target domains are fed into the network for adversarial training; the accuracy of the domain discriminator is monitored, along with the source domain supervised classification loss; training is terminated when the discriminator accuracy drops to a random level or the source domain supervised loss continues to rise, resulting in a completed well selection model. The training loss curve is shown below. Figure 4 As shown, the source domain loss and the target domain loss decrease simultaneously; Figure 2 Principal component visualization shows that the feature spaces of the source and target domains are aligned after adversarial training is completed.
[0078] Let the source domain sample pair be ( , ),in, Assign corresponding fracturing production labels or well selection classification labels. Optimize the following loss function over the source domain:
[0079] In the formula, To optimize the following loss function over the source domain, The loss is either mean squared error or cross-entropy, depending on the target form of the prediction task (regression or classification).
[0080] Let the target domain input be It only contains historical production data and construction parameters, without label information. The model uses an adversarial mechanism to force the target domain encoder to output features. Output features of the source domain encoder It appears indistinguishable to the discriminator. Define adversarial loss. for:
[0081] In the formula, This represents the total number of samples in the source domain. This represents the total number of samples in the target domain. Input tensors to the source domain samples. For the first The source domain encoder outputs features, For the first The target domain encoder outputs features.
[0082] The encoder training objective is to minimize the prediction loss and maximize the discriminator loss, thereby "tricking" the discriminator so that it cannot determine which domain the input comes from.
[0083] Combining source domain supervised training with target domain adversarial learning into a joint optimization objective:
[0084] In the formula, Weight hyperparameters to combat loss.
[0085] 5. Input the 15 oil wells to be tested in Block B into the model and output the recommended score for repeated fracturing for each well; sort the scores in descending order, take K=5, output the top 5 target operating wells, and simultaneously output the model inference confidence and the ranking of the key feature weights of fracturing parameters.
[0086] 6. Subsequently, after collecting production feedback from 5 wells after fracturing in Block B, periodic offline fine-tuning is triggered, freezing the front of the Transformer encoder and fine-tuning the decoder and the neighborhood discriminator; whenever a new well fracturing response is received, an online incremental update is performed to adjust the weight of the neighborhood discriminator.
[0087] After each iteration, a validation process is performed: the domain class discrimination accuracy of the domain discriminator on the validation set is monitored in real time; when the discriminator's discrimination accuracy drops to a preset random level, it means that the discriminator can no longer distinguish whether the input features come from the source domain or the target domain, and it is determined that the source domain and target domain mapping features have completed feature space alignment, thus exiting the adversarial iterative training; the source domain supervised classification loss is monitored simultaneously and continuously. If the supervised classification loss continues to rise over multiple iterations, it indicates that the encoder has compromised the performance of the well selection prediction task in order to obfuscate the domain features. In this case, the early stop mechanism is immediately triggered to terminate the adversarial training process and save the current model weights as the completed training model.
[0088] Figure 2 The results shown are principal component visualizations. Before adaptation, the target domain features are far from the source domain distribution, while after adaptation, the two domains are highly mixed. Figure 3 The results demonstrate that the domain-independent adaptation method and the method of this invention significantly improve accuracy, precision, recall, and F1 score in the target domain; Figure 4 The loss curves during the training process are shown, with the loss values in both the source and target domains decreasing steadily without significant oscillations.
[0089] Given the complete absence of fracturing labeling samples in the target domain B, well selection was accomplished using knowledge from the source domain A; from Figure 3 The performance comparison bar chart shows that, compared with the model that does not use domain adversarial transfer, this solution achieves significant improvements in all indicators, including accuracy, precision, recall, and F1 score, making it suitable for repeated fracturing well selection in newly developed blocks in old areas.
[0090] As one example, in the adversarial training feature alignment verification stage, in addition to monitoring the accuracy of the domain discriminator, the maximum mean difference of MMD is introduced as the source domain. An auxiliary evaluation metric for the feature distribution distance in the target domain is provided; when the MMD distance drops to a set threshold, feature alignment can be determined to be complete, and adversarial training can be terminated. An additional set of distribution distance quantification metrics is added, which, along with the discriminator accuracy, serves as a cross-validation, further improving the robustness of the training termination condition.
[0091] As one example, whenever the target domain accumulates... Post-fracturing production feedback of wells (e.g.) It automatically triggers an offline fine-tuning, freezing all parameters of the front network layer of the Transformer encoder without updating them, and only using the new production feedback data to perform offline fine-tuning updates on the network parameters of the decoder and the domain discriminator, retaining the cross-domain general oil well features that have been learned, and avoiding catastrophic forgetting; online incremental learning mechanism: every time fracturing response data of a new well is received, a small-scale incremental update is performed, only the weight parameters of the domain discriminator are incrementally updated, maintaining the ability to perceive domain drift.
[0092] Example 4 Figure 5 An internal structural diagram of a computer device is shown in one embodiment. This computer device may specifically be a terminal or a server. Figure 5As shown, the computer device includes a processor, memory, network interface, display, camera, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an adaptive migration selection method. The display can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.
[0093] As will be understood by those skilled in the art, computer equipment Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0094] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.
[0095] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.
[0096] In summary, the adaptive migration well selection method, system, computer equipment, and storage medium provided in this application's embodiments are based on the core of obtaining historical well production and fracturing operation parameters from the source and target domains, constructing a well selection model including a Transformer encoder, decoder, and domain discriminator; supervised training is conducted in the source domain, and the encoder... The discriminator conducts adversarial game to realize the source domain. The system aligns the target domain feature space; it feeds the oil well data of the target domain to be tested into the trained model, outputs the recommended score for repeated fracturing, and filters the target operating wells; it also sets up regular offline fine-tuning and single-well online incremental learning to realize dynamic iterative updates of the model; the corresponding system modularizes the process and the computer equipment executes the entire algorithm.
[0097] The comprehensive technical benefits of this solution include: mitigating the distribution offset between different oilfield blocks, enabling unsupervised migration well selection in target domains lacking labeled samples; enhancing the physical interpretability of the model by using broadcast splicing to fuse production time series and fracturing physical parameters; possessing early training stop protection and a two-level dynamic update mechanism, making it highly practical for engineering implementation; and outputting recommendation scores, confidence levels, and feature weights to meet the business review needs of oilfield engineers and improve the accuracy of cross-block repeated fracturing well selection.
[0098] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.
[0099] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.
Claims
1. An adaptive migration well selection method, characterized in that, include: Acquire source domain data and target domain data. The source domain data includes historical production sequence data and fracturing operation parameter data of source domain oil wells. The target domain data includes historical production sequence data and fracturing operation parameter data of target domain oil wells. Construct a well selection model, which includes a Transformer encoder, a decoder, and a neighborhood discriminator; The source domain data and the target domain data are respectively input into the well selection model. The well selection model is trained under supervision using the source domain data. Adversarial training is then conducted between the Transformer encoder and the neighborhood discriminator using the source domain data and the target domain data until the source domain data and the target domain data are aligned in the feature space, thus obtaining the trained well selection model. The actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested are obtained, and the actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested are input into the well selection model that has been trained. The repeated fracturing recommendation score of the oil wells in the target area to be tested is output, and the target operating well is determined based on the repeated fracturing recommendation score. Adversarial training is performed between the Transformer encoder and the neighborhood discriminator using the source domain data and the target domain data until the source domain data and the target domain data are aligned in the feature space, including: Construct a domain discriminator loss function for the domain discriminator, wherein the optimization objective of the domain discriminator loss function is to maximize the accuracy of the domain discriminator in identifying the domain category to which the input features belong; Construct an encoder back optimization loss function, the optimization objective of which is to minimize the discrimination accuracy of the domain discriminator in determining the domain category to which the target domain input features belong; Obtain the supervised classification loss for the source domain data, and then combine the supervised classification loss with the encoder's back optimization loss function in a weighted manner to construct a joint optimization objective function; The source domain data and the target domain data are input into the Transformer encoder to generate source domain mapping features and target domain mapping features respectively. The source domain mapping features and the target domain mapping features are then input into the domain discriminator for domain category discrimination. The network parameters of the Transformer encoder are updated based on the joint optimization objective function, and the network parameters of the neighborhood discriminator are updated based on the neighborhood discriminator loss function; The above parameter update steps are executed alternately, so that the domain discriminator and the Transformer encoder form an adversarial game until the source domain mapping features and the target domain mapping features are aligned in the feature space.
2. The adaptive migration well selection method according to claim 1, characterized in that, Before inputting the source domain data and the target domain data into the well selection model, the process further includes processing the fracturing operation parameter data to construct the input tensor, specifically including: Extract the actual fluid injection volume, actual sand addition volume, and actual sand addition intensity from the fracturing construction parameter data; The actual liquid inflow, actual sand addition, and actual sand addition intensity are broadcast along the time dimension to generate a parameter matrix of the same length as the historical production sequence data. The historical production sequence data and the parameter matrix are concatenated along the feature dimension to generate an input tensor.
3. The adaptive migration well selection method according to claim 1, characterized in that, The steps of acquiring source domain data and target domain data also include an outlier cleaning step based on the 3σ principle: Calculate the mean and variance of the source domain data and the target domain data; Determine whether the data points corresponding to the source domain data and the target domain data meet the first condition. If the first condition is met, the corresponding data points are determined to be outliers and are cleaned. If the first condition is not met, the corresponding data points are filtered. The first condition is: In the formula, For data indexing, For indexing source domain data, An index for the target domain data; These are the data points corresponding to the source domain data and the target domain data. The mean of the source domain data or the target domain data. The variance of the source domain data or the target domain data.
4. The adaptive migration well selection method according to claim 1, characterized in that, The loss function of the domain discriminator is: In the formula, The output of the loss function for the neighborhood discriminator. The number of data points in the source domain data. The number of data points in the target domain data. For data points in the target domain, For data points in the source domain, For Transformer encoders, ( ) represents the output of the neighborhood discriminator; The encoder inverse optimization loss function is: In the formula, The loss function is optimized inversely for the encoder; The joint optimization objective function is: In the formula, To jointly optimize the objective function, To monitor the classification of losses, The weight hyperparameters are used to inversely optimize the loss function for the encoder.
5. The adaptive migration well selection method according to claim 1, characterized in that, The process of alternately executing the above parameter update steps also includes feature alignment verification and early stopping steps: Monitor the discrimination accuracy of the domain discriminator; If the discrimination accuracy of the domain discriminator drops to a preset random level, it is determined that the source domain mapping features and the target domain mapping features are aligned in the feature space. If the supervised classification loss is detected to be continuously increasing, an early stop mechanism is triggered to terminate adversarial training.
6. The adaptive migration well selection method according to claim 1, characterized in that, The determination of target operating wells based on the repeated fracturing recommendation score specifically includes: The recommended scores for repeated fracturing of the oil wells in the target area to be tested are sorted from high to low; The top K ranked oil wells in the target domain to be tested are selected as the target operating wells, where K is a positive integer; While outputting the target operating well, the confidence level of the trained well selection model and the ranking of key feature weights of the actual fracturing parameters are also output simultaneously.
7. The adaptive migration well selection method according to claim 1, characterized in that, After determining the target operating well based on the repeated fracturing recommendation score, the method further includes updating the trained well selection model, specifically including: When the target domain accumulates a preset number of new oil well fracturing production feedback data, a periodic fine-tuning mechanism is triggered to freeze the parameters of the front network layer of the Transformer encoder and use the production feedback data to perform offline fine-tuning and updates of the decoder and the domain discriminator. When fracturing response data from a new single oil well is received, an online incremental learning mechanism is triggered to perform incremental updates based on the fracturing response data to adjust the weight parameters of the domain discriminator.
8. An adaptive migration well selection system, characterized in that, To implement the adaptive migration well selection method according to any one of claims 1 to 7, comprising: The data acquisition module is used to acquire source domain data and target domain data. The source domain data includes historical production sequence data and fracturing operation parameter data of source domain oil wells; the target domain data includes historical production sequence data and fracturing operation parameter data of target domain oil wells. The model building module is used to build a well selection model, which includes a Transformer encoder, a decoder, and a neighborhood discriminator. The model training module is used to input the source domain data and the target domain data into the well selection model respectively, perform supervised training on the well selection model using the source domain data, and perform adversarial training between the Transformer encoder and the neighborhood discriminator using the source domain data and the target domain data until the source domain data and the target domain data are aligned in the feature space, thereby obtaining the trained well selection model. The well selection recommendation module is used to obtain the actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested, and input the actual fracturing parameters and historical production sequence of the oil wells in the target area to be tested into the trained well selection model, output the repeated fracturing recommendation score of the oil wells in the target area to be tested, and determine the target operating well based on the repeated fracturing recommendation score.
9. A computer device, characterized in that, include: Processor and computer-readable storage media; A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program that, when executed by the processor, implements the adaptive migration well selection method as described in any one of claims 1 to 7.
