Tight sandstone reservoir evaluation method and system based on knowledge graph and storage medium
By constructing a dense sandstone reservoir evaluation method based on knowledge graphs, using the knowledge extraction module and inference module for multi-task collaborative reasoning, the problems of data fragmentation and model staticization are solved, and efficient and accurate evaluation of dense sandstone reservoirs are achieved.
Patent Information
- Application Number
- CN202511009527.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-22
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2045-07-22
AI Technical Summary
The existing tight sandstone reservoir evaluation methods have problems with data fragmentation and model staticization, and are easily affected by subjective experience in manual decision-making, resulting in low evaluation accuracy and insufficient stability.
Construct a dense sandstone reservoir evaluation method based on knowledge graph, including data preprocessing, knowledge extraction modules and knowledge inference modules, perform multi-task collaborative reasoning through a structure-aware convolution network embedded in triplet extraction and rule-embedded structure-aware convolution network, integrate multi-source data such as geology, well logging, and experiments, and design an intelligent reservoir evaluation system.
The rapid and accurate classification and parameter prediction of tight sandstone reservoirs are achieved, which reduces artificial experience dependence and subjective errors, improves work efficiency, and enhances the unity and accuracy of evaluation.
Smart Images

Figure CN120509473A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of geological exploration and development, and in particular to a knowledge graph-based tight sandstone reservoir evaluation method, system, and storage medium. Background Art
[0002] Tight reservoirs are an important exploration area for tight oil and gas resources. Due to their strong heterogeneity and complex pore structure, the knowledge system and exploration methods summarized in conventional oil and gas reservoir production operations make it difficult to identify the geological type of tight sandstone reservoirs. During the evaluation process, due to the large span of reservoir characteristics in different regions and even within the same basin, different parameters must be configured and different models must be selected to improve the compliance of reservoir evaluation. In addition, the evaluation methods and standards adopted by different scholars and research institutions vary, making it difficult to uniformly compare and weigh the evaluation results. Therefore, how to integrate multi-source data such as geology, well logging, and experiments to build a comprehensive reservoir characteristic knowledge base, and combine technologies such as artificial intelligence to design intelligent tight sandstone reservoir evaluation methods has become a key research topic at present.
[0003] As knowledge graphs continue to expand in size and their application scenarios deepen, graph neural networks employ a unique node description method to produce a state that incorporates both neighbor node information and graph topology features. Applying a corresponding convolutional learning approach to the graph and aggregating relevant features and information on the graph provides greater interpretability during prediction. This approach effectively considers both semantic and structural information within the knowledge graph, particularly capturing the structural information within the knowledge graph. Existing evaluation methods often suffer from data fragmentation and static modeling. Because the evaluation of tight sandstone reservoirs requires comprehensive consideration of reservoir dynamics data and actual conditions, manual decision-making is easily limited by subjective experience, resulting in low evaluation accuracy and stability. Summary of the Invention
[0004] The technical problems to be solved by the present invention are:
[0005] Existing evaluation methods often have problems with data fragmentation and model staticness, and the manual evaluation process is easily affected by subjective experience.
[0006] The present invention is to solve the above technical problems using the following technical solutions:
[0007] The present invention provides a tight sandstone reservoir evaluation method based on a knowledge graph, comprising the following steps:
[0008] Step S100: Collect tight sandstone reservoir evaluation data, pre-process the tight sandstone reservoir evaluation data, build an ontology knowledge base, and build a tight sandstone reservoir domain ontology layer based on the knowledge and its associations;
[0009] Step S200: Construct a tight sandstone reservoir evaluation model, comprising a knowledge extraction module and a knowledge reasoning module. The knowledge extraction module is used to extract triples from entity relationships annotated in the tight sandstone reservoir sequence and construct a tight sandstone reservoir knowledge graph with entity relationships combined. The knowledge reasoning module is based on a rule-embedded structure-aware convolutional network and is used to perform multi-task collaborative reasoning.
[0010] Step S300: Perform multi-task collaborative reasoning of tight sandstone reservoirs based on the tight sandstone reservoir evaluation model.
[0011] Furthermore, the tight sandstone reservoir evaluation data in step S100 includes text, numerical values, and image data.
[0012] Furthermore, the construction of the main layer in the tight sandstone reservoir area includes:
[0013] Step S110: Clarify the scope of the tight reservoir field: The core information sources of the reservoir field include: a tight reservoir basic information database, a terminology set and academic literature related to tight reservoir evaluation industry standards, and a systematic compilation and summary of the practical experience of reservoir evaluation experts and their professional teams;
[0014] Step S120, Entity Concept and Attribute Definition: There are two methods for concept extraction. One is to extract concepts from a macro to micro level and in a step-by-step manner based on a pre-defined business process architecture. The other is to conduct in-depth mining of information from resources such as document materials and basic information databases through collection and retrieval to extract the required concepts to be defined.
[0015] Step S130, relationship definition: Use the attribute ontology architecture and clustering attribute method to describe the characteristics of the classification relationship between attributes, and clarify the relationship between each category based on the extracted concepts and the results obtained from the demand analysis.
[0016] Furthermore, the knowledge extraction module in step S200 includes: an embedding module, an encoding module and a decoding module;
[0017] The embedding module is constructed with a BERT module for converting sentences into dynamic semantic information at the character level and word level;
[0018] The encoding module is constructed with a BiLSTM module and a CNN module. The BiLSTM module is used to extract context information of dense reservoir sequence data. The CNN module is used to extract detailed features of the data using multiple convolution kernels of different sizes, and select the maximum value of the feature map through a pooling layer to further aggregate features.
[0019] The decoding layer is constructed with a CRF module for obtaining the global optimal label sequence of the entity;
[0020] Finally, triple extraction of entity relationships is achieved, and a knowledge graph of tight sandstone reservoirs with entity relationships is constructed.
[0021] Furthermore, the CNN module uses 1×3 convolution to capture the entity features of single geological terms and reservoir texts, and uses 5×5 convolution to capture geological term combinations and their contextual relationships.
[0022] Furthermore, the knowledge reasoning module in step S200 includes a rule-enhanced graph convolution layer and a rule-constrained prediction layer. The rule-enhanced graph convolution layer is used in the encoding stage, and the matching degree between the modeling node and the geological rule is dynamically modeled through a differentiable rule encoder, and the rule weight gating in the multi-head attention mechanism is used to realize the directional screening and aggregation of neighborhood node information; in the decoding stage, the rule-constrained prediction layer introduces a two-stage correction mechanism, first filtering out illegal predictions through hard constraints, and then combining soft rules to construct residual correction terms, and embedding the physical equations in a learnable form into the fully connected layer, and finally realizing the full-link closed-loop constraint of geological knowledge from the feature space to the decision space.
[0023] Furthermore, the functional implementation process of the rule-enhanced graph convolution layer is as follows:
[0024] In the reservoir knowledge graph G = (V, E, R), node feature updates use a reservoir rule-aware graph convolution operation, and the calculation formula is:
[0025]
[0026] in, For the Nodes in the layer The feature representation of is the activation function, R rule Represents a set of relations related to the rules, Indicates that in the relationship Next, node The set of neighbor nodes of Indicates that in the relationship Next node right Regularized attention weights, Indicates in Layers and relationships The transformation matrix under Indicates the Nodes in the layer The feature representation of Indicates the The rule-related transformation matrix in the layer, Represents the feature representation related to rule triggering, The calculation formula is as follows:
[0027]
[0028] in, represents the exponential function, 、 is the rule matching function, represents the weight vector in the attention mechanism, Represents the node i and the representation of node j To splice, Represents the node i and the representation of node k To splice, represents the set of neighbor nodes of node i, Represents an activation function used to introduce nonlinear transformation;
[0029] By defining the rule matching function Adjust the importance of neighboring nodes. The calculation formula is:
[0030]
[0031] in, represents the atomic formula input to the rule encoder, is a rule encoder that maps atomic formulas into computable conditions in a vector space:
[0032]
[0033] in, and represents the weight matrix, and represents the bias term.
[0034] Furthermore, the loss function of the rule-constrained prediction layer is a dynamic weighted loss function based on classification cross entropy, regression mean square error and rule constraints:
[0035]
[0036]
[0037] in, is the soft rule error term, is the model’s predicted output for the sample, is the reference output associated with the soft rule, is the rule-constrained prediction layer loss function, is the basic prediction loss of the model, represents a set of hard rules, represents a set of soft rules, Penalty for violating hard rules. is the soft rule error term, and is a weight parameter used to control the influence of the rule constraint.
[0038] The present invention provides a tight sandstone reservoir evaluation system based on a knowledge graph. The system has a program module corresponding to the steps of the method described in any one of the above technical solutions, and executes the steps in the above-mentioned tight sandstone reservoir evaluation method based on a knowledge graph during operation.
[0039] The present invention also provides a computer-readable storage medium, which stores a computer program. The computer program is configured to implement the steps of the knowledge graph-based tight sandstone reservoir evaluation method described in any one of the above technical solutions when called by a processor.
[0040] Compared with the prior art, the present invention has the following beneficial effects:
[0041] In view of the characteristics of dense professional terms, complex entity relationships and strong implicitness in tight reservoir text data, the present invention constructs an entity relationship joint extraction module based on a sequence annotation mode, breaking through the error accumulation problem caused by the separation of entity recognition and relationship classification in traditional pipeline extraction methods, and synchronously realizes end-to-end extraction of key reservoir entities and their associated relationships through a unified coding framework. The present invention condenses expert experience knowledge and formulates standard expressions for reasoning model recognition, constructs a structure-aware network reasoning module based on rule embedding, realizes multi-task reasoning in the field of tight sandstone reservoir evaluation, and designs a tight reservoir evaluation system based on a knowledge graph in combination with business background. It can quickly and accurately classify tight reservoirs and predict reservoir parameters, which not only improves work efficiency, but also reduces the problems of human experience dependence, subjective errors and multi-source data redundancy, and has high application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 Schematic diagram of the structure of a tight sandstone reservoir evaluation system based on knowledge graph in an embodiment of the present invention;
[0043] Figure 2 Schematic diagram of entity relationship extraction based on the knowledge extraction module in an embodiment of the present invention;
[0044] Figure 3 This is a network structure diagram of a knowledge reasoning module in an embodiment of the present invention;
[0045] Figure 4: is a diagram of a network structure of a rule encoder in an embodiment of the present invention;
[0046] Figure 5 A flowchart of multi-task collaborative processing in an embodiment of the present invention;
[0047] Figure 6 This is a diagram of the tight reservoir domain ontology in an embodiment of the present invention. DETAILED DESCRIPTION
[0048] In order to enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below with reference to the accompanying drawings. Obviously, the described embodiments or examples are only some of the embodiments or examples of the present invention, and not all of them. Based on the embodiments or examples of the present invention, all other embodiments or examples obtained by those skilled in the art without creative work should fall within the scope of protection of the present invention.
[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0050] Specific implementation scheme 1: The present invention provides a tight sandstone reservoir evaluation method based on knowledge graph, comprising the following steps:
[0051] Step S100: Collect tight sandstone reservoir evaluation data, pre-process the tight sandstone reservoir evaluation data, build an ontology knowledge base, and build a tight sandstone reservoir domain ontology layer based on the knowledge and its associations;
[0052] Step S200: Construct a tight sandstone reservoir evaluation model, comprising a knowledge extraction module and a knowledge reasoning module. The knowledge extraction module is used to extract triples from entity relationships annotated in the tight sandstone reservoir sequence and construct a tight sandstone reservoir knowledge graph with entity relationships combined. The knowledge reasoning module is based on a rule-embedded structure-aware convolutional network and is used to perform multi-task collaborative reasoning.
[0053] Step S300: Perform multi-task collaborative reasoning of tight sandstone reservoirs based on the tight sandstone reservoir evaluation model.
[0054] Specific implementation plan 2: The tight sandstone reservoir evaluation data includes: text, numerical and image data. Specifically including:
[0055] Regional geological data: including the stratigraphic sequence and structural characteristics, rock types and distribution, mineral resources, scale, as well as relevant geological maps and historical exploration results data of the region.
[0056] Core experimental data: including physical, chemical and structural property data of core samples obtained through laboratory analysis.
[0057] Curve data: formation information obtained using well logging technology, expressed in the form of curves.
[0058] Pore information: Pore images are obtained using cast thin section methods, scanning electron microscopy, and CT scanning. Pore characteristics are obtained from image analysis, including but not limited to pore structure, type, morphology, size, distribution, mineral composition and distribution, particle structure, and quantitative parameters extracted through image analysis.
[0059] Well logging data: including well logging documents and raw logging data
[0060] Oil and gas reservoir evaluation data: including reservoir geological and engineering evaluation documents and basic data. This implementation plan is otherwise the same as the first implementation plan.
[0061] Specific implementation plan three: image data preprocessing includes:
[0062] Standardize the image data: unify the image size to 512×512 pixels, remove noise, and enhance the visibility of cracks or pores;
[0063] Extract features from the image: Use the pre-trained ResNet50 feature extraction network, add the SE module after the residual block of ResNet, focus on the key feature channels, and pass FPN Fuse features at different levels to extract image features;
[0064] Cross-modal mapping: Extract reservoir information from image-attached labels, embed image feature vectors and text descriptions into the same space, and establish a relationship between the image and text. The rest of this implementation is the same as Specific Implementation Plan 2.
[0065] Specific implementation plan 4: The preprocessing of text data includes:
[0066] Standardize the text data: Convert the original reservoir data PDF file into TXT format, remove line breaks, spaces, punctuation marks in sentences, and remove numerical description text that does not have entity relationships.
[0067] Text filtering: Filter out irrelevant text and filter the text.
[0068] Text annotation: For entity boundary labels, each word in the text is marked with multiple labels, including the position of each word in the entity, the relationship type, and the entity location information. This implementation plan is otherwise the same as the specific implementation plan three.
[0069] Specific implementation plan 5: Construction of the main layer in the tight sandstone reservoir area, including:
[0070] Step S110: Clarify the scope of the tight reservoir field: The core sources of reservoir information include: the primary source is the tight reservoir basic information database, followed by the terminology and academic literature related to tight reservoir evaluation industry standards, including the classic works in the field of oil and gas evaluation, "Comprehensive Evaluation Technology of Oil and Gas Reservoirs" and "Tight Oil and Gas"; finally, the systematic organization and summary of the practical experience of reservoir evaluation experts and their professional teams;
[0071] A top-down domain ontology knowledge representation approach was used to construct the tight reservoir domain ontology layer. OWL DL was selected as the modeling tool, striving to ensure strong descriptive expressiveness while fully considering and achieving effective compatibility with reasoning mechanisms. The "entity-attribute-attribute value" relationship was mined from tight sandstone reservoir information to construct the tight sandstone reservoir domain knowledge ontology layer.
[0072] Step S120: Entity Concept and Attribute Definition: The ontology layer of the tight reservoir domain knowledge graph consists of a set of concept nodes and a set of concept relationship edges. It represents concepts and relationships between concepts in the geological field. Prior knowledge is used to define the concept hierarchy, attribute relationships, and semantic relationships between concepts for each element.
[0073] Concept extraction methods include two strategies. The first is to extract concepts from a macroscopic to microscopic, step-by-step, approach based on a pre-defined business process architecture. This process must closely adhere to the specific operational procedures of the reservoir evaluation field. For example, when classifying reservoir performance, reservoirs are divided into Class I, Class II, and Class III based on permeability differentials and porosity. To classify reservoirs based on this, the first step is to construct a tight sandstone reservoir concept ontology, including: reservoir, pore structure, fractures, minerals, porosity, and permeability. Well logging experiments reveal the relationships between the four properties: physical properties, oil properties, electrical properties, and lithology. Therefore, an ontology for these relationships is required. The lithology ontology includes sandstone type and cementation type, while the electrical ontology includes resistivity and natural potential.
[0074] Another strategy involves deep mining information from document sources and basic information databases through collection and retrieval, thereby extracting the desired concepts to be defined. For example, a search for the keyword "lithology" will retrieve rock types such as "dolomite," "gypsum," and "granite." These results are directly incorporated into the ontology as concepts for clear definition. After establishing these basic concepts, they must be refined and contextualized, analyzing the connotations of the concepts themselves and their associated attributes. This is then used to optimize the hierarchical structure between categories. The model concepts constructed include mathematical expressions such as Archie's formula and Darcy's law, as well as standard interpretation maps and visualizations of regionally adapted interpretation maps. The concepts and attributes extracted through these two approaches can broaden the scope of reservoir evaluation knowledge to the greatest extent possible and provide a strong foundation for the subsequent practical application of the ontology and attributes.
[0075] This implementation plan analyzes the entire tight sandstone reservoir domain knowledge base and related reservoir data. Through the above two detailed concept analysis methods, it is obtained that the tight reservoir domain knowledge includes 14 entity types: region, oil field, well, logging experiment, four properties, pore structure, mineralization, formula, plate, formation and geological data, as shown in Table 1.
[0076] Table 1 Reservoir entity type information
[0077] Entity Type Chinese meaning Example Area area Sanzhao Depression OilField oilfield Daqing Oilfield Well well Tai 27 well area Logging Well logging projects Caliper logging; Neutron logging Physical physical properties Porosity; Permeability Oiliness Oily Oil and gas saturation Electric Electrical properties Resistivity; natural potential Lithology Lithology mudstone pore structure Pore structure intergranular pores; dissolution pores; microcracks Mineral mineralization quartz; feldspar; kaolinite Formula formula Archie formula Plate plate Logging parameter cross plot; Stratum Strata conglomerate layer, sandstone layer Sdata geological data Reservoir literature; geological reports
[0078] Step S130, relationship definition: Use the attribute ontology architecture and clustering attribute method to describe the characteristics of the classification relationship between attributes. Use the attribute ontology architecture and clustering attribute method to describe the characteristics of the classification relationship between attributes. According to the extracted concepts and the results obtained from the demand analysis, clarify the relationship between each category. For example, oil saturation, permeability, and mud content are all obtained through calculation and analysis of model formulas, so there is a calculation relationship between the model formula and oiliness, electrical properties, and physical properties. At the same time, each oil field is located in a specific geographical area. For example, the relationship between the entity objects "Fuyi Oil Group Structural High Point" and "Tai 27 Well Area" is "located in". It is inferred that there is a containment relationship between the oil field and the oil well, and there is a geographical location relationship between the oil field and the region where it is located.
[0079] Finally, the well logging domain knowledge obtained includes: location, calculation, division, related formulas and usage relationships, as shown in Table 2.
[0080] Table 2
[0081] Relationship Type Chinese meaning Example Include Include <Greyish-black middle layer - massive limestone, containing, flint bands> Located lie in <The structural high point of the Fuyi oil group is located in the Tai 27 well area> Calculate calculate <Archie formula, calculation, oil saturation> Divide Division <Hole diameter logging, division, lithology> Relate_formu Related formulas <SP Logging, Related Formulas, Diffusion Electromotive Force Formula> Use belong <Survey area, belonging to, alpine region of the Qinghai-Tibet Plateau>
[0082] The ontology layer information of tight reservoir area is obtained as follows Figure 6The rest of this embodiment is the same as the specific embodiment four.
[0083] Specific implementation plan six: Figure 2 As shown, the knowledge extraction module includes: an embedding module, an encoding module and a decoding module;
[0084] The embedding module is constructed with a BERT module to convert sentences into dynamic semantic information at the character level and word level; the semantic information is then input into the encoding layer.
[0085] The encoding module is constructed with a BiLSTM module and a CNN module. The BiLSTM module is used to extract context information of dense reservoir sequence data. The CNN module is used to extract detailed features of the data using multiple convolution kernels of different sizes, and select the maximum value of the feature map through the pooling layer to further aggregate features.
[0086] The encoding module, for input X = [Sanzhao depression area, permeability, porosity, ..., larger], first inputs the vector representation of each word into the LSTM in sequence according to the natural order of the sentence to obtain a set of sequential hidden layer representations.
[0087] The forget gate is used to determine whether to ignore certain historical information that is irrelevant to the current reservoir characteristics. The calculation is:
[0088] (1)
[0089] in For the Gate of Forgetfulness, represents the recursive weight shared by the previous state, represents the hidden state of the previous time step, represents the feature vector input at time t, represents the bias term, is the sigmoid activation function.
[0090] The input gate selectively introduces new geological characteristics or rock type information based on the current reservoir attribute description. Its output is a vector between 0 and 1. The calculation formula of the input gate is:
[0091] (2)
[0092] in is the output of the input gate, is the sigmoid activation function, represents the recursive weight shared by the previous state, represents the hidden state of the previous time step, represents the feature vector input at time t, represents the bias term.
[0093] It is a new reservoir feature representation generated based on current information and historical experience. The formula for updating the cell state is:
[0094] (3)
[0095] in For the Gate of Forgetfulness, represents the cell state at time step t-1, is the output of the input gate, represents the hyperbolic tangent activation function, represents the recursive weight shared by the previous state, represents the hidden state of the previous time step, represents the feature vector input at time t, represents the bias term.
[0096] Among them, the output gate selectively outputs the predicted value of reservoir quality according to the current processing result. The calculation formula is shown in (4), and the calculation formula of the hidden state is shown in (5);
[0097] (4)
[0098] (5)
[0099] in represents the hidden state at time step t, represents the output of the output gate at time step t, represents the hyperbolic tangent activation function, It is a new reservoir characteristic representation generated based on current information and historical experience. represents the recursive weight shared by the previous state, represents the hidden state of the previous time step, represents the feature vector input at time t, represents the bias term.
[0100] Reverse the order of the words in the sentence and then input them into the LSTM to obtain a set of reverse order hidden layer representations. Finally, concatenate the sequential hidden layer representation and the reverse order hidden layer representation to obtain the final hidden layer representation of each sentence. .
[0101] The CNN module uses 1×3 convolution to capture the entity features of single geological terms and reservoir texts, uses 5×5 convolution to capture geological term combinations and their contextual relationships, and uses the ReLU function to generate higher-level feature representations to reflect the complex relationship between geological terms and attributes.
[0102] For each convolution kernel, the sliding window is used to Slide word or character by word on the image to get the feature map In order to obtain the most valuable features of the matrix obtained after convolution, the maximum pooling operation is performed on the feature map and the maximum value is selected. As the feature corresponding to a specific filter.
[0103] Finally, the fully connected layer is used for regularization and the dropout strategy is adopted to randomly discard hidden units to obtain the final feature representation of the input sentence. , and enter the decoding layer operation.
[0104] The decoding layer is constructed with a CRF module to obtain the global optimal label sequence. , let the corresponding actual label sequence be , the score function calculation formula of the label sequence is:
[0105]
[0106] in, is the tight reservoir characteristic representation, is the tight reservoir characteristic representation, It is the initial score matrix obtained by linear operation after the output of the BiLSTM layer. represents the state transition matrix, is a continuous word label Transfer to label The score, Indicates the first The first word The score of the label. The corresponding output label sequence Calculate the scores, and the sequence with the highest score is the final predicted label sequence.
[0107] Triples are extracted from entity relationships to construct a knowledge graph of tight sandstone reservoirs with entity relationships.
[0108] The resulting relationship classification results are converted into triples according to specific matching rules. First, based on entity boundaries, entity relationships, and annotated entity role labels, a relatively complete set of entities and relationships is obtained, including reservoir type, rock type, porosity, and permeability. Furthermore, the corresponding relationship type, entity role information, and location information are obtained. Entity-relationship triples are extracted only when the entity boundary labels meet the requirements of starting with "B," non-heading "I," and maintaining consistency between the relationship and role labels of each character within the entity. Secondly, for the extracted entities with common relationships, forward and backward searches are used to directly search for matching entities in the text. The following rules apply to entity matching: matching entities must have the same relationship label as the entity, while their entity role labels must be different. Finally, after obtaining matching entities from forward and backward searches, the distance between the current entity and the entity to be matched is determined, and the closest pair of entities is selected to form a matching pair, forming an entity-relationship triple.
[0109] First, entities related to tight sandstone reservoirs are treated as nodes of the graph, and each node is assigned a unique ID. Next, each triple is converted into a directed edge, with the direction from the head entity to the tail entity, and the triples are stored in the Neo4j graph database. Finally, it is converted into an adjacency matrix. , used for reservoir relationship modeling, and adopted the TransE embedding method to map entity relationships into vector space.
[0110] In the intelligent tight reservoir evaluation system, a rule base based on expert experience forms the foundation for tight sandstone reservoir evaluation reasoning. 263 domain experiences found in unstructured text, including geological data and core analysis reports from the Fuyu oil reservoir, were validated through multiple rounds of verification by geological experts, ultimately resulting in an evaluation system containing 217 valid rules. To enhance the machine interpretability and computability of knowledge, a generative rule representation method was used to formally encode implicit experience. At the semantic level, the expert experience was deconstructed into triples of (precondition, conclusion, confidence parameter), as shown in Table 3.
[0111] Table 3
[0112]
[0113] In the knowledge graph conversion stage, the visualization expression and reasoning calculation of rules are achieved through the mapping of predicate logic to graph structure:
[0114] For simple rules This means that all entities x that satisfy condition P(x) must be associated with Q(x). In this implementation, the rule "Reservoir porosity > 8% → sweet spot" is expressed as: [Reservoir A porosity > 8%] → [sweet spot] (w = 0.88).
[0115] For compound logic conjunction (∧) and disjunction (∨) processing: For conjunction rules , a dual-path joint reasoning mechanism is adopted: P(x)-[Rule1]→R(x) and Q(x)-[Rule1]→R(x), and the weights are jointly controlled by logical relationships. In this implementation scheme, in the rule "(Mercury injection median pore throat radius>1 )∧(CT pore coordination number>3)→ pore structure type = Type I” is expressed as: [mercury intrusion radius>1 ]→(AND_Gate)→[Pore structure type I](w=0.95)、[CT coordination number>3]→(AND_Gate)→[Pore structure type I](w=0.95);For the disjunctive rule , constructing independent reasoning paths with individually assigned weights. Activation occurs when any condition is met. In this implementation plan, the rule "(crack density > 2 cracks / m) ∨ (crack length > 10m) → crack development level = medium-high" is represented as: [crack density > 2 cracks / m] → [crack development level = medium-high] (w = 0.78), [crack length > 10m] → [crack development level = medium-high] (w = 0.78). This implementation plan is otherwise identical to Specific Implementation Plan 5.
[0116] Specific implementation plan seven: Figure 3 As shown, the knowledge reasoning module in step S200 includes a rule-enhanced graph convolution layer and a rule-constrained prediction layer. The rule-enhanced graph convolution layer is used in the encoding stage, and the matching degree between the modeling node and the geological rule is dynamically modeled through a differentiable rule encoder, and the rule weight gating in the multi-head attention mechanism is used to realize the directional screening and aggregation of neighborhood node information; in the decoding stage, the rule-constrained prediction layer introduces a two-stage correction mechanism, first filtering out illegal predictions through hard constraints, and then combining soft rules to construct residual correction terms, and embedding the physical equations in a learnable form into the fully connected layer, finally realizing the full-link closed-loop constraint of geological knowledge from feature space to decision space.
[0117] The functional implementation process of the knowledge reasoning module is as follows:
[0118] Step 1: Construct a rule-enhanced graph convolution layer, design a multi-layer graph convolution structure, define the convolution kernel size, dilation factor, and number of layers to capture the multi-scale topological relationships in the reservoir knowledge graph; embed a rule attention module in the convolution layer, and dynamically adjust the feature weights based on the rule matching degree.
[0119] Step 2: Train the tight sandstone reservoir knowledge reasoning module. Entity relationships in the tight sandstone reservoir knowledge graph are loaded as rule constraints. The processed reservoir data is fed into the knowledge reasoning module for a training loop. The graph convolution layer extracts reservoir topological features, and the rule attention module optimizes the feature graph, enhancing high-contribution features relevant to the current rule. The optimized features are then concatenated with the original features to form a joint semantic representation.
[0120] Step 3: Perform model optimization and dynamically adjust the weight coefficient of the rule penalty term based on the rule violation rate of the test set prediction results; according to the above criteria, optimize the network weight through the gradient descent algorithm, and give priority to retaining the parameter combination that conforms to the geological rules.
[0121] Step 4: Finally, output the reservoir classification label, simultaneously predict continuous parameters such as porosity and permeability, and modify the original rules to complete the inference of tight sandstone reservoirs. The rest of this implementation plan is the same as Specific Implementation Plan 6.
[0122] Specific implementation scheme eight: The functional implementation process of the rule-enhanced graph convolution layer is as follows:
[0123] In the reservoir knowledge graph G = (V, E, R), node feature updates use a reservoir rule-aware graph convolution operation, calculated as follows:
[0124]
[0125] in, For the Nodes in the layer The feature representation of is the activation function, R rule Represents a set of rule-related relationships, which contains various relationships used for reservoir rule-aware graph convolution , Indicates that in the relationship Next, node The set of neighbor nodes of Indicates that in the relationship Next node right The regularized attention weight is determined by the reservoir rule matching degree Determined together with the topology, Indicates in Layers and relationships The transformation matrix under Indicates the Nodes in the layer The feature representation of Indicates the The rule-related transformation matrix in the layer, Represents the feature representation related to rule triggering, The calculation formula is as follows:
[0126]
[0127] in represents the exponential function, 、 is the rule matching function, represents the weight vector in the attention mechanism, Represents the node i and the representation of node j To splice, Represents the node i and the representation of node k To splice, represents the set of neighbor nodes of node i, Represents an activation function used to introduce nonlinear transformation.
[0128] Rule matching is a measure of the degree to which a node complies with a specific rule. This includes: under the "fracture validity rule", some fractures contribute more to seepage, while other fractures may be ineffective. By defining the rule matching function , adjust the importance of neighboring nodes so that nodes that comply with the rules obtain higher weights during information propagation, while the influence of nodes that do not comply with the rules is weakened, ensuring that the learned representation conforms to the reservoir domain knowledge. The calculation formula is as follows:
[0129]
[0130] in, represents the atomic formula input to the rule encoder, It is a rule encoder whose goal is to map atomic formulas into computable conditions in the vector space. The calculation formula is as follows:
[0131]
[0132] in, and represents the weight matrix, and represents the bias term.
[0133] Rule Encoder The network structure is a shared MLP, such as Figure 4 As shown, the nodes participating in the rule judgment , concatenate its hidden state and relation embedding. Assuming that the feature dimension of each node is d and the dimension of relation embedding is k, the total dimension of the input is 2d + k. Output scalar value, and then use Sigmoid to get the rule matching degree. , so that its output value range [0,1] reflects the degree of compliance with the rules.
[0134] During the tight reservoir evaluation reasoning process, a collaborative reasoning framework is constructed by integrating the well logging data, image features, geological text descriptions, and domain knowledge rules required for intelligent tight reservoir evaluation. Based on the reservoir evaluation knowledge system, semantic association rules between "reservoir characteristics-reservoir attributes" entity nodes are defined, a weighted adjacency matrix is constructed, and a weighted graph convolutional network (WGCN) is used to process the graph structure data of reservoir samples to generate entity embedding vectors. During the GCN iteration process, the semantic propagation weights between "reservoir characteristics-reservoir attributes" type entity nodes are dynamically adjusted to address the oversmoothing problem caused by the heterogeneity of geological relationships in traditional GCNs. Entity relationship embedding vectors with geological significance are generated, enabling better training and optimization of deep networks in tight reservoir classification tasks. The rest of this implementation plan is the same as Specific Implementation Plan 7.
[0135] Specific implementation scheme nine: The functional implementation process of the rule-constrained prediction layer is as follows:
[0136] During the decoding process, for the tight sandstone reservoir evaluation results and parameter prediction values initially generated by the Conv-TransE layer, the results are first screened through constraints constructed based on logical rules and expert knowledge to remove predictions that violate hard constraints. Then, the results that meet the hard constraints but have deviations are weighted and adjusted to make them more consistent with the reservoir rules. For known empirical relationships, the prediction results are adjusted using weighting factors to make them more consistent with domain knowledge. Finally, in the training phase, the rule constraints are explicitly introduced into the loss function so that the model can learn prediction patterns that are consistent with domain knowledge. For prediction results that violate hard rule constraints, a large penalty term is used to significantly increase their losses to avoid generating illegal predictions. The error between the calculated prediction value and the rule expectation value is minimized to improve the model's adaptability to soft rules. The calculation formula is:
[0137]
[0138]
[0139] in, is the soft rule error term, is the model’s predicted output for the sample, is the reference output associated with the soft rule, is the overall loss function, is the basic prediction loss of the model, represents a set of hard rules, represents a set of soft rules, Penalty for violating hard rules. is the soft rule error term, and is a weight parameter used to control the influence of the rule constraint. The rest of this implementation is the same as the specific implementation plan eight.
[0140] Specific implementation plan 10: Figure 5 As shown, the collaborative reasoning of multiple tasks in step S300 includes the following steps:
[0141] Reservoir classification task:
[0142] Based on traditional reservoir classification, a static parameter classification system was constructed based on the experience of geological experts and the parameter association classification standard system in the knowledge graph: Class I reservoir: porosity >8%, permeability >0.1mD; Class II reservoir: porosity 5%-8%, permeability 0.05-0.1mD; Class III reservoir: porosity <5%, permeability <0.05mD. The reservoir quality grade is output based on the input reservoir characteristics, and a rule-constrained cross-entropy loss is used to ensure that the classification results are consistent with expert experience. The calculation formula is:
[0143]
[0144] in, is the embedding vector of the reservoir evaluation rule in the knowledge graph;
[0145] The loss function integrates data-driven loss and knowledge-constrained loss, and the calculation formula is:
[0146]
[0147] Among them, when the rules are met, the indicator function Otherwise, it is recorded as 0.
[0148] Parameter prediction task:
[0149] Simultaneously predict continuous physical parameters including porosity and permeability, and reduce the risk of overfitting by sharing feature layers among multiple tasks. The parameter-specific projection is calculated as:
[0150]
[0151] in, The formula dynamically generated through the relationship path is:
[0152]
[0153] Representation parameters All associated paths in the knowledge graph;
[0154] Physical equation constraint: Adjust the K value based on the Kozeny-Carman equation. The calculation formula is:
[0155]
[0156] In the reservoir classification task, a fully connected layer maps joint features with geological semantics to a reservoir classification probability space. This is then hard-verified in a rule-constrained layer, ultimately outputting a reservoir type label. In the parameter prediction task, after regressing initial physical property values through a residual network, the Kozeny-Carman equation physical constraints are introduced for adaptive correction, simultaneously verifying that the predicted values conform to the empirical porosity-permeability relationship. A dynamic weighted loss function, comprising classification cross entropy, regression mean square error, and a rule penalty term, is then designed to achieve multi-objective collaborative optimization.
[0157] In a multi-task reasoning framework, reservoir classification and parameter prediction achieve a two-way reinforcement effect through a knowledge-driven collaborative mechanism. The reservoir classification task provides geological semantic context for parameter prediction through rule-constrained hierarchical classification, constraining the physically reasonable range of physical property parameters. The parameter prediction task, on the other hand, uses high-precision regression results, including continuous values of porosity and permeability, to feed back into classification decisions. Differential equations are used to reveal theoretical correlations between parameters and correct classification ambiguities caused by data noise. Dynamic weight allocation enables gradient collaborative optimization, effectively resolving the semantic conflict problem present in traditional multi-task learning.
[0158] During the evaluation of tight sandstone reservoirs, a knowledge-driven, bidirectional enhancement mechanism forms a closed-loop optimization system through the forward constraints of "reservoir classification → parameter prediction" and the reverse correction of "parameter prediction → reservoir classification." Tight reservoir evaluation criteria derived from geological laws and expert experience are converted into logical rules or physical equations to constrain the output range of the parameter prediction model. Data-driven models are used to mine implicit associations, dynamically adjusting the strength of knowledge constraints and local deviations in expert knowledge based on data quality and knowledge confidence. The remainder of this implementation plan is identical to Specific Implementation Plan Nine.
[0159] The knowledge graph-based tight sandstone reservoir evaluation method (algorithm) proposed in the present invention is the underlying technical core of the present invention, and various products can be derived based on the algorithm.
[0160] Based on the method proposed in the present invention, a knowledge graph-based tight sandstone reservoir evaluation system is developed using a programming language. The system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned knowledge graph-based tight sandstone reservoir evaluation method during operation.
[0161] The overall framework of the tight sandstone reservoir evaluation model based on the knowledge graph is as follows: Figure 1 Shown, including:
[0162] Data layer: used for various data collection and storage, providing the required structured and unstructured data for building the knowledge graph in the field of tight sandstone reservoirs, and performing data preprocessing operations on the collected multi-source heterogeneous data.
[0163] Service Layer: This layer, the model's middle layer and the system's core processing module, is responsible for extracting entity relationships within the tight sandstone reservoir domain and implementing a rule-enhanced, structure-aware network-based reservoir quality inference method. In the tight reservoir entity relationship joint extraction model, entity relationship extraction is transformed into a sequence labeling task, focusing on identifying and extracting nested entities and complex overlapping relationships within the tight reservoir text, ultimately achieving a globally optimal label sequence. The extracted triples are stored in a Neo4j graph database system, and a rule base is constructed based on domain expert experience. Combined with the rule-enhanced, structure-aware network-based reservoir quality inference method, tight sandstone reservoir evaluation results are obtained.
[0164] Application layer: The application layer is the top layer of the model, including the knowledge graph management module, reservoir evaluation module and system function configuration.
[0165] The developed system (software) computer program is stored on a computer-readable storage medium. When called by a processor, the computer program is configured to implement the steps of the aforementioned knowledge graph-based tight sandstone reservoir evaluation method. This materializes the present invention on a carrier, becoming a computer program product.
[0166] Various implementations of the systems and techniques described herein can be realized in digital electronic circuitry, integrated circuitry, dedicated ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system comprising at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.
[0167] The computer programs (also referred to as programs, software, software applications, or code) herein comprise machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., a magnetic disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.
[0168] Although the present invention is disclosed as above, the scope of protection disclosed by the present invention is not limited thereto. Those skilled in the art of the present invention may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the scope of protection of the present invention.
Claims
1. A tight sandstone reservoir evaluation method based on knowledge graph, characterized by: The steps include: Step S100: Collect tight sandstone reservoir evaluation data, pre-process the tight sandstone reservoir evaluation data, build an ontology knowledge base, and build a tight sandstone reservoir domain ontology layer based on the knowledge and its associations; Step S200: Construct a tight sandstone reservoir evaluation model, the evaluation model comprising a knowledge extraction module and a knowledge reasoning module. The knowledge extraction module is used to extract triples from entity relationships annotated in the tight sandstone reservoir sequence and construct a tight sandstone reservoir knowledge graph with entity relationships combined. The knowledge reasoning module is based on a rule-embedded structure-aware convolutional network and is used to perform multi-task collaborative reasoning. Step S300: Perform multi-task collaborative reasoning of tight sandstone reservoirs based on the tight sandstone reservoir evaluation model.
2. The method for evaluating tight sandstone reservoirs based on knowledge graph according to claim 1, characterized in that: The tight sandstone reservoir evaluation data in step S100 includes text, numerical values, and image data.
3. The tight sandstone reservoir evaluation method based on knowledge graph according to claim 2 is characterized in that: The construction of the main layer in the tight sandstone reservoir field includes: Step S110: Clarify the scope of the tight reservoir field: The core information sources of the reservoir field include: a tight reservoir basic information database, a terminology set and academic literature related to tight reservoir evaluation industry standards, and a systematic compilation and summary of the practical experience of reservoir evaluation experts and their professional teams; Step S120, Entity Concept and Attribute Definition: There are two methods for concept extraction. One is to extract concepts from a macro to micro level and in a step-by-step manner based on a pre-defined business process architecture. The other is to conduct in-depth mining of information from resources such as document materials and basic information databases through collection and retrieval to extract the required concepts to be defined. Step S130, relationship definition: Use the attribute ontology architecture and clustering attribute method to describe the characteristics of the classification relationship between attributes, and clarify the relationship between each category based on the extracted concepts and the results obtained from the demand analysis.
4. The method for evaluating tight sandstone reservoirs based on knowledge graph according to claim 3, characterized in that: The knowledge extraction module in step S200 includes: an embedding module, an encoding module and a decoding module; The embedding module is constructed with a BERT module for converting sentences into dynamic semantic information at the character level and word level; The encoding module is constructed with a BiLSTM module and a CNN module. The BiLSTM module is used to extract context information of dense reservoir sequence data. The CNN module is used to extract detailed features of the data using multiple convolution kernels of different sizes, and select the maximum value of the feature map through a pooling layer to further aggregate features. The decoding layer is constructed with a CRF module for obtaining the global optimal label sequence of the entity; Finally, triple extraction of entity relationships is achieved, and a knowledge graph of tight sandstone reservoirs with entity relationships is constructed.
5. The method for evaluating tight sandstone reservoirs based on knowledge graph according to claim 4, characterized in that: The CNN module uses 1×3 convolution to capture the entity features of single geological terms and reservoir texts, and uses 5×5 convolution to capture geological term combinations and their contextual relationships.
6. The method for evaluating tight sandstone reservoirs based on knowledge graph according to claim 5, characterized in that: The knowledge reasoning module in step S200 includes a rule-enhanced graph convolution layer and a rule-constrained prediction layer. The rule-enhanced graph convolution layer is used in the encoding stage to dynamically model the matching degree between the node and the geological rule through a differentiable rule encoder, and utilizes the rule weight gating in the multi-head attention mechanism to achieve targeted screening and aggregation of neighborhood node information. In the decoding stage, the rule-constrained prediction layer introduces a two-stage correction mechanism. First, illegal predictions are filtered through hard constraints, and then residual correction terms are constructed in combination with soft rules. The physical equations are embedded in the fully connected layer in a learnable form, and finally the full-link closed-loop constraint of geological knowledge from feature space to decision space is realized.
7. The method for evaluating tight sandstone reservoirs based on knowledge graph according to claim 6, characterized in that: The functional implementation process of the rule-enhanced graph convolution layer is as follows: In the reservoir knowledge graph G = (V, E, R), node feature updates use a reservoir rule-aware graph convolution operation, and the calculation formula is: ; in, For the Nodes in the layer The feature representation of is the activation function, R rule Represents a set of relations related to the rules, Indicates that in the relationship Next, node The set of neighbor nodes of Indicates that in the relationship Next node right Regularized attention weights, Indicates in Layers and relationships The transformation matrix under Indicates the Nodes in the layer The feature representation of Indicates the The rule-related transformation matrix in the layer, Represents the feature representation related to rule triggering, The calculation formula is as follows: ; in, represents the exponential function, 、 is the rule matching function, represents the weight vector in the attention mechanism, Represents the node i and the representation of node j To splice, Represents the node i and the representation of node k To splice, represents the set of neighbor nodes of node i, Represents an activation function used to introduce nonlinear transformation; By defining the rule matching function Adjust the importance of neighboring nodes. The calculation formula is: ; in, represents the atomic formula input to the rule encoder, is a rule encoder that maps atomic formulas into computable conditions in a vector space: ; in, and represents the weight matrix, and represents the bias term.
8. The method for evaluating tight sandstone reservoirs based on knowledge graph according to claim 7, characterized in that: The loss function of the rule-constrained prediction layer is a dynamic weighted loss function based on classification cross entropy, regression mean square error and rule constraints: ; ; in, is the soft rule error term, is the model’s predicted output for the sample, is the reference output associated with the soft rule, is the rule-constrained prediction layer loss function, is the basic prediction loss of the model, represents a set of hard rules, represents a set of soft rules, Penalty for violating hard rules. is the soft rule error term, and is a weight parameter used to control the influence of the rule constraint.
9. A tight sandstone reservoir evaluation system based on knowledge graph, characterized by: The system has a program module corresponding to the steps of the method described in any one of claims 1 to 8 above, and executes the steps in the above-mentioned knowledge graph-based tight sandstone reservoir evaluation method when running.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps in the tight sandstone reservoir evaluation method based on the knowledge graph according to any one of claims 1 to 8 when called by a processor.
Citation Information
Patent Citations
Fracture-vuggy carbonate reservoir reasoning method and device based on knowledge graph
CN119721230A
Knowledge graph-based semantic association and logic rule reasoning method
CN120011368A
Cited By
Engineering cost intelligent evaluation method based on graph neural network
CN121258187A
Carbonate reservoir geological sweet spot prediction method, device, equipment and medium
CN121703922A
