TL-AEAT-BIGRU post-compression yield prediction method based on data joint driving under knowledge constraint

Abstract

The invention relates to a TL-AEAT-BIGRU post-compression yield prediction method based on data joint driving under knowledge constraint, which adopts a CWGAN-GP model based on conditional constraint to perform data enhancement on a small amount of multi-source data, adopts a Pearson + mRmR correlation analysis algorithm to determine main control factors influencing post-compression yield, and performs prediction on the post-compression yield by using a TL-AEAT-BIGRU model. And classifying reservoir categories through a main component analysis method based on the reservoir classification standard established by the former. Correlation experience knowledge between the post-compression yield and the input main control factors is combined with the TL-AEAT-BiGRU post-compression yield prediction model, and the screened main control factors are used as input for post-compression yield prediction. The result shows that compared with other yield prediction models, the model has the best performance in the post-pressure yield prediction of the research area. An ablation experiment shows that each module of the model contributes to improvement of the yield prediction effect, so that the yield prediction accuracy of the model is effectively improved, the problem that the yield prediction precision of the reservoir after pressure is not high under the condition that the sample size is small is effectively solved, and the method has guiding significance on yield analysis of an oil and gas well.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of geological exploration, in particular to a TL-AEAT-BIGRU post-frac production prediction method based on data joint driving under knowledge constraints. BACKGROUND

[0002] At present, the exploitation of oil and gas resources plays a crucial role in energy regeneration and recycling, and is also an important link of resource replacement. The rational exploitation of tight oil and gas reservoirs not only can increase energy supply and affect the supply and demand pattern, but also can promote technological innovation and promote energy transformation. Precise prediction of post-frac production can provide a scientific basis for the exploitation of tight oil reservoirs, optimize well deployment and fracturing design, improve recovery efficiency, and control production cost and resource allocation.

[0003] For the M block in the Junggar Basin, the internal reservoir of the tight sandstone and conglomerate reservoir presents low permeability and strong heterogeneity, and the research on the post-frac production prediction model of the vertical fracture well of the sandstone and conglomerate reservoir in the M block is less, and most of the research methods have certain limitations, so it is necessary to establish a reasonable and high generalization performance post-frac production prediction model. SUMMARY

[0004] The application provides a TL-AEAT-BIGRU post-frac production prediction method based on data joint driving under knowledge constraints to solve the above technical problems.

[0005] The application is realized through the following technical scheme: The TL-AEAT-BIGRU post-frac production prediction method based on data joint driving under knowledge constraints comprises the following steps: the application considers that the collected data includes reservoir geology, reservoir geomechanics and engineering data, etc., first determines the reservoir classification evaluation index by using the principal component analysis method to accurately identify the production conditions of different reservoirs, and guides the optimization of perforation position. The reservoir geology parameters and reservoir geomechanics parameter data corresponding to each reservoir are converted into corresponding perforation section data, combined with engineering data and post-frac production data to jointly construct the post-frac production prediction data set of this study. In view of the problem of small sample size existing in the actual work area data set, in order to accurately predict the post-frac production of the sandstone and conglomerate reservoir, the CWGAN-GP method is used for data enhancement on the geology and engineering parameter data set, and the main control factors affecting the post-frac production of the reservoir are selected by using the "mrmr+PEARSON" feature selection algorithm. And a TL-AEAT-BIGRU post-frac production prediction model based on data joint driving under knowledge constraints is designed, the correlation experience knowledge constraint term is introduced into the model, the experience knowledge and the deep fusion model are jointly driven to predict the post-frac production, and the transfer learning strategy is applied to the post-frac production prediction model.

[0006] Compared with the prior art, the application has at least the following beneficial effects: This application employs a conditionally constrained CWGAN-GP model to augment limited, multi-source data. A Pearson+mRmR correlation analysis algorithm is used to identify the main controlling factors influencing post-compression production. Principal component analysis is used to classify reservoirs based on previously established reservoir classification standards. Furthermore, empirical knowledge of the correlation between post-compression production and the input main controlling factors is combined with the TL-AEAT-BiGRU post-compression production prediction model, using the selected main controlling factors as input for post-compression production prediction. Results show that this model performs best in post-compression production prediction in the study area compared to other production prediction models. Ablation experiments demonstrate that each module of the model contributes to improving production prediction accuracy, effectively addressing the problem of low accuracy in post-compression production prediction when sample size is limited. This provides guidance for production analysis of oil and gas wells. Attached Figure Description

[0007] 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.

[0008] Figure 1 This is a diagram illustrating the overall framework for predicting post-pressed output in the embodiments. Figure 2 This is a flowchart of the transfer learning process in the embodiment; Figure 3 This is a flowchart of the CWGAN-GP training process in the embodiment. Figure 4 This is a comparison chart of kurtosis before and after enhancement of some feature data in the embodiment; Figure 5 This is a comparison chart of statistical indicators before and after feature enhancement in some examples; Figure 6 This is a graph showing the correlation between reservoir geological parameters and production in the examples. Figure 7 This is a graph showing the correlation between reservoir geomechanical parameters and production in the examples. Figure 8 This is a graph showing the correlation between engineering parameters and production output in the embodiment. Figure 9 This is a feature importance ranking diagram of the mRmR algorithm in the embodiment; Figure 10 This is a Pearson correlation matrix analysis diagram from the embodiment; Figure 11This is a comparison chart of prediction results before and after the introduction of transfer learning in the embodiment. Figure 12 This is a comparison chart of the output prediction results of different models in the embodiments; Figure 13 This is a comparison chart of well logging interpretation and production prediction for well a in the embodiment; Figure 14 This is a comparison chart of well logging interpretation and production prediction for well b in the embodiment; Figure 15 This is a comparison chart of well logging interpretation and production prediction for well c in the embodiment. Detailed Implementation

[0009] 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 some embodiments of the present invention, but not all embodiments.

[0010] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. It should also be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0011] This embodiment takes Block M of the Junggar Basin as an example. Considering that post-compression production of sandstone and conglomerate reservoirs in Block M is influenced by multiple factors including reservoir geological parameters, reservoir geomechanical parameters, and engineering parameters, this application, taking into account that the collected data includes reservoir geology, reservoir geomechanical, and engineering data, firstly uses principal component analysis to determine reservoir classification evaluation indicators to accurately identify the production status of different reservoirs and guide the optimal selection of perforation locations. The reservoir geological and geomechanical parameters corresponding to each reservoir are converted into corresponding perforation segment data, which are then combined with engineering data and post-compression production data to form the post-compression production prediction dataset for this study. Addressing the issue of limited sample size in the actual work area dataset, to accurately predict post-compression production of sandstone and conglomerate reservoirs, the CWGAN-GP method is used to augment the geological and engineering parameter datasets. The "mrmr+PEARSON" feature selection algorithm is then used to screen the main controlling factors affecting reservoir post-compression production. Furthermore, a TL-AEAT-BIGRU post-press yield prediction model based on knowledge constraints and data joint driving was designed. By introducing relevant empirical knowledge constraints into the model, empirical knowledge and deep fusion model are jointly driven to predict post-press yield. Transfer learning strategy is applied to the yield prediction model.

[0012] like Figure 1As shown in this embodiment, the TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving includes the following steps: Step 1: Acquire data and perform data preprocessing. This embodiment takes the M block of the Junggar Basin as an example. The post-compression production of sandstone and conglomerate reservoirs in the M block is affected by a variety of factors, including reservoir geological parameters, reservoir geomechanical parameters, and engineering parameters.

[0013] In this embodiment, the data for predicting post-fracturing production in conglomerate reservoirs comes from the project "Research on Classification and Evaluation Methods and Key Engineering Geological Parameters of Low-Permeability Reservoirs in the Junggar Basin". The study analyzes the impact of reservoir geological parameters (porosity, permeability, pore structure index, reservoir quality factor, resistivity, sonic transit time, clay content, formation pressure, density), reservoir geomechanical parameters (brittleness index, Poisson's ratio, tensile strength, horizontal stress difference, minimum horizontal principal stress, maximum horizontal principal stress), and engineering parameters (post-flowback fluid deficit ratio, sand addition intensity per meter, instantaneous pump shutdown pressure, discharge rate, average sand ratio) on post-fracturing production in the conglomerate oil development process of the study block. Reservoir geological parameters are determined by the reservoir's inherent properties, while reservoir geomechanical and engineering parameters are parameters used during drilling and fracturing. Low permeability and strong heterogeneity of the reservoir, as well as the effectiveness of reservoir fracturing, all affect the production of horizontal wells. In this embodiment, some reservoir geological parameters (porosity, permeability, pore structure index, and clay content) are predicted based on machine learning models, while some reservoir geological parameters and reservoir geomechanical parameters are calculated based on existing drilling data, well logging data, and synchronous indoor rock physics and rock mechanics experiments to establish reservoir geological parameters and reservoir geomechanical parameters formula models.

[0014] The process of obtaining each parameter in the multi-source data in this embodiment is as follows.

[0015] (1) Reservoir geological parameters Pore ​​structure index: Based on previous calculations and analyses of the same study area, a multiple regression model for the pore structure index was established, as shown in formula (1): (1); In the above formula, PI is the porosity index, AC is the acoustic transit time (µs / ft), CNL is the neutron value (v / v), DEN is the density (g / cm3), and SH is the clay content.

[0016] Reservoir quality factors: A calculation model is constructed based on porosity (POR), permeability (PERM), oil saturation (SW), and the calculated pore structure index, as shown in formula (2): (2); In the above formula, RQI is the reservoir quality factor. This is the normalized pore structure index.

[0017] (2) Reservoir geomechanical parameters The results were calculated based on well logging data from the key well reservoir sections in the study area: elastic modulus The calculation formula is: , The formula for calculating the brittleness index B is: , Tensile strength The calculation formula is: , Maximum and minimum principal stresses: First, the effective stress method is used to calculate the formation pore pressure to avoid the low calculation accuracy caused by methods such as Eaton's method and equivalent depth method due to the selection of only one logging parameter. The formula for calculating the effective stress is: (3); In the above formula, The effective stress is MPa.

[0018] Based on density logging data and effective stress theory, a formation pressure calculation model is established, as shown in equation (6): (4); In the above formula, Formation pressure, MPa; The vertical stress is measured in MPa. The density at point h in the untested section. ; ρ represents the depth value at the initial logging point. The density at point h in the well-logged section. .

[0019] During drilling and fracturing, the formation fracturing pressure and closure pressure are calculated using the fracturing operation curve. Based on this, the combined spring model calculates the geostress profile of a single well by obtaining the tectonic strain coefficient at the corresponding depth. The calculation formula for the combined spring model is as follows: (5), (6), In the above formula, , These represent the maximum and minimum principal stresses at the horizontal level, respectively. , These are the structural strain coefficients along the directions of the maximum and minimum horizontal principal stresses, respectively. This is the boit coefficient.

[0020] Different reservoir types possess varying physical properties such as porosity and permeability, which directly influence the reservoir's fluid flow capacity and production potential. Furthermore, the heterogeneity between reservoirs significantly impacts fracturing effectiveness and production. Therefore, it is necessary to analyze the reservoir types in the study block before predicting post-fracturing production in sandstone and conglomerate reservoirs. This embodiment follows the reservoir classification criteria established based on reservoir geological parameters in the study "Research on Classification and Evaluation Methods and Key Engineering Geological Parameters of Low-Permeability Reservoirs in the Junggar Basin." The reservoirs in the M block dataset are sequentially classified into Class I, Class II, Class III, and Class IV, with Class I reservoirs exhibiting the best fluid production capacity, followed by Class II, and Class IV reservoirs exhibiting the worst. The reservoir classification criteria based on reservoir geological parameters are shown in Table 1.

[0021] Table 1: Criteria for Classifying Reservoirs

[0022] Based on the multidimensional reservoir classification system, a geological evaluation index reflecting reservoir properties was constructed using principal component analysis. The classification evaluation criteria established based on this reservoir geological parameter classification evaluation index are shown in Table 2. Table 3 shows the coefficients of independent variables and the results of principal component analysis for each principal component, and Table 4 shows the coefficients of independent variables. As can be seen from Tables 3 and 4, the cumulative variance contribution rate of the first two principal components reaches 86.19%, which can replace most parameters for reservoir category discrimination. Formula (9) is the multivariate regression reservoir category evaluation index established by the first two principal components and the reservoir category label. The accuracy of this evaluation index for the reservoir category label after data augmentation reaches 91.3%, that is, this index can effectively classify reservoirs based on the reservoir classification criteria.

[0023] (7), Table 2: Classification Standards for Geological Evaluation Indicators

[0024] Table 3: Principal Component Analysis Results

[0025] Table 4: Coefficients of Independent Variables

[0026] Since the engineering parameter data in this embodiment were directly obtained during the perforation process at each well location for each perforated section, the reservoir geological parameters (porosity, permeability, pore structure index, sonic transit time, resistivity, clay content, formation pressure, density), reservoir geomechanical parameters (brittleness index, Poisson's ratio, tensile strength, horizontal stress difference, maximum horizontal principal stress, minimum horizontal principal stress), and post-compression production values ​​of each sublayer were converted into corresponding values ​​for the corresponding perforated sections. The aforementioned reservoir geological parameters, reservoir geomechanical parameters, and engineering parameters (drainage ratio after backflow, sand addition intensity per meter, average sand ratio, discharge rate, instantaneous pump shutdown pressure) were considered as influencing factors affecting post-compression production in conglomerate reservoirs. A multi-source dataset was constructed to jointly study its influence mechanism on post-compression production in conglomerate reservoirs.

[0027] Step 2: Define the transfer learning objectives Traditional machine learning algorithms require independent training in various domains, while transfer learning can directly transfer learned knowledge and feature extraction capabilities from one domain to another related new domain. In the task of predicting post-compression production in the conglomerate reservoir of the study area, a transfer learning strategy was adopted. Initial model parameters were generated through pre-training on geological and engineering data from Block M, ensuring these parameters contained rich feature representations. The parameters of the pre-trained knowledge-constrained, data-driven TL-AEAT-BiGRU post-compression production prediction model were transferred to the model for the new task. First, the parameters of the first few layers of the model were frozen, and then fine-tuned on the target domain geological and engineering dataset to improve the usability of target domain data from mature blocks. By employing transfer learning, the model can better adapt to the specific geological conditions of Block M. With limited data, it reduces the number of training parameters while obtaining a high-quality prediction model, making it an effective alternative method for predicting post-compression production in the same block. The training process of transfer learning is as follows: Figure 2 As shown.

[0028] In this embodiment, given that the original data contains multi-source data on the post-compression production of sandstone and conglomerate reservoirs from 7 wells in the "Jinlong Formation" and multi-source data on the post-compression production of sandstone and conglomerate reservoirs from 9 wells in the "Mahu Formation," and considering that the wells in the "Jinlong Formation" and "Mahu Formation" belong to the Upper Urho Formation in the Mahu Block and have strong geological correlation (such as similar sedimentary backgrounds (e.g., alluvial fans, fan deltas, etc.), resulting in comparable lithological assemblages (gravel, sand, gravel, interstitial ratio), mineral composition, etc.), and that the source rocks, migration paths, and capping conditions of the Mahu and Jinlong areas are similar, resulting in regular differences in fluid properties, and that regional heterogeneity is easily controlled by similar sedimentary microfacies, thus the "Jinlong" and "Mahu" Formation areas have relatively similar heterogeneity), the 31 data points from the 7 wells in the "Jinlong Formation" will be used as the source domain data for migration, and the 38 data points from the 9 wells in the "Mahu Formation" will be used as the target domain data for migration.

[0029] like Figure 11 As shown in the figures, Figure (a) presents the yield prediction results obtained after training the knowledge-constrained data-driven AEAT-BiGRU model on the post-pressed yield data of the Mahu Formation without employing a transfer learning strategy; Figure (b) presents the yield prediction results obtained after training the knowledge-constrained data-driven TL-AEAT-BiGRU model on the post-pressed yield data of the Mahu Formation. By comparing the prediction results of the knowledge-constrained data-driven AEAT-BiGRU model on the post-pressed yield of the Mahu Formation before and after adopting a transfer learning strategy, it is demonstrated that by employing a transfer learning strategy, the sharing of geological patterns between the source and target domains can be effectively achieved, enabling the model to learn more extensive and universal geological patterns in the study area. Furthermore, by fine-tuning the model in the target domain, the model can effectively adapt to local differences between the source and target domains, improving the prediction accuracy of the target domain data.

[0030] Step 3: Use CWGAN-GP to augment the data.

[0031] In this embodiment, CWGAN-GP is used to augment the reservoir geological parameters, reservoir geomechanical parameters, and engineering parameters.

[0032] The reservoir geological parameters include porosity, permeability, pore structure index, resistivity, sonic transit time, clay content, formation pressure, and density; the reservoir geomechanical parameters include brittleness index, Poisson's ratio, resistance strength, horizontal stress difference, minimum horizontal principal stress, maximum horizontal principal stress, pore structure index, and reservoir quality factor; the engineering parameters include the liquid shortage ratio after backflow, sand addition intensity per meter, instantaneous pump stop pressure, discharge rate, and average sand ratio.

[0033] like Figure 3As shown, the CWGAN-GP generative adversarial network (GAN) consists of two neural networks: a generator and a discriminator. Through multiple rounds of adversarial training, it generates distributions similar to the original samples. WGAN-GP is an improved GAN model, using JS divergence to measure the difference between distributions. The discriminator must satisfy the Lipsehitz condition to ensure the effectiveness of the Wasserstein distance and the stability of training. The goal of network training is to maximize the discriminator's probability of judging real samples (D(x)) to better distinguish between real and generated samples. Minimizing the generator loss function makes the generated samples closer to real samples, thus deceiving the discriminator. To further improve the quality of sample generation, this embodiment uses conditional constraints and feature extraction capabilities. However, if the constraints only contain the sample label information, it is difficult to learn the inherent distribution patterns of the feature data during adversarial training, limiting further optimization iterations of the generator. Therefore, in the discriminator and generator models based on the original model of WGAN-GP, in addition to using the lithology label corresponding to the sample as a constraint, additional dimensionless skewness and kurtosis characteristic indicators are added to realize conditional generation, which constitutes the CWGAN-GP network structure of the embodiment. The skewness and kurtosis calculation formulas are shown in Equation (8) and Equation (9) to stably reflect and limit the distribution of geological and engineering fracturing data.

[0034] (8), (9), In the above formula, T represents the kurtosis of the dataset, and S represents the skewness of the dataset. This represents a sample in the dataset, where Std is the standard deviation, M is the sample mean, and N is the total number of samples in the dataset.

[0035] Two conditional values, derived from the formulas for kurtosis and skewness, are concatenated with the sample lithology label L to obtain a conditional vector C=[L,K,S]. This conditional vector is then input into the WGAN-GP model to aid training. The generator generates new samples, and the discriminator distinguishes between the generated and real samples. The parameters of the generator and discriminator are updated by calculating the loss function.

[0036] The loss function of the CGAN-GP discriminator is: (10) In the above formula, Z represents the discriminator loss function; Z represents the real sample. The data distribution represents the real samples; x represents the samples generated by the generator; y represents the conditional constraints added during the training of the discriminator. The prior distribution representing the noise; Linear interpolation representing real samples and generated samples; The data distribution representing the linear interpolation samples; The expected score of the discriminator on the real data. The expected score of the discriminator on the generated data; This is the gradient penalty term for the discriminator. Lambda The coefficient of the gradient penalty term, To introduce a gradient penalty term for linear interpolation, the gradient penalty term is added to the generated linear interpolation samples to ensure that the discriminator applies the gradient penalty uniformly throughout the data space. This ensures that the discriminator satisfies the Lipschitz condition across the entire data distribution, which helps the discriminator to have a smooth gradient not only on real and generated data points, but on all points, thus effectively preventing gradient vanishing or gradient exploding.

[0037] The loss function of the generator is: (11), In the above formula, represents the discriminator loss function; x represents the samples generated by the generator; y represents the constraints added during the generator training process. Let $\frac{ ...

[0038] When using the CWGAN-GP method for data augmentation, seven key hyperparameters were set, and the parameter selection results are shown in Table 5.

[0039] Table 5: CWGAN-GP Parameter Settings Table

[0040] In this embodiment, data collected from key wells in the study area were selected as source domain data, and data from 40 geological and engineering perforation sections of 8 wells in a key reservoir were selected as target domain data. The CWGAN-GP method was used to perform data augmentation processing on the data from both types of wells, and GAN was simultaneously used for data augmentation experiments for comparison.

[0041] Figure 4Figures (a) and (b) show the changes in skewness and kurtosis values ​​before and after enhancement of some feature data. The comparative analysis shows that after processing with the CWGAN-GP data enhancement method, the enhanced data can still effectively maintain the positive and negative values ​​of skewness and kurtosis of each feature parameter, indicating that this method has a significant advantage in preserving the data distribution pattern. Compared with traditional GAN ​​enhancement methods, the data enhanced by CWGAN-GP can not only more accurately fit the skewness and kurtosis statistical characteristics of the original data, but also effectively improve the diversity and reliability of the generated data through lithological labels, skewness and kurtosis constraint information, and gradient penalty mechanisms, avoiding model collapse problems. This result shows that the CWGAN-GP method has better stability and generalization ability in reservoir parameter data enhancement, and can provide a higher quality data foundation for subsequent post-pressurization production prediction.

[0042] Figure 5 (a)-(c) show the evaluation results of statistical indicators before and after feature enhancement. From the perspective of standard deviation, GAN failed to fully learn the significance of feature fluctuations, such as the poor learning of tensile strength and resistivity features. This is mainly because the feature samples with extreme standard deviations have large dispersion differences, resulting in poor generalization ability of GAN. Combining the mean and median, it can be seen that GAN basically learns the trend of data distribution and can fit the overall distribution pattern of the data, but its learning ability for individual samples is insufficient. The quality of the generated data is not as good as CWGAN-GP. The distribution pattern of the data generated by CWGAN-GP can better fit the original data and effectively learn feature information beyond the original data. Furthermore, it can preserve the depth sequence and neighborhood information of the sandstone and conglomerate reservoir data, enhancing the interpretability of the model predictions.

[0043] Table 8 compares the maximum mean difference (MMD) values ​​of the original and generated data from the source and target domains, as well as the data from the source and target domains. It can be seen that the MMD values ​​for all three data categories are less than 1. Figure 5 The statistical comparison of the parameters shows that the data generated by the WGAN-GP model, which incorporates the dimensionless formulas for skewness and kurtosis, can fit the distribution pattern of the original data well, and the geological and engineering data of the source and target domains have high similarity.

[0044] Table 8: Maximum mean difference among data in block M

[0045] Step 4: Screening of Controlling Factors For the perforated and fractured sections of 8 wells in the study area, the corresponding reservoir geological parameters, reservoir geomechanical parameters, and engineering parameters were calculated and obtained sequentially. Table 7 records the main statistical parameters of each parameter.

[0046] Table 7: Statistical Indicators for Each Parameter

[0047] To analyze the correlation between various parameters and post-press yield, this embodiment calculates correlation coefficients and plots scatter plots to preliminarily clarify the degree of influence of different parameters on post-press yield, providing reasonable data support for subsequent prediction of post-press yield. The analysis results are as follows: Figure 6 to Figure 8 As shown.

[0048] Among reservoir geological parameters, porosity index, permeability, and reservoir quality factor all show a positive correlation with production. Specifically, in the fracturing and production of sandstone and conglomerate reservoirs, a higher porosity index indicates better fracture connectivity, higher permeability, and smoother oil flow channels, typically resulting in higher post-fracturing production. The reservoir quality factor, a key indicator of reservoir oil saturation, directly relates to the reservoir's storage space and permeability. In high-quality reservoirs, fracture development is usually better, making it easier for fracturing to penetrate the reservoir and effectively support the fractures, thereby improving reservoir permeability. Porosity, clay content, deep resistivity, and density did not show a significant correlation with post-fracturing production.

[0049] Among the reservoir geomechanical parameters, the brittleness index is positively correlated with post-compression production, while the minimum horizontal principal stress, horizontal stress difference, and tensile strength show a negative correlation with post-compression production. A higher minimum horizontal principal stress increases the difficulty of fracture initiation and reduces fracture conductivity. The horizontal stress difference directly affects the opening and closing state of fractures, impacting production stability. A reasonable tensile strength helps ensure fracture ductility and reservoir integrity. A higher brittleness index facilitates the formation of artificial fracture networks, improving hydrocarbon permeability. The remaining parameters did not show a significant correlation with production.

[0050] Engineering parameters directly affect the volume and effectiveness of large-scale fracturing. In this embodiment, the flow rate and sand-addition intensity per meter show a positive correlation with production output; generally, the higher these two parameter values, the better the fracturing effect. However, the correlation between parameters such as the flowback fluid ratio and average sand ratio and production output is not significant.

[0051] Combination Figure 6 to Figure 8 As can be seen, although there is a linear correlation between the various influencing factors and production, the correlation coefficient is low. Furthermore, the linear correlation analysis mainly focuses on the qualitative evaluation of the impact of each influencing factor on production, lacking quantitative analysis. Therefore, it is necessary to adopt a reasonable and more precise method to determine the main controlling factors affecting the post-compression production of sandstone and conglomerate reservoirs in the study area.

[0052] Different reservoir geological parameters, engineering geological factors, and engineering parameters reflect different reservoir characteristics. In order to accurately predict post-pressurization production, the linear correlation between each parameter and production was analyzed. The positive and negative correlation between each parameter and production was qualitatively obtained. However, selecting appropriate geological and engineering data to construct the model input data is crucial for the accurate prediction of post-pressurization production. Therefore, further quantitative analysis is needed on the complex relationship between each parameter and production.

[0053] This embodiment uses the Pearson + mR / mR correlation analysis method to screen the main controlling factors affecting post-press yield. Pearson analysis is used to identify characteristic parameters with a significant linear correlation to post-press yield. The mR / mR correlation coefficient is then used to further analyze the nonlinear correlation between each parameter and post-press yield, as well as the redundancy among the parameters. Figure 9 This is a feature importance graph for the mRmR algorithm. Figure 10 The Pearson correlation matrix of each factor and output is shown.

[0054] The mR correlation analysis method is based on mutual information metrics to evaluate the correlation between parameters. This metric can be used to measure the nonlinear relationship between feature variables. The maximum correlation between features and target variables is determined by calculating the mutual information between features, so as to minimize the correlation between features and reduce feature redundancy. The calculation formulas are shown in equations (12)-(14), thereby accurately achieving feature selection.

[0055] (12) (13) (14) In the above formula, This represents the mutual information value between two random variables. Represents the target variable. A feature dataset representing the predicted yield after compression. This indicates the degree of maximum correlation between the feature and the output variable. This represents the minimum redundancy between two feature variables.

[0056] Pearson correlation coefficient: a linear correlation function used to analyze the linear correlation between two features. Suppose there are two feature sets X and Y: , The formula for calculating the Pearson correlation coefficient is: (15) In the above formula, and represents the average values ​​of x and y, respectively, and r represents the Pearson correlation coefficient.

[0057] This embodiment refers to the feature selection criteria established by predecessors for the fusion correlation coefficient method. Reservoir geological parameters, reservoir engineering parameters, and engineering parameters with a Pearson coefficient value greater than 0.5 are classified as main control factors. If the main control factors selected are only geological or engineering parameters, or if the number of main control factors is too small, then factors with both Pearson correlation coefficient and MrMr correlation coefficient values ​​greater than 0.45 are taken as main control factors. Figure 9 The results show that the Pearson correlation coefficients for permeability, porosity index, reservoir quality factor, horizontal stress difference, and brittleness index are all greater than 0.5. Among them, the Pearson correlation coefficient for permeability is the highest at 0.725. This means that reservoir permeability directly affects fluid flow, fracture propagation, and connectivity, thus altering production. The Pearson correlation coefficient for reservoir quality factor is the second highest at 0.686. This parameter is usually calculated from key reservoir geological parameters such as oil saturation and other heterogeneous parameters. A higher reservoir quality factor generally indicates that the reservoir has better permeability, is more likely to form efficient fracture networks, and effectively prevents proppant settling and blockage, thus maintaining the long-term conductivity of the reservoir. Reservoirs with low quality factors require optimization of fracturing parameters and development technologies to maintain conductivity. The Pearson correlation coefficient for porosity index is 0.598. Since the porosity index directly determines the original permeability of the reservoir. Reservoirs with a high porosity index allow for efficient connectivity between artificial fractures and the natural pore network, expanding the drainage area and giving the reservoir excellent permeability characteristics, enabling rapid release of oil and gas after fracturing. Reservoirs with low porosity typically have poor connectivity, making them prone to proppant blockage or sedimentation, requiring fracturing stimulation to improve permeability and obtain stable initial production capacity.

[0058] Among the reservoir geomechanical parameters, the Pearson correlation coefficients for horizontal stress difference and brittleness index are 0.622 and 0.620, respectively, indicating that both parameters have a significant impact on post-fracturing production. Horizontal stress difference affects reservoir stimulation and post-fracturing production by controlling fracture morphology, propagation direction, and propagation effect. A larger horizontal stress difference causes fractures to tend to propagate along the direction of the minimum horizontal principal stress. This unidirectional fracture propagation means that higher closure pressure and stronger proppant are needed to maintain the fracture conductivity and thus maintain the fracturing effect to achieve stable post-fracturing production. Conversely, reservoirs with higher brittleness are more conducive to inducing microfractures to facilitate the flow rate of oil and gas from the reservoir.

[0059] Furthermore, the correlation coefficients for tensile strength and the proppant addition rate and displacement rate in the engineering parameters are both greater than 0.45, indicating that these parameters are closely related to production. Tensile strength leads to increased fracturing operation costs and reduced fracture complexity, thus affecting reservoir post-fracturing production. Displacement rate directly affects the injection rate of fracturing fluid into the reservoir, fracture propagation scale, and energy transfer efficiency. A reasonable displacement rate ensures that fracturing fluid effectively enters the reservoir and carries proppant to fill fractures, improving the overall conductivity of the reservoir and thus increasing the initially estimated post-fracturing production. The proppant addition intensity per meter is also crucial for improving oil and gas well production. With a reasonable proppant content, a higher proppant addition intensity can promote the formation of a wider proppant zone and improve fracture conductivity. Based on the above analysis, permeability, reservoir quality factor, pore structure index, tensile strength, horizontal stress difference, brittleness index, proppant addition intensity per meter, and displacement rate are selected as the main controlling factors affecting post-fracturing production in conglomerate reservoirs.

[0060] This embodiment uses the "mRmR+Pearson" method for feature extraction to remove redundancy and multicollinearity among features, and screens out the main controlling factors affecting the post-compression production of sandstone and conglomerate reservoirs, which is beneficial for subsequent model prediction.

[0061] Step 5: Data normalization.

[0062] This embodiment employs a maximum-minimum standardization method, mapping all data to a (0, 1) sphere to avoid dimensional differences between data points. The data standardization formula is as follows: (16) In the above formula, X represents the original values ​​of each column feature, Xmin and Xmax represent the maximum and minimum values ​​of the column feature, respectively; Xnorm represents the normalized value (range [0,1]).

[0063] The steps of data normalization can be summarized as follows: For each feature column, the maximum and minimum values ​​of the data in each column are calculated.

[0064] For each data point X in each column of features, the above formula is used to calculate the normalized value of Xnorm.

[0065] The normalized data is divided into training and testing sets for model training and accuracy evaluation. Step 6: Reservoir category analysis and positive / negative correlation analysis Four reservoir geological parameters were obtained for each well's corresponding sub-layer: clay content, porosity, permeability, and oil saturation. Based on previously established reservoir parameter calculation models, two other reservoir geological parameters were obtained: tensile strength and brittleness index, as well as pore structure index and reservoir quality factor. Engineering parameters for the corresponding sub-layers were also obtained from the fracturing operation site. Reservoir category evaluation indicators were established based on previously established reservoir classification standards to accurately classify sub-layer reservoirs. The reservoir geological and geomechanical parameter data obtained from each sub-layer were converted to values ​​for the corresponding perforated sections after fracturing, and combined with engineering data obtained from the engineering site for each perforated section. This multi-source data was fused to form a post-fracturing production prediction dataset. The correlation between the obtained multi-source data and production was analyzed, and the positive and negative correlations of each factor on post-fracturing production were evaluated.

[0066] Step 7: Construct a data-driven TL-AEAT-BiGRU post-press yield prediction model based on knowledge constraints.

[0067] Bidirectional gated recurrent neural networks (Bi-GRU) predict outputs based on bidirectional input sequences. By capturing upper and lower reservoir information through forward and backward propagation, they more effectively obtain the relationship between yield and reservoir, improving the model's generalization performance. The forward and backward hidden layer neuron structures of Bi-GRU are consistent, containing two gating structures: reset gate and update gate. The reset gate regulates the dependence on historical data, while the update gate is mainly used to adjust the balance between forgotten and new information.

[0068] The autoencoder module learns a constant function in an unsupervised manner, transforming the original input into an efficient compressed representation. This compressed representation captures potential key features with high quality and reconstructs the original input. Leveraging its significant feature extraction capabilities, the autoencoder can learn data-generalizable feature representations during the transfer learning pre-training stage, enhancing cross-domain generalization ability. Residual connections are added to the Bi-Gru layer to alleviate the vanishing gradient problem, reduce the impact of noise on prediction results, better capture complex patterns in reservoir depth sequences, learn the correlations between different well locations, and the connectivity between different regions within the reservoir, thus better understanding the spatial structure of the M-block sandstone-conglomerate reservoir and the reservoir connectivity and fluid flow paths after fracturing. A multi-head attention mechanism is introduced to perform multi-head attention calculations on the output of the Bi-GRU layer, allocating weights to different features to enable the model to process information from different feature spaces in parallel. This allows the model to automatically select key features, helping it to more accurately predict reservoir post-fracturing production.

[0069] Furthermore, to ensure that the model's predictions better align with objective understanding, this embodiment incorporates empirical knowledge regarding the correlation between "geological + engineering" data and post-compression yield as a penalty constraint term into the neural network model, serving as the boundary constraint for the model's prediction process. Here, the boundary constraint refers to the empirical knowledge constraint that expresses the positive and negative correlation between the main control factors influencing post-compression yield and the yield in the form of partial derivatives. Specifically, if yield is positively correlated with a main control factor, the partial derivative of yield with respect to that main control factor is greater than 0, thus determining the sign of the partial derivative of yield with respect to each main control parameter. Since the ReLU activation function preserves the original positive input values ​​and converts all negative input values ​​to 0, combined with the previous analysis of the positive and negative correlation between yield and each main control factor, if yield is positively correlated with a main control factor, the negative of the partial derivative is input to the ReLU activation function; if yield is negatively correlated with some main control factors, the partial derivative value is directly used as the input to the ReLU activation function.

[0070] The calculation result of the ReLU activation function is added to the loss function as a penalty constraint. The selected main control factor data is input into the post-pressure production prediction model, and a transfer learning strategy is introduced. Augmented data from Class A wells is used as the source domain dataset, and augmented data from Class B wells is used as the target domain dataset. The model is fine-tuned based on the target domain input data. Ultimately, accurate prediction of post-pressure production is achieved. The overall framework for source domain post-pressure production prediction is as follows: Figure 1 As shown.

[0071] The formula for calculating the loss function is shown in (17): (17) In the above formula, LOSS represents the total loss value of the model, and f represents the predicted value of the yield after pressing. , These represent the first and nth controlling factors, respectively. ... Let represent the constant penalty coefficients of the loss function for the partial derivatives of the output after compression with respect to each major controlling factor; to ensure that the errors generated by the neural network itself are of consistent importance with the errors generated under knowledge constraints, the penalty coefficients of the first type of loss function are defined. The value is 2, the second type of loss function ... The sum of the penalty items is 1.

[0072] Among them, based on Figure 6 to Figure 8 Based on the analysis results and previous correlation analyses of geology, engineering, and fracturing effects in the study area of ​​this embodiment, correlation coefficients were calculated and scatter plots were drawn to clarify the correlation between the main control parameters and production output. The correlation between the main control parameters and production output is shown in Table 8. Table 8: Correlation Relationship between Main Control Parameters and Output

[0073] The knowledge-constrained, data-driven TL-AEAT-BiGru model combines autoencoders, BiGBU, and multi-head attention mechanisms in transfer learning tasks. This allows for the effective capture of dependencies between features during fine-tuning, dynamically adjusting the focus to further enhance model stability. This overcomes the differences in geological features and fracturing effects across different data sets, preventing knowledge transfer from the source domain of block M from causing a performance degradation in the target domain and reducing the risk of negative transfer.

[0074] In some embodiments, model training mainly includes the following steps: 8.1 The multi-source data (including reservoir geological parameter data, reservoir geomechanical commitment data, and engineering parameter data) are divided into source domain and target domain data according to Class A well data and Class B well data, with Class A well data as source domain data and Class B well data as target domain data.

[0075] Then, the divided source domain data, after data augmentation, selection of key control factors, and data normalization, is input into the fusion deep learning output prediction model.

[0076] In this embodiment, Class A well data refers to multi-source data from the "Jinlong" Formation wells, and Class B well data refers to multi-source data from the "Mahu" Formation wells. Since the "Jinlong" and "Mahu" Formation wells both belong to the Upper Urho Formation in the study area, and have similar sedimentary backgrounds and heterogeneity (i.e., the geological characteristics of the two areas are similar but there are distribution offsets (such as differences in the mean porosity and permeability), which meets the requirements for dividing source and target domain data in transfer learning, the Class A wells from the "Jinlong" Formation and the Class B wells from the "Mahu" Formation are respectively divided into source and target domains.

[0077] 8.2 The multi-source data from the source domain first enters the autoencoder module. The autoencoder consists of an input layer, a hidden layer (64 units), and an output layer, which is used to reduce the dimensionality of the input data and extract features. After processing by the autoencoder, the data undergoes a linear transformation through the ReLU activation function to enhance the expressive power of the model.

[0078] Simultaneously, the source domain data enters the BiGru module through dimensionality transformation. BiGru contains multiple GRU units (64 units per layer), which can capture the sequential dependencies of production data from each well with depth sequence characteristics; the first layer of BiGru and the subsequent Nth layer of BiGru process the data sequentially, with each layer undergoing nonlinear transformation through the ReLU activation function; then the BiGru module outputs the feature representation after dimensionality transformation.

[0079] In section 8.3, the feature vectors from the autoencoder and BiGru modules are fed into the multi-head attention mechanism module. In this module, the feature vectors undergo linear transformation, dot product calculation, feature weighting, and a softmax operation to achieve mutual attention and weight allocation among different features. The output of the attention mechanism is then fused with the output of BiGru through a multiply operation to further enhance the correlation and importance of the features.

[0080] 8.4 The fused feature vectors enter the prediction layer and undergo operations such as Flattening, Linear (fully connected layer), and ReLU activation function in sequence. Finally, the prediction performance of the model is further optimized through residual learning blocks, and the model outputs the final output prediction yield of the source domain. After the model reaches the maximum number of iterations, the optimal model and optimal hyperparameters are saved.

[0081] 8.5 Transfer the trained and saved optimal source domain model to the target domain. Freeze the first two layers of BiGru in the model with the data preprocessed and enhanced data of the target domain. Fine-tune the last BiGru layer during the training of the target domain model.

[0082] 8.6 During the training of the target domain model, the error between the model prediction and the true value is quantified by the loss function characterized by mean squared error (MSE). This provides a clear direction for optimization algorithms (such as gradient descent) to adjust the model parameters, so as to minimize the error and improve the model performance.

[0083] Since the ReLU activation function preserves positive input values ​​and converts all negative inputs to 0, combined with the previous analysis of the positive and negative correlations between output and various controlling factors, if output is positively correlated with the controlling factors, the negative of the partial derivative is input into the ReLU activation function; if output is negatively correlated with some controlling factors, the partial derivative value is directly used as the input to the ReLU activation function. The calculation result of the ReLU activation function is added to the loss function as a penalty constraint term. At the end of each training session, the partial derivative values ​​of output with respect to each controlling parameter are calculated. If the sign of the partial derivative value satisfies the empirical constraint relationship between the predicted label and the corresponding feature, the penalty term is 0; if the penalty term is greater than 0, it will cause the loss function to increase during model training, allowing the model to better satisfy the empirical constraint relationship in subsequent training processes. This method achieves data-driven intelligent prediction of post-pressure output under knowledge constraints.

[0084] 8.7 Finally, the performance of post-press yield prediction is evaluated using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute prediction.

[0085] This embodiment uses the positive and negative correlations between each key control factor and yield as partial derivatives to serve as empirical knowledge constraints. Based on empirical knowledge constraints and transfer learning theory, a multi-module fusion-based, knowledge-constrained, data-driven TL-AEAT-BIGRU post-pressing yield prediction model is constructed to uncover the complex nonlinear mapping relationship between input key control factor data and reservoir post-pressing yield.

[0086] For the TL-AEAT-BIGRU post-pressure yield prediction model based on knowledge-constrained data joint drive, a three-layer BiGRU structure was designed, with the autoencoder and the number of BiGRU units in each layer set to 64. An Adam optimizer was employed to dynamically adjust the learning rate based on the historical gradient information and current gradient of each parameter. The ReLU activation function was used to reduce noise interference in the data. By setting appropriate batch size and training epochs, the model was ensured to fully capture the intrinsic features of the dataset. Parameters such as the number of attention heads and key dimensions in the multi-head attention mechanism were reasonably set to improve the model's ability to capture complex features. In the loss function, the penalty term coefficient... ... Appropriate values ​​should be selected to control the strength of the knowledge constraint. Parameter settings are shown in Table 6.

[0087] Table 6: Model Parameter Settings Table

[0088] Simultaneously, comparative experiments were conducted, referencing innovative methods used for post-compression production prediction. This embodiment applies knowledge constraints to the following four models to predict post-compression production in conglomerate reservoirs: 1) a CNN-LSTM hybrid model; 2) a masked autoencoder-based convolutional neural network (M-CNN); 3) a CNN-BiGru-AM model composed of CNN, bidirectional gated recurrent neural units, and an attention mechanism; and 4) a T-BiLSTM model combining BiLSTM with transfer learning algorithms. Figure 12 As shown in (a)-(d).

[0089] Table 9 evaluates the predictive performance of different models, with evaluation metrics including coefficient of determination (R²), mean squared error (MSE), and mean absolute error (MAE).

[0090] Table 9: Comparison of Experimental Results

[0091] The results show that the post-compression yield prediction method based on knowledge-constrained data-driven TL-AEAT-BIGRU, as proposed in this embodiment, outperforms other yield prediction models (such as CNN-LSTM, M-CNN, TCN, etc.) in post-compression yield prediction. These improvements ranged from 10% to 21%. Other models also demonstrated superior performance in predicting post-fracturing production capacity, but their feature extraction effects on geological data and engineering fracturing data from Block M were not as good as the model in this embodiment. The knowledge-constrained, data-driven TL-AEAT-BIGRU model achieved higher accuracy and stronger generalization ability in post-fracturing production prediction, validating its advantages in handling complex feature relationships and improving prediction accuracy.

[0092] For different production wells, Table 10 compares the actual mean and predicted mean of reservoir post-furlough production for each well in the dataset of this embodiment. The mean relative error (MAPE) is used to evaluate the prediction effect of the average reservoir post-furlough production for each well.

[0093] Table 10: Comparison of reservoir post-furlough production prediction results for different wells in Block M

[0094] It can be seen that the average relative error of the prediction for the 8 wells is less than 5%, which means that the data-driven TL-AEAT-BiGru sandstone and conglomerate reservoir post-compression production prediction model based on knowledge constraints has good applicability, and well control strategies can be reasonably optimized based on the prediction results.

[0095] To verify the effectiveness of the fusion model, ablation experiments were conducted, removing physical constraints, data augmentation, residual connections, attention mechanisms, autoencoders, and transfer learning, respectively, and analyzing the impact of these modules on model performance. Table 11 shows the comparison results of the ablation experiments.

[0096] Table 11: Comparison of Ablation Experiment Results

[0097] It can be seen that each module improves the model performance and the prediction effect of reservoir post-pressurization yield. The data augmentation module, residual connection module, attention mechanism module, autoencoder module, transfer learning, and knowledge constraint conditions improve the prediction accuracy of the model by 18%, 3%, 4%, 6%, 7%, and 9%, respectively. The fine-tuned complete model has the best performance, with an R² of 0.94. Ablation experiments show that the fusion model alleviates the gradient vanishing problem, enhances the model's feature representation ability, and achieves the best prediction effect.

[0098] Figure 13 to Figure 15The diagram shows the logging interpretation and production evaluation results for different well sections. Section a was perforated and fracturing in three stages, primarily consisting of Class III and IV reservoir sections, with some thinner Class I and II reservoir sections. Section b, during perforation and fracturing, mainly consisted of Class IV reservoir sections. Section c was perforated and fracturing in two stages, primarily consisting of Class III and IV reservoir sections, with some Class I reservoir sections interspersed. The actual production of the corresponding perforated sections in wells a, b, and c were 28.13 (…). ), 17.56 ), 33.81 ( The predicted production of the perforated section is 29.46 ( ), 16.78 ( ), 34.54 The knowledge-constrained TL-AEAT-BiGru deep learning model can accurately predict the production situation after fracturing in each perforation section, and this method has guiding significance for the prediction of post-fracturing production in the study area.

[0099] This embodiment proposes a multi-module, knowledge-constrained, data-driven TL-AEAT-BiGru post-pile yield prediction model. Geological and engineering data from block M are selected for neural network training. Data augmentation and feature filtering are performed on the geological and engineering data of the work area. Based on the filtered key control factor data, the key control factor dataset is input into the post-pile yield prediction model to achieve accurate prediction of post-pile yield.

[0100] Comparative experiments show that the model used in this embodiment outperforms conventional machine learning models in terms of prediction performance. Ablation experiments demonstrate that the data augmentation module, residual connection module, attention mechanism module, autoencoder module, transfer learning module, and knowledge constraint module all contribute to improving the overall prediction performance of the model, and the model exhibits better prediction results and performance when all modules work together. Furthermore, the comparison between the actual and predicted production of perforated sections containing different reservoir types in the study area was analyzed, indicating that the model used in this embodiment can accurately predict the production of each perforated section after fracturing.

[0101] This application employs a conditionally constrained CWGAN-GP model to augment limited, multi-source data. A Pearson+mRmR correlation analysis algorithm is used to identify the main controlling factors influencing post-compression production. Principal component analysis is used to classify reservoirs based on previously established reservoir classification standards. Furthermore, empirical knowledge of the correlation between post-compression production and the input main controlling factors is combined with the TL-AEAT-BiGRU post-compression production prediction model, using the selected main controlling factors as input for post-compression production prediction. Results show that this model performs best in post-compression production prediction in the study area compared to other production prediction models. Ablation experiments demonstrate that each module of the model contributes to improving production prediction accuracy, resulting in a production prediction accuracy of 0.94. This effectively solves the problem of low accuracy in reservoir post-compression production prediction when sample size is small, providing guidance for oil and gas well production analysis.

[0102] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving, characterized in that, Includes the following steps: Construct a TL-AEAT-BIGRU post-press yield prediction model based on knowledge-constrained data joint drive; The main controlling factors affecting post-press yield are input into the TL-AEAT-BIGRU post-press yield prediction model to train the model and obtain the trained TL-AEAT-BIGRU post-press yield prediction model.

2. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving as described in claim 1, characterized in that, The main controlling factors include permeability, reservoir quality factor, pore structure index, tensile strength, horizontal stress difference, brittleness index, sand addition intensity per meter, and discharge rate.

3. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving as described in claim 1, characterized in that, Methods for determining the main controlling factors affecting reservoir post-furlition production include: S1, Obtaining multi-source datasets: 1.1 Obtain reservoir geological parameters, reservoir geomechanical parameters, engineering parameters, and post-pressurization production data for each reservoir; 1.2 Establish reservoir classification and evaluation indicators to classify reservoirs into different categories; 1.3 The reservoir geological parameters, reservoir geomechanical parameters and post-pressurization production values ​​of each layer are converted into corresponding values ​​under the corresponding perforation section, and combined with the engineering data obtained by each perforation section at the engineering site to form a multi-source dataset. S2 uses the "mrmr+PEARSON" feature selection algorithm to analyze the correlation between reservoir geological parameters, reservoir geomechanical parameters, engineering parameters and post-pressing production, and screens out the main controlling factors affecting reservoir post-pressing production.

4. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving as described in claim 3, characterized in that, The reservoir geological parameters include porosity, permeability, pore structure index, resistivity, sonic transit time, clay content, formation pressure, and density. The reservoir geomechanical parameters include brittleness index, Poisson's ratio, tensile strength, horizontal stress difference, minimum horizontal principal stress, maximum horizontal principal stress, pore structure index, and reservoir quality factor. The engineering parameters include the liquid shortage ratio after backflow, sand addition intensity per meter, instantaneous pump stop pressure, discharge rate, and average sand ratio.

5. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving as described in claim 3, characterized in that, Before step S2, the reservoir geological parameters, reservoir geomechanical parameters and engineering parameters are enhanced by using a CWGAN-GP network that incorporates dimensionless skewness and kurtosis calculation formulas as constraints.

6. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving as described in claim 1, characterized in that, The method for constructing the TL-AEAT-BIGRU post-press yield prediction model includes: Input feature extraction is performed in the autoencoder network layer; Residual connections are added to the Bi-GRU layer of a bidirectional gated recurrent neural network; A multi-head attention mechanism is introduced to perform multi-head attention calculation on the output of the autoencoder layer, and then feature fusion is performed with the input result of the Bi-GRU layer with residual connection added. The positive and negative correlation between the main controlling factors affecting the yield after compression and the yield is introduced as a penalty constraint term into the bidirectional gated recurrent neural network.

7. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving as described in claim 6, characterized in that, The method for introducing the penalty constraint into the bidirectional gated recurrent neural network is as follows: if a certain controlling factor is positively correlated with output, the negative of the partial derivative of the controlling factor is input into the ReLU activation function; if a certain controlling factor is negatively correlated with output, the partial derivative of the controlling factor is directly used as the input of the ReLU activation function; the calculation result of the ReLU activation function is added to the loss function as the penalty constraint.

8. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving as described in claim 7, characterized in that, The formula for calculating the loss function of the TL-AEAT-BIGRU post-pressurization yield prediction model is as follows: ， In the above formula, LOSS represents the total loss value of the model, and f represents the predicted value of the yield after pressing. , These represent the first and nth controlling factors, respectively. ... Let represent the constant penalty coefficients of the loss function for the partial derivatives of the output after compression with respect to each major controlling factor; to ensure that the errors generated by the neural network itself are of consistent importance with the errors generated under knowledge constraints, the penalty coefficients of the first type of loss function are defined. The value is 2, the second type of loss function ... The sum of the penalty items is 1.

9. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving according to any one of claims 1-8, characterized in that, The transfer learning strategy is applied to model training, and the transfer learning strategy includes the following steps: The multi-source dataset is divided into source domain data and target domain data for transfer learning; Select key control factors for source and target domain data; Standardize the data of the main control factors; The TL-AEAT-BIGRU post-press yield prediction model is trained based on the source domain master control factor data. After reaching the maximum number of iterations, the optimal model and the optimal hyperparameter combination are saved to obtain the optimal model corresponding to the optimal hyperparameters trained by the source domain master control factor data. The optimal model corresponding to the optimal hyperparameters trained on the source domain master control factor data is transferred to the target domain master control factor dataset. During the model training process, the first two layers of the autoencoder network and the Bi-GRU network are frozen, and the last layer of the network is trained based on the target domain master control factor data.

10. The TL-AEAT-BIGRU post-press yield prediction method based on knowledge-constrained data joint driving as described in claim 9, characterized in that, Based on the TL-AEAT-BIGRU post-pressure output prediction model driven by data under knowledge constraints, the model is fine-tuned for the target domain input data. At the end of each training session, the partial derivative values ​​of output with respect to each main control parameter are calculated. If the sign of the partial derivative value satisfies the empirical constraint relationship between the prediction label and the main control factor, then the penalty term is 0. If the penalty term is greater than 0, it will cause the loss function to increase during model training, so that the model can better satisfy the empirical constraints in subsequent training processes. The empirical constraint relationship is based on the positive and negative correlation between the main controlling factors affecting the yield after pressing and the yield.

Images