A method and system for masking irregular production data of oil well repeated fracturing

Abstract

The present application belongs to the field of reservoir stimulation and reconstruction, and discloses a mask prediction method and system for irregular production data of oil well repeated fracturing. The method obtains oil well repeated fracturing production time series data and generates time series embedding representation, retaining original time series characteristic information. The method calculates the production change rate at each time point, screens out data high change points, performs mask processing on the time series embedding representation corresponding to the high change points, and reconstructs the embedding representation after the mask to obtain reconstructed time series data. According to the difference between the original production time series data and the reconstructed time series data, the reconstruction error is calculated and minimized, the total loss is constructed by combining self-supervised learning to optimize the model parameters, and the self-supervised training of the time series prediction model is completed. Finally, the to-be-predicted time series data is input into the trained model, and the production prediction value is output. The present application effectively improves the irregular time series data feature extraction capability and production prediction accuracy by fully mining the effective features of irregular production data, and is suitable for oil well repeated fracturing production dynamic prediction scene.

Description

Technical Field

[0001] This invention belongs to the field of reservoir enhancement and stimulation technology, specifically relating to a masking prediction method and system for irregular production data from repeated fracturing of oil wells. Background Technology

[0002] Oil well repeated fracturing production data are often collected and recorded at irregular time intervals, with varying observation frequencies, sometimes daily and sometimes monthly. This irregular time-series recording method can indirectly reflect on-site production control behavior and contains exploitable feature information, making it highly valuable for engineering prediction scenarios such as final recovery rate prediction, production decline prediction, and production risk assessment.

[0003] Existing research on irregular time series modeling mainly falls into two categories: developing specialized architectures to handle irregularities, or designing effective pre-training tasks for self-supervised learning. Despite some progress, processing irregular time series data still faces numerous challenges. First, due to the uneven recording time intervals, time series observations are often asynchronous, making it difficult for models to identify patterns and capture temporal changes. Second, even if data features are recorded synchronously, uneven observation intervals can still hinder pattern recognition. Finally, the scarcity of labeled data further limits the reliability of model training results.

[0004] In summary, existing modeling methods are difficult to adapt to the irregular time-series data of repeated fracturing production in oil wells, and cannot fully exploit the effective information carried by uneven observation intervals. They also suffer from technical problems such as difficulty in extracting time-series features, limited prediction accuracy, and insufficient model generalization ability under conditions of limited annotation. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of difficulty in effectively extracting the time-series characteristics of irregular production in the prior art, and insufficient prediction accuracy and generalization ability, and to provide a mask prediction method and system for irregular production data of repeated fracturing in oil wells.

[0006] To achieve the above objectives, the present invention employs the following technical solution: This invention proposes a masking prediction method for irregular production data from repeated fracturing in oil wells, comprising the following steps: The production time series data of repeated fracturing of oil wells is obtained as the original production time series data. The production time series data is then embedded to generate a time-series embedded representation. Based on the rate of change at each time point in the production time series data, identify the points of high change from the production time series data; The temporal embedding representation of the time step corresponding to the high change point is masked, and the masked embedding representation is reconstructed to obtain the reconstructed temporal data. The reconstruction error is calculated and minimized based on the difference between the reconstructed time series data and the original production time series data; the total loss of the time series prediction model is constructed by combining self-supervised learning, while optimizing the parameters of the time series prediction model to complete the self-supervised training of the time series prediction model; Input the time series data to be predicted into the trained time series prediction model, output the production prediction value, and complete the mask prediction of irregular production data of repeated fracturing of oil wells.

[0007] Preferably, the step of generating a time-series embedded representation from the production time series data specifically involves: The output data at each time step is mapped to a high-dimensional embedding vector through a linear layer, and the time series order information is preserved by encoding the superposition position of each time step.

[0008] Preferably, the absolute change in output at each time point in the output time series data is specifically as follows:

[0009] in, For the first The original output value at the time step; For the first The original output value at the time step; For the first The absolute change in output at each observation point relative to the previous observation point; The first The absolute change in output at each observation point relative to the previous observation point Compared to the preset threshold used to filter high-change points, if Then the first The observation points were identified as points of high change.

[0010] Preferably, the step of masking the temporal embedding representation of the time step corresponding to the high change point specifically involves: The timing data after masking is as follows:

[0011] The embedded representation after masking is as follows:

[0012] in, The timing data after masking; For mask marking; This is a preset threshold used to filter points with high variation; when the first... The absolute change in output at each observation point relative to the previous observation point When the value exceeds the preset threshold used to filter high-change points, the corresponding time step value is replaced with a mask. This is the embedded representation after masking. For the first The mask indicator for each observation point , Indicator functions, This is a preset threshold used to filter points with high variation. Trainable vectors with the same dimension as the original temporal embeddings; unmasked original temporal embedding representations , No. Standardized yield values ​​at each observation point , and These are the weights and biases of the linear mapping, respectively. For time-location encoding function, the time interval between adjacent observations , For the first The actual observation time of each observation point This refers to the actual observation time at the previous observation point. This is the sequence number of the time step. For the first The raw output value at the time step. and These are the mean and standard deviation of the original yield series, respectively.

[0013] Preferably, the reconstruction of the embedded representation after masking to obtain reconstructed time-series data specifically involves: Embedded representation after masking The input is to the decoder, which uses the built-in network parameters to reverse-engineer the output dimension, time step and reconstructed time series data that correspond one-to-one with the original output data. The reconstructed time series data is as follows:

[0014] in, To reconstruct time series data; For refactoring functions; These are the parameters for the decoder. This is the embedded representation after masking.

[0015] Preferably, the step of calculating and minimizing the reconstruction error based on the difference between the reconstructed time-series data and the original production time-series data specifically involves: The reconstruction error is calculated using the mean square error index. ; The reconstruction error is minimized using the backpropagation gradient descent algorithm: ; in, The length of the time series. For the first The raw output value at the time step. To reconstruct time series data; These are the network parameters before the decoder and embedding layer are updated. These are the network parameters updated via gradient descent. To train the learning rate, The number of masked observations is used to reconstruct the gradient of the error relative to the network parameters. The set of indexes of the masked observation points , For the first The mask indicator for each observation point For the first The mask indicator for each observation point To express summation; Indicates belonging to, The number of elements in the set. For reconstruction error, This is the time step number.

[0016] Preferably, the step of combining self-supervised learning to construct the total loss of the time series prediction model, while simultaneously optimizing the parameters of the time series prediction model, to complete the self-supervised training of the time series prediction model, specifically involves: Total loss of time series prediction model ; To minimize the total loss of the time series prediction model, all network parameters are updated synchronously to complete the self-supervised training of the time series prediction model. in, To balance the coefficients of reconstruction loss and self-supervised loss, To reconstruct the error, time-series comparison self-supervised loss. , This represents the number of sequence windows in the batch. For cosine similarity, the first Anchor point representation of each original sequence window Positive sample representation formed by mask reconstruction views of the same window Negative sample representations formed by other windows in the same batch ; For projection function, For the time encoder and its sequence characterization aggregation process, the first i Original view feature representation obtained after encoding of original sequence windows The reconstructed view feature representation obtained after the reconstructed sequences of the same sequence window are processed by the same encoder. , No. kOriginal view feature representation obtained after encoding of original sequence windows , For the first The raw output value at the time step. To reconstruct time series data, For the first One original sequence window, For the reconstructed view of this window, For the same batch One original sequence window, To express summation, The anchor point window number. It is the natural logarithm function. It is a natural exponential function. The temperature coefficient is greater than zero. This is the negative sample window number.

[0017] This invention proposes a masking prediction system for irregular production data from repeated fracturing in oil wells, comprising: The data embedding module is used to acquire the production time series data of repeated fracturing of oil wells as the original production time series data, and to generate a time-series embedded representation by embedding the production time series data. A high-change-point identification module is used to identify high-change-points from the production time series data based on the rate of change of each time point in the production time series data. An embedding reconstruction module is used to perform masking processing on the time-series embedding representation corresponding to the time step of the high change point, and to reconstruct the embedded representation after masking processing to obtain reconstructed time-series data. The model training module is used to calculate and minimize the reconstruction error based on the difference between the reconstructed time series data and the original production time series data; it combines self-supervised learning to construct the total loss of the time series prediction model, while optimizing the parameters of the time series prediction model, and completing the self-supervised training of the time series prediction model. The production prediction output module is used to input the time series data to be predicted into the trained time series prediction model and output the production prediction value to complete the mask prediction of irregular production data of repeated fracturing of oil wells.

[0018] Preferably, the absolute change in output at each time point in the output time series data is specifically as follows:

[0019] in, For the first The original output value at the time step; For the first The original output value at the time step; For the first The absolute change in output at each observation point relative to the previous observation point; The first The absolute change in output at each observation point relative to the previous observation point Compared to the preset threshold used to filter high-change points, if Then the first The observation points were identified as points of high change.

[0020] Preferably, the reconstruction of the embedded representation after masking to obtain reconstructed time-series data specifically involves: Embedded representation after masking The input is to the decoder, which uses the built-in network parameters to reverse-engineer the output dimension, time step and reconstructed time series data that correspond one-to-one with the original output data. The reconstructed time series data is as follows:

[0021] in, To reconstruct time series data; For refactoring functions; These are the parameters for the decoder. This is the embedded representation after masking.

[0022] Preferably, the step of calculating and minimizing the reconstruction error based on the difference between the reconstructed time-series data and the original production time-series data specifically involves: The reconstruction error is calculated using the mean square error index. ; The reconstruction error is minimized using the backpropagation gradient descent algorithm: ; in, The length of the time series. For the first The raw output value at the time step. To reconstruct time series data; These are the network parameters before the decoder and embedding layer are updated. These are the network parameters updated via gradient descent. To train the learning rate, The number of masked observations is used to reconstruct the gradient of the error relative to the network parameters. The set of indexes of the masked observation points , For the first The mask indicator for each observation point To express summation; Indicates belonging to, The number of elements in the set. For reconstruction error, This is the time step number.

[0023] Preferably, the step of combining self-supervised learning to construct the total loss of the time series prediction model, while simultaneously optimizing the parameters of the time series prediction model, to complete the self-supervised training of the time series prediction model, specifically involves: Total loss of time series prediction model ; To minimize the total loss of the time series prediction model, all network parameters are updated synchronously to complete the self-supervised training of the time series prediction model. in, To balance the coefficients of reconstruction loss and self-supervised loss, To reconstruct the error, time-series comparison self-supervised loss. , This represents the number of sequence windows in the batch. For cosine similarity, the first Anchor point representation of each original sequence window Positive sample representation formed by mask reconstruction views of the same window Negative sample representations formed by other windows in the same batch ; For projection function, For the time encoder and its sequence characterization aggregation process, the first i Original view feature representation obtained after encoding of original sequence windows The reconstructed view feature representation obtained after the reconstructed sequences of the same sequence window are processed by the same encoder. , No. k Original view feature representation obtained after encoding of original sequence windows , For the first The raw output value at the time step. To reconstruct time series data, For the first One original sequence window, For the reconstructed view of this window, For the same batch One original sequence window, To express summation, The anchor point window number. It is the natural logarithm function. It is a natural exponential function. The temperature coefficient is greater than zero. This is the negative sample window number.

[0024] A terminal device includes 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 a masking prediction method for irregular production data from repeated fracturing of oil wells.

[0025] A computer-readable storage medium storing a computer program, wherein when executed by a processor, the computer program implements the steps of a masking prediction method for irregular production data from repeated fracturing of oil wells.

[0026] A computer program product includes a computer program that, when executed by a processor, implements steps for a masking prediction method for irregular production data from repeated fracturing of oil wells.

[0027] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a masking prediction method for irregular production data from repeated fracturing in oil wells. First, it collects irregular production time-series data from the field and generates a time-series embedded representation, fully preserving the time-series characteristics and variation information of the original production, thus avoiding the distortion problem in feature extraction from non-uniformly spaced and asynchronously observed data in traditional models. Second, it accurately locates high-change points in production by calculating the change rate of the time-series data, specifically capturing the dynamic mutation characteristics of fracturing production, solving the problem that existing technologies struggle to identify key change patterns in irregular time series. Subsequently, it masks the embedded features of high-change points and reconstructs the time-series data, simulating scenarios of missing and abrupt changes in irregular data, fully mining the effective production information carried by irregular observation intervals. Based on the error constraints between the original and reconstructed data, it constructs a joint loss to optimize model parameters using self-supervised learning, completing self-supervised training of the model without requiring a large number of manually labeled samples, effectively improving the poor generalization ability and weak prediction stability caused by scarce labeled data. Finally, it outputs production prediction results through a dedicated time-series prediction model that accurately adapts to the characteristics of irregular production data from repeated fracturing in oil wells, effectively improving the time-series feature extraction capability and production prediction accuracy, and meeting the actual production prediction needs in the field.

[0028] This invention proposes a masking prediction system for irregular production data from repeated fracturing in oil wells. By dividing the system into a data embedding module, a high-variance point identification module, an embedding reconstruction module, a model training module, and a production prediction output module, it obtains production prediction values ​​and completes the masking prediction of irregular production data from repeated fracturing in oil wells. The modular approach ensures that each module is independent, facilitating unified management of all modules. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of the masking prediction method for irregular production data from repeated fracturing of oil wells according to the present invention.

[0031] Figure 2 This is a detailed flowchart of the masking prediction method for irregular production data from repeated fracturing of oil wells according to the present invention.

[0032] Figure 3 This is the irregular production data and highly variable mask point map of the present invention.

[0033] Figure 4 This is a comparison image of the mask point reconstruction before and after the present invention.

[0034] Figure 5 This is a graph showing the decreasing trend of mask reconstruction loss during the model training process of this invention.

[0035] Figure 6 This is a diagram of the mask prediction system for irregular production data from repeated fracturing of oil wells according to the present invention.

[0036] Figure 7 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0038] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0039] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0040] This invention provides a masking prediction method for irregular production data from repeated fracturing in oil wells, including data preprocessing and embedding representation generation; identifying high-variability points and generating masks; mask embedding representation and mask embedding reconstruction; calculating and minimizing reconstruction error; self-supervised learning and training optimization; model evaluation and field application. Unlike directly masking the original input, this invention masks the time, numerical, and feature embedding representations, aiming to improve model robustness by reconstructing these embeddings. Compared to masking the original data, this invention helps the model develop more generalizable features, thereby improving the model's ability to process time-series data.

[0041] The present invention will now be described in further detail with reference to the accompanying drawings: This invention proposes a masking prediction method for irregular production data from repeated fracturing in oil wells, such as... Figure 1 and Figure 2 As shown, it includes the following steps: The well data, thresholds, and model parameters given below are used to illustrate a specific execution process of this method.

[0042] S1. Obtain the production time series data of repeated fracturing of oil wells as the original production time series data, and generate the time series embedded representation by embedding the production time series data to obtain the time series embedded representation. Collect data on the production output of oil wells subjected to repeated fracturing from oil production plants; This embodiment selects the field production record of Well A in an oilfield after repeated fracturing as a representative sample, and the well number and block information are anonymized. This well formed 14 observation points at unequal intervals within 50 days, with observation times at 1, 2, 4, 7, 11, 12, 13, 18, 20, 25, 30, 36, 43, and 50 days, corresponding to production values ​​of 2.10, 2.40, 1.90, 2.60, 3.00, 2.70, 2.20, 3.30, 2.90, 2.60, 2.20, 1.80, 1.50, and 1.30 t, respectively. The interval between adjacent observations is 1–7 days. The above data is consistent with… Figure 3 , Figure 4 The complete production sequence is consistent with that in the data.

[0043] Field records are sorted in ascending order of observation time, duplicate records at the same time are deleted, and missing dates are not interpolated at equal intervals to preserve the information carried by the original sampling interval. The yield value at each time step is used for linear mapping, and the observation time and adjacent observation intervals are used to generate time location codes. The yield value is standardized according to the sample mean and standard deviation before entering the model, and then destandardized to tons after the model output.

[0044] The process of generating a time-series embedded representation from the production time series data is as follows: The output data at each time step is mapped to a high-dimensional embedding vector through a linear layer, and the time series order information is preserved by encoding the superposition position of each time step.

[0045] Specifically, the Transformer architecture is used to map production time series data into a high-dimensional embedding representation: The raw data at each time step is mapped to a high-dimensional embedding through a linear layer, and then position encoding is added to the embedding at each time step to capture the sequential information of the time series.

[0046] In this embodiment, the output value is mapped to a 32-dimensional embedding vector, and the time position encoding formed by the observation time and the interval between adjacent observations is superimposed on the embedding vector. The Transformer encoder adopts a two-layer structure, with four attention heads in each layer, and the decoder outputs a univariate output sequence of the same length as the input sequence. The embedding dimension, number of layers, and number of attention heads can be adjusted according to the sample size.

[0047] S2. Based on the rate of change at each time point in the production time series data, identify the high-change points from the production time series data; The rate of change is represented by numerical difference or gradient of the time series, specifically the absolute change in output at each time point in the output time series data:

[0048] in, For the first The original output value at the time step; For the first The original output value at the time step; For the first The absolute change in output at each observation point relative to the previous observation point; The first The absolute change in output at each observation point relative to the previous observation point Compared to the preset threshold used to filter high-change points, if Then the first The observation points were identified as points of high change.

[0049] The rate of change for each point in the above sequence was calculated. The rate of change for day 7 was |2.60 - 1.90| = 0.70t, and the rate of change for day 18 was |3.30 - 2.20| = 1.10t. The rate of change for all other adjacent observation points was no greater than 0.50t. In this embodiment, the threshold was set to 0.60t, thus determining day 7 and day 18 as high change points, with their specific locations... Figure 3 The crosses in the text are consistent.

[0050] S3. Perform masking on the time series embedding representation corresponding to the time step of the high change point, and reconstruct the embedded representation after masking to obtain the reconstructed time series data. The masking process for the temporal embedding representation of the time step corresponding to the high change point is specifically as follows: Points with high rates of change are masked to conceal their numerical values ​​and feature information, and then replaced with special mask markers. The masked time series data is as follows:

[0051] The embedded representations of time, numerical values, and features are masked while preserving the feature information of the masked points. The masked embedded representations are as follows:

[0052] in, The timing data after masking; For mask marking; This is a preset threshold used to filter points with high variation; when the first... The absolute change in output at each observation point relative to the previous observation point When the value exceeds the preset threshold used to filter high-change points, the corresponding time step value is replaced with a mask. This is the embedded representation after masking. For the first The mask indicator for each observation point , Indicator functions, This is a preset threshold used to filter points with high variation. For trainable vectors with the same dimension as the original temporal embeddings, the mask value is used to influence the embedding computation; the unmasked original temporal embedding representation , No. Standardized yield values ​​at each observation point , and These are the weights and biases of the linear mapping, respectively. For time-location encoding function, the time interval between adjacent observations , For the first The actual observation time of each observation point This refers to the actual observation time at the previous observation point. This is the sequence number of the time step. For the first The raw output value at the time step. and These are the mean and standard deviation of the original yield series, respectively.

[0053] When generating the mask, the output values ​​and corresponding embeddings of days 7 and 18 are replaced with the same trainable mask vector, while the remaining time steps retain their original values. The unmasked time steps are used as context input to the decoder. Therefore, when reconstructing these two time steps, the model cannot directly read the true output; it can only recover the masked information using the preceding and following observations, the observation times, and the sampling interval.

[0054] The process of reconstructing the embedded representation after masking to obtain reconstructed time-series data is as follows: Embedded representation after masking The input is to the decoder, which uses the built-in network parameters to reverse-engineer the output dimension, time step and reconstructed time series data that correspond one-to-one with the original output data. The reconstructed time series data is as follows:

[0055] in, To reconstruct time series data; For refactoring functions; These are the parameters for the decoder. This is the embedded representation after masking.

[0056] S4. Calculate and minimize the reconstruction error based on the difference between the reconstructed time series data and the original production time series data; combine self-supervised learning to construct the total loss of the time series prediction model, while optimizing the parameters of the time series prediction model to complete the self-supervised training of the time series prediction model. The process of calculating and minimizing the reconstruction error based on the difference between the reconstructed time-series data and the original production time-series data specifically involves: The reconstruction error is calculated using the mean square error index. ; The reconstruction error is minimized using the backpropagation gradient descent algorithm: ; in, The length of the time series. For the first The raw output value at the time step. To reconstruct time series data; These are the network parameters before the decoder and embedding layer are updated. These are the network parameters updated via gradient descent. To train the learning rate, The number of masked observations is used to reconstruct the gradient of the error relative to the network parameters. The set of indexes of the masked observation points , To express summation; Indicates belonging to, The number of elements in the set. For reconstruction error, This is the time step number.

[0057] The process of combining self-supervised learning to construct the total loss of the time series prediction model, while simultaneously optimizing the parameters of the time series prediction model, and completing the self-supervised training of the time series prediction model, specifically involves: Total loss of time series prediction model ; To minimize the total loss of the time series prediction model, all network parameters are updated synchronously to complete the self-supervised training of the time series prediction model. in, To balance the coefficients of reconstruction loss and self-supervised loss, To reconstruct the error, time-series comparison self-supervised loss. , This represents the number of sequence windows in the batch. For cosine similarity, the first Anchor point representation of each original sequence window Positive sample representation formed by mask reconstruction views of the same window Negative sample representations formed by other windows in the same batch ; For projection function, For the time encoder and its sequence characterization aggregation process, the first i Original view feature representation obtained after encoding of original sequence windows The reconstructed view feature representation obtained after the reconstructed sequences of the same sequence window are processed by the same encoder. , No. k Original view feature representation obtained after encoding of original sequence windows , For the first The raw output value at the time step. To reconstruct time series data, For the first One original sequence window, For the reconstructed view of this window, For the same batch One original sequence window, To express summation, The anchor point window number. It is the natural logarithm function. It is a natural exponential function. The temperature coefficient is greater than zero. This is the negative sample window number.

[0058] In this embodiment, the aforementioned sequence is used to construct training samples using a sliding window of length 8 and a window step size of 1. The Adam optimizer is employed with an initial learning rate of 0.001 and λ of 0.5, and training is performed for 100 epochs. In each training epoch, the reconstruction error and temporal contrastive self-supervised loss are calculated, and the parameters of the embedding layer, encoder, and decoder are updated synchronously. Figure 5 As can be seen, the reconstruction loss decreased continuously from the initial value of about 0.48, fell below 0.05 after about 60 rounds, and stabilized at about 0.01 after 100 rounds, indicating that the model has converged.

[0059] S5. Input the time series data to be predicted into the trained time series prediction model, output the production prediction value, and complete the mask prediction of irregular production data of repeated fracturing of oil wells.

[0060] Sequence prediction using embedding representations generated by time-series prediction models:

[0061] in, This is the predicted value for the next time step; This is the prediction function.

[0062] The predictive performance of the model is evaluated using the mean absolute error, and adjustments and optimizations are made accordingly.

[0063] in, The mean absolute error, To evaluate the set of locations, To evaluate the number of positions, This represents absolute value operations.

[0064] After training, without providing the model with true values, the two high-change points on days 7 and 18 were reconstructed, with output values ​​of approximately 2.65 t and 3.12 t, corresponding to true values ​​of 2.60 t and 3.30 t, and absolute errors of approximately 0.05 t and 0.18 t, respectively. To test the model's recovery ability in the yield decay phase, the 1.50 t point on day 43 was set as an independent reserved mask point during the evaluation phase. This point was not involved in the high-change point selection and model training; the model output was approximately 1.30 t, with an absolute error of approximately 0.20 t. The average absolute error of the three test points was approximately 0.14 t, which is consistent with... Figure 4 The cross and triangle in the image are in the same position.

[0065] This embodiment encapsulates the trained model into an internal prediction program for the oilfield. Input fields include well number, observation date, and production value. The program outputs the reconstructed value of the masked time step, the predicted production value for the next time step, and corresponding error indicators. Field technicians jointly verify the model output with well intervention records and production decline trends to generate a ranking of candidate wells for repeated fracturing, serving as an auxiliary basis for well selection decisions.

[0066] Figure 3 Corresponding to the above 14 field observation points, the solid line represents the complete yield sequence, and the cross marks represent the high change points on the 7th and 18th days selected by threshold=0.60 t. Figure 4 Using the same complete sequence, crosses represent true values ​​and triangles represent model reconstruction values; days 7 and 18 show the reconstruction results for high-change points, while day 43 is an independent hold-out point set during the evaluation phase. This hold-out point is only used to verify the recovery capability of the decay segment and does not change the model's performance. Figure 3 The results of screening high-change points. Figure 5 The reconstruction loss curve is shown for a training process of 100 rounds. The loss decreases from about 0.48 to less than 0.02 and tends to stabilize as the number of training rounds increases.

[0067] Therefore, this embodiment completes the entire process of data processing, embedding generation, threshold calculation, mask generation, model training, reconstruction verification, and on-site output based on the on-site unequal interval production records, and provides a method that can be compared with... Figures 3-5 Each input value, intermediate result, and evaluation metric was checked item by item. Figure 4 The average absolute error calculated for the three test points shown is approximately 0.14t. Figure 5 This indicates that the training process can converge stably. The results demonstrate that this method can recover key production points and generate production prediction results for field personnel to verify without forcibly converting the original records into an equally spaced sequence.

[0068] In one embodiment of this application, such as Figure 6 As shown, a masking prediction system for irregular production data from repeated fracturing of oil wells is provided, comprising: The data embedding module is used to acquire the production time series data of repeated fracturing of oil wells as the original production time series data, and to generate a time-series embedded representation by embedding the production time series data. A high-change-point identification module is used to identify high-change-points from the production time series data based on the rate of change of each time point in the production time series data. An embedding reconstruction module is used to perform masking processing on the time-series embedding representation corresponding to the time step of the high change point, and to reconstruct the embedded representation after masking processing to obtain reconstructed time-series data. The model training module is used to calculate and minimize the reconstruction error based on the difference between the reconstructed time series data and the original production time series data; it combines self-supervised learning to construct the total loss of the time series prediction model, while optimizing the parameters of the time series prediction model, and completing the self-supervised training of the time series prediction model. The production prediction output module is used to input the time series data to be predicted into the trained time series prediction model and output the production prediction value to complete the mask prediction of irregular production data of repeated fracturing of oil wells.

[0069] Specific limitations regarding the masking prediction system for irregular production data from repeated fracturing of oil wells can be found in the limitations of the masking prediction method for such data described above; the corresponding technical effects are equivalent and will not be repeated here. Each module in the aforementioned masking prediction system for irregular production data from repeated fracturing of oil wells 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.

[0070] Figure 7 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 7 As 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 in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a masking prediction method for irregular production data from repeated fracturing in oil wells. The display screen can be an LCD screen or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0071] As will be understood by those skilled in the art, computer equipment Figure 7 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. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0072] 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.

[0073] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. 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.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A masking prediction method for irregular production data from repeated fracturing of oil wells, characterized in that, Includes the following steps: The production time series data of repeated fracturing of oil wells is obtained as the original production time series data. The production time series data is then embedded to generate a time-series embedded representation. Based on the rate of change at each time point in the production time series data, identify the points of high change from the production time series data; The temporal embedding representation of the time step corresponding to the high change point is masked, and the masked embedding representation is reconstructed to obtain the reconstructed temporal data. The reconstruction error is calculated and minimized based on the difference between the reconstructed time series data and the original production time series data; the total loss of the time series prediction model is constructed by combining self-supervised learning, while optimizing the parameters of the time series prediction model to complete the self-supervised training of the time series prediction model; Input the time series data to be predicted into the trained time series prediction model, output the production prediction value, and complete the mask prediction of irregular production data of repeated fracturing of oil wells. The absolute change in output at each time point in the output time series data is specifically as follows: in, For the first The original output value at the time step; For the first The original output value at the time step; For the first The absolute change in output at the nth observation point relative to the previous observation point; The absolute change in output at each observation point relative to the previous observation point Compared to the preset threshold used to filter high-change points, if Then the first The observation points were identified as points of high change.

2. The masking prediction method for irregular production data from repeated fracturing of oil wells according to claim 1, characterized in that, The process of generating a time-series embedded representation from the production time series data is as follows: The output data at each time step is mapped to a high-dimensional embedding vector through a linear layer, and the time series order information is preserved by encoding the superposition position of each time step.

3. The masking prediction method for irregular production data from repeated fracturing of oil wells according to claim 1, characterized in that, The masking process for the temporal embedding representation of the time step corresponding to the high change point is specifically as follows: The timing data after masking is as follows: The embedded representation after masking is as follows: in, The timing data after masking; For mask marking; This is a preset threshold used to filter points with high variation; when the first... The absolute change in output at each observation point relative to the previous observation point When the value exceeds the preset threshold used to filter high-change points, the corresponding time step value is replaced with a mask. This is the embedded representation after masking. For the first The mask indicator for each observation point , Indicator functions, This is a preset threshold used to filter points with high variation. Trainable vectors with the same dimension as the original temporal embeddings; unmasked original temporal embedding representations , No. Standardized yield values ​​at each observation point , and These are the weights and biases of the linear mapping, respectively. For time-location encoding function, the time interval between adjacent observations , For the first The actual observation time of each observation point This refers to the actual observation time at the previous observation point. This is the sequence number of the time step. For the first The raw output value at the time step. and These are the mean and standard deviation of the original yield series, respectively.

4. The masking prediction method for irregular production data from repeated fracturing of oil wells according to claim 1, characterized in that, The process of reconstructing the embedded representation after masking to obtain reconstructed time-series data is as follows: Embedded representation after masking The input is to the decoder, which uses the built-in network parameters to reverse-engineer the output dimension, time step and reconstructed time series data that correspond one-to-one with the original output data. The reconstructed time series data is as follows: in, To reconstruct time series data; For refactoring functions; These are the parameters for the decoder. This is the embedded representation after masking.

5. The masking prediction method for irregular production data from repeated fracturing of oil wells according to claim 1, characterized in that, The process of calculating and minimizing the reconstruction error based on the difference between the reconstructed time-series data and the original production time-series data specifically involves: The reconstruction error is calculated using the mean square error index. ; The reconstruction error is minimized using the backpropagation gradient descent algorithm: ; in, The length of the time series. For the first The raw output value at the time step. To reconstruct time series data; These are the network parameters before the decoder and embedding layer are updated. These are the network parameters updated via gradient descent. To train the learning rate, The number of masked observations is used to reconstruct the gradient of the error relative to the network parameters. The set of indexes of the masked observation points , For the first The mask indicator for each observation point To express summation; Indicates belonging to, The number of elements in the set. For reconstruction error, This is the time step number.

6. The masking prediction method for irregular production data from repeated fracturing of oil wells according to claim 1, characterized in that, The process of combining self-supervised learning to construct the total loss of the time series prediction model, while simultaneously optimizing the parameters of the time series prediction model, and completing the self-supervised training of the time series prediction model, specifically involves: Total loss of time series prediction model ; To minimize the total loss of the time series prediction model, all network parameters are updated synchronously to complete the self-supervised training of the time series prediction model. in, To balance the coefficients of reconstruction loss and self-supervised loss, To reconstruct the error, time-series comparison self-supervised loss. , This represents the number of sequence windows in the batch. For cosine similarity, the first Anchor point representation of each original sequence window Positive sample representation formed by mask reconstruction views of the same window Negative sample representations formed by other windows in the same batch ; For projection function, For the time encoder and its sequence characterization aggregation process, the first i Original view feature representation obtained after encoding of original sequence windows The reconstructed view feature representation obtained after the reconstructed sequences of the same sequence window are processed by the same encoder. , No. k Original view feature representation obtained after encoding of original sequence windows , For the first The raw output value at the time step. To reconstruct time series data, For the first One original sequence window, For the reconstructed view of this window, For the same batch One original sequence window, To express summation, The anchor point window number. It is the natural logarithm function. It is a natural exponential function. The temperature coefficient is greater than zero. This is the negative sample window number.

7. A masking prediction system for irregular production data from repeated fracturing of oil wells, characterized in that, The masking prediction method for irregular production data from repeated fracturing of oil wells according to any one of claims 1 to 6 includes: The data embedding module is used to acquire the production time series data of repeated fracturing of oil wells as the original production time series data, and to generate a time-series embedded representation by embedding the production time series data. A high-change-point identification module is used to identify high-change-points from the production time series data based on the rate of change of each time point in the production time series data. An embedding reconstruction module is used to perform masking processing on the time-series embedding representation corresponding to the time step of the high change point, and to reconstruct the embedded representation after masking processing to obtain reconstructed time-series data. The model training module is used to calculate and minimize the reconstruction error based on the difference between the reconstructed time series data and the original production time series data; it combines self-supervised learning to construct the total loss of the time series prediction model, while optimizing the parameters of the time series prediction model, and completing the self-supervised training of the time series prediction model. The production prediction output module is used to input the time series data to be predicted into the trained time series prediction model and output the production prediction value to complete the mask prediction of irregular production data of repeated fracturing of oil wells.

8. The masking prediction system for irregular production data from repeated fracturing of oil wells according to claim 7, characterized in that, The process of reconstructing the embedded representation after masking to obtain reconstructed time-series data is as follows: Embedded representation after masking The input is to the decoder, which uses the built-in network parameters to reverse-engineer the output dimension, time step and reconstructed time series data that correspond one-to-one with the original output data. The reconstructed time series data is as follows: in, To reconstruct time series data; For refactoring functions; These are the parameters for the decoder. This is the embedded representation after masking.

9. The masking prediction system for irregular production data from repeated fracturing of oil wells according to claim 7, characterized in that, The process of calculating and minimizing the reconstruction error based on the difference between the reconstructed time-series data and the original production time-series data specifically involves: The reconstruction error is calculated using the mean square error index. ; The reconstruction error is minimized using the backpropagation gradient descent algorithm: ; in, The length of the time series. For the first The raw output value at the time step. To reconstruct time series data; These are the network parameters before the decoder and embedding layer are updated. These are the network parameters updated via gradient descent. To train the learning rate, The number of masked observations is used to reconstruct the gradient of the error relative to the network parameters. The set of indexes of the masked observation points , For the first The mask indicator for each observation point To express summation; Indicates belonging to, The number of elements in the set. For reconstruction error, This is the time step number.

10. The masking prediction system for irregular production data from repeated fracturing of oil wells according to claim 7, characterized in that, The process of combining self-supervised learning to construct the total loss of the time series prediction model, while simultaneously optimizing the parameters of the time series prediction model, and completing the self-supervised training of the time series prediction model, specifically involves: Total loss of time series prediction model ; To minimize the total loss of the time series prediction model, all network parameters are updated synchronously to complete the self-supervised training of the time series prediction model. in, To balance the coefficients of reconstruction loss and self-supervised loss, To reconstruct the error, time-series comparison self-supervised loss. , This represents the number of sequence windows in the batch. For cosine similarity, the first Anchor point representation of each original sequence window Positive sample representation formed by mask reconstruction views of the same window Negative sample representations formed by other windows in the same batch ; For projection function, For the time encoder and its sequence characterization aggregation process, the first i Original view feature representation obtained after encoding of original sequence windows The reconstructed view feature representation obtained after the reconstructed sequences of the same sequence window are processed by the same encoder. , No. k Original view feature representation obtained after encoding of original sequence windows , For the first The raw output value at the time step. To reconstruct time series data, For the first One original sequence window, For the reconstructed view of this window, For the same batch One original sequence window, To express summation, The anchor point window number. It is the natural logarithm function. It is a natural exponential function. The temperature coefficient is greater than zero. This is the negative sample window number.

11. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the mask prediction method for irregular production data of repeated fracturing of oil wells as described in any one of claims 1 to 6.

12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the mask prediction method for irregular production data of repeated fracturing of oil wells as described in any one of claims 1 to 6.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the mask prediction method for irregular production data of repeated fracturing of oil wells as described in any one of claims 1 to 6.

Images