A method and system for constructing a reservoir evaluation model based on deep learning

Abstract

The present invention relates to the field of reservoir evaluation technology, and more particularly to a method and system for constructing a reservoir evaluation model based on deep learning. The method comprises the following steps: acquiring well logging data and core analysis data from a target work area, performing outlier processing and missing value filling, respectively, to obtain complete well logging data and complete core data; calculating reservoir pore structure parameters based on the complete core data; and inverting reservoir macroporosity based on the complete well logging data. By fusing core and logging data, combining pore structure parameters, macroporosity, and a geological attention mechanism, the present invention significantly improves the accuracy, stability, and robustness of reservoir evaluation, addresses data loss and small sample size issues, and enhances the model's interpretability and geological consistency.

Description

Technical Field

[0001] The present invention relates to the field of reservoir evaluation technology, and in particular to a method and system for constructing a reservoir evaluation model based on deep learning. Background Art

[0002] During oil and gas exploration and development, reservoir evaluation is a crucial tool for identifying and quantifying oil and gas reservoirs. Its accuracy is directly related to resource utilization efficiency and development benefits. Traditional reservoir evaluation techniques have evolved from core experiments and empirical formulas for well logging interpretation to statistical regression and machine learning models. Early methods, such as Archie's formula and Wyllie's time-averaged equation, were based on simplified physical assumptions and were suitable for specific lithologies, but they lacked adaptability under complex geological conditions. With the advancement of data mining technology, methods such as support vector machines (SVMs) and random forests (RFs) have been applied to reservoir evaluation, improving their ability to handle multiple variables. However, these methods generally ignore the depth sequence characteristics and geological context of well logging data during modeling, limiting the models' predictive accuracy and generalization capabilities.

[0003] One-dimensional convolutional neural networks (1D-CNNs) can effectively capture local patterns in well logs, but they still struggle to incorporate prior geological knowledge. Introducing an attention mechanism can enhance the model's focus on key features. However, existing attention mechanisms are mostly data-driven and lack geological knowledge constraints, which can easily lead to attention drift. Furthermore, models trained on small sample reservoirs lack sufficient training data, which affects prediction stability. Summary of the Invention

[0004] Based on this, it is necessary for the present invention to provide a method and system for constructing a reservoir evaluation model based on deep learning to solve at least one of the above technical problems.

[0005] To achieve the above objectives, a method for constructing a reservoir evaluation model based on deep learning includes the following steps:

[0006] Step S1: Acquire well logging data and core analysis data of the target work area, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data;

[0007] Step S2: Calculate reservoir pore structure parameters based on complete core data; invert reservoir macroporosity based on complete logging data;

[0008] Step S3: Analyze reservoir lithofacies physical properties based on reservoir pore structure parameters and reservoir macroporosity, determine pore connectivity index, reservoir quality factor, lithologic control coefficient, and heterogeneity index, and generate a reservoir evaluation index dataset;

[0009] Step S4: Divide the reservoir evaluation index dataset into a training set, a validation set, and a test set according to the layer interval; use the training set to construct the basic structure of the one-dimensional convolutional neural network model, and introduce the geological attention mechanism to obtain a geological attention enhanced neural network;

[0010] Step S5: Weakly supervise the geological attention enhanced neural network to obtain a geological knowledge constrained attention network; synthesize and expand the training set by accumulating small sample reservoirs through a generative adversarial network to obtain a balanced training data set; use the geological knowledge constrained attention network and the balanced training data set to perform deep learning model training, and perform hyperparameter optimization and performance monitoring based on the validation set to obtain a candidate model;

[0011] Step S6: Use the test set to evaluate and optimize the performance of the candidate model to obtain an initial reservoir evaluation model; integrate the preset SZ3 framework with the initial reservoir evaluation model to generate a final reservoir evaluation model.

[0012] This invention improves the accuracy and data integrity of reservoir evaluation by systematically integrating core data and well logging data, solving the problem of error accumulation caused by missing data and improper anomaly handling in traditional methods. The introduction of multi-scale fusion of pore structure parameters and macro-porosity enhances the comprehensive analysis capability of reservoir physical properties, which helps to improve the accuracy of reservoir quality identification. The geological attention mechanism is combined with a one-dimensional convolutional neural network to enhance the model's perception of key geological features and expression of geological context, overcoming the problem of attention drift in pure data-driven attention mechanisms. At the same time, through weak supervision signal guidance and geological knowledge embedding, the interpretability and geological consistency of the model are effectively improved. To address the problem of small sample reservoirs, a generative adversarial network is used for data expansion, which enhances the training stability and generalization ability of the model in the case of insufficient samples. The final model has stronger robustness, adaptability and prediction accuracy, significantly improving the efficiency of reservoir evaluation and development decision support capabilities under complex geological conditions.

[0013] Preferably, the present invention further provides a system for constructing a reservoir evaluation model based on deep learning, which is used to execute the above-mentioned method for constructing a reservoir evaluation model based on deep learning. The system for constructing a reservoir evaluation model based on deep learning includes:

[0014] The data preprocessing module is used to obtain well logging data and core analysis data of the target work area, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data;

[0015] The pore parameter inversion module is used to calculate reservoir pore structure parameters based on complete core data; and to invert reservoir macroporosity based on complete well logging data;

[0016] The reservoir attribute analysis module is used to analyze reservoir lithofacies physical properties based on reservoir pore structure parameters and reservoir macroporosity, determine pore connectivity index, reservoir quality factor, lithologic control coefficient and heterogeneity index, and generate a reservoir evaluation index data set;

[0017] The geological attention model construction module is used to divide the reservoir evaluation index dataset into training set, validation set and test set according to the layer segment; the training set is used to build the basic structure of the one-dimensional convolutional neural network model, and the geological attention mechanism is introduced to obtain the geological attention enhanced neural network;

[0018] The model training and enhancement module is used to guide the geological attention enhancement neural network with weak supervision signals to obtain a geological knowledge constrained attention network. The training set is expanded by data synthesis of small sample reservoirs through generative adversarial networks to obtain a balanced training data set. The geological knowledge constrained attention network and the balanced training data set are used to perform deep learning model training, and hyperparameter optimization and performance monitoring are performed based on the validation set to obtain a candidate model.

[0019] The model evaluation and integration module is used to evaluate and optimize the performance of candidate models using the test set to obtain an initial reservoir evaluation model; the preset SZ3 framework is integrated with the initial reservoir evaluation model to generate the final reservoir evaluation model.

[0020] The present invention effectively improves the accuracy and reliability of reservoir evaluation through the collaborative work of multiple modules. In the data preprocessing stage, the integrity of logging data and core data is ensured by outlier processing and missing value filling, providing high-quality input data for subsequent analysis. The pore parameter inversion module combines core data and logging data to achieve accurate inversion of reservoir pore structure and macroporosity, laying the foundation for the analysis of reservoir physical properties. The reservoir attribute analysis module further analyzes pore connectivity, reservoir quality, lithologic control and heterogeneity, providing a comprehensive reservoir indicator data set for comprehensive evaluation. The geological attention model construction module optimizes the model's attention to key geological features by introducing a geological attention mechanism, making the model more interpretable and accurate when processing complex geological data. The model training and enhancement module enhances the model's learning and generalization capabilities through weak supervision guidance and small sample data expansion, ensuring that the model can still be efficiently trained and optimized when the data is incomplete or unbalanced. Finally, the model evaluation and integration module generated a final reservoir evaluation model that was accurate and consistent with geological reality through test set evaluation and optimization, combined with geological knowledge constraints. This comprehensively improved the intelligence level and application effect of reservoir evaluation, and provided strong technical support for oil and gas exploration and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0022] Figure 1 Schematic diagram of the steps of a method for constructing a reservoir evaluation model based on deep learning according to the present invention;

[0023] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0024] Figure 3 for Figure 1 Detailed step flow chart of step S3 in FIG. DETAILED DESCRIPTION

[0025] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0026] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0027] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0028] To achieve this, please refer to Figures 1 to 3 The present invention provides a method for constructing a reservoir evaluation model based on deep learning, the method comprising the following steps:

[0029] Step S1: Acquire well logging data and core analysis data of the target work area, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data;

[0030] Step S2: Calculate reservoir pore structure parameters based on complete core data; invert reservoir macroporosity based on complete logging data;

[0031] Step S3: Analyze reservoir lithofacies physical properties based on reservoir pore structure parameters and reservoir macroporosity, determine pore connectivity index, reservoir quality factor, lithologic control coefficient, and heterogeneity index, and generate a reservoir evaluation index dataset;

[0032] Step S4: Divide the reservoir evaluation index dataset into a training set, a validation set, and a test set according to the layer interval; use the training set to construct the basic structure of the one-dimensional convolutional neural network model, and introduce the geological attention mechanism to obtain a geological attention enhanced neural network;

[0033] Step S5: Weakly supervise the geological attention enhanced neural network to obtain a geological knowledge constrained attention network; synthesize and expand the training set by accumulating small sample reservoirs through a generative adversarial network to obtain a balanced training data set; use the geological knowledge constrained attention network and the balanced training data set to perform deep learning model training, and perform hyperparameter optimization and performance monitoring based on the validation set to obtain a candidate model;

[0034] Step S6: Use the test set to evaluate and optimize the performance of the candidate model to obtain an initial reservoir evaluation model; integrate the preset SZ3 framework with the initial reservoir evaluation model to generate a final reservoir evaluation model.

[0035] In the embodiment of the present invention, reference Figure 1 FIG. 1 is a flow chart showing the steps of a method for constructing a reservoir evaluation model based on deep learning according to the present invention. In this example, the method for constructing a reservoir evaluation model based on deep learning includes the following steps:

[0036] Step S1: Acquire well logging data and core analysis data of the target work area, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data;

[0037] In the embodiment of the present invention, natural gamma ray data is first collected using a sodium iodide crystal detector with an operating voltage of 1200V, a current of 10mA, a crystal size of 5cm×5cm×5cm, and a sampling interval of 0.1s. At the same time, a comprehensive logging instrument is used to collect acoustic transit time, density, neutron porosity, and resistivity data. The acoustic transit time sensor has an operating frequency of 20kHz and a measurement range of 30μs / m to 200μs / m; the density sensor has a source distance of 35cm and a measurement accuracy of ±0.01g / cm³; the neutron source intensity of the neutron porosity sensor is The resistivity sensor operates at a frequency of 1 MHz and has a measurement range of 0.1 Ω·m to 2000 Ω·m. All logging data was collected at a depth interval of 0.1524 m. Next, porosity, permeability, and water saturation data were collected using a core porosity and permeability meter. This meter uses a gas permeability method, has an operating pressure range of 0 MPa to 70 MPa, and a temperature control range of 20°C to 300°C. The porosity measurement accuracy is ±0.5%, the permeability measurement accuracy is ±10%, and the water saturation measurement accuracy is ±3%. The core sample size is 3.8 cm in diameter and 10 cm in length. The collected logging data and core analysis data were then depth-matched using a bicubic interpolation algorithm, achieving a matching accuracy of ±0.1 m. For well logging data, the 3σ principle is used to handle outliers. That is, data outside the range of the mean plus or minus 3 times the standard deviation are considered outliers, and missing values ​​are filled using linear interpolation. For core analysis data, outliers are identified using the box-and-whisker plot method, and missing values ​​are filled using the mean method, thus obtaining complete well logging data and complete core data.

[0038] Step S2: Calculate reservoir pore structure parameters based on complete core data; invert reservoir macroporosity based on complete logging data;

[0039] In this embodiment of the present invention, complete core data was scanned using X-ray computed tomography (CT) equipment. The CT equipment's X-ray tube voltage was 120 kV, the current was 100 μA, the detector pixel size was 100 μm × 100 μm, the scanning slice thickness was 0.5 mm, and the scanning range covered the entire length of the core. The scanned image data was imported into image processing software and processed using an adaptive threshold segmentation algorithm, where the threshold was calculated using the Otsu method. The segmented three-dimensional distribution data of the pore space was stored in voxel form. A pore geometry and topology feature extraction tool based on the Marching Cubes algorithm was used to calculate geometric features such as pore volume, surface area, and connectivity. The extracted results were then subjected to K-means cluster analysis with a K value set to 3 to obtain pore type classification results. A lithology-dependent scaling box dimension calculation tool based on fractal theory was used to calculate fractal dimension feature data of the pore space within a range of 2 to 3. Based on the pore type classification results, an integral calculation method was used to quantify the volume of each pore type to obtain microporosity distribution data. The maximum inscribed sphere algorithm was used to calculate the pore type classification results and obtain pore throat radius distribution data. The tensor field analysis tool was used to analyze the pore geometry and obtain reservoir pore geometry parameters. Furthermore, the complete logging data was calibrated for multi-sensor acquisition accuracy and standardized. The least squares method was used to fit the calibration curve and convert the data collected by different sensors into standardized logging curve data. Principal component analysis was used to perform multivariate nonlinear covariance analysis on the standardized logging curve data and extract logging response characteristic parameters. Based on the logging response characteristic parameters, a spectral analysis method based on the Fast Fourier Transform (FFT) was used to analyze reservoir fluid properties. Reservoir fluid phase separation was performed based on the spectral analysis results to obtain fluid type distribution data. The standardized logging curve data was combined with the fluid type distribution data, and the initial effective porosity data was obtained using a neutron-density porosity joint inversion method based on Bayesian inversion. Based on the initial effective porosity data, the clay mineral content in the reservoir was measured using an X-ray diffractometer. The diffraction angle range was 2θ = 3° to 60°, the scanning step was 0.02°, and the diffraction intensity was recorded in counting rate mode. The clay mineral content was calculated based on the diffraction peak intensity and position. The initial effective porosity data was corrected in combination with the correction coefficient to obtain the reservoir macroporosity data.

[0040] Step S3: Analyze reservoir lithofacies physical properties based on reservoir pore structure parameters and reservoir macroporosity, determine pore connectivity index, reservoir quality factor, lithologic control coefficient, and heterogeneity index, and generate a reservoir evaluation index dataset;

[0041] In an embodiment of the present invention, first, the pore geometry parameters and pore throat radius distribution data are input into an analysis tool based on a graph theory algorithm. The tool adopts a breadth-first search algorithm and uses pore connectivity conditions as search rules to construct a pore network topology relationship diagram. Based on this diagram, the Dijkstra shortest path algorithm is used to track the seepage path and extract effective fluid channel data such as path length, number and connectivity probability. These data are input into a numerical integration tool and numerically integrated using the Simpson integral method. The results are statistically normalized based on the maximum and minimum normalization method to obtain a pore connectivity index. Secondly, the pore fractal dimension feature data is analyzed using a scale qualitative analysis tool based on fractal geometry theory to obtain an initial pore complexity score. This score and the microporosity distribution data are input into a weighted fusion tool and fused using the weighted average method with weight coefficients of 0.6 and 0.4, respectively, to obtain an initial pore structure complexity index. This index was input into the experimental verification and calibration module of the porous media mercury intrusion instrument, using a pressure calibration method with a calibration pressure range of 0 MPa to 70 MPa to obtain a modified pore structure complexity index. Next, the reservoir macroporosity data and pore connectivity index were input into an assessment tool based on a reservoir physical model. A linear regression method was used to construct a reservoir fluid storage capacity assessment model and generate a reservoir capacity factor. Simultaneously, the permeability data and pore connectivity index from the complete core data were input into a correlation analysis tool, and a correlation analysis was performed using the Pearson correlation coefficient method to obtain a seepage capacity factor. The reservoir capacity factor and seepage capacity factor were then input into a multi-objective optimization tool with the optimization objectives of maximizing the reservoir capacity factor and the seepage capacity factor, respectively. A weighted sum method was used for integration, with weight coefficients of 0.7 and 0.3, respectively, to obtain a reservoir quality factor. Then, based on the fluid type distribution data and reservoir macroporosity data, the data were input into a rock-fluid interaction analysis tool, and an analysis method based on fluid mechanics and thermodynamic equilibrium was used to obtain wettability assessment results. This result, along with the modified pore structure complexity index, was input into the Multivariate Regression Tool using multiple linear regression to obtain the lithologic control coefficient. Reservoir pore geometry parameters and pore type classification results were input into the Spatial Variability Identification Tool using semivariogram analysis to identify spatial variability and generate the spatial variation coefficient. This coefficient was then input into the Sequence Constraint Processing Tool using a sequence stratigraphic constraint algorithm to obtain the heterogeneity index. Finally, the reservoir quality factor, lithologic control coefficient, and heterogeneity index were input into the Analytic Hierarchy Process (AHP) Weight Assignment Tool. Using the AHP calculation process, a judgment matrix was constructed, and the index weight matrix was calculated using the eigenvalue method. This matrix and the evaluation indicators were then input into the Weighted Fusion Tool using the weighted summation method to obtain the comprehensive reservoir evaluation score. This score was then input into the Quantitative Classification Tool, which used an equidistant segmentation method to classify the score range into five levels, resulting in the reservoir evaluation index dataset.

[0042] Step S4: Divide the reservoir evaluation index dataset into a training set, a validation set, and a test set according to the layer interval; use the training set to construct the basic structure of the one-dimensional convolutional neural network model, and introduce the geological attention mechanism to obtain a geological attention enhanced neural network;

[0043] In this embodiment of the present invention, the reservoir evaluation index dataset is first input into a feature correlation analysis tool. This tool uses the Pearson correlation coefficient method with a threshold set to 0.7 to detect multicollinearity, and the variance inflation factor (VIF) method with a threshold set to 10 to generate a standardized feature dataset. Then, based on preset geological stratification information, the geological stratification identification tool employs a deep learning-based convolutional neural network (CNN) algorithm, with a training set ratio of 75%, a validation set ratio of 15%, and a test set ratio of 15%. The standardized feature dataset is then segmented into reservoir layers to obtain a labeled layer segment dataset. The training set was used to construct the basic structure of a one-dimensional convolutional neural network model using the TensorFlow framework. The model had 128 neurons in the input layer, a 3×3 kernel size with a stride of 1 and padding of the same, and a Reluctant Unit (ReLU) activation function. A combination of max and average pooling was used, with a pooling window size of 2×2 and a stride of 2. The fully connected layer had 64 neurons, the output layer had 8 neurons, and a softmax activation function. Nonlinear transformations were performed to obtain the basic network structure. Dropout and L2 weight regularization were applied to the basic network structure, with a dropout probability of 0.5 and an L2 regularization coefficient of 0.01, to obtain an enhanced network structure that prevents overfitting. A self-attention computation unit was introduced, using the Scaled Dot-ProductAttention mechanism with 8 attention heads and a hidden layer dimension of 512. Weights were initialized to obtain the geologically sensitive feature enhancement component. Weights were assigned to the geologically sensitive feature enhancement component, and a gradient descent-based optimization algorithm with a learning rate of 0.001 was used to generate key geological features. Based on key geological features, the original convolutional features were integrated at multiple scales using a feature pyramid network (FPN) architecture, integrating convolution kernels of 1×1, 3×3, and 5×5 scales to generate geological enhancement features. Skip connections were constructed based on the geological enhancement features, using the skip connection method in the U-Net architecture with four connection layers, resulting in a gradient transfer optimization network structure. An attention-guided loss combination was designed, combining the cross-entropy loss function and the Dice loss function with weights of 0.7 and 0.3, respectively, to obtain a geologically aware comprehensive loss function. Using the distributed deep learning framework Horovod, the basic network structure, the geologically sensitive feature enhancement component, the gradient transfer optimization network structure, and the geologically aware comprehensive loss function were integrated into a geological attention enhancement neural network.

[0044] Step S5: Weakly supervise the geological attention enhanced neural network to obtain a geological knowledge constrained attention network; synthesize and expand the training set by accumulating small sample reservoirs through a generative adversarial network to obtain a balanced training data set; use the geological knowledge constrained attention network and the balanced training data set to perform deep learning model training, and perform hyperparameter optimization and performance monitoring based on the validation set to obtain a candidate model;

[0045] In an embodiment of the present invention, first, a geological constraint rule set is obtained, which includes 10 geological constraint rules such as reservoir type, pore structure, and permeability. Each rule sets a clear threshold, such as a porosity threshold of 15%-30% and a permeability threshold of 0.1-1000mD. Based on this rule set, the training set is matched rule by rule, and an exact matching algorithm is used with a matching accuracy of 100% to obtain a weak supervision signal label. The weak supervision signal label is applied to a geological attention enhanced neural network, and an attention weight optimization algorithm is used. The optimization goal is to minimize the prediction error, the learning rate is 0.001, and the number of iterations is 1000 times to generate a weak supervision guidance architecture. Based on the weak supervision guidance architecture, a geological rule consistency loss term is designed, and the loss value is calculated using the mean square error method with a weight coefficient of 0.8 to generate a multi-objective loss function. The multi-objective loss function is constrained optimization using the Lagrange multiplier method. The constraint condition is that the loss value is less than 0.01, and a geological knowledge constraint optimizer is obtained. A geological attention-enhanced neural network and a geological knowledge-constrained optimizer were integrated and reconstructed using a modular integration approach and a TCP / IP interface protocol, resulting in a geological knowledge-constrained attention network. The training set was subjected to statistical and cluster analysis of the sample distribution of each reservoir type. A K-means clustering algorithm was used with a K value of 5 and a clustering accuracy of 95% to identify small sample paragraphs and obtain a small sample paragraph label set. Feature-preserving data generation was performed on the small sample paragraph label set using a generative adversarial network (GAN). Both the generator and discriminator of the GAN were five-layer neural networks with Reluctant Unit (ReLU) activation functions, an Adam optimization algorithm, a learning rate of 0.002, and a batch size of 32. Synthetic reservoir small sample data was generated. The training set and synthetic reservoir small sample data were then distributed resampled using a stratified sampling method with a sampling rate of 1.5. Based on the distribution resampling results, the weight ratio of each reservoir type was adjusted with a correction factor of 0.8-1.2 to obtain a balanced training dataset. A geologically constrained attention network was trained on a balanced training dataset using forward and backpropagation, employing stochastic gradient descent with a learning rate of 0.01, a momentum coefficient of 0.9, a batch size of 64, and 2000 iterations to obtain the initial network weight parameter set. The loss function values, validation set prediction errors, and gradient trends during training were statistically analyzed. Dynamic performance indicators were calculated using a moving average method with a window size of 50, generating a dynamic performance indicator dataset. An automated hyperparameter search was performed on the geologically constrained attention network using the dynamic performance indicator dataset. Using a Bayesian optimization algorithm, the search space ranged from a learning rate of 0.001 to 0.1, a batch size of 16 to 128, and an iteration count of 1000 to 3000, resulting in the optimal hyperparameter combination. Based on this optimal hyperparameter combination, batch training was performed on the balanced training dataset using parallel computing and GPU acceleration for 3000 iterations to obtain the candidate model parameter set.Based on the validation set, early stopping and Dropout validation are performed on the candidate model parameter set. The early stopping condition is that the validation loss does not decrease for 10 epochs and the Dropout probability is 0.5 to obtain the candidate model.

[0046] Step S6: Use the test set to evaluate and optimize the performance of the candidate model to obtain an initial reservoir evaluation model; integrate the preset SZ3 framework with the initial reservoir evaluation model to generate a final reservoir evaluation model.

[0047] In this embodiment of the present invention, a candidate model was used for prediction using a test set consisting of 1,000 reservoir samples, each with 10 features and one true reservoir parameter label. The test set data was fed into the candidate model and a forward propagation algorithm was used to generate a set of reservoir parameter predictions. These predictions included five reservoir parameters, including porosity and permeability, with a prediction accuracy of 95%. Error comparison analysis was performed between the predicted reservoir parameter values ​​and the true core labels, using root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) as error evaluation metrics. Based on the model error evaluation metric set, parameter and structural fine-tuning was performed on the candidate model. Parameter fine-tuning was performed using gradient descent with a learning rate of 0.01, 100 iterations, and a batch size of 32. Structural fine-tuning involved adding one hidden layer with 64 neurons and a Reinforced Lu (ReLU) activation function. This resulted in an initial reservoir evaluation model with a prediction accuracy of 97% for the test set. The initial reservoir evaluation model was integrated with the pre-set SZ3 framework using a modular approach. The interface protocol was TCP / IP, and the data transmission format was JSON. The SZ3 framework provided data preprocessing, feature extraction, and model optimization. The initial reservoir evaluation model sent its prediction results in JSON format to the SZ3 framework, which then performed post-processing on the results, including error correction and visualization, ultimately producing the final reservoir evaluation model.

[0048] This invention improves the accuracy and data integrity of reservoir evaluation by systematically integrating core data and well logging data, solving the problem of error accumulation caused by missing data and improper anomaly handling in traditional methods. The introduction of multi-scale fusion of pore structure parameters and macro-porosity enhances the comprehensive analysis capability of reservoir physical properties, which helps to improve the accuracy of reservoir quality identification. The geological attention mechanism is combined with a one-dimensional convolutional neural network to enhance the model's perception of key geological features and expression of geological context, overcoming the problem of attention drift in pure data-driven attention mechanisms. At the same time, through weak supervision signal guidance and geological knowledge embedding, the interpretability and geological consistency of the model are effectively improved. To address the problem of small sample reservoirs, a generative adversarial network is used for data expansion, which enhances the training stability and generalization ability of the model in the case of insufficient samples. The final model has stronger robustness, adaptability and prediction accuracy, significantly improving the efficiency of reservoir evaluation and development decision support capabilities under complex geological conditions.

[0049] Preferably, step S1 includes the following steps:

[0050] Step S11: collecting natural gamma rays through a sodium iodide crystal detector, and collecting acoustic wave transit time, density, neutron porosity and resistivity through a comprehensive logging instrument to generate logging data;

[0051] Step S12: collecting porosity, permeability and water saturation using a core porosity and permeability measuring instrument to generate core analysis data;

[0052] Step S13: perform depth matching on the well logging data and the core analysis data, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data.

[0053] In the embodiment of the present invention, natural gamma ray data is first collected by a sodium iodide crystal detector. The working voltage of the detector is 1200V, the current is 10mA, the crystal size is 5cm×5cm×5cm, and the sampling interval is 0.1s. At the same time, a comprehensive logging instrument is used to collect acoustic transit time, density, neutron porosity and resistivity data. The operating frequency of the acoustic transit time sensor is 20kHz, and the measurement range is 30μs / m to 200μs / m; the source distance of the density sensor is 35cm, and the measurement accuracy is ±0.01g / cm³; the neutron source intensity of the neutron porosity sensor is The resistivity sensor operates at a frequency of 1 MHz and has a measurement range of 0.1 Ω·m to 2000 Ω·m. All logging data was collected at a depth interval of 0.1524 m. Next, porosity, permeability, and water saturation data were collected using a core porosity and permeability meter. This meter uses a gas permeability method with an operating pressure range of 0 MPa to 70 MPa and a temperature control range of 20°C to 300°C. The porosity measurement accuracy is ±0.5%, the permeability measurement accuracy is ±10%, and the water saturation measurement accuracy is ±3%. The core sample dimensions are 3.8 cm in diameter and 10 cm in length. The collected logging data and core analysis data were then depth-matched using a bicubic interpolation algorithm, achieving a matching accuracy of ±0.1 m. For well logging data, the 3σ principle is used to handle outliers. That is, data outside the range of the mean plus or minus 3 times the standard deviation are considered outliers, and missing values ​​are filled using linear interpolation. For core analysis data, outliers are identified using the box-and-whisker plot method, and missing values ​​are filled using the mean method, thus obtaining complete well logging data and complete core data.

[0054] This invention integrates multiple logging and experimental equipment to comprehensively collect multi-physical logging data such as natural gamma, acoustic wave transit time, density, neutron porosity, and resistivity, as well as key physical parameters such as core porosity, permeability, and water saturation, thereby achieving multi-source complementarity of reservoir information. By deeply matching logging and core data, the consistency and spatial correspondence accuracy of the data are improved, laying a high-quality data foundation for subsequent attribute inversion and model training. Outlier processing and missing value filling further enhance the integrity and usability of the data, reduce the impact of noise interference and information loss on analysis results, ensure the stability and credibility of model input data, and effectively improve the accuracy and reliability of reservoir evaluation.

[0055] Preferably, calculating reservoir pore structure parameters based on complete core data in step S2 includes:

[0056] Perform CT scanning on the complete core data to generate a three-dimensional digital model of the core;

[0057] Based on the adaptive threshold value of the three-dimensional digital model of the core, the reservoir pore space is segmented to obtain the three-dimensional distribution data of the pore space;

[0058] The pore geometry-topology feature extraction is performed on the 3D distribution data of the pore space, and the cluster analysis of the pore geometry-topology feature extraction results is performed to obtain the pore type classification results;

[0059] The lithology-dependent scaling box dimension is calculated for the three-dimensional distribution data of pore space to obtain the pore fractal dimension characteristic data;

[0060] Based on the pore type classification results, the volume of each type of pore is quantitatively integrated to obtain the microporosity distribution data;

[0061] The maximum inscribed sphere is calculated for the pore type classification results to obtain the pore throat radius distribution data;

[0062] Based on the pore type classification results, the pore geometry is analyzed by tensor field to obtain the reservoir pore geometry parameters.

[0063] In this embodiment of the present invention, complete core data was scanned using X-ray computed tomography (CT) equipment. The X-ray tube voltage was set to 120 kV, the current was 100 μA, the detector pixel size was 100 μm × 100 μm, the scanning slice thickness was 0.5 mm, and the scanning range covered the entire length of the core. The scanned image data was imported into image processing software and processed using an adaptive threshold segmentation algorithm, where the threshold was calculated using the Otsu method. The segmented three-dimensional distribution data of the pore space was stored in voxel form. A pore geometry and topology feature extraction tool based on the Marching Cubes algorithm was used to calculate geometric features such as pore volume, surface area, and connectivity. The extracted results were then subjected to K-means cluster analysis with a K value set to 3 to obtain pore type classification results. A lithology-dependent scaling box dimension calculation tool based on fractal theory was used to calculate fractal dimension feature data of the pore space within a range of 2 to 3. Based on the pore type classification results, an integral calculation method was used to quantify the volume of each pore type to obtain microporosity distribution data. The maximum inscribed sphere algorithm was used to calculate the pore type classification results and obtain pore throat radius distribution data. The tensor field analysis tool was used to analyze pore geometry and obtain reservoir pore geometry parameters. Specifically, the image data generated by the CT scan was segmented using an adaptive threshold. A three-dimensional pore geometry model was constructed using the MarchingCubes algorithm, and characteristic parameters such as pore volume, surface area, and connectivity were calculated. K-means cluster analysis classified the pores into three categories, and characteristic parameters for each pore category were extracted and recorded. Fractal dimension calculation used the box dimension method to determine the complexity of the pore space within a range of 2 to 3. Microporosity distribution data was obtained based on integral calculations, which quantified the volume of different pore types. The calculation of pore throat radius distribution data relied on the maximum inscribed sphere algorithm to determine the radius of the largest inscribed sphere in the pores, reflecting the size distribution of pore throats. The tensor field analysis tool analyzed the anisotropic characteristics of the pore geometry to obtain reservoir pore geometry parameters, which together constitute the reservoir pore structure parameters.

[0064] The present invention achieves high-resolution visualization and digital representation of pore structure through core CT scanning and three-dimensional reconstruction, significantly improving the perception of reservoir microstructure. Adaptive threshold segmentation and geometric-topological feature extraction enable accurate separation and quantification of pore space information, enhancing the accuracy of the description of complex pore systems. Automatic identification of pore types is achieved through cluster analysis, providing a clear grouping basis for subsequent quantitative analysis; fractal dimension calculation reveals the multi-scale nonlinear characteristics of pore structure, enriching the characterization dimension of pore space structure. The precise acquisition of key parameters such as microporosity distribution and pore throat radius helps to more comprehensively reflect the reservoir seepage capacity and storage space characteristics. Tensor field analysis further captures the anisotropy and spatial orientation characteristics of pore geometry, providing important support for the construction of high-reliability reservoir physical models, and overall improving the scientificity and accuracy of reservoir fine evaluation and modeling.

[0065] Preferably, in step S2, inverting the reservoir macro-porosity based on the complete well logging data includes:

[0066] Perform multi-sensor acquisition accuracy calibration and unified standardization processing on the complete logging data to obtain standardized logging curve data;

[0067] Perform multivariate nonlinear covariance analysis on standardized logging curve data to generate logging response characteristic parameters;

[0068] Analyze reservoir fluid properties based on the spectrum of logging response characteristic parameters, and perform reservoir fluid phase separation based on the spectrum analysis results to obtain fluid type distribution data;

[0069] Perform fluid-constrained neutron-density porosity joint inversion on the standardized well logging curve data and fluid type distribution data to obtain initial effective porosity data;

[0070] Based on the initial effective porosity data, the clay mineral content in the reservoir is analyzed by X-ray diffraction and corrected to obtain the clay correction coefficient;

[0071] The initial effective porosity data and the clay correction coefficient are numerically integrated, and the numerical integration results are volume averaged to obtain the reservoir macro-porosity data.

[0072] In an embodiment of the present invention, the complete logging data is first subjected to multi-sensor acquisition accuracy calibration and unified standardization. The least squares method is used to fit the calibration curve, and the data collected by different sensors is uniformly converted into standard logging curve data. The standardized logging curve data is subjected to multivariate nonlinear covariance analysis using principal component analysis to extract logging response characteristic parameters. Based on the logging response characteristic parameters, a spectral analysis method based on Fast Fourier Transform (FFT) is used to analyze reservoir fluid properties. Reservoir fluid phase separation is performed based on the spectral analysis results to obtain fluid type distribution data. The standardized logging curve data is combined with the fluid type distribution data, and a neutron-density porosity joint inversion method based on Bayesian inversion is used to obtain initial effective porosity data. Based on the initial effective porosity data, the clay mineral content in the reservoir was measured using an X-ray diffractometer. The diffraction angle range was 2θ = 3° to 60°, with a scanning step size of 0.02°. The diffraction intensity was recorded in count rate mode. The clay mineral content was calculated based on the diffraction peak intensity and position. The initial effective porosity data was corrected using a correction factor to obtain the reservoir macroporosity data. Specifically, the well logging data was first calibrated and standardized. The calibration curve was fitted using the least squares method, and the data collected by different sensors were uniformly converted into standard well logging data. Then, principal component analysis was used to extract the characteristic parameters of the well logging response. FFT spectral analysis was used to analyze the reservoir fluid properties and perform fluid phase separation to obtain fluid type distribution data. The standardized well logging data were then combined with the fluid type distribution data, and a Bayesian inversion method was used to perform a joint neutron-density porosity inversion to obtain the initial effective porosity data. Finally, the clay mineral content was determined by X-ray diffractometer, and the initial effective porosity data was corrected with the correction coefficient to obtain the reservoir macro-porosity data.

[0073] The present invention improves the consistency and comparability of multi-source logging information by standardizing and precision-calibrating logging data, providing high-quality input for subsequent inversion. Multivariate nonlinear covariance analysis enhances the ability to model the nonlinear relationship between logging responses and geological properties, and helps to accurately extract key parameters that characterize pore structure and fluid characteristics. Spectral analysis and fluid phase separation effectively identify the distribution of different types of fluids in the reservoir, improving the adaptability and constraint of the inversion process to fluid influences. Neutron-density porosity joint inversion combined with fluid type information significantly improves the accuracy and physical consistency of porosity inversion. The introduction of X-ray diffraction for clay effect correction solves the interference of clay minerals on porosity measurement and enhances the geological authenticity of the inversion results. The macro-porosity data finally obtained by numerical integration and volume averaging have higher representativeness and engineering applicability, providing scientific and reliable basic parameters for quantitative characterization and development of reservoirs.

[0074] Preferably, step S3 includes the following steps:

[0075] Step S31: performing a three-dimensional network connectivity assessment on the reservoir pore geometry parameters and pore throat radius distribution data to obtain a pore network topology diagram; tracing the seepage path based on the pore network topology diagram to obtain effective fluid channel data; performing numerical integration and statistical normalization on the effective fluid channel data to obtain a pore connectivity index;

[0076] Step S32: Qualitatively analyzing the microscopic pore structure based on the pore fractal dimension characteristic data scale to obtain an initial pore complexity score; weightedly fusing the initial pore complexity score with the microporosity distribution data to obtain an initial pore structure complexity index; experimentally verifying and calibrating the initial pore structure complexity index using a porous media mercury intrusion porosimeter to obtain a revised pore structure complexity index;

[0077] Step S33: constructing a reservoir fluid storage capacity assessment model based on the reservoir macro-porosity data and pore connectivity index to generate a reservoir capacity factor; performing a correlation analysis between the permeability data in the complete core data and the pore connectivity index to obtain a seepage capacity factor; performing a multi-objective optimization integration on the reservoir capacity factor and the seepage capacity factor to obtain a reservoir quality factor;

[0078] Step S34: Analyze rock-fluid interaction based on fluid type distribution data and reservoir macro-porosity data to obtain wettability assessment results; perform multivariate regression on the wettability assessment results and the modified pore structure complexity index to obtain the lithology control coefficient; identify spatial variability based on reservoir pore geometry parameters and pore type classification results to generate a spatial variation coefficient; perform sequence constraint processing on the spatial variation coefficient to obtain a heterogeneity index;

[0079] Step S35: Perform analytic hierarchy process weight assignment on the reservoir quality factor, lithologic control coefficient, and heterogeneity index to obtain an index weight matrix; weightedly fuse the evaluation indicators based on the index weight matrix to obtain a comprehensive reservoir evaluation score; quantitatively grade the comprehensive reservoir evaluation score to obtain a reservoir evaluation index data set.

[0080] In this embodiment of the present invention, the pore connectivity index is first calculated. Pore geometry parameters and pore throat radius distribution data are input into an analysis tool based on a graph theory algorithm. This tool uses a breadth-first search algorithm and pore connectivity criteria as search criteria to construct a pore network topology diagram. The connectivity threshold between pore nodes is set to a pore throat radius ratio greater than 0.7. Based on the constructed pore network topology diagram, the Dijkstra shortest path algorithm is used to trace the seepage path and extract effective fluid channel data such as path length, number, and connectivity probability. This data is then input into a numerical integration tool, which numerically integrates the effective fluid channel data using the Simpson integral method to obtain a raw quantitative value of pore connectivity. The integration result is then statistically normalized using the maximum-minimum normalization method to constrain the pore connectivity index to a range between 0 and 1, ultimately obtaining the pore connectivity index. The pore fractal dimension feature data is then input into a qualitative scaling analysis tool based on fractal geometry theory. After qualitative scaling analysis, an initial pore complexity score is obtained, ranging from 1 to 10. This score and the microporosity distribution data obtained in step S2 are input into the weighted fusion tool and fused using the weighted average method with weight coefficients of 0.6 and 0.4, respectively, to obtain the initial pore structure complexity index. This index is input into the experimental verification and calibration module of the porous media mercury intrusion instrument, and the pressure calibration method is used. The calibration pressure range is 0MPa to 70MPa. By comparing the pore structure data obtained from the mercury intrusion experiment with the initial pore structure complexity index, a calibration model is constructed to correct the initial index to obtain the corrected pore structure complexity index. The reservoir macroporosity data and pore connectivity index are input into the evaluation tool based on the reservoir physical model, and the linear regression method is used to construct a reservoir fluid storage capacity evaluation model to generate the reservoir capacity factor. At the same time, the permeability data and pore connectivity index in the complete core data are input into the correlation analysis tool, and the Pearson correlation coefficient method is used for correlation analysis to obtain the seepage capacity factor. The reservoir capacity factor and seepage capacity factor were input into the multi-objective optimization tool, with the optimization objectives being to maximize the reservoir capacity factor and the seepage capacity factor, respectively. The weighted sum method was used for integration, with weight coefficients of 0.7 and 0.3, respectively, to obtain the reservoir quality factor. Fluid type distribution data and reservoir macroporosity data were input into the rock-fluid interaction analysis tool, and an analysis method based on fluid mechanics and thermodynamic equilibrium was used to obtain the wettability assessment results, expressed as a wettability index ranging from 0 to 1. This result and the modified pore structure complexity index were input into the multivariate regression tool, and the lithologic control coefficient was obtained using the multivariate linear regression method. Reservoir pore geometry parameters and pore type classification results were input into the spatial variability identification tool, and the semivariogram analysis method was used to identify spatial variability and generate the spatial variation coefficient.The spatial variation coefficient was subjected to sequence constraints, and a sequence stratigraphic constraint algorithm was used to obtain a heterogeneity index. The index value ranges from 0 to 1, with larger values ​​indicating greater heterogeneity. The reservoir quality factor, lithologic control coefficient, and heterogeneity index were input into the analytic hierarchy process (AHP) weight assignment tool. A judgment matrix was constructed using the analytic hierarchy process (AHP) calculation process, and the index weight matrix was calculated using the eigenvalue method. Based on the index weight matrix, the weighted summation method was used to weight the evaluation indicators and combine them to obtain a comprehensive reservoir evaluation score, ranging from 0 to 100. Finally, the comprehensive reservoir evaluation score was quantitatively graded using an equidistant segmentation method, with the score range divided into five levels: 0-20 for extremely poor reservoirs, 21-40 for poor reservoirs, 41-60 for moderate reservoirs, 61-80 for good reservoirs, and 81-100 for excellent reservoirs. This resulted in a reservoir evaluation index dataset.

[0081] The present invention accurately depicts the reservoir seepage path and connectivity characteristics through three-dimensional pore network modeling and fluid channel tracking, thereby enhancing the depth of understanding of the reservoir's microscopic seepage capacity. The integration of fractal dimension and microporosity enables quantitative evaluation of pore structure complexity, and calibration through mercury injection experiments ensures the physical consistency and engineering feasibility of the evaluation results. The comprehensive modeling of reservoir capacity and seepage capacity provides a quality factor that comprehensively characterizes reservoir quality, reflecting the synergistic characteristics of the reservoir in terms of storage and conductivity. Rock-flow interaction analysis reveals the regulatory effect of wettability on pore structure, and the establishment of lithologic control coefficients enhances the model's ability to interpret geological attribute variations. The heterogeneity index captures the internal structural differences of the geological body through spatial variation coefficient and sequence constraint modeling, providing support for fine reservoir division. Finally, the scientific weight distribution of each indicator is achieved through the hierarchical analysis method. The reservoir evaluation index system generated by comprehensive weighting has high accuracy, systematicity and scalability, providing a strong decision-making basis for subsequent reservoir classification, development optimization and risk prediction.

[0082] Preferably, step S4 includes the following steps:

[0083] Step S41: performing feature correlation analysis on the complete data set and performing multicollinearity detection to generate a standardized feature data set;

[0084] Step S42: Identify reservoir segments in the standardized feature dataset based on the preset geological stratification information, and divide the reservoir segments based on the identification results to obtain a labeled segment dataset; divide the labeled segment dataset into 75% as a training set, 15% as a validation set, and 15% as a test set;

[0085] Step S43: constructing a one-dimensional convolutional neural network model using the training set, and performing nonlinear transformation configuration to obtain a basic network structure;

[0086] Step S44: performing Dropout and L2 weight regularization on the basic network structure to obtain an anti-overfitting enhanced network structure;

[0087] Step S45: introducing a self-attention calculation unit into the anti-overfitting enhancement network structure and performing weight initialization to obtain a geological sensitive feature enhancement component;

[0088] Step S46: weighting the geological sensitive feature enhancement components to generate geological key features; performing multi-scale integration of the original convolution feature application features based on the geological key features to obtain geological enhancement features;

[0089] Step S47: constructing skip connections based on geological enhancement features to obtain a gradient transfer optimization network structure;

[0090] Step S48: performing an attention-guided loss combination design on the gradient transfer optimization network structure to obtain a geological perception comprehensive loss function;

[0091] Step S49: Integrate the basic network structure, geological sensitive feature enhancement components, gradient transfer optimization network structure and geological perception comprehensive loss function into a geological attention enhancement neural network through a distributed deep learning framework.

[0092] This embodiment of the present invention first performs feature correlation analysis on the complete dataset, using the Pearson correlation coefficient method with a threshold of 0.7 to screen for highly correlated features. Multicollinearity detection is then performed using the variance inflation factor (VIF) method with a threshold of 10 to remove features exhibiting multicollinearity, generating a standardized feature dataset. Based on pre-set geological stratification information, a geological stratification recognition tool employs a convolutional neural network (CNN) algorithm, with a training set ratio of 75%, a validation set ratio of 15%, and a test set ratio of 15%. The standardized feature dataset is then partitioned into reservoir segments, generating a labeled segment dataset. The labeled segment dataset is then partitioned proportionally: 75% for training, 15% for validation, and 15% for testing. Next, the requirements for the reservoir evaluation model architecture are analyzed based on the complete dataset, and a basic architecture diagram is drawn to clarify the model's input, output, and intermediate layer structures. The input layer structure is designed based on the optimal feature subset, with parameters set to 128 neurons and a data type of 64-bit floating-point numbers, resulting in the model input layer. A convolutional layer group was designed using multi-scale convolutional kernels (sizes of 1×1, 3×3, and 5×5, respectively) and 64-bit floating-point precision computational units. Each layer had 64 convolution kernels and the activation function was Reinforced Luminance (ReLU). A combination of max pooling and average pooling was applied to the convolutional layer group, with a pooling window size of 2×2 and a stride of 2, resulting in a dimensionally reduced feature map pooling layer. A fully connected layer architecture was designed based on the dimensionally reduced feature map pooling layer, with 256 neurons, dropout regularization, and a dropout probability of 0.5, resulting in a fully connected feature representation layer. The output layer structure was designed based on the preset reservoir parameter prediction task, with 8 neurons and a softmax activation function, resulting in the multi-task learning output layer. Nonlinear transformations were applied to the model input layer, convolutional layer group, dimensionally reduced feature map pooling layer, fully connected feature representation layer, and multi-task learning output layer, using the Reinforced Luminance (ReLU) activation function to obtain the basic network architecture. Dropout and L2 weight regularization were applied to the base network structure, with a dropout probability of 0.5 and an L2 regularization coefficient of 0.01, resulting in an enhanced network structure that prevents overfitting. A self-attention computation unit was introduced, using the scaled dot-product attention mechanism. The number of attention heads was set to 8, the hidden layer dimension was set to 512, and weight initialization was performed using He regularization to obtain the geologically sensitive feature enhancement component. Weights were assigned to the geologically sensitive feature enhancement component, and a gradient descent-based optimization algorithm with a learning rate of 0.001 was used to generate geological key features. Based on the geological key features, multi-scale integration of the original convolutional features was performed using a feature pyramid network (FPN) structure with convolution kernels of 1×1, 3×3, and 5×5 scales to obtain the geologically enhanced features.Based on geological enhancement features, skip connections were constructed using the skip connection method in the U-Net architecture with four connection layers, resulting in a gradient transfer optimization network structure. A combination of attention-guided losses was designed, using a cross-entropy loss function and a Dice loss function with weights of 0.7 and 0.3, respectively, to obtain a geological perception comprehensive loss function. Using the distributed deep learning framework Horovod, the basic network structure, geologically sensitive feature enhancement components, the gradient transfer optimization network structure, and the geological perception comprehensive loss function were integrated into a geological attention enhancement neural network. A data parallel strategy was implemented, with each GPU responsible for computing a portion of the data. The results from each component were then aggregated to complete the construction of the geological attention enhancement neural network.

[0093] This paper effectively improves the independence and expressiveness of model input features through feature correlation analysis and multicollinearity detection, providing a high-quality data foundation for subsequent modeling. Combined with geological stratification, it enables precise identification and labeling of reservoir segments, enhancing the model's perception of geological distribution patterns. The construction and nonlinear configuration of a one-dimensional convolutional neural network enables the model to extract deep features from well logging sequence data. The introduction of dropout and L2 regularization effectively mitigates overfitting and improves model stability and generalization. The integration of a self-attention mechanism with a geologically sensitive feature enhancement component allows the model to focus more closely on key geological features, enhancing its ability to capture important information. Multi-scale feature integration combined with skip connections optimizes information flow and gradient propagation paths, improving the training efficiency and robustness of deep networks. The design of an attention-guided loss function enhances the model's sensitivity to geological changes and helps capture the complex connections between reservoir nonlinear characteristics and geological background. Ultimately, the geological attention-enhanced neural network, constructed through a distributed deep learning framework, exhibits enhanced geological adaptability, feature expressiveness, and prediction accuracy, providing a high-performance intelligent solution for complex reservoir evaluation.

[0094] It is particularly important that step S43 includes the following steps:

[0095] Step S431: Analyze the reservoir evaluation model architecture requirements based on the complete data set to obtain a basic architecture diagram;

[0096] Step S432: Designing an input layer structure based on the optimal feature subset and configuring parameters to obtain a model input layer;

[0097] Step S433: Designing a convolution layer based on the model input layer using a multi-scale convolution kernel and a 64-bit floating-point precision computing unit to obtain a convolution layer group;

[0098] Step S434: performing a maximum pooling and average pooling combination design on the convolutional layer group to obtain a dimension reduction feature map pooling layer;

[0099] Step S435: performing a fully connected layer architecture design and dropout regularization configuration based on the dimension reduction feature map pooling layer to obtain a fully connected feature expression layer;

[0100] Step S436: designing an output layer structure based on the preset reservoir parameter prediction task to obtain a multi-task learning output layer;

[0101] Step S437: Perform nonlinear transformation configuration on the model input layer, convolution layer group, dimensionality reduction feature map pooling layer, fully connected feature expression layer and multi-task learning output layer to obtain the basic network structure.

[0102] This embodiment of the present invention first analyzes the reservoir evaluation model architecture requirements based on a complete dataset, draws a basic architecture diagram, and clarifies the model input, output, and intermediate layer structures. Next, the input layer structure is designed based on the optimal feature subset, with parameters set to 128 neurons and a data type of 64-bit floating-point numbers, resulting in the model input layer. Next, convolutional layers are designed using multi-scale convolutional kernels (sizes of 1×1, 3×3, and 5×5, respectively) and 64-bit floating-point precision computational units. Each layer has 64 convolution kernels and uses the Reluctant Unit (ReLU) activation function, resulting in a convolutional layer group. A combination of maximum pooling and average pooling is performed on the convolutional layer group, with a pooling window size of 2×2 and a stride of 2, resulting in a dimensionality reduction feature map pooling layer. Based on the dimensionality reduction feature map pooling layer, a fully connected layer architecture is designed, with 256 neurons, dropout regularization, and a dropout probability of 0.5, resulting in a fully connected feature representation layer. Next, the output layer structure was designed based on the preset reservoir parameter prediction task, with the number of neurons set to 8 and the activation function set to softmax, resulting in the multi-task learning output layer. Finally, nonlinear transformations were performed on the model input layer, convolutional layer group, dimensionality reduction feature map pooling layer, fully connected feature expression layer, and multi-task learning output layer, using the Reluctant Unit (ReLU) activation function to obtain the basic network structure. Specifically, the reservoir evaluation model architecture requirements were analyzed, a basic architecture diagram was drawn, and the model input, output, and intermediate layer structures were determined. The input layer structure was designed based on the optimal feature subset, with the number of neurons and data type set to obtain the model input layer. A convolutional layer group was designed using multi-scale convolutional kernels and 64-bit floating-point precision computation units, with the number of kernels and activation function set for each layer to obtain the convolutional layer group. Pooling was performed on the convolutional layer group, with the pooling window size and stride set to obtain the dimensionality reduction feature map pooling layer. Based on the dimensionality reduction feature map pooling layer, the fully connected layer architecture was designed, with the number of neurons set and dropout regularization configured to obtain the fully connected feature expression layer. Based on the reservoir parameter prediction task, the output layer structure is designed, the number of neurons and the activation function are set, and the multi-task learning output layer is obtained. Nonlinear transformations are performed on each layer, and the ReLU activation function is used to finally obtain the basic network structure.

[0103] The present invention screens and standardizes high-quality input features through feature correlation analysis and collinearity detection, effectively reducing redundant information and improving model stability and learning efficiency. Layer identification guided by geological stratification improves the model's adaptability to geological structures and achieves accurate labeling and scientific data division. The network architecture built based on task requirements integrates multi-scale convolution and high-precision computing units in design, enhancing the model's ability to recognize reservoir features at different scales. The combined design of the pooling layer retains key information while reducing the dimension, enhancing the model's ability to express local and global features. The fully connected layer and dropout configuration further improve the feature fusion effect and effectively prevent overfitting. The output layer is based on the concept of multi-task learning, taking into account the needs of multi-dimensional reservoir attribute prediction and enhancing the model's generalization ability. Nonlinear transformation and regularization mechanisms are implemented throughout the various layers of the model to ensure efficient information flow transmission and deep feature extraction. The overall architecture design is scientific and reasonable, with good scalability, learning ability and prediction accuracy, providing strong support for complex reservoir modeling and multi-parameter joint evaluation.

[0104] Preferably, in step S5, guiding the geological attention enhancement neural network with weak supervision signals includes:

[0105] Obtain a geological constraint rule set; perform rule matching on the training set based on the geological constraint rule set to obtain a weak supervision signal label;

[0106] Apply weak supervision signal labeling to the geological attention-enhanced neural network and optimize the attention weights to generate a weak supervision guided architecture;

[0107] Based on the weak supervision guidance architecture, a geological rule consistency loss term is designed to generate a multi-objective loss function.

[0108] Perform constrained optimization on the multi-objective loss function to obtain a geological knowledge constrained optimizer;

[0109] The geological attention enhanced neural network and geological knowledge constraint optimizer are integrated and reconstructed to obtain the geological knowledge constrained attention network.

[0110] In the embodiments of the present invention, first, a predefined set of geological constraint rules is extracted from the reservoir geology expert knowledge base. This rule set contains 120 clear geological rules, covering 5 categories: porosity-permeability relationship rules (e.g., when the porosity < 5%, the permeability must < 0.1 mD; when the porosity > 15%, the permeability must > 10 mD), pore-throat structure constraint rules (e.g., when the average throat radius < 2 microns, the connectivity index must < 0.4; when the throat-pore ratio < 0.1, the reservoir heterogeneity index must > 0.6), mineral composition influence rules (e.g., when the clay content > 12%, the lithology control coefficient must > 0.7; when the quartz content > 85%, the lithology control coefficient must < 0.3), sedimentation-related rules (e.g., in the delta front facies area, the change rate of the vertical heterogeneity index must < 0.05 / m; in the shoreface sandbar facies area, the quality factor must > 0.65), and fluid property constraint rules (e.g., when the water saturation > 65%, the reservoir quality factor must < 0.5; when the crude oil viscosity > 50 mPa·s, the pore connectivity index must < 0.6); then, rule matching is performed on the training set based on these rules. The specific operation is as follows: Each rule is converted into a deterministic condition-result statement. The condition part is expressed as a parameter range constraint (e.g., 5% ≤ porosity ≤ 10%), and the result part is expressed as a value range constraint (e.g., 0.1 mD ≤ permeability ≤ 1 mD). The parameters in the condition part are accurate to two decimal places, and the parameter range in the result part is accurate to the same precision level as the measured physical quantity; for each sample point in the training set, the probability value P_cond that meets the rule condition is calculated. P_cond follows a Gaussian function distribution on the interval [0, 1] (centered at the midpoint of the condition interval, with a standard deviation of 1 / 6 of the interval width). When the sample is completely inside the condition interval, P_cond = 1; when it is completely outside the condition interval, P_cond = 0; when it is near the boundary, 0 < P_cond < 1; at the same time, the probability value P_result that meets the rule result is calculated, and the calculation method is the same as that of P_cond; the rule matching degree M = P_cond × P_result is defined. For each sample point, the matching degrees M_i of all applicable rules are calculated, and the highest 5 M_i values are retained; these 5 M_i values are weighted and averaged according to the weights 1.0, 0.8, 0.6, 0.4, 0.2 to obtain the comprehensive rule confidence C of this sample point; according to the C value, the training set samples are divided into strong rule samples (C ≥ 0.85), medium rule samples (0.5 ≤ C < 0.85), weak rule samples (0.2 ≤ C < 0.5), and no rule samples (C < 0.2) Four categories; for each category of samples, the corresponding weak supervision signal labels are generated: strong rule samples are labeled as [1, 0, 0, 0], medium rule samples are labeled as [0, 1, 0, 0], weak rule samples are labeled as [0, 0, 1, 0], and no rule samples are labeled as [0, 0, 0, 1]; these weak supervision signal labels are added as auxiliary labels to the training set to obtain an enhanced training set with rule confidence labels; then the weak supervision signal labels are applied to the geological attention enhancement neural network, by inserting a rule perception layer in the middle layer of the network (after the convolution layer and before the fully connected layer), the rule perception layer sets an attention Channel, each channel contains 64 attention units, each attention unit is implemented as a W·tanh(Ux+b) structure, where W is a 1×64-dimensional weight vector, U is a 64×n-dimensional transformation matrix (n is the feature dimension), b is a 64-dimensional bias vector, and tanh is a hyperbolic tangent activation function; the attention mechanism is applied to the network intermediate feature map F, the attention weight α=softmax (rule perception layer (F)) is calculated, and then the rule-enhanced feature representation is obtained through α·F; the attention weight is L1 regularized, and the regularization coefficient is set to 0.01 to prevent excessive concentration of attention; an attention update step is introduced during training, and each The attention weights are updated once every 100 batches, with an update step of 0.003. A weakly supervised guided architecture is generated based on the updated attention structure, combining the original classification or regression task with the rule category prediction task to form a multi-task learning framework. Based on the multi-task learning framework, a geological rule consistency loss term is designed, defined as Lrule=λ1·BCE(ypred_rule,ytrue_rule)+λ2·KL(αpred,αtarget), where BCE is the binary cross entropy loss, used to evaluate the accuracy of rule category prediction, and KL is the KL divergence, used to measure the predicted attention distribution. The difference from the target attention distribution is considered, with λ1 and λ2 being weight coefficients, set to 0.2 and 0.15, respectively. The geological rule consistency loss term is combined with the original task loss (mean squared error loss) to generate a multi-objective loss function, Ltotal = Lorigin + Lrule. Constrained optimization of the multi-objective loss function is performed using the Adam optimizer with momentum. The initial learning rate is set to 0.001, the weight decay coefficient is set to 0.0005, and the momentum parameter is set to 0.9. A cosine annealing learning rate scheduling strategy is used, with a minimum learning rate of 1 / 10 of the initial learning rate and a 30-epoch cycle. A gradient clipping mechanism is added, with a clipping threshold of 5.0 to prevent gradient explosion; implement an early stopping mechanism to monitor the rule consistency indicator on the validation set, stop training if there is no improvement for five consecutive rounds, and save the optimal model parameters; finally, modularly integrate the backbone structure of the geological attention-enhanced neural network with the geological knowledge constraint optimizer, reconstruct the network computation graph, and optimize the memory layout to obtain the geological knowledge constraint attention network, which includes a rule embedding layer, an attention enhancement layer, and a rule consistency verification layer.

[0111] The present invention realizes weak supervision labeling of training data by introducing a set of geological constraint rules, effectively improving the learning ability of the model under conditions of incomplete sample labels, and enhancing its responsiveness to geological prior knowledge. Attention weight optimization guides the model to focus on geological key feature areas, improving the accuracy and geological consistency of feature selection. The geological rule consistency loss term is introduced to explicitly integrate geological knowledge into the model training objectives, so that the model not only relies on data-driven, but also has geological logical constraints, thereby improving generalization ability and prediction reliability. The design of the geological knowledge constraint optimizer realizes coordinated optimization among multi-objective training tasks, taking into account both accuracy and geological rationality. Finally, through the integrated reconstruction of network structure and optimization strategy, the constructed geological knowledge constraint attention network has stronger knowledge fusion ability and geological interpretability, providing a stable and controllable model foundation for intelligent identification and fine evaluation in complex reservoir environments.

[0112] Of particular importance is that the data synthesis expansion of the training set by small sample reservoir accumulation in step S5 through the generative adversarial network includes:

[0113] Perform statistical and cluster analysis on the sample distribution of each reservoir type during training, identify small sample segments, and obtain a small sample segment label set;

[0114] Generate feature-preserving data from a small sample paragraph label set using a generative adversarial network to obtain synthetic reservoir small sample data.

[0115] The training set and the synthetic reservoir small sample data are distributed resampled, and the weight proportions of various types of reservoirs are corrected based on the distribution resampling results to obtain a balanced training data set.

[0116] In this embodiment of the present invention, a statistical and cluster analysis of the sample distribution of each reservoir type was first performed on the training set. A K-means clustering algorithm was used, with a K value of 5 and a clustering accuracy of 95%. Small sample paragraphs were identified and a small sample paragraph label set was obtained. Feature-preserving data generation was then performed on the small sample paragraph label set using a generative adversarial network (GAN). Both the generator and discriminator of the GAN were five-layer neural networks, with a Reluctant Linear Unit (ReLU) activation function, an Adam optimization algorithm, a learning rate of 0.002, and a batch size of 32. Synthetic reservoir small sample data was generated. Distribution resampling was performed on the training set and the synthetic reservoir small sample data using a stratified sampling method with a sampling rate of 1.5. Based on the distribution resampling results, the weight ratio of each reservoir type was adjusted with a correction factor of 0.8-1.2 to obtain a balanced training dataset.

[0117] The present invention identifies small sample segments through statistical and cluster analysis of reservoir type samples, thereby accurately locating the problem of data imbalance and effectively improving the model's learning ability on minority class samples. The generative adversarial network generates feature-preserving data to ensure that the synthesized reservoir small sample data retains key features while enhancing data diversity and representativeness, providing richer sample information for the training set. Through distribution resampling and reservoir weight ratio correction, the problem of uneven distribution of various types of samples in the training data is solved, effectively avoiding the model's bias towards majority class samples, and ensuring that the learning ability of different reservoir types is evenly improved. Overall, the synthetically expanded small sample data not only improves the generalization ability of the model in small sample scenarios, but also enhances the diversity and representativeness of the training set, providing more stable and efficient data support for the subsequent training of deep learning models.

[0118] Preferably, in step S5, using geological knowledge to constrain the attention network and the balanced training data set to perform deep learning model training, and performing hyperparameter optimization and performance monitoring based on the validation set includes:

[0119] Based on the geological knowledge constrained attention network, forward propagation and backpropagation training are performed on the balanced training data set to obtain the initial network weight parameter set;

[0120] Statistically analyze the loss function value, validation set prediction error, and gradient change trend during the training process to generate a dynamic performance indicator dataset;

[0121] Perform automated hyperparameter search on the geological knowledge-constrained attention network based on a dynamic performance indicator dataset to obtain the optimal hyperparameter combination set;

[0122] Based on the optimal combination of hyperparameters, batch training is performed on the balanced training dataset to obtain a candidate model parameter set;

[0123] Based on the validation set, early stopping and Dropout validation are performed on the candidate model parameter set to obtain the candidate model.

[0124] In the embodiment of the present invention, a geological knowledge constrained attention network is first deployed on a distributed computing cluster (32 nodes, each node is configured with an 8-core CPU, an NVIDIA A100 GPU, and 64 GB of system memory). The network parameters are initialized by the Xavier method, the standard deviation of the weight matrix initialization is set to 0.01, the initial value of the bias term is set to 0.001, the number of input layer nodes is the complete feature set dimension (including 10 well logging curve parameters, 5 core physical property parameters, and 8 reservoir structure parameters), the first convolution layer uses 32 convolution kernels of size 3×1, and the activation function is LeakyReLU (negative slope is 0.01), the second convolution layer uses 64 convolution kernels of size 5×1, and the activation function is also LeakyReLU, and the pooling layer uses maximum pooling and average pooling in parallel. The structure is as follows: the pooling kernel size is 2×1, the number of nodes in the fully connected layer is 512, 256, and 128, respectively; the attention mechanism layer contains 8 attention heads, the attention head dimension is 32, and the number of nodes in the output layer is the number of reservoir evaluation indicators (including pore connectivity index, reservoir quality factor, lithology control coefficient, and heterogeneity index); then the balanced training data set (containing 15,000 samples, each sample containing an input feature vector and a target label value) is read, and the data is divided into mini-batches of size 32. During the data loading process, random horizontal flipping (with a probability of 0.3) and Gaussian noise perturbation (with a mean of 0 and a standard deviation of 0.02) are implemented online. Enhancement; then set the training parameters, where the momentum parameter of the batch normalization layer is 0.99, the epsilon value is 1e-5, the dropout rate of the Dropout layer is 0.3, and the loss function uses a weighted average combined loss, including mean square error loss (weight 0.5), Huber loss (weight 0.3, delta value 1.0), and geological rule consistency loss (weight 0.2); start the forward propagation training process, first input the feature data into the network through the input layer, extract local features through the convolution layer, reduce the dimension and retain key information through the pooling layer, reweight the feature map through the geological constraint attention layer, and calculate the attention formula. The formula is Attention(Q,K,V)=softmax(QK^T / √d)V, where Q, K, and V are the query, key, and value matrices, respectively, and d is the dimension of the attention head. The output of the attention layer is further extracted through a fully connected layer to extract high-order features, and finally the output layer predicts the reservoir evaluation index value. After the forward propagation is completed, the loss value is calculated, and performance indicators such as the loss value, gradient norm, and prediction error are recorded for each mini-batch. Backpropagation is then performed to calculate the gradient of the loss function with respect to each layer parameter. The Adam optimizer is used to update the parameters, with an initial learning rate of 0.001, β1=0.9, and β2=0.999, ε=1e-8, weight decay coefficient is 3e-5, learning rate scheduling adopts CosineAnnealingLR strategy, cycle is 10 rounds, and the minimum learning rate is 1 / 20 of the initial learning rate; after each round of training, the model performance is evaluated on the validation set (3000 samples), and the four indicators of root mean square error (RMSE), mean absolute error (MAE), R² determination coefficient, and geological rule consistency score (rule compliance rate) are calculated. These indicators are recorded in the dynamic performance indicator dataset; an early stopping strategy is implemented during the training process, and the patience value is set to 15 rounds, that is, if there is no improvement in the performance of the validation set after 15 consecutive rounds, the training is stopped, and the maximum number of training rounds is 2 00 rounds; after completing the basic training, the dynamic performance indicator dataset is imported into the Bayesian hyperparameter optimizer to define the hyperparameter search space, including the learning rate range [1e-4, 1e-2], the dropout rate range [0.1, 0.5], the L2 regularization coefficient range [1e-6, 1e-3], the convolution kernel size combination {(3, 5), (3, 7), (5, 7), (5, 9)}, the number of attention heads range [4, 8, 12, 16], and the loss function weight ratio range {(0.4-0.6, 0.2-0.4, 0.1-0.3)}; the hyperparameter optimization adopts the Gaussian process regression model, and the acquisition function is the expected improvement (Expected Improvement), the maximum number of evaluations is 50, each evaluation includes 5 full rounds of training, and the hyperparameter combination with the best average performance on the validation set is selected; after determining the optimal hyperparameter combination, the network parameters are reset, and the optimized hyperparameters are used to perform final training on the balanced training data set. During the training process, gradient clipping (threshold is 5.0), weight normalization (performed once every 5 rounds), learning rate decay (starting from the 50th round, decaying by 10% every 10 rounds) and other technologies are implemented to ensure training stability; after the training is completed, a candidate model parameter set is generated, including 5 training rounds (final round, best validation performance round, best training performance round, highest verification rule consistency round, average Finally, the candidate model parameter sets are evaluated on the validation set, and the final model parameters are selected using an ensemble voting mechanism. The specific method is to weight each model's prediction results according to the performance metric (RMSE weighted by 0.4, MAE weighted by 0.2, R² weighted by 0.2, and rule consistency weighted by 0.2). The model parameters with the highest weighted scores are retained as candidate model parameters. Dropout validation techniques are also applied (50 forward propagations, each using a different random seed) to calculate the confidence interval of the predicted value. The width of the confidence interval does not exceed 15% of the standard deviation of the target value. The candidate model retained must meet the conditions of both optimal average performance and reasonable confidence interval.

[0125] The present invention ensures the integration of geological knowledge and balanced data distribution of deep learning models during the training process through the combination of geological knowledge-constrained attention networks and balanced training data sets, thereby enhancing the model's ability to accurately capture reservoir characteristics. During the training process, through statistical analysis of loss function values, validation set errors, and gradient change trends, the model performance can be monitored in real time and potential overfitting or underfitting problems can be quickly identified, ensuring stability and convergence during the training process. Automated hyperparameter search makes the adjustment of model hyperparameters more efficient, finds the optimal combination, and thus improves the performance and prediction accuracy of the model. Batch training iterations further improve the model's ability to handle large-scale data, while early stopping and Dropout verification ensure the generalization ability and stability of the model. Overall, combining geological knowledge constraints with automated optimization processes, this method significantly improves the accuracy, efficiency, and reliability of the model, providing strong support for accurate assessment and development decisions of complex reservoirs.

[0126] Preferably, step S6 includes the following steps:

[0127] Step S61: using the test set to perform the prediction task on the candidate model to obtain a reservoir parameter prediction value set;

[0128] Step S62: performing error comparison analysis of real core labels based on the reservoir parameter prediction value set to obtain a model error evaluation index set;

[0129] Step S63: fine-tuning parameters and structures of the candidate model based on the model error evaluation index set to obtain an initial reservoir evaluation model;

[0130] Step S64: performing integrated nesting on the initial reservoir evaluation model and the preset SZ3 framework to obtain the final reservoir evaluation model.

[0131] In the embodiment of the present invention, first, a three-layer cross-validation method is used to perform the reservoir parameter prediction task on the candidate model, and the test set is input into the candidate model. The predicted value set of reservoir porosity, permeability, water saturation and reservoir quality score is obtained by forward propagation calculation. CUDA is used to accelerate the calculation during the prediction process, the batch size is set to 32, and the output result is saved as a 64-bit floating-point value; the root mean square error (RMSE), mean absolute error (MAE) and R² determination coefficient are used to quantitatively compare the predicted value with the actual core label, and the error index is calculated for each reservoir segment to generate a model error evaluation index set containing the value of each index for each segment; based on the error evaluation results, parameter fine-tuning is performed in a targeted manner, and the Adam optimizer with a learning rate of 0.0001 is used to perform 1000 iterations of optimization for the relevant weights of the segment with RMSE>0.05, and the number of attention heads (range 4-16) and the convolution kernel size in the model structure are optimized by the Bayesian optimization algorithm. The initial reservoir evaluation model was fine-tuned with the number of neurons in the validation set (ranging from 3 to 11) and the number of neurons in the fully connected layer (ranging from 128 to 512). An early stopping strategy was adopted during the fine-tuning process. If the validation set loss function showed no improvement after 10 consecutive rounds, the model was stopped and the optimized initial reservoir evaluation model was obtained. The initial reservoir evaluation model was integrated and nested with the preset SZ3 framework (including the lithofacies classification module, the physical property prediction module and the comprehensive evaluation module). Specifically, the reservoir feature vector extracted by the initial model was used as the input of the lithofacies classification module of the SZ3 framework. The attention weight of the initial model was fused with the physical property prediction module of the SZ3 framework at the tensor level with a fusion coefficient of 0.6:0.4. The prediction results of the initial model and the evaluation results of the SZ3 framework were merged into the final output through weighted averaging (weight of 0.7:0.3) to generate the final reservoir evaluation model. The output format of the model is a standardized evaluation report containing porosity (accuracy 0.01), permeability (accuracy 0.1mD) and reservoir quality (1-5 points).

[0132] The present invention predicts candidate models using a test set and compares the results with real core labels for error analysis, effectively quantifying the model's prediction accuracy and error distribution, thereby providing a reliable basis for further optimization. Based on the error evaluation index set, the model can perform parameter fine-tuning and structural fine-tuning to ensure higher accuracy and stronger robustness of the prediction results. The initial reservoir evaluation model obtained through preliminary tuning is then integrated and nested with the preset SZ3 framework, which not only integrates the model's efficient prediction capabilities, but also further strengthens the logical consistency of geological constraints and reservoir evaluation, ultimately generating a more adaptable, accurate, and interpretable reservoir evaluation model. This method effectively improves the intelligence level and prediction accuracy of reservoir evaluation, and provides efficient and reliable technical support for development decisions in complex reservoirs.

[0133] Preferably, the present invention further provides a system for constructing a reservoir evaluation model based on deep learning, which is used to execute the above-mentioned method for constructing a reservoir evaluation model based on deep learning. The system for constructing a reservoir evaluation model based on deep learning includes:

[0134] The data preprocessing module is used to obtain well logging data and core analysis data of the target work area, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data;

[0135] The pore parameter inversion module is used to calculate reservoir pore structure parameters based on complete core data; and to invert reservoir macroporosity based on complete well logging data;

[0136] The reservoir attribute analysis module is used to analyze reservoir lithofacies physical properties based on reservoir pore structure parameters and reservoir macroporosity, determine pore connectivity index, reservoir quality factor, lithologic control coefficient and heterogeneity index, and generate a reservoir evaluation index data set;

[0137] The geological attention model construction module is used to divide the reservoir evaluation index dataset into training set, validation set and test set according to the layer segment; the training set is used to build the basic structure of the one-dimensional convolutional neural network model, and the geological attention mechanism is introduced to obtain the geological attention enhanced neural network;

[0138] The model training and enhancement module is used to guide the geological attention enhancement neural network with weak supervision signals to obtain a geological knowledge constrained attention network. The training set is expanded by data synthesis of small sample reservoirs through generative adversarial networks to obtain a balanced training data set. The geological knowledge constrained attention network and the balanced training data set are used to perform deep learning model training, and hyperparameter optimization and performance monitoring are performed based on the validation set to obtain a candidate model.

[0139] The model evaluation and integration module is used to evaluate and optimize the performance of candidate models using the test set to obtain an initial reservoir evaluation model; the preset SZ3 framework is integrated with the initial reservoir evaluation model to generate the final reservoir evaluation model.

[0140] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is not limited by the above description. Therefore, it is intended that all changes that fall within the meaning and scope of the equivalent elements of the application documents are included in the present invention.

[0141] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a reservoir evaluation model based on deep learning, characterized in that: The following steps are involved: Step S1: Acquire well logging data and core analysis data of the target work area, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data; Step S2: Calculating reservoir pore structure parameters based on complete core data; Invert reservoir macro-porosity based on complete well logging data; Step S3: Analyze reservoir lithofacies physical properties based on reservoir pore structure parameters and reservoir macroporosity, determine pore connectivity index, reservoir quality factor, lithologic control coefficient, and heterogeneity index, and generate a reservoir evaluation index dataset. Step S3 includes the following steps: Step S31: performing a three-dimensional network connectivity assessment on the reservoir pore geometry parameters and pore throat radius distribution data to obtain a pore network topology diagram; tracing the seepage path based on the pore network topology diagram to obtain effective fluid channel data; performing numerical integration and statistical normalization on the effective fluid channel data to obtain a pore connectivity index; Step S32: Qualitatively analyzing the microscopic pore structure based on the pore fractal dimension characteristic data scale to obtain an initial pore complexity score; weightedly fusing the initial pore complexity score with the microporosity distribution data to obtain an initial pore structure complexity index; experimentally verifying and calibrating the initial pore structure complexity index using a porous media mercury intrusion porosimeter to obtain a revised pore structure complexity index; Step S33: constructing a reservoir fluid storage capacity assessment model based on the reservoir macro-porosity data and pore connectivity index to generate a reservoir capacity factor; performing a correlation analysis between the permeability data in the complete core data and the pore connectivity index to obtain a seepage capacity factor; performing a multi-objective optimization integration on the reservoir capacity factor and the seepage capacity factor to obtain a reservoir quality factor; Step S34: Analyze rock-fluid interaction based on fluid type distribution data and reservoir macro-porosity data to obtain wettability assessment results; perform multivariate regression on the wettability assessment results and the modified pore structure complexity index to obtain the lithology control coefficient; identify spatial variability based on reservoir pore geometry parameters and pore type classification results to generate a spatial variation coefficient; perform sequence constraint processing on the spatial variation coefficient to obtain a heterogeneity index; Step S35: Performing analytic hierarchy process weight assignment on the reservoir quality factor, lithologic control coefficient, and heterogeneity index to obtain an index weight matrix; weightedly integrating the evaluation indexes based on the index weight matrix to obtain a reservoir comprehensive evaluation score; quantitatively grading the reservoir comprehensive evaluation score to obtain a reservoir evaluation index dataset; Step S4: Divide the reservoir evaluation index dataset into a training set, a validation set, and a test set according to the layer interval; use the training set to construct the basic structure of the one-dimensional convolutional neural network model, and introduce the geological attention mechanism to obtain a geological attention enhanced neural network; Step S5: Weakly supervise the geological attention enhanced neural network to obtain a geological knowledge constrained attention network; synthesize and expand the training set by accumulating small sample reservoirs through a generative adversarial network to obtain a balanced training data set; use the geological knowledge constrained attention network and the balanced training data set to perform deep learning model training, and perform hyperparameter optimization and performance monitoring based on the validation set to obtain a candidate model; Step S6: Use the test set to evaluate and optimize the performance of the candidate model to obtain an initial reservoir evaluation model; integrate the preset SZ3 framework with the initial reservoir evaluation model to generate a final reservoir evaluation model.

2. The method for constructing a reservoir evaluation model based on deep learning according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: collecting natural gamma rays through a sodium iodide crystal detector, and collecting acoustic wave transit time, density, neutron porosity and resistivity through a comprehensive logging instrument to generate logging data; Step S12: collecting porosity, permeability and water saturation using a core porosity and permeability measuring instrument to generate core analysis data; Step S13: perform depth matching on the well logging data and the core analysis data, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data.

3. The method for constructing a reservoir evaluation model based on deep learning according to claim 1, characterized in that: Calculation of reservoir pore structure parameters based on complete core data in step S2 includes: Perform CT scanning on the complete core data to generate a three-dimensional digital model of the core; Based on the adaptive threshold value of the three-dimensional digital model of the core, the reservoir pore space is segmented to obtain the three-dimensional distribution data of the pore space; The pore geometry-topology feature extraction is performed on the 3D distribution data of the pore space, and the cluster analysis of the pore geometry-topology feature extraction results is performed to obtain the pore type classification results; The lithology-dependent scaling box dimension is calculated for the three-dimensional distribution data of pore space to obtain the pore fractal dimension characteristic data; Based on the pore type classification results, the volume of each type of pore is quantitatively integrated to obtain the microporosity distribution data; The maximum inscribed sphere is calculated for the pore type classification results to obtain the pore throat radius distribution data; Based on the pore type classification results, the pore geometry is analyzed by tensor field to obtain the reservoir pore geometry parameters.

4. The method for constructing a reservoir evaluation model based on deep learning according to claim 1, characterized in that: In step S2, inverting the reservoir macro-porosity based on the complete well logging data includes: Perform multi-sensor acquisition accuracy calibration and unified standardization processing on the complete logging data to obtain standardized logging curve data; Perform multivariate nonlinear covariance analysis on standardized logging curve data to generate logging response characteristic parameters; Analyze reservoir fluid properties based on the spectrum of logging response characteristic parameters, and perform reservoir fluid phase separation based on the spectrum analysis results to obtain fluid type distribution data; Perform fluid-constrained neutron-density porosity joint inversion on the standardized well logging curve data and fluid type distribution data to obtain initial effective porosity data; Based on the initial effective porosity data, the clay mineral content in the reservoir is analyzed by X-ray diffraction and corrected to obtain the clay correction coefficient; The initial effective porosity data and the clay correction coefficient are numerically integrated, and the numerical integration results are volume averaged to obtain the reservoir macro-porosity data.

5. The method for constructing a reservoir evaluation model based on deep learning according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: performing feature correlation analysis on the complete data set and performing multicollinearity detection to generate a standardized feature data set; Step S42: Identify reservoir segments in the standardized feature dataset based on the preset geological stratification information, and divide the reservoir segments based on the identification results to obtain a labeled segment dataset; divide the labeled segment dataset into 75% as a training set, 15% as a validation set, and 15% as a test set; Step S43: constructing a one-dimensional convolutional neural network model using the training set, and performing nonlinear transformation configuration to obtain a basic network structure; Step S44: performing Dropout and L2 weight regularization on the basic network structure to obtain an anti-overfitting enhanced network structure; Step S45: introducing a self-attention calculation unit into the anti-overfitting enhancement network structure and performing weight initialization to obtain a geological sensitive feature enhancement component; Step S46: weighting the geological sensitive feature enhancement components to generate geological key features; performing multi-scale integration of the original convolution feature application features based on the geological key features to obtain geological enhancement features; Step S47: constructing skip connections based on geological enhancement features to obtain a gradient transfer optimization network structure; Step S48: performing an attention-guided loss combination design on the gradient transfer optimization network structure to obtain a geological perception comprehensive loss function; Step S49: Integrate the basic network structure, geological sensitive feature enhancement components, gradient transfer optimization network structure and geological perception comprehensive loss function into a geological attention enhancement neural network through a distributed deep learning framework.

6. The method for constructing a reservoir evaluation model based on deep learning according to claim 1, characterized in that: In step S5, the weak supervision signal guidance of the geological attention enhancement neural network includes: Obtain a geological constraint rule set; perform rule matching on the training set based on the geological constraint rule set to obtain a weak supervision signal label; Apply weak supervision signal labeling to the geological attention-enhanced neural network and optimize the attention weights to generate a weak supervision guided architecture; Based on the weak supervision guidance architecture, a geological rule consistency loss term is designed to generate a multi-objective loss function. Perform constrained optimization on the multi-objective loss function to obtain a geological knowledge constrained optimizer; The geological attention enhanced neural network and geological knowledge constraint optimizer are integrated and reconstructed to obtain the geological knowledge constrained attention network.

7. The method for constructing a reservoir evaluation model based on deep learning according to claim 1, characterized in that: In step S5, the deep learning model training is performed using geological knowledge to constrain the attention network and the balanced training dataset, and hyperparameter optimization and performance monitoring are performed based on the validation set, including: Based on the geological knowledge constrained attention network, forward propagation and backpropagation training are performed on the balanced training data set to obtain the initial network weight parameter set; Statistically analyze the loss function value, validation set prediction error, and gradient change trend during the training process to generate a dynamic performance indicator dataset; Perform automated hyperparameter search on the geological knowledge-constrained attention network based on a dynamic performance indicator dataset to obtain the optimal hyperparameter combination set; Based on the optimal combination of hyperparameters, batch training is performed on the balanced training dataset to obtain a candidate model parameter set; Based on the validation set, early stopping and Dropout validation are performed on the candidate model parameter set to obtain the candidate model.

8. The method for constructing a reservoir evaluation model based on deep learning according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: using the test set to perform the prediction task on the candidate model to obtain a reservoir parameter prediction value set; Step S62: performing error comparison analysis of real core labels based on the reservoir parameter prediction value set to obtain a model error evaluation index set; Step S63: fine-tuning parameters and structures of the candidate model based on the model error evaluation index set to obtain an initial reservoir evaluation model; Step S64: performing integrated nesting on the initial reservoir evaluation model and the preset SZ3 framework to obtain the final reservoir evaluation model.

9. A system for constructing a reservoir evaluation model based on deep learning, characterized in that: The method for constructing a reservoir evaluation model based on deep learning according to claim 1 is used to execute the system for constructing a reservoir evaluation model based on deep learning, comprising: The data preprocessing module is used to obtain well logging data and core analysis data of the target work area, and perform outlier processing and missing value filling respectively to obtain complete well logging data and complete core data; The pore parameter inversion module is used to calculate reservoir pore structure parameters based on complete core data; and to invert reservoir macroporosity based on complete well logging data; The reservoir attribute analysis module is used to analyze reservoir lithofacies physical properties based on reservoir pore structure parameters and reservoir macroporosity, determine pore connectivity index, reservoir quality factor, lithologic control coefficient and heterogeneity index, and generate a reservoir evaluation index data set; The geological attention model construction module is used to divide the reservoir evaluation index dataset into training set, validation set and test set according to the layer segment; the training set is used to build the basic structure of the one-dimensional convolutional neural network model, and the geological attention mechanism is introduced to obtain the geological attention enhanced neural network; The model training and enhancement module is used to guide the geological attention enhancement neural network with weak supervision signals to obtain a geological knowledge constrained attention network. The training set is expanded by data synthesis of small sample reservoirs through generative adversarial networks to obtain a balanced training data set. The geological knowledge constrained attention network and the balanced training data set are used to perform deep learning model training, and hyperparameter optimization and performance monitoring are performed based on the validation set to obtain a candidate model. The model evaluation and integration module is used to evaluate and optimize the performance of candidate models using the test set to obtain an initial reservoir evaluation model; the preset SZ3 framework is integrated with the initial reservoir evaluation model to generate the final reservoir evaluation model.

Images