Method and device for grading reserves of oil and gas fields

Abstract

The application discloses an oil and gas field reserve grading processing method and device, and relates to the field of oil and gas field reserve grading processing. The method comprises the following steps: obtaining original data of an oil field to be evaluated; the original data comprises data of multiple oil reservoir development dimensions; inputting the original data into a pre-trained automatic encoder network, and outputting a cost level difference feature vector from an encoder in the automatic encoder network; the automatic encoder network comprises an encoder and a decoder, and the automatic encoder network is trained to minimize the error between the encoder input data and the decoder reconstruction output data; inputting the cost level difference feature vector into a value rating prediction model to output the reserve grade of the oil field to be evaluated; the value rating prediction model is trained by using the cost level difference feature vector and the known reserve grade of a historical oil field by ridge regression. The application can improve the objectivity, accuracy and generalization ability of the automatic reserve grading evaluation of the oil and gas field.

Description

Technical Field

[0001] This invention relates to the fields of oil and gas field reserve evaluation, development technology and artificial intelligence technology, and in particular to a method and apparatus for classifying oil and gas field reserves. Background Technology

[0002] The grading and evaluation of oil and gas field reserves is a core component of oil and gas companies' investment decisions, asset acquisitions, and corporate planning. The scientific validity and accuracy of the evaluation results directly affect the returns on investments worth billions or even tens of billions of dollars. Scientific, accurate, and efficient reserve evaluation methods are of paramount importance for optimizing resource allocation, reducing investment risks, and improving efficiency.

[0003] Currently, the most commonly used reserve valuation methods in the industry mainly rely on traditional multi-index comprehensive evaluation systems. These methods typically follow this process: First, a system of indicators encompassing geology, engineering, and reservoir aspects is constructed; then, linear dimensionality reduction methods such as principal component analysis (PCA) are used to compress multiple original indicators into a few comprehensive indicators (often referred to as "cost differentials"); finally, the oil and gas field reserve value level is determined based on the variance contribution rate of each comprehensive indicator.

[0004] However, as oil and gas exploration and development expands into unconventional, deep-water, and complex structural areas, traditional evaluation methods have revealed several inherent defects and limitations in practical applications. First, the linear assumption cannot capture the complex nonlinear interactions between geological, engineering, and cost indicators, leading to distorted information extraction. Second, the accuracy is low, and the variance contribution rate cannot represent the true predictive importance of value. Third, the model has poor generalization ability, resulting in insufficient evaluation accuracy and reliability under complex geological conditions.

[0005] Valuation of unconventional and complex oil and gas reservoirs often involves high-dimensional, nonlinear, and strongly coupled data characteristics, which are difficult to handle effectively by traditional linear methods. Therefore, there is an urgent need for a classification and evaluation method for oil and gas field reserves that can overcome these shortcomings. Summary of the Invention

[0006] This invention provides a method for classifying oil and gas field reserves, which accurately characterizes the complex relationships among multiple factors affecting reserve value, and improves the objectivity, accuracy, and generalization ability of automated oil and gas field reserve classification. The method includes:

[0007] Obtain raw data for the oilfield to be evaluated; the raw data includes data from multiple reservoir development considerations.

[0008] The raw data is input into a pre-trained autoencoder network, and the encoder in the autoencoder network outputs a cost level difference feature vector. The autoencoder network includes an encoder and a decoder, and the autoencoder network is trained with the goal of minimizing the error between the encoder input data and the decoder reconstructed output data.

[0009] The cost difference feature vector is input into the value rating prediction model, and the reserve level of the oilfield to be evaluated is output. The value rating prediction model is obtained by training the ridge regression model in advance using the cost difference feature vector and known reserve level of historical oilfields.

[0010] In one embodiment, the encoder architecture in the autoencoder network includes an input layer, a hidden layer, and a bottleneck layer; the decoder architecture in the autoencoder network is symmetrical to the encoder architecture in the autoencoder network; wherein the hidden layer uses the ReLU activation function, and the bottleneck layer uses a linear activation function.

[0011] In one embodiment, the autoencoder network is trained using unsupervised learning, with mean squared error as the loss function and Adam as the optimization algorithm.

[0012] In one embodiment, after training is complete, the decoder is removed from the autoencoder network, and the encoder is retained for feature extraction.

[0013] In one embodiment, the value rating prediction model is pre-trained as follows:

[0014] Collect raw data from multiple historical oilfields;

[0015] By using a pre-trained autoencoder network to extract features from the raw data of multiple historical oilfields, the cost difference feature vector of each historical oilfield is obtained.

[0016] By analyzing the correlation between the historical cost difference feature vector and the original data of each oilfield, the cost difference feature vector of each oilfield is given physical meaning, resulting in a cost difference feature vector with physical meaning. The cost difference feature vector with physical meaning includes external cost difference, gathering and transportation cost difference, reserve cost difference, extraction cost difference, and production cost difference.

[0017] By using the physically meaningful cost difference feature vectors of each historical oilfield and the known reserve levels, a ridge regression model is trained to obtain a value rating prediction model.

[0018] In one embodiment, each cost level feature vector consists of multiple cost level features;

[0019] By analyzing the correlation between the historical cost difference feature vector and the original data for each oilfield, physical meaning is assigned to the cost difference feature vector for each historical oilfield, resulting in a cost difference feature vector with physical meaning, including:

[0020] The Pearson correlation coefficient between the cost difference feature vector and the original data for each historical oilfield is calculated. Combined with perturbation analysis, the correlation between the cost difference feature vector and the original data for each historical oilfield is determined, and the cost difference feature vector for each historical oilfield is given physical meaning, resulting in a cost difference feature vector with physical meaning. The perturbation analysis means: applying a preset percentage increase or decrease to the original data of the historical oilfield, observing the rate of change of each cost difference feature, and associating the cost difference feature with the largest rate of change with the corresponding original data with the preset percentage increase or decrease.

[0021] In one embodiment, the ridge regression model has the following objective function:

[0022] ;

[0023] In the formula, T represents the known reserve level; Z train is the cost difference feature vector; w is the weight of the cost difference feature vector; is the regularization parameter; the objective function selects the optimal regularization parameter through cross-validation and solves for the optimal weight w.

[0024] In one embodiment, before inputting the raw data into the pre-trained autoencoder network, the method further includes:

[0025] The raw data of the oilfield to be evaluated is preprocessed to obtain preprocessed raw data; the preprocessing includes one or any combination of eigenvalue mapping, data cleaning, and standardization.

[0026] The raw data is fed into a pre-trained autoencoder network, including:

[0027] The preprocessed raw data is input into a pre-trained autoencoder network.

[0028] This invention also provides an oil and gas field reserve grading processing device to accurately characterize the complex relationships among multiple factors affecting reserve value, thereby improving the objectivity, accuracy, and generalization ability of automated oil and gas field reserve grading evaluation. The device includes:

[0029] The data acquisition module is used to acquire raw data of the oilfield to be evaluated; the raw data includes data from multiple reservoir development consideration dimensions.

[0030] The cost difference feature vector processing module is used to input the raw data into a pre-trained autoencoder network, and the encoder in the autoencoder network outputs the cost difference feature vector; the autoencoder network includes an encoder and a decoder, and the autoencoder network is trained with the goal of minimizing the error between the encoder input data and the decoder reconstructed output data.

[0031] The reserve level output module is used to input the cost difference feature vector into the value rating prediction model and output the reserve level of the oilfield to be evaluated. The value rating prediction model is obtained by training a ridge regression model in advance using the cost difference feature vector of historical oilfields and the known reserve level.

[0032] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for classifying and processing oil and gas field reserves.

[0033] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described oil and gas field reserve classification processing method.

[0034] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described oil and gas field reserve classification processing method.

[0035] In this embodiment of the invention, data from multiple reservoir development consideration dimensions are acquired, and feature extraction is performed using an autoencoder network to obtain cost difference feature vectors. Relying on the powerful nonlinear mapping capability of the autoencoder network, low-dimensional essential features containing complex relationships—intelligent cost difference—are automatically extracted from the original indicators, accurately depicting the complex relationships of multiple factors affecting reserve value. Finally, a supervised training value rating prediction model is used to accurately determine the reserve level of the oilfield to be evaluated, significantly improving the objectivity, accuracy, and generalization ability of automated oil and gas field reserve level evaluation. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0037] Figure 1 This is a schematic flowchart of the oil and gas field reserve classification and processing method in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of a data preprocessing process in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram of the automatic encoder network topology in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the encoder feature extraction process in an embodiment of the present invention;

[0041] Figure 5 This is a dimensionality-reduced visualization of the cost difference feature vector in an embodiment of the present invention;

[0042] Figure 6 This is a heatmap showing the correlation between cost differential characteristics and original indicator characteristics in an embodiment of the present invention.

[0043] Figure 7 This is a performance comparison diagram between the oil and gas field reserve classification method in this embodiment of the invention and existing methods;

[0044] Figure 8 This is a schematic diagram of an example of the oil and gas field reserve classification and processing method in this invention.

[0045] Figure 9 This is a schematic diagram of an oil and gas field reserve grading and processing device in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0047] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0048] Currently, the most commonly used reserve valuation methods in the industry mainly rely on traditional multi-index comprehensive evaluation systems. These methods typically follow this process: First, a multi-dimensional index system encompassing geology, engineering, and reservoir aspects is constructed; then, linear dimensionality reduction methods such as principal component analysis (PCA) are used to compress multiple original indices into a few comprehensive indices (often referred to as "cost gradients"); finally, the weights of each comprehensive index are determined based on its variance contribution rate, and a final value score is synthesized through linear weighting, which is then used for grading.

[0049] However, as oil and gas exploration and development expands into unconventional, deep-water, and complex structural areas, traditional evaluation methods have revealed several inherent defects and limitations in practical applications, mainly in the following aspects:

[0050] One issue is the limitation of the linear assumption.

[0051] Linear PCA, as a linear transformation method based on the eigenvalue decomposition of the covariance matrix, assumes a linear relationship between variables. However, there are numerous complex and nonlinear interactions among the factors influencing the value of oil and gas field reserves. Specifically:

[0052] There is a significant nonlinear relationship between reservoir physical properties (such as porosity and permeability) and development costs. For example, when permeability falls below a certain critical value, development costs increase exponentially, and traditional linear models cannot accurately characterize this abrupt change.

[0053] There is a complex coupling relationship between geological conditions and cost parameters. Taking offshore oil fields as an example, the relationship between increased water depth and rising development costs is not a simple linear one, but rather exhibits a step-like growth pattern.

[0054] Significant interaction effects exist among the various indicators. For example, effective thickness and crude oil viscosity jointly affect oil recovery, and this interaction is difficult to express effectively within the linear framework of PCA.

[0055] Due to the prevalence of the aforementioned nonlinear relationships, the comprehensive information extracted by PCA is often distorted, failing to accurately reflect the complex characteristics of oil and gas reservoirs, leading to a systematic deviation between the evaluation results and the actual situation. Limitations of the assumptions: As a linear transformation method, PCA's core assumption is the existence of linear relationships between variables. However, there are numerous complex and nonlinear interactions among the factors influencing the value of oil and gas field reserves (such as reservoir properties and development costs, geological conditions and cost parameters). PCA cannot effectively capture these nonlinear relationships, resulting in distorted comprehensive information that fails to accurately reflect the complex characteristics of oil and gas reservoirs.

[0056] Second, there is the issue of subjectivity and rigidity in determining weights.

[0057] In traditional evaluation methods, the weight of a comprehensive index is usually determined by its variance contribution rate. This method has obvious theoretical flaws:

[0058] The variance contribution rate only represents the extent to which the composite index explains the variation in the original data, not its predictive importance for the final reserve value. Statistically, a variable with a large explanatory power for variance is not necessarily a variable with strong predictive power.

[0059] The weighting allocation is disconnected from the actual drivers of business value. For example, in some cases, although a certain cost increment may have a low variance contribution rate, its marginal impact on the cost value of an oilfield may be very significant.

[0060] The weights are fixed and lack a dynamic adjustment mechanism. In reality, the importance of each cost factor changes significantly under different oil price environments and at different stages of development, and the fixed weights of traditional methods cannot adapt to this dynamic nature.

[0061] This method of determining weights makes the cost significance of the evaluation results unclear and fails to truly reflect the actual contribution of each factor to the value of reserves.

[0062] Third, there is the problem of feature interpretation relying on human experience.

[0063] The principal components obtained after PCA dimensionality reduction are linear combinations of the original indices, and their feature interpretation presents many difficulties:

[0064] The physical meaning is vague. Each principal component is usually a mixture of multiple original variables, making it difficult to correspond to specific cost categories, such as investment costs, operating costs, and other concrete concepts.

[0065] The interpretation process relies heavily on the professional experience and subjective judgment of the analysts. Different experts may provide completely different interpretations of the same principal component, leading to a lack of consistency and comparability in the evaluation results.

[0066] The interpretation process is cumbersome and inefficient. Analysts need to infer by observing the factor loading matrix and combining it with their professional knowledge, a process that is both time-consuming and prone to human bias.

[0067] The lack of a standardized interpretive framework makes it difficult to directly compare evaluation results from different projects and at different times, hindering the practical application of the methodology.

[0068] Fourth, there is the problem of insufficient model generalization ability and accuracy.

[0069] Due to the simplicity of the model structure, traditional methods face significant challenges when processing modern oil and gas field data:

[0070] PCA has limited capabilities for processing high-dimensional data. As evaluation metrics become more refined and data dimensionality increases, PCA performs poorly in the face of the curse of dimensionality.

[0071] The model has insufficient ability to fit nonlinear relationships. The nonlinear interactions between various factors in complex oil and gas reservoirs cannot be fully represented in linear models.

[0072] Sensitive to outliers. PCA is based on the L2 norm, which is quite sensitive to outliers. A single outlier sample may cause the entire model to become distorted.

[0073] The prediction accuracy is insufficient. In practical applications, the prediction error of traditional methods often exceeds the acceptable range, affecting their credibility in investment decisions.

[0074] It is particularly noteworthy that with the large-scale development of unconventional oil and gas resources, the evaluation data of resources such as shale oil and tight gas exhibit stronger nonlinearity, higher dimensionality, and more complex coupling characteristics, which makes the limitations of traditional methods more prominent.

[0075] To overcome the above problems, academia and industry have made various attempts to improve the existing technology:

[0076] Some studies have attempted to introduce nonlinear PCA methods to handle nonlinear problems through kernel functions, but this method has significant shortcomings in terms of theoretical complexity and computational efficiency, and still cannot solve the problems of weight determination and feature interpretation.

[0077] Other studies have used subjective weighting methods such as the Analytic Hierarchy Process (AHP) to determine weights. Although these methods take into account the differences in importance of each factor to some extent, they are highly subjective and cannot achieve data-driven dynamic weight adjustment.

[0078] Other studies have attempted to introduce machine learning methods such as neural networks, but due to the limited number of oil and gas field samples, overfitting is likely to occur, and the interpretability of the models is poor, making it difficult to promote their application in oil and gas investment decisions that require high credibility.

[0079] In summary, existing technologies suffer from systemic defects in addressing the problem of grading and evaluating the value of oil and gas field reserves, such as limitations of linear assumptions, rigid weight determination, difficulties in interpreting features, and insufficient model generalization. There is an urgent need to develop a new evaluation method that can simultaneously solve these problems.

[0080] Based on this, this invention provides a method and apparatus for classifying oil and gas field reserves, which addresses the problems of traditional PCA linear methods, such as inability to capture complex nonlinear relationships, subjective and rigid weight determination, difficulty in interpreting physical features, and insufficient model generalization ability. It accurately characterizes the complex relationships among multiple factors affecting reserve value, improving the objectivity, accuracy, and generalization ability of automated oil and gas field reserve classification. Figure 1 This is a schematic flowchart of the oil and gas field reserve classification and processing method in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0081] Step 101: Obtain the raw data of the oilfield to be evaluated; the raw data includes data from multiple reservoir development consideration dimensions;

[0082] Step 102: Input the raw data into the pre-trained autoencoder network, and the encoder in the autoencoder network outputs the cost level difference feature vector; the autoencoder network includes an encoder and a decoder, and the autoencoder network is trained with the goal of minimizing the error between the encoder input data and the decoder reconstructed output data;

[0083] Step 103: Input the cost difference feature vector into the value rating prediction model and output the reserve level of the oil field to be evaluated; the value rating prediction model is obtained by training the ridge regression model in advance using the cost difference feature vector and known reserve level of historical oil fields.

[0084] The method for classifying and processing oil and gas field reserves in the embodiments of the present invention will be explained in detail below.

[0085] In step 101, the raw data of the oilfield to be evaluated is obtained. The raw data includes data from multiple reservoir development consideration dimensions, such as external development conditions, gathering and transportation capacity, reserve level, internal reservoir exploitation conditions, and production conditions. Each dimension may include specific data.

[0086] During implementation, specific data on multiple reservoir development considerations of the oilfield to be evaluated can be obtained to construct the original data matrix.

[0087] For example, in this embodiment of the invention, a comprehensive evaluation system is constructed that includes 5 primary dimensions (external development conditions, gathering and transportation capacity, reserve level, reservoir internal exploitation conditions, and production conditions) and 21 specific indicators, wherein each primary dimension includes multiple indicator data:

[0088] External development conditions (4 items): geomorphological features, oil and gas reservoir structure, offshore oil reservoir water depth, and reservoir lithology;

[0089] Gathering and transportation capacity (4 items): gathering and transportation distance, gathering and transportation method, gathering and transportation equipment conditions, and surface crude oil quality;

[0090] Reserve grade (3 items): proven reserve level, proven geological reserve size, and reserve abundance;

[0091] Internal oil reservoir exploitation conditions (4 items): resource type, reservoir type, driving type-development method, and oil production method;

[0092] Production conditions (6 items): oil and gas reservoir burial depth, effective thickness, permeability, heterogeneity, underground crude oil viscosity, and saturation pressure.

[0093] In one embodiment, before inputting the raw data into a pre-trained autoencoder network, the method may further include: preprocessing the raw data of the oilfield to be evaluated to obtain preprocessed raw data; the preprocessing includes feature value mapping, data cleaning, and standardization, or any combination thereof.

[0094] Inputting raw data into a pre-trained autoencoder network can include: inputting pre-processed raw data into a pre-trained autoencoder network.

[0095] Figure 2 This is a schematic diagram of a data preprocessing process in an embodiment of the present invention, such as... Figure 2 As shown, it specifically includes:

[0096] Step 201: Data Acquisition and Quantification: Acquire raw data from multiple oilfields to be evaluated, including the data of the 21 indicators mentioned above;

[0097] Step 202, Feature value mapping: Based on the degree of impact on development costs, each indicator data is mapped to a numerical range of 0 to 1 to form a feature value; the mapping principle is: the closer the feature value is to 1, the higher the development cost and the lower the economic value.

[0098] Step 203: Construct the original data matrix: Combine the 21 eigenvalues ​​of each oilfield sample into a 21-dimensional eigenvector; the eigenvectors of all samples together constitute the original data matrix;

[0099] Step 204: Data cleaning and standardization.

[0100] The following example, using the XC oilfield, illustrates the data preprocessing process.

[0101] First, a eigenvalue quantification matrix of 21 indicators for the XC oilfield was established. Each indicator was classified into levels and assigned an eigenvalue of 0 to 1 (1 represents the highest cost and the lowest value). The actual parameters of each indicator (such as "plain", "fault block", "developed") were mapped to a unified 0 to 1 eigenvalue scale according to their extraction difficulty or cost impact.

[0102] Assignment principle: 1 represents the highest cost and the lowest development value; 0 represents the lowest cost and the highest development value.

[0103] For example, "offshore oil reservoir water depth" is assigned the value of "onshore," which has extremely low cost and is set to 0. "Oil and gas reservoir structure" is assigned the value of "fault block," which has high complexity and high cost, and is set to 0.68. During implementation, a mapping relationship between index keywords and feature values ​​is established. Keywords are extracted from the acquired raw data and assigned feature values. Table 1 shows the mapping process of feature values ​​for 21 indicators across 5 dimensions in the XC oilfield.

[0104] Table 1: Characteristic values ​​of original indicators for XC oilfield

[0105]

[0106] Eigenvalue quantization: Using the original data of XC oilfield, the corresponding indicator levels are mapped to eigenvalues ​​to form a 21-dimensional eigenvector of XC oilfield. Table 2 shows the 21 original indicator eigenvalues ​​of XC oilfield, which can form the original data matrix.

[0107] Table 2. Original Data Matrix of XC Oilfield

[0108]

[0109] Constructing the original data matrix Then, the original data matrix Data cleaning was performed, and each feature was processed using the Z-score normalization method.

[0110] To eliminate the influence of unit dimensions, Z-score standardization is performed on each feature (each column). This step unifies the scale of all features, generating a standardized data matrix with a mean of 0 and a standard deviation of 1, which serves as the input to the subsequent model. The processing formula is as follows:

[0111] ;

[0112] in:

[0113] x is the feature value obtained in the previous stage;

[0114] It is the average value of this feature across all n samples;

[0115] It is the standard deviation of this feature over all n samples.

[0116] The 21-dimensional vector above is Z-score normalized, and global values ​​are calculated using multiple oilfield samples. and Assume we only have the "effective thickness" data for the above 5 samples: [25.0, 12.5, 45.0, 8.0, 30.0].

[0117] Maximum value = 45.0;

[0118] Minimum value = 8.0;

[0119] Range = 45.0 - 8.0 = 37.0.

[0120] The "effective thickness" feature value for each sample is calculated as follows:

[0121] Oilfield A: (45.0 - 25.0) / 37.0 = 20.0 / 37.0 ≈ 0.54;

[0122] Oilfield B: (45.0 - 12.5) / 37.0 = 32.5 / 37.0 ≈ 0.88;

[0123] Oilfield C: (45.0 - 45.0) / 37.0 = 0.0 / 37.0 = 0.00 (Greatest thickness, lowest cost);

[0124] Oilfield D: (45.0 - 8.0) / 37.0 = 37.0 / 37.0 = 1.00 (smallest thickness, highest cost);

[0125] Oilfield E: (45.0 - 30.0) / 37.0 = 15.0 / 37.0 0.41.

[0126] For negative indicators (the larger the value, the worse, such as heterogeneity), use:

[0127] .

[0128] The Z-score normalization is demonstrated using the "effective thickness" feature values ​​[0.54, 0.88, 0.00, 1.00, 0.41] of 5 samples:

[0129] Calculate the mean (0.54+0.88+0.00+1.00+0.41) / 5=2.83 / 5=0.566;

[0130] Calculate the standard deviation :

[0131]

[0132] Standard deviation .

[0133] Z-score normalization is applied to each value:

[0134]

[0135] For a single XC oilfield sample, the final preprocessed raw data matrix is ​​as follows:

[0136]

[0137] Furthermore, in step 102, the preprocessed original data matrix can be input into a pre-trained autoencoder network, and the encoder in the autoencoder network outputs multiple cost difference features to form a cost difference feature vector.

[0138] The autoencoder network consists of an encoder and a decoder, and is trained with the goal of minimizing the error between the encoder input data and the decoder reconstructed output data. After training, the encoder part is used to nonlinearly reduce the high-dimensional original data to a low-dimensional bottleneck layer, and extract the intelligent cost difference feature vector that represents the core features of the oilfield.

[0139] During implementation, a symmetric autoencoder neural network is pre-built, and then forward propagation is defined, core parameters are set, etc.

[0140] The encoder architecture in the autoencoder network includes an input layer, a hidden layer, and a bottleneck layer; the decoder architecture in the autoencoder network is symmetrical to the encoder architecture in the autoencoder network.

[0141] The encoder consists of 21 neurons in the input layer, followed by at least two hidden layers using the ReLU activation function, and a bottleneck layer of 5 neurons using a linear activation function. After training, it outputs intelligent cost differential features.

[0142] Decoder: As a mirror-symmetric structure of the encoder, it is designed to reconstruct data from the bottleneck layer, and the final output layer has 21 neurons.

[0143] Figure 3 This is a schematic diagram of the automatic encoder network topology in an embodiment of the present invention, with reference to... Figure 3 The encoder compresses the high-dimensional input into a low-dimensional latent representation, and the decoder reconstructs the input from the latent representation. The encoder part is a multi-layer feedforward neural network, and its output is the bottleneck layer. Its topology is as follows:

[0144] Input layer: 21 neurons, receiving standardized 21-dimensional data;

[0145] Encoder hidden layers: 12 neurons in the hidden layer, ReLU activation function; 8 neurons in the hidden layer, ReLU activation function; 5 neurons in the bottleneck layer, linear activation function, outputting intelligent cost difference features.

[0146] Decoder hidden layers: 8 neurons in the hidden layer, ReLU activation function; 12 neurons in the hidden layer, ReLU activation function; output layer 21 neurons, linear activation function, reconstructing the original input.

[0147] Define forward propagation: clarify the computation method of each layer of the encoder and decoder, that is, through the linear transformation of the weight matrix and bias vector, and then the application of the activation function (ReLU or linear), to realize the complete reconstruction process from input to bottleneck layer encoding, and then to output.

[0148] Core parameter settings: The bottleneck layer dimension is strictly set to 5 dimensions, corresponding to the five types of intelligent cost difference features to be extracted.

[0149] Among them, the encoder forward propagation, for the input vector :

[0150] First hidden layer:

[0151] h1 represents the output of the first hidden layer. b1 represents the connection weights between the neurons in the first hidden layer and all neurons in the input layer, and b1 represents the bias vector of the first hidden layer.

[0152] Second hidden layer:

[0153] h2 represents the output of the second hidden layer. b1 represents the connection weights between the neurons in the second hidden layer and all neurons in the input layer, and b2 represents the bias vector of the first hidden layer.

[0154] Bottleneck layer:

[0155] ; b1 represents the connection weights between the bottleneck layer neurons and all neurons in the input layer, and b2 represents the bottleneck layer bias vector.

[0156] in This is the cost difference feature vector.

[0157] Decoder forward propagation:

[0158] Third hidden layer:

[0159] h4 represents the output of the third hidden layer. b4 represents the connection weights between the neurons in the third hidden layer and all neurons in the input layer, and b4 represents the bias vector of the third hidden layer.

[0160] Fourth hidden layer:

[0161] h5 represents the output of the fourth hidden layer. b5 represents the connection weights between the neurons in the fourth hidden layer and all neurons in the input layer, and b5 represents the bias vector of the fourth hidden layer.

[0162] Output layer:

[0163] ; This represents the output of the final decoder output layer. b represents the connection weights of the output layer, and b6 represents the bias vector of the output layer.

[0164] After training, the encoder outputs This refers to the intelligent cost difference feature vector, where each cost difference feature vector consists of multiple cost difference features. Table 3 shows the intelligent cost difference feature vectors extracted from the XC oilfield by the automatic encoder. Example.

[0165] Table 3. Smart cost difference feature vector Z extracted from XC oilfield by automatic encoder.

[0166]

[0167] The reconstructed output should be as close as possible to the original input, and the reconstruction error is used for training and optimization.

[0168] In one embodiment, the autoencoder network is trained using unsupervised learning, with mean squared error as the loss function and Adam as the optimization algorithm.

[0169] During implementation, raw sample data from multiple oilfields within the target evaluation area were collected, covering 21 indicators across five dimensions: external development conditions, gathering and transportation capacity, reserve level, internal reservoir exploitation conditions, and production conditions. After preprocessing, a standardized dataset was formed and divided into training, validation, and test sets. For example, the data was divided into three sets: 70% for training, 15% for validation, and 15% for test. The training set was used for model parameter learning, the validation set was used to monitor the training process and implement early stopping to prevent overfitting, and the test set was used for final model performance evaluation.

[0170] The autoencoder is pre-trained in an unsupervised manner with the goal of minimizing reconstruction error. During training, the model learns to compress high-dimensional input data to a bottleneck layer before reconstruction, forcing the network to extract the most representative features from the data.

[0171] Training configuration:

[0172] - Optimization algorithm: Adam optimizer;

[0173] - Learning rate: 0.001;

[0174] - Batch size: 32;

[0175] - Number of training rounds: 500;

[0176] - Early stopping mechanism: Training is terminated when the validation set loss does not decrease for 20 consecutive rounds; an early stopping mechanism is used to prevent overfitting;

[0177] - Loss function: Mean squared error (MSE) is used:

[0178] ;

[0179] In the formula:

[0180] This represents the original input data for the i-th sample.

[0181] Output data reconstructed from the automatic encoder;

[0182] Reconstruction loss is used to measure the difference between the reconstructed output of the autoencoder and the original input.

[0183] n is the number of samples in the batch.

[0184] The training process is divided into the initial, mid-convergence, late-convergence, and steady-state phases. The loss decreases steadily. Key indicators during the training process are shown in Table 4, which illustrates the changing trends of these key indicators.

[0185] Table 4. Trends in key indicators during the training process

[0186]

[0187] The encoder serves as a feature extractor: After training, the decoder is removed from the autoencoder network, leaving only the encoder for feature extraction. As a feature extractor, the encoder maps any input oilfield sample Y to its low-dimensional latent representation Z:

[0188] ;

[0189] In the formula:

[0190] Z is the extracted 5-dimensional intelligent cost differential feature vector;

[0191] These are the encoder's weights and bias parameters;

[0192] This is the ReLU activation function.

[0193] Figure 4 This is a schematic diagram of the encoder feature extraction process in an embodiment of the present invention, as shown below. Figure 4 As shown, the processing flow includes: input data (21-dimensional standardized data). Encoder forward propagation (nonlinear transformation compression 15) 12 8 5) Bottleneck layer output (5-dimensional latent representation) ) Cost differential features (feature vector Z).

[0194] Table 5 shows the model evaluation performance of the autoencoder network in the embodiments of the present invention.

[0195] Table 5

[0196]

[0197] In one embodiment, each cost level feature in the cost level feature vector output by the encoder has physical meaning.

[0198] In one embodiment, before inputting the cost differential feature vector into the value rating prediction model, the method further includes:

[0199] By analyzing the correlation between the cost differential feature vector and the original data of the oilfield to be evaluated, each cost differential feature is given a physical meaning, resulting in cost differential features with physical meaning. The cost differential feature vector with physical meaning includes external cost differential, gathering and transportation cost differential, reserve cost differential, extraction cost differential, and production cost differential.

[0200] Furthermore, multiple cost difference feature vectors are input into the value rating prediction model, including inputting cost difference feature vectors with physical meaning into the value rating prediction model.

[0201] For example, Table 6 shows the intelligent cost differential characteristics Z of the XC oilfield output by the autoencoder network.

[0202] Table 6

[0203]

[0204] In this embodiment, correlation analysis includes calculating the Pearson correlation coefficient between the cost differential feature vector and the original data. Based on the absolute value of the correlation coefficient (a threshold is set, e.g., |absolute value of correlation coefficient| > 0.6 indicates a strong correlation), the analysis identifies which clusters of original indicators each cost differential feature is primarily associated with. Based on the analysis results and combined with oilfield development expertise, these five potential dimensions are interpreted and named as follows: external cost differential, gathering and transportation cost differential, reserve cost differential, extraction cost differential, and production cost differential, giving them clear physical meanings.

[0205] In one embodiment, t-SNE (t-distributed stochastic neighbor embedding) is used to reduce the dimensionality of the cost difference feature vector for visualization, which can effectively maintain the local proximity relationship between data points. For example, the 5-dimensional Z is reduced to 2-dimensional for visualization, as shown in Table 7, which displays the t-SNE dimensionality-reduced coordinates of the XC oilfield.

[0206] Table 7

[0207]

[0208] Figure 5 This is a dimensionality-reduced visualization of the cost difference feature vector in an embodiment of the present invention. Figure 5 The table shows the first and second dimensions of t-SNE in Table 7.

[0209] in:

[0210] The horizontal axis (t-SNE dimensionality 1): the first dimension after t-SNE dimensionality reduction.

[0211] Vertical axis (t-SNE dimensionality 2): The second dimension after t-SNE dimensionality reduction.

[0212] Each point represents the projection of a sample's cost difference feature vector into a two-dimensional space.

[0213] The color of a point is represented by the color bar on the right, corresponding to the magnitude of the eigenvalue. The closer the color is to red, the higher the eigenvalue; the closer the color is to green, the lower the eigenvalue.

[0214] Figure 5 Points marked with a star represent a specific reference sample; for example, the eigenvalue of the XC oil field is approximately 0.32.

[0215] High value (low eigenvalue): corresponding to Figure 5 The data points are tinged with green. These points represent samples with low eigenvalues ​​in the cost differential eigenvector. Low eigenvalues ​​usually mean that these samples perform well on cost-related features, and may have lower costs or higher cost-effectiveness.

[0216] Low value (high eigenvalue): Corresponds to data points that are reddish in the graph. These points represent samples with high eigenvalues ​​in the cost differential eigenvector. High eigenvalues ​​usually mean that these samples perform poorly on cost-related features, and may have high costs or low cost-effectiveness.

[0217] In step 103, the cost difference feature vector is input into the value rating prediction model, and the reserve level of the oil field to be evaluated is output. The value rating prediction model is obtained by training the ridge regression model in advance using the cost difference feature vector and known reserve level of historical oil fields.

[0218] In one embodiment, the value rating prediction model can be pre-trained as follows:

[0219] Collect raw data from multiple historical oilfields;

[0220] By using a pre-trained autoencoder network to extract features from the raw data of multiple historical oilfields, the cost difference feature vector of each historical oilfield is obtained.

[0221] By analyzing the correlation between the historical cost difference feature vector and the original data of each oilfield, the cost difference feature vector of each oilfield is given physical meaning, resulting in a cost difference feature vector with physical meaning. The cost difference feature vector with physical meaning includes external cost difference, gathering and transportation cost difference, reserve cost difference, extraction cost difference, and production cost difference.

[0222] By using the physically meaningful cost difference feature vectors of each historical oilfield and the known reserve levels, a ridge regression model is trained to obtain a value rating prediction model.

[0223] This process involves analyzing the correlation between the historical cost difference feature vector and the original data for each oilfield, assigning physical meaning to the cost difference feature vector of each historical oilfield, and obtaining a physically meaningful cost difference feature vector. This can include: calculating the Pearson correlation coefficient between the historical cost difference feature vector and the original data for each oilfield, and combining this with perturbation analysis to determine the correlation between the historical cost difference feature vector and the original data, thus assigning physical meaning to the cost difference feature vector of each historical oilfield and obtaining a physically meaningful cost difference feature vector. The perturbation analysis involves applying a preset percentage increase or decrease to the original data of the historical oilfield, observing the rate of change of each cost difference feature, and associating the cost difference feature with the largest rate of change with the corresponding original data to which the preset percentage increase or decrease was applied.

[0224] During implementation, raw data from multiple historical oilfields are collected, and after training the autoencoder network, the ridge regression model is further trained to obtain the value rating prediction model.

[0225] After extracting the cost differential feature vector for each historical oilfield, the potential features (i.e., cost differential features) are interpreted and named. The correlation between the potential feature Z and the original index data is analyzed by calculating the Pearson correlation coefficient matrix. The correlation coefficient calculation formula is as follows:

[0226] ;

[0227] In the formula,

[0228] Let i be the i-th potential feature (i=1,2,3,4,5);

[0229] For the j-th original index (j=1,2,...,21);

[0230] It is the value of the i-th potential feature (cost difference feature) in the k-th oilfield sample;

[0231] It is the value of the j-th original index in the k-th oilfield sample;

[0232] n is the number of samples;

[0233] These are the means of the corresponding features.

[0234] Based on the correlation analysis results, each potential feature was interpreted and named separately as external cost difference, gathering and transportation cost difference, reserve cost difference, mining cost difference and production cost difference. The Pearson correlation coefficient between Z and the original data matrix Y was calculated (Table 8), and perturbation analysis was performed. Table 8 shows the correlation coefficient between Z and the original data matrix Y.

[0235] Table 8

[0236]

[0237] The factor loadings in Table 8 are approximate values ​​of the correlation coefficients because the nonlinear features learned by the autoencoder should have similar physical meanings to the principal components of PCA.

[0238] Perturbation sensitivity analysis verification, applying perturbation sensitivity to each original index Observe the changes in latent features by observing minute perturbations:

[0239] ;

[0240] In the formula, This represents the change in the i-th cost differential feature;

[0241] Indicates the amount of disturbance;

[0242] This represents the original index value.

[0243] Perturbation analysis involves applying a small perturbation to a raw indicator and observing the changes in all potential features. The potential dimension with the largest change is considered most correlated with that indicator. Table 9 shows the results of perturbation analysis. For example, by applying a perturbation to the raw indicator of landform features, the changes in each feature in the cost difference feature vector are observed to determine whether the most sensitive feature is consistent with Pearson correlation analysis. If inconsistent, a data review is conducted to check for errors or outliers in the data used for both analyses, confirm whether the data distribution meets the assumptions of Pearson correlation analysis (normal distribution, linear relationship, etc.), or review whether the magnitude and method of the perturbation in the perturbation analysis are reasonable.

[0244] Table 9

[0245]

[0246] Figure 6 This is a heatmap illustrating the correlation between cost differential features and original indicator features in an embodiment of the present invention. The heatmap shows the correlation between each intelligent cost differential feature (i.e., each dimension z1 to z5 in the Z vector) and the original indicator. Such heatmaps typically use color intensity to represent the degree of correlation between different dimensions, where red may indicate a strong positive correlation, while blue or green may indicate a weak or negative correlation.

[0247] Figure 6 middle:

[0248] The rows represent different intelligence cost tiers, such as z1 to z5.

[0249] The column represents 21 original indicators, or, for simplicity, representative indicators after these indicators have been categorized.

[0250] Color: Represents the absolute value of the correlation coefficient. The redder the color, the stronger the correlation between the two variables; conversely, a lighter color indicates a weaker or no correlation.

[0251] Figure 6 Content Interpretation:

[0252] z1 (external) is highly correlated with geographical factors such as geomorphological features, seawater depth, and reservoir lithology, which suggests that z1 may represent external environmental factors that affect the difficulty of oilfield development.

[0253] z2 (gathering and transportation) is closely related to certain indicators of surface extraction, such as gathering and transportation distance, crude oil quality, and equipment conditions, which indicates that z2 focuses on how to efficiently transport oil from the production site to the processing site.

[0254] z3 (reserves) is strongly correlated with indicators such as reserve abundance, reserve size, and degree of exploration, indicating key factors that need to be considered when determining reserve scale and resource endowment.

[0255] z4 (exploitation) is directly related to the burial depth of oil and gas reservoirs, drive type, and oil production method, revealing the technical path to improve extraction efficiency.

[0256] z5 (production) is related to effective thickness, permeability, heterogeneity, etc., reflecting the various conditions required to maintain high productivity.

[0257] Perturbation sensitivity analysis verification: Applying perturbation sensitivity analysis to the original indicators We observed the rate of change of the latent feature Z under small perturbations. Each latent dimension was most sensitive to perturbations of its highly correlated original indices, validating the rationality of the naming and interpretation. For example, the perturbation of "topographic features" showed the largest change (8.2%), confirming that z1 is "external cost differential".

[0258] t-SNE is used to reduce the 5-dimensional Z to 2-dimensional, which can then be visualized on a scatter plot. For example... Figure 5 In the t-SNE dimensionality reduction, high-value oilfields (green dots) and low-value oilfields (red dots) show a clear clustering separation trend, proving that the intelligent cost difference feature has good discriminative ability.

[0259] When constructing a value rating prediction model, the extracted low-dimensional features are used as input and the known reserve level is used as output to train a ridge regression model, thereby learning the optimal weight coefficient vector of each intelligent cost difference feature when predicting the reserve level.

[0260] In this embodiment of the invention, the core consideration for choosing ridge regression is:

[0261] - Feature Dimensions: 5-dimensional intelligent cost differential features;

[0262] - Number of samples: n training samples;

[0263] - Regularization requirement: to prevent possible multicollinearity among the 5-dimensional features;

[0264] - Stability: L2 regularization improves the model's generalization ability.

[0265] In this embodiment, the ridge regression model has the following objective function, which is solved by introducing an L2 regularization term to obtain the weight coefficients:

[0266] ;

[0267] In the formula, T represents the known reserve level; Z train is the cost difference feature vector; w is the weight of the cost difference feature vector; is the regularization parameter; the objective function selects the optimal regularization parameter through cross-validation and solves for the optimal weight w.

[0268] Hyperparameter tuning: Using cross-validation, select the regularization coefficient that optimizes the performance of the validation set within the candidate value range. ).

[0269] Obtaining weight coefficients: After training, the weight coefficient vector of the ridge regression model is obtained. This vector reflects the contribution of each intelligent cost level difference to the final value level prediction.

[0270] The specific training steps are as follows:

[0271] 1. Input Preparation: Intelligent Cost Differential Feature Matrix Z train ;

[0272] 2. Output label: Known reserve value level T;

[0273] 3. Parameter optimization: Select the optimal parameter through cross-validation. ;

[0274] 4. Weight learning: Finding the w that minimizes the objective function;

[0275] 5. Model Validation: Evaluate the prediction performance on the validation set.

[0276] Referring to Table 10, the optimal regularization parameter is selected through grid search and cross-validation. .

[0277] Table 10

[0278]

[0279] After the ridge regression model is trained, a weight coefficient vector is obtained. This weight coefficient vector reflects the contribution of each intelligent cost difference feature to the final reserve level prediction. The weights trained by the ridge regression model are then allocated to each intelligent cost difference feature for oil and gas field reserve level prediction. In this embodiment of the invention, the weights of the cost difference features are dynamically generated based on data, replacing the traditional fixed weight allocation method based on the principal component variance contribution rate.

[0280] For example, using Z and a known reserve value level T to train a ridge regression, the ridge regression learns the weights. Since Z is a standardized latent feature, the predicted value needs to be de-standardized or mapped to the [0,1] interval by setting a threshold.

[0281] Table 11 shows a comparison between the traditional PCA weights and the ridge regression weights of the present invention.

[0282] Table 11

[0283]

[0284] Train the value rating prediction model as described above and use it for new oilfield evaluation and decision support.

[0285] In step 103, the cost difference feature vector is input into the value rating prediction model, and the reserve level of the oil field to be evaluated is output.

[0286] For example, the aforementioned standardization process is applied to the new oilfield sample Snew, and then input into the trained encoder. The intelligent cost differential feature Z is obtained. new , will Z new The data is input into the value rating prediction model to obtain the predicted value of its reserves. :

[0287] ;

[0288] According to the predicted value Oil and gas reserves are classified into three levels: high-value Level I (0-0.3), medium-value Level II (0.3-0.6), and low-value Level III (0.6-0.9), achieving accurate reserve classification. The above classification thresholds (0-0.3, 0.3-0.6, 0.6-0.9) are applicable to the final predicted value T_pred mapped to the [0,1] interval after destandardization.

[0289] Using the XC oilfield prediction example as an illustration:

[0290] Use its smart cost tier feature Heling Regression Weights Make a prediction:

[0291]

[0292] Due to Z and It is a standardized value, which needs to be destandardized and mapped back to the original level range [0,1].

[0293] Assuming the mean of the training set labels Standard deviation :

[0294] ;

[0295] The predicted value of 0.3502 is very close to the actual value of 0.32, which proves the effectiveness of the model.

[0296] Figure 7 This is a performance comparison diagram between the oil and gas field reserve classification method in this embodiment of the invention and existing methods, such as... Figure 7 As shown, the X-axis represents the name of the performance metric (such as MSE, MAE, R). 2 The Y-axis represents the index value (the smaller the better or the larger the better), green represents the index value of the oil and gas field reserve classification method in the embodiment of the present invention, and red represents the existing PCA linear regression method. The specific value is marked at the top of the column for easy and accurate comparison.

[0297] In this embodiment of the invention, MSE = 0.008, and the prediction is very close to the true value;

[0298] The PCA method has an MSE of 0.012, which is 50% higher than the error in the embodiment of this invention.

[0299] MSE represents the average absolute value of the difference between the predicted value and the actual value.

[0300] In this embodiment of the invention, the MAE is 0.06, and the average prediction deviation is only 6%.

[0301] The PCA method has an MAE of 0.08 and an average deviation of 8%, indicating a larger error.

[0302] R² (Coefficient of Determination) is the proportion of the variance of the target variable explained by the model, reflecting the goodness of fit of the model.

[0303] The meaning of R²: the closer the value is to 1, the better (maximum value is 1, minimum value can be negative).

[0304] Embodiment R of the present invention 2 = 0.92, the model explains 92% of the value level changes;

[0305] PCA method R 2 = 0.85 Only 85% was explained, and 15% could not be explained.

[0306] Calculate MSE, MAE, and R on the test set. 2 Table 12 shows a performance comparison diagram.

[0307] Table 12

[0308]

[0309] Figure 8 This is a schematic diagram of an example of a method for classifying and processing oil and gas field reserves according to an embodiment of the present invention. The method includes:

[0310] Step 801: Data preparation and preprocessing;

[0311] Step 802: Design and training of the autoencoder network;

[0312] Step 803: Feature extraction and dimensionality reduction visualization based on autoencoder network;

[0313] Step 804: Explanation and naming of cost differential characteristics;

[0314] Step 805: Construct a value rating prediction model based on ridge regression;

[0315] Step 806: System Integration and Model Validation;

[0316] Step 807: Evaluation and application of new oilfields.

[0317] During implementation, the system can simultaneously perform oil and gas reserve classification for multiple new oilfields, outputting predicted reserve levels for each new oilfield, and then sorting and classifying them. This provides quantitative data support for investment prioritization and development decisions. Furthermore, the predicted reserve levels and cost differentials of new oilfields can be visualized, and the key cost factors influencing the reserve level can be identified through the analysis of each component, providing precise guidance for cost reduction and efficiency improvement.

[0318] This invention addresses the limitations of existing PCA methods, such as linearity, subjective weighting, and difficulty in interpretation. Based on the multi-dimensional characteristics of oilfield value assessment, it implements a process including data standardization preprocessing, unsupervised feature extraction via autoencoder, intelligent cost differential physical interpretation, ridge regression weight learning, and new oilfield value prediction. This method leverages the powerful nonlinear mapping capabilities of deep learning to automatically extract low-dimensional essential features (intelligent cost differentials) containing complex relationships from the original indicators. Supervised learning is then used to dynamically determine their objective weights, ultimately achieving accurate value grading. This invention not only outputs rating results but also diagnoses key cost drivers affecting value, providing in-depth insights for cost reduction, efficiency improvement, and investment decisions. The technical solution of this invention comprehensively covers the entire process from data input, intelligent feature extraction, model training to decision support, combining theoretical advancement with engineering practicality. It can be widely applied to reserve value grading and evaluation in conventional and unconventional oil and gas fields at different development stages.

[0319] The embodiments of this invention overcome the limitations of traditional linear methods in capturing nonlinear relationships, determining the objectivity of weights, and adapting to complex geological conditions. In practical applications in multiple oilfields, the classification consistency rate is improved by more than 15% compared with the traditional PCA method, providing a scientific basis for oil and gas field investment decisions and cost reduction and efficiency improvement.

[0320] This invention also provides an oil and gas field reserve grading and processing device, as described in the following embodiments. Since the principle by which this device solves the problem is similar to that of the oil and gas field reserve grading and processing method, the implementation of this device can refer to the implementation of the oil and gas field reserve grading and processing method; repeated details will not be elaborated further.

[0321] Figure 9 This is a schematic diagram of an oil and gas field reserve grading and processing device in an embodiment of the present invention, as shown below. Figure 9 As shown, the device 900 includes a data acquisition module 901, a cost difference feature vector processing module 902, and a reserve level output module 903.

[0322] The data acquisition module 901 is used to acquire the raw data of the oilfield to be evaluated; the raw data includes data from multiple reservoir development consideration dimensions.

[0323] The cost difference feature vector processing module 902 is used to input the original data into a pre-trained autoencoder network, and the encoder in the autoencoder network outputs the cost difference feature vector; the autoencoder network includes an encoder and a decoder, and the autoencoder network is trained with the goal of minimizing the error between the encoder input data and the decoder reconstructed output data.

[0324] The reserve level output module 903 is used to input the cost difference feature vector into the value rating prediction model and output the reserve level of the oilfield to be evaluated. The value rating prediction model is obtained by training the ridge regression model in advance using the cost difference feature vector of historical oilfields and the known reserve level.

[0325] In one embodiment, the encoder architecture in the autoencoder network includes an input layer, a hidden layer, and a bottleneck layer; the decoder architecture in the autoencoder network is symmetrical to the encoder architecture in the autoencoder network; wherein the hidden layer uses the ReLU activation function, and the bottleneck layer uses a linear activation function.

[0326] In one embodiment, the autoencoder network is trained using unsupervised learning, with mean squared error as the loss function and Adam as the optimization algorithm.

[0327] In one embodiment, after training is complete, the decoder is removed from the autoencoder network, and the encoder is retained for feature extraction.

[0328] In one embodiment, the value rating prediction model is pre-trained as follows:

[0329] Collect raw data from multiple historical oilfields;

[0330] By using a pre-trained autoencoder network to extract features from the raw data of multiple historical oilfields, the cost difference feature vector of each historical oilfield is obtained.

[0331] By analyzing the correlation between the historical cost difference feature vector and the original data of each oilfield, the cost difference feature vector of each oilfield is given physical meaning, resulting in a cost difference feature vector with physical meaning. The cost difference feature vector with physical meaning includes external cost difference, gathering and transportation cost difference, reserve cost difference, extraction cost difference, and production cost difference.

[0332] By using the physically meaningful cost difference feature vectors of each historical oilfield and the known reserve levels, a ridge regression model is trained to obtain a value rating prediction model.

[0333] In one embodiment, each cost level feature vector consists of multiple cost level features;

[0334] By analyzing the correlation between the historical cost difference feature vector and the original data for each oilfield, physical meaning is assigned to the cost difference feature vector for each historical oilfield, resulting in a cost difference feature vector with physical meaning, including:

[0335] The Pearson correlation coefficient between the cost difference feature vector and the original data for each historical oilfield is calculated. Combined with perturbation analysis, the correlation between the cost difference feature vector and the original data for each historical oilfield is determined, and the cost difference feature vector for each historical oilfield is given physical meaning, resulting in a cost difference feature vector with physical meaning. The perturbation analysis means: applying a preset percentage increase or decrease to the original data of the historical oilfield, observing the rate of change of each cost difference feature, and associating the cost difference feature with the largest rate of change with the corresponding original data with the preset percentage increase or decrease.

[0336] In one embodiment, the ridge regression model has the following objective function:

[0337] ;

[0338] In the formula, T represents the known reserve level; Z train is the cost difference feature vector; w is the weight of the cost difference feature vector; is the regularization parameter; the objective function selects the optimal regularization parameter through cross-validation and solves for the optimal weight w.

[0339] In one embodiment, the device 900 further includes: a preprocessing module;

[0340] The preprocessing module is used to preprocess the original data of the oilfield to be evaluated before the cost difference feature vector processing module 902 inputs the original data into the pre-trained autoencoder network, so as to obtain the preprocessed original data; the preprocessing includes one or any combination of feature value mapping, data cleaning, and standardization.

[0341] The cost difference feature vector processing module 902 is specifically used for:

[0342] The preprocessed raw data is input into a pre-trained autoencoder network.

[0343] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for classifying and processing oil and gas field reserves.

[0344] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described oil and gas field reserve classification processing method.

[0345] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described oil and gas field reserve classification processing method.

[0346] This invention provides a method and apparatus for oil and gas field reserve grading and evaluation based on an autoencoder and ridge regression. This method is particularly suitable for processing high-dimensional, nonlinear, and highly coupled oilfield value evaluation data. It can accurately characterize the complex relationships between multiple factors affecting reserve value, achieving objective, accurate, and efficient automated grading and evaluation. Compared with existing linear evaluation methods based on principal component analysis, this invention obtains raw multi-dimensional evaluation index data from multiple oilfield samples within the target evaluation area; performs data cleaning and standardization preprocessing; constructs and trains an autoencoder to extract nonlinear intelligent cost difference features; provides physical interpretation and naming of the features; trains a ridge regression model to learn weights in a data-driven manner; and performs value prediction and grading of new oilfields. This effectively overcomes the linear limitations and subjectivity of traditional methods, automatically captures complex nonlinear relationships between indicators, and provides objective, accurate, and insightful reserve value grading results, significantly improving the scientific rigor, accuracy, and decision support efficiency of the evaluation work.

[0347] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0348] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0349] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0350] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0351] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for classifying and processing oil and gas field reserves, characterized in that, include: Obtain raw data for the oilfield to be evaluated; the raw data includes data from multiple reservoir development considerations. The raw data is input into a pre-trained autoencoder network, and the encoder in the autoencoder network outputs a cost level difference feature vector. The autoencoder network includes an encoder and a decoder, and the autoencoder network is trained with the goal of minimizing the error between the encoder input data and the decoder reconstructed output data. The cost difference feature vector is input into the value rating prediction model, and the reserve level of the oil field to be evaluated is output. The value rating prediction model is obtained by training the ridge regression model in advance using the cost difference feature vector and known reserve level of historical oil fields. The value rating prediction model is pre-trained in the following manner: Collect raw data from multiple historical oilfields; By using a pre-trained autoencoder network to extract features from the raw data of multiple historical oilfields, the cost difference feature vector of each historical oilfield is obtained. By analyzing the correlation between the historical cost difference feature vectors of each oilfield and the original data, the cost difference feature vectors of each historical oilfield are given physical meaning, resulting in cost difference feature vectors with physical meaning. These cost difference feature vectors with physical meaning include external cost difference, gathering and transportation cost difference, reserve cost difference, extraction cost difference, and production cost difference. By using the physically meaningful cost difference feature vectors of each historical oilfield and the known reserve levels, a ridge regression model is trained to obtain a value rating prediction model.

2. The method as described in claim 1, characterized in that, The encoder architecture in the autoencoder network includes an input layer, a hidden layer, and a bottleneck layer; the decoder architecture in the autoencoder network is symmetrical to the encoder architecture; wherein the hidden layer uses the ReLU activation function, and the bottleneck layer uses the linear activation function.

3. The method as described in claim 1, characterized in that, The autoencoder network was trained using an unsupervised learning approach, with mean squared error as the loss function and Adam as the optimization algorithm.

4. The method as described in claim 3, characterized in that, After training is complete, the decoder is removed from the autoencoder network, and the encoder is retained for feature extraction.

5. The method as described in claim 1, characterized in that, Each cost difference feature vector consists of multiple cost difference features; By analyzing the correlation between the historical cost difference feature vector and the original data for each oilfield, physical meaning is assigned to the cost difference feature vector for each historical oilfield, resulting in a cost difference feature vector with physical meaning, including: The Pearson correlation coefficient between the cost difference feature vector and the original data for each historical oilfield is calculated. Combined with perturbation analysis, the correlation between the cost difference feature vector and the original data for each historical oilfield is determined, and the cost difference feature vector for each historical oilfield is given physical meaning, resulting in a cost difference feature vector with physical meaning. The perturbation analysis means: applying a preset percentage increase or decrease to the original data of the historical oilfield, observing the rate of change of each cost difference feature, and associating the cost difference feature with the largest rate of change with the corresponding original data with the preset percentage increase or decrease.

6. The method as described in claim 1, characterized in that, The ridge regression model has the following objective function: ； In the formula, T indicates a known reserve level; Z train This represents the cost-level feature vector; w represents the weight of the cost differential feature vector; For regularization parameters; The objective function selects the optimal regularization parameter through cross-validation and solves for the optimal weight w.

7. The method as described in claim 1, characterized in that, Before feeding the raw data into the pre-trained autoencoder network, the following steps are also included: The raw data of the oilfield to be evaluated is preprocessed to obtain preprocessed raw data; the preprocessing includes one or any combination of eigenvalue mapping, data cleaning, and standardization. The raw data is fed into a pre-trained autoencoder network, including: The preprocessed raw data is then input into a pre-trained autoencoder network.

8. A device for classifying and processing oil and gas field reserves, characterized in that, include: The data acquisition module is used to acquire raw data of the oilfield to be evaluated; the raw data includes data from multiple reservoir development consideration dimensions. The cost difference feature vector processing module is used to input the raw data into a pre-trained autoencoder network, and the encoder in the autoencoder network outputs the cost difference feature vector; the autoencoder network includes an encoder and a decoder, and the autoencoder network is trained with the goal of minimizing the error between the encoder input data and the decoder reconstructed output data. The reserve level output module is used to input the cost difference feature vector into the value rating prediction model and output the reserve level of the oilfield to be evaluated. The value rating prediction model is obtained by training a ridge regression model in advance using the cost difference feature vector of historical oilfields and the known reserve level. The value rating prediction model is pre-trained in the following manner: Collect raw data from multiple historical oilfields; By using a pre-trained autoencoder network to extract features from the raw data of multiple historical oilfields, the cost difference feature vector of each historical oilfield is obtained. By analyzing the correlation between the historical cost difference feature vectors of each oilfield and the original data, the cost difference feature vectors of each historical oilfield are given physical meaning, resulting in cost difference feature vectors with physical meaning. These cost difference feature vectors with physical meaning include external cost difference, gathering and transportation cost difference, reserve cost difference, extraction cost difference, and production cost difference. By using the physically meaningful cost difference feature vectors of each historical oilfield and the known reserve levels, a ridge regression model is trained to obtain a value rating prediction model.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

11. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

Images