Power transmission and transformation project ecological risk identification method based on multi-granularity semantic understanding

Abstract

The invention discloses a power transmission and transformation project ecological risk identification method based on multi-granularity semantic understanding, and relates to the crossing field of power transmission and transformation projects and ecological protection. Engineering ecological risks are recognized through macroscopic, mesoscopic and microscopic multi-granularity semantics, and macroscopic risk rules are extracted through analysis of laws and regulations and environmental assessment texts; remote sensing images and unmanned aerial vehicle videos are analyzed through an improved RT-DETRv2 model, and mesoscopic risks such as habitat conflicts are recognized; microcosmic layer real-time data such as noise and soil humidity are collected through a sensor; and constructing a dynamic knowledge graph based on the three-layer data, and updating according to a period to realize multi-dimensional node association. Constructing a risk identification large model through Phi-4 layered fine tuning, and generating a risk judgment result; the atlas consistency is verified through a verification agent, and a result is optimized in combination with a confidence coefficient threshold value formula. Efficient, multi-dimensional and accurate power transmission and transformation project ecological risk identification is realized, and reliable technical support is provided for power transmission and transformation project full life cycle ecological risk management.

Description

An Ecological Risk Identification Method for Power Transmission and Transformation Projects Based on Multi-Granularity Semantic Understanding Technical Field

[0001] This invention relates to the intersection of power transmission and transformation engineering and ecological protection, specifically to a method for identifying ecological risks in power transmission and transformation engineering based on multi-granularity semantic understanding. Background Technology

[0002] With the accelerated construction of my country's ultra-high-voltage (UHV) power transmission network, the scale and coverage of these projects continue to expand, placing higher demands on the accuracy, real-time nature, and intelligence of ecological risk identification. As a core component of the national energy infrastructure, power transmission and transformation projects cover diverse terrains, including plains, mountains, and coastal areas, inevitably interacting with the ecological environment along their routes. This can potentially trigger various ecological risks such as soil erosion, vegetation destruction, water pollution, and biodiversity disturbance. Currently, the industry has conducted relevant research on the identification and assessment technologies for ecological risks in power transmission and transformation projects, and has published several patented achievements. However, existing technologies still have significant limitations and are insufficient to meet the risk management needs of complex engineering scenarios.

[0003] Chinese patent (publication number CN108229543A) discloses a method for ecological risk assessment of power transmission and transformation projects. This method collects remote sensing image data of the project area and combines it with a GIS system to perform overlay analysis of factors such as terrain and vegetation coverage, thereby classifying risk levels. However, this method relies on periodic remote sensing imagery, which is usually updated every 1-3 months, resulting in poor dynamic adaptability and difficulty in matching risk changes throughout the entire project lifecycle.

[0004] Chinese patent (publication number CN112115183B) discloses an ecological environment monitoring system for power transmission and transformation projects. This system achieves real-time early warning by deploying single-type sensors such as noise and dust sensors at the project site, combined with preset threshold rules. However, this method only monitors and focuses on a single risk factor, making it unable to identify complex risks with coupled characteristics, and easily leading to missed risk assessments.

[0005] Chinese patent (publication number CN113658724A) discloses an ecological risk classification model for power transmission and transformation projects. This model extracts textual features from the project's environmental impact assessment report and historical risk event data, then uses traditional machine learning algorithms such as Support Vector Machine (SVM) and Random Forest to construct the classification model. However, this method is limited to shallow feature extraction and a single model architecture, failing to dynamically adjust according to the type of project area. This results in poor model generalization ability and scenario adaptability, necessitating manual intervention.

[0006] In recent years, the rapid development of artificial intelligence technology has provided a new path to solve the above problems. Breakthroughs in semantic understanding by large language models and the advantages of graph neural networks in relation modeling have laid the technical foundation for building an intelligent ecological risk identification system with deep analysis and dynamic recognition capabilities. Summary of the Invention

[0007] To address the aforementioned technical issues, this application discloses a method for identifying ecological risks in power transmission and transformation projects based on multi-granularity semantic understanding, specifically including:

[0008] By analyzing relevant data on power transmission and transformation projects, macro-risk categories, constraints, and evaluation criteria are extracted to form top-level rule nodes in the knowledge graph.

[0009] An improved RT-DETRv2 model is trained based on the ECA attention mechanism. This model is used to periodically process satellite remote sensing images and drone aerial videos along the power transmission and transformation project to identify meso-level risk factors such as land cover changes, water disturbances, and vegetation damage along the project route, forming geographic entities and event nodes in the knowledge graph.

[0010] Real-time collection of micro-environmental data at the engineering site, including noise, water quality, soil moisture and dust concentration, to form real-time monitoring nodes in the knowledge graph;

[0011] Based on the top-level rule nodes, geographic entities and event nodes, and real-time monitoring nodes, a multi-granular semantic understanding module containing macro, meso, and micro semantics is constructed to form an ecological risk knowledge graph.

[0012] The Phi-4 model learns terminology and knowledge in the field of power transmission and transformation ecological risk through a masked language modeling task, and is fine-tuned in layers to obtain a large-scale risk identification model. Based on the ecological risk knowledge graph, the large-scale risk identification model is used to determine the ecological risks in power transmission and transformation projects.

[0013] Based on the topological structure and semantic information of the ecological risk knowledge graph, a dual-review intelligent agent is constructed. The first review is risk self-checking and improvement. Through graph content query and data confidence test, when data is missing or the data confidence is low in the knowledge graph, the knowledge graph is iteratively optimized.

[0014] The second level of review is causal deduction and decision optimization. Multiple independent risk identification models are used to judge the same risk structure. When the judgment results are inconsistent, the judgment results of each model and the relevant evidence chain are input into the big language model, which performs comprehensive analysis and outputs the final ruling result, and automatically triggers early warning for high-risk events.

[0015] Preferably, the improved RT-DETRv2 model specifically involves: enhancing feature extraction at the Backbone layer using an ECA attention mechanism on the basic architecture of the RT-DETRv2 model, as shown in the formula:

[0016]

[0017] in, For the first Channel weight, For the output features of Stage 4 or Stage 5 in the Backbone of RT-DETRv2, For 1D convolution kernel size, The Sigmoid activation function is used to enhance the distinction between habitat texture features and construction machinery structural features;

[0018] The fusion process is optimized in the Neck layer using a BiFPN bidirectional feature pyramid mechanism, with the following formula:

[0019]

[0020] in, For the first Path weight, The total number of paths, This is the scaling factor;

[0021] The loss functions in the Head layer include CIoU regression loss and Focal classification loss;

[0022] After training, the improved RT-DETRv2 model is used to identify meso-level engineering risk factors and output the risk factor identification results. At regular intervals, the identification results are converted into geographic entities and event nodes in the knowledge graph to form geographic entities and event nodes in the knowledge graph.

[0023] Preferably, the loss function is specifically: CIoU regression loss ensures that the predicted bounding box accurately covers the target region, providing an accurate regional basis for classification, as shown in the formula:

[0024]

[0025] in, For intersection, union, and comparison, For prediction boxes With real frame The square of the Euclidean distance from the center point, The length of the diagonal of the minimum bounding rectangle that encloses the predicted bounding box and the ground truth bounding box. For balance coefficient, A parameter used to measure the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box;

[0026] Focal classification loss ensures accurate classification of the target category within the precisely selected region. The formula is as follows:

[0027]

[0028] in, For category weights, Predict the probability for the target category. For focusing parameters.

[0029] Preferably, the construction of the multi-granularity semantic understanding module specifically involves: providing high-quality multi-dimensional data for the construction of the knowledge graph through the coordinated linkage of three levels of semantics; macro-level rules guiding the key directions of meso-level identification and micro-level monitoring; and meso-level and micro-level data feeding back into the dynamic optimization of macro-level rules.

[0030] Three types of node data are processed using entity alignment and relationship fusion algorithms. Entity alignment is achieved by calculating node name similarity and attribute similarity, using the following formula:

[0031]

[0032] in, To assess overall similarity, , These are the name similarity weight and the attribute similarity weight, respectively. For name similarity, For attribute similarity; when If the nodes are identified as the same entity, they are merged. The preset judgment threshold;

[0033] The relationship fusion algorithm uses a rule base in the substation ecosystem to fill in missing relationships between nodes, ensuring that the relationship coverage of the knowledge graph is not lower than the preset minimum coverage.

[0034] Preferably, the process of hierarchical fine-tuning of the Phi-4 model specifically includes:

[0035] Collect a corpus of ecological risk in the field of power transmission and transformation, which includes texts of ecological protection regulations, historical risk event reports, and environmental impact assessment technical documents;

[0036] Input the corpus of the field of power transmission and transformation ecological risk, and fine-tune the Phi-4 small parameter model through the mask language model task with preset training parameters to build a preliminary model;

[0037] Based on the labeled risk data and the sample set of judgment results, the preliminary model is optimized with the cross-entropy loss function. Then, through the incremental pre-training strategy, the underlying weights are frozen on the domain corpus, and only the top layer is fine-tuned to build a large risk identification model.

[0038] Preferably, the process of the risk identification big data model determining the ecological risks in power transmission and transformation projects specifically involves: extracting the node feature matrix of the target project from the ecological risk knowledge graph using a dual-input mode that combines graph features with domain knowledge. Adjacency matrix ,in For the number of nodes, For feature dimensions;

[0039] The node embedding vectors are generated using a Graph Convolutional Neural Network (GCN), and the calculation formula is as follows:

[0040]

[0041] in, for Layer node characteristics, ( (the identity matrix) for The degree matrix, For the first The weight matrix of the layer, It is the ReLU activation function;

[0042] The DPO algorithm fuses node embedding vectors with domain knowledge learned by a large model to output risk levels and corresponding judgment criteria. The formula is as follows:

[0043]

[0044] in For temperature parameters, For engineering command input, High-quality risk recommendations are manually labeled. Recommendations based on low quality and risk For the Phi-4 model, the input... Generate output The reward points.

[0045] Preferably, the process of fusing the node embedding vector with the domain knowledge learned by the large model is as follows:

[0046] By standardizing to eliminate dimensional differences in node embedding vectors, the vectors are made consistent with the parameter distribution of the domain knowledge in the large model. The formula is as follows:

[0047]

[0048] in, After eliminating dimensional differences, the first Layer node feature vectors for The standard deviation matrix;

[0049] Extract the association weights of risk factors and their impact on consequences learned during the pre-training stage of the risk identification model and store them in the multi-head attention matrix of the Transformer layer; extract the mapping weights between the indicator exceeding the standard and the risk level and store them in the fully connected weight matrix of the output layer.

[0050] Integration is achieved through a three-tiered, progressive knowledge injection process:

[0051] The first layer, based on the domain terminology association parameters, combines the annotation results of power transmission and transformation engineering risk samples during the GCN backpropagation process to make the transformation direction of the node embedding vector conform to the domain knowledge logic.

[0052] The second layer highlights key risk factors in domain knowledge through attention weight allocation, transforming the risk judgment threshold parameters learned by the large model into bias terms for attention calculation, as shown in the formula:

[0053]

[0054] in, , , These are the query vector, key vector, and value vector generated based on node embedding vectors, respectively. , They are query vectors Key vector The bias term, for Feature dimensions;

[0055] Through this layer of fusion, the feature proportion of key risk factors in the global feature vector is consistent with the risk priority in the domain knowledge;

[0056] In the third layer, when the global feature vector is mapped to the risk level probability distribution in the fully connected layer, the fusion is completed by binding the output layer weights with the judgment logic of the large model.

[0057] Preferably, the active query verification of the review agent specifically employs a hybrid search strategy combining depth-first search (DFS) and breadth-first search (BFS): DFS is used to traverse the association between top-level rule nodes and geographic entity nodes; BFS is used to traverse the matching relationship between real-time monitoring nodes and geographic entity nodes to ensure the integrity of the logical chain; the frequency of query verification is synchronized with the knowledge graph update frequency, and each query generates a verification report containing passed items, items to be optimized, and abnormal items.

[0058] Preferably, the data confidence level is specifically:

[0059] The reliability of data acquisition is calculated based on a multi-factor evaluation of sensor status and data quality, using the following formula:

[0060]

[0061] in To ensure the credibility of data collection, Sensor calibration status. For data integrity, The degree of environmental disturbance;

[0062] The confidence level is identified by mapping confidence levels by model type and fusion calculation, using the following formula:

[0063]

[0064] in, To enhance the credibility of the model, The output confidence score of the computer vision model RT-DETRv2. The output confidence score of the large-scale risk identification model based on Phi-4 fine-tuning;

[0065] The semantic matching reliability assessment based on the domain rule base calculates the rule matching credibility using the following formula:

[0066]

[0067] in, To improve the credibility of rule matching, This represents the number of rules in the domain rule base that are relevant to the current data. For the first The semantic matching degree between the rule and the current data To preset the minimum rule credibility;

[0068] The data confidence score is calculated by considering the confidence scores of data collection, model recognition, and rule matching. The formula is as follows:

[0069]

[0070] in, To assess the overall data confidence level, These are data collection weights, model recognition weights, and rule matching weights, respectively; when When the preset confidence threshold is reached, the model retraining or data acquisition process will be automatically triggered.

[0071] Preferably, the process of iteratively optimizing the knowledge graph specifically involves: setting a preset threshold for missing data. Through iterative optimization using a strategy combining incremental and full updates, when the percentage of missing data is ≤ Incremental updates are used, supplementing only missing nodes and relationships; the percentage of missing data is greater than [a certain percentage]. If the overall confidence level of the data is lower than the preset confidence threshold, a full update is adopted, and the node generation and fusion process of the multi-granularity semantic understanding module is re-executed; after the optimized knowledge graph passes the consistency check and integrity check, it is put into the subsequent risk assessment process.

[0072] Preferably, the second review specifically involves: deployment Each independent risk identification model has a preset consistency threshold. The consistency of the model's judgment results is calculated using the following formula:

[0073]

[0074] in To determine the consistency of the results, Don't be the first , The judgment results of each model, The semantic similarity between the two results is when < If the result is deemed inconsistent, the model's judgment result and the corresponding evidence chain are input into the pre-fine-tuned large language model, and the final ruling result is output, including the unified risk level, key credible evidence, explanation of the disputed points, and an early warning is triggered according to the risk level.

[0075] Compared with the prior art, the technical solution of this application has the following technical effects:

[0076] This invention improves the recognition accuracy of mesoscopic risk factors such as vegetation destruction, water disturbance, and land cover change along power transmission and transformation projects by using an improved RT-DETRv2 model. This solves the problem of low recognition accuracy of existing computer vision models for multi-scale and easily obscured risk factors in power transmission and transformation scenarios.

[0077] This invention constructs an ecological risk knowledge graph through a multi-granularity semantic understanding module, enabling real-time association of macro-level rules, meso-level geographical entities, and micro-level monitoring node data. This overcomes the limitations of existing technologies, such as the inability of static analysis to match dynamic changes in engineering projects and low data update efficiency.

[0078] This invention achieves automatic identification of complex risks by using a large-scale risk identification model combined with graph convolutional neural networks and attention mechanisms, thus overcoming the shortcomings of existing technologies, such as the lack of modeling of the correlation between multiple risk factors and insufficient ability to identify complex risks.

[0079] This invention implements a hierarchical fine-tuning strategy on the Phi-4 model, combined with reinforcement learning and iterative optimization of the review agent, enabling the model to automatically adapt to different power transmission and transformation engineering scenarios such as mountainous areas, urban areas, and coastal areas. Moreover, the accuracy of risk assessment in new scenarios does not require manual annotation, thus solving the problems of poor generalization ability and reliance on a large amount of manual intervention in existing technology models.

[0080] This invention constructs a dual-verification intelligent agent and adopts a hybrid search strategy of DFS and BFS, combined with multi-dimensional confidence, to achieve automatic verification of the consistency between risk identification results and knowledge graph topology. It can also trigger data supplementation or model retraining to ensure the reliability of data and results, avoiding the risks of existing technologies lacking dual verification and being prone to misjudgment due to single model output.

[0081] In summary, this invention, through the collaborative innovation of a multi-granularity semantic understanding module, an ecological risk knowledge graph, a large-scale risk identification model, and a dual review mechanism, has achieved the construction of a comprehensive, real-time, and intelligent ecological risk identification system for power transmission and transformation projects. It comprehensively overcomes the limitations of existing technologies in terms of semantic understanding, dynamic response, complex risk identification, scenario adaptation, and result reliability, providing efficient and accurate technical support for the management and control of ecological risks in power transmission and transformation projects.

[0082] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0083] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0084] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0085] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0086] Figure 1 is a flowchart of the method for identifying ecological risks in power transmission and transformation projects based on multi-granularity semantic understanding;

[0087] Figure 2 shows the overall architecture of the method for identifying ecological risks in power transmission and transformation projects based on multi-granularity semantic understanding;

[0088] Figure 3 shows the neural network architecture of the improved RT-DETRv2 model;

[0089] Figure 4 shows the overall architecture of the hierarchical fine-tuning Phi-4 large language model;

[0090] Figure 5 is a diagram of the overall structure of an experiment conducted along the route of an ultra-high voltage power transmission and transformation project through a nature reserve.

[0091] Figure 6 is a comparison of the identification accuracy of the three methods for meso- and micro-level risk factors in the experiment;

[0092] Figure 7 is a comparison of the risk discovery time and the time to provide recommendations for the three methods in the experiment;

[0093] Figure 8 is a comparison of the confidence levels of the risk assessment results of the three methods in the experiment. Detailed Implementation

[0094] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0095] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0096] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0097] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0098] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0099] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0100] Example 1 describes a method for identifying ecological risks in power transmission and transformation projects based on multi-granularity semantic understanding, as shown in Figures 1 and 2. Specifically, it includes:

[0101] By analyzing relevant data on power transmission and transformation projects, macro-risk categories, constraints, and evaluation criteria are extracted to form top-level rule nodes in the knowledge graph.

[0102] An improved RT-DETRv2 model is trained based on the ECA attention mechanism. This model is used to periodically process satellite remote sensing images and drone aerial videos along the power transmission and transformation project to identify meso-level risk factors such as land cover changes, water disturbances, and vegetation damage along the project route, forming geographic entities and event nodes in the knowledge graph.

[0103] Real-time collection of micro-environmental data at the engineering site, including noise, water quality, soil moisture and dust concentration, to form real-time monitoring nodes in the knowledge graph;

[0104] Based on the top-level rule nodes, geographic entities and event nodes, and real-time monitoring nodes, a multi-granular semantic understanding module containing macro, meso, and micro semantics is constructed to form an ecological risk knowledge graph.

[0105] The Phi-4 model learns terminology and knowledge in the field of power transmission and transformation ecological risk through a masked language modeling task, and is fine-tuned in layers to obtain a large-scale risk identification model. Based on the ecological risk knowledge graph, the large-scale risk identification model is used to determine the ecological risks in power transmission and transformation projects.

[0106] Based on the topological structure and semantic information of the ecological risk knowledge graph, a dual-review intelligent agent is constructed. The first review is risk self-checking and improvement. Through graph content query and data confidence test, when data is missing or the data confidence is low in the knowledge graph, the knowledge graph is iteratively optimized.

[0107] The second level of review is causal deduction and decision optimization. Multiple independent risk identification models are used to judge the same risk structure. When the judgment results are inconsistent, the judgment results of each model and the relevant evidence chain are input into the big language model, which performs comprehensive analysis and outputs the final ruling result, and automatically triggers early warning for high-risk events.

[0108] Furthermore, the extraction of macro-risk categories, constraints, and evaluation criteria specifically includes: parsing unstructured texts such as environmental impact assessment reports, ecological protection regulations, and design specifications for power transmission and transformation projects using a pre-trained large language model, and constructing a macro-semantic layer with multi-granularity semantics.

[0109] Furthermore, the improved RT-DETRv2 model is specifically as follows:

[0110] As shown in Figure 3, the improved RT-DETRv2 model architecture, based on the RT-DETRv2 model, enhances feature extraction in the Backbone layer through an ECA attention mechanism, as shown in the formula:

[0111]

[0112] in, For the first Channel weight, For the output features of Stage 4 or Stage 5 in the Backbone of RT-DETRv2, For 1D convolution kernel size, The Sigmoid activation function is used to enhance the distinction between habitat texture features and construction machinery structural features;

[0113] The fusion process is optimized in the Neck layer using a BiFPN bidirectional feature pyramid mechanism, with the following formula:

[0114]

[0115] in, For the first Path weight, The total number of paths, This is the scaling factor;

[0116] The loss functions in the Head layer include CIoU regression loss and Focal classification loss;

[0117] After training, the improved RT-DETRv2 model is used to identify meso-level engineering risk factors and output the risk factor identification results. At regular intervals, the identification results are converted into geographic entities and event nodes in the knowledge graph to form geographic entities and event nodes in the knowledge graph.

[0118] Furthermore, the loss function is specifically defined as follows: CIoU regression loss ensures that the predicted bounding box accurately covers the target region, providing an accurate regional basis for classification. The formula is:

[0119]

[0120] in, For intersection, union, and comparison, For prediction boxes With real frame The square of the Euclidean distance from the center point, The length of the diagonal of the minimum bounding rectangle that encloses the predicted bounding box and the ground truth bounding box. For balance coefficient, A parameter used to measure the consistency of the aspect ratio between the predicted bounding box and the ground truth bounding box;

[0121] Focal classification loss ensures accurate classification of the target category within the precisely selected region. The formula is as follows:

[0122]

[0123] in, For category weights, Predict the probability for the target category. For focusing parameters.

[0124] Furthermore, real-time environmental data, including noise, water quality, and soil moisture, is collected through an IoT sensor array. After 3σ outlier removal, the data is used to construct micro-semantics in multi-granularity semantics, which is then linked with the quantitative rules extracted from the macro-level to achieve integrated data collection, rule matching, and over-limit early warning.

[0125] Furthermore, the construction of the multi-granularity semantic understanding module specifically includes:

[0126] Through the coordinated linkage of three levels of semantics, high-quality multi-dimensional data is provided for the construction of knowledge graphs. Macro rules guide the key directions of meso-level identification and micro-level monitoring, and meso- and micro-level data feed back into the dynamic optimization of macro rules.

[0127] Three types of node data are processed using entity alignment and relationship fusion algorithms. Entity alignment is achieved by calculating node name similarity and attribute similarity, using the following formula:

[0128]

[0129] in, To assess overall similarity, , These are the name similarity weight and the attribute similarity weight, respectively. For name similarity, For attribute similarity; when If the nodes are identified as the same entity, they are merged. The preset judgment threshold;

[0130] The relationship fusion algorithm uses a rule base in the substation ecosystem to fill in missing relationships between nodes, ensuring that the relationship coverage of the knowledge graph is not lower than the preset minimum coverage.

[0131] Furthermore, the process of hierarchical fine-tuning of the Phi-4 model specifically includes:

[0132] As shown in Figure 4, the model fine-tuning architecture collects a corpus of ecological risk in the field of power transmission and transformation, which includes ecological protection regulations, historical risk event reports, and environmental impact assessment technical documents.

[0133] Input the corpus of the field of power transmission and transformation ecological risk, and fine-tune the Phi-4 small parameter model through the mask language model task with preset training parameters to build a preliminary model;

[0134] Based on the labeled risk data and the sample set of judgment results, the preliminary model is optimized with the cross-entropy loss function. Then, through the incremental pre-training strategy, the underlying weights are frozen on the domain corpus, and only the top layer is fine-tuned to build a large risk identification model.

[0135] Furthermore, the process by which the large-scale risk identification model determines the ecological risks in power transmission and transformation projects is specifically as follows: Through a dual-input mode combining graph features and domain knowledge, the node feature matrix of the target project is extracted from the ecological risk knowledge graph. Adjacency matrix ,in For the number of nodes, For feature dimensions;

[0136] The node embedding vectors are generated using a Graph Convolutional Neural Network (GCN), and the calculation formula is as follows:

[0137]

[0138] in, for Layer node characteristics, ( (the identity matrix) for The degree matrix, For the first The weight matrix of the layer, It is the ReLU activation function;

[0139] The DPO algorithm fuses node embedding vectors with domain knowledge learned by a large model to output risk levels and corresponding judgment criteria. The formula is as follows:

[0140]

[0141] in For temperature parameters, For engineering command input, High-quality risk recommendations are manually labeled. Recommendations based on low quality and risk For the Phi-4 model, the input... Generate output The reward points.

[0142] Furthermore, the node embedding vectors are fused with the domain knowledge learned by the large model, specifically as follows:

[0143] By standardizing to eliminate dimensional differences in node embedding vectors, the vectors are made consistent with the parameter distribution of the domain knowledge in the large model. The formula is as follows:

[0144]

[0145] in, After eliminating dimensional differences, the first Layer node feature vectors for The standard deviation matrix;

[0146] Extract the association weights of risk factors and their impact on consequences learned during the pre-training stage of the risk identification model and store them in the multi-head attention matrix of the Transformer layer; extract the mapping weights between the indicator exceeding the standard and the risk level and store them in the fully connected weight matrix of the output layer.

[0147] Integration is achieved through a three-tiered, progressive knowledge injection process:

[0148] The first layer, based on the domain terminology association parameters, combines the annotation results of power transmission and transformation engineering risk samples during the GCN backpropagation process to make the transformation direction of the node embedding vector conform to the domain knowledge logic.

[0149] The second layer highlights key risk factors in domain knowledge through attention weight allocation, transforming the risk judgment threshold parameters learned by the large model into bias terms for attention calculation, as shown in the formula:

[0150]

[0151] in, , , These are the query vector, key vector, and value vector generated based on node embedding vectors, respectively. , They are query vectors Key vector The bias term, for Feature dimensions;

[0152] Through this layer of fusion, the feature proportion of key risk factors in the global feature vector is consistent with the risk priority in the domain knowledge;

[0153] In the third layer, when the global feature vector is mapped to the risk level probability distribution in the fully connected layer, the fusion is completed by binding the output layer weights with the judgment logic of the large model.

[0154] Furthermore, the active query verification of the review agent is specifically a hybrid search strategy combining depth-first search (DFS) and breadth-first search (BFS): DFS is used to traverse the association between top-level rule nodes and geographic entity nodes; BFS is used to traverse the matching relationship between real-time monitoring nodes and geographic entity nodes to ensure the integrity of the logical chain; the frequency of query verification is synchronized with the knowledge graph update frequency, and each query generates a verification report containing passed items, items to be optimized, and abnormal items.

[0155] Furthermore, according to the method of claim 1, the data confidence level is specifically:

[0156] The reliability of data acquisition is calculated based on a multi-factor evaluation of sensor status and data quality, using the following formula:

[0157]

[0158] in To ensure the credibility of data collection, Sensor calibration status. For data integrity, The degree of environmental disturbance;

[0159] The confidence level is identified by mapping confidence levels by model type and fusion calculation, using the following formula:

[0160]

[0161] in, To enhance the credibility of the model, The output confidence score of the computer vision model RT-DETRv2. The output confidence score of the large-scale risk identification model based on Phi-4 fine-tuning;

[0162] The semantic matching reliability assessment based on the domain rule base calculates the rule matching credibility using the following formula:

[0163]

[0164] in, To improve the credibility of rule matching, This represents the number of rules in the domain rule base that are relevant to the current data. For the first The semantic matching degree between the rule and the current data To preset the minimum rule credibility;

[0165] The data confidence score is calculated by considering the confidence scores of data collection, model recognition, and rule matching. The formula is as follows:

[0166]

[0167] in, To assess the overall data confidence level, These are data collection weights, model recognition weights, and rule matching weights, respectively; when When the preset confidence threshold is reached, the model retraining or data acquisition process will be automatically triggered.

[0168] Furthermore, the iterative optimization of the knowledge graph specifically involves setting a preset threshold for missing data. Through iterative optimization using a strategy combining incremental and full updates, when the percentage of missing data is ≤ Incremental updates are used, supplementing only missing nodes and relationships; the percentage of missing data is greater than [a certain percentage]. If the overall confidence level of the data is lower than the preset confidence threshold, a full update is adopted, and the node generation and fusion process of the multi-granularity semantic understanding module is re-executed; after the optimized knowledge graph passes the consistency check and integrity check, it is put into the subsequent risk assessment process.

[0169] Furthermore, the second review specifically involves: deployment Each independent risk identification model has a preset consistency threshold. The consistency of the model's judgment results is calculated using the following formula:

[0170]

[0171] in To determine the consistency of the results, Don't be the first , The judgment results of each model, The semantic similarity between the two results is when < If the result is deemed inconsistent, the model's judgment result and the corresponding evidence chain are input into the pre-fine-tuned large language model, and the final ruling result is output, including the unified risk level, key credible evidence, explanation of the disputed points, and an early warning is triggered according to the risk level.

[0172] This implementation details a method for identifying ecological risks in power transmission and transformation projects. It provides technical support for ecological risk management throughout the entire lifecycle of the project through a multi-granularity semantic understanding module, an ecological risk knowledge graph, a large-scale risk identification model, and a dual-review mechanism. The multi-granularity module processes multi-source data at the macro, meso, and micro levels, forming three types of nodes. The graph integrates these three types of nodes and their relationships. The large-scale model is fine-tuned hierarchically using Phi-4 to generate judgment results. The graph is verified by a dual-review intelligent agent, and the model is optimized based on thresholds to achieve targeted and precise risk control for power transmission and transformation projects.

[0173] Example 2 details an experiment using this method in a nature reserve along the route of an ultra-high voltage power transmission and transformation project. The experiment involved the erection of a 12.8-kilometer line and the construction of three substations. Key considerations included controlling the risk of disturbance to migratory bird habitats during construction. Specifically:

[0174] As shown in Figure 5, through three-level semantic extraction, macro-semantic layer information in policy documents such as the "Measures for the Protection and Management of Migratory Birds in the Yellow Sea Wetlands" is identified, and 18 quantitative constraint rules are automatically extracted.

[0175] Using a DJI Matrice 350RTK with an aerial resolution of 0.05 meters per pixel and a flight altitude of 100 meters, images were captured every 2 hours. Targets identified included migratory bird habitats (reed wetlands and shallow waters), temporary sheds, and construction machinery. An improved RT-DETRv2 model was deployed, using an ECA attention module to enhance the distinction between wetland vegetation textures and the steel structure features of the sheds. A BiFPN layer fused multi-scale features, ultimately outputting the meso-level semantics of the temporary shed's location coordinates, habitat range, and the distance between the two.

[0176] Six AWA5688 noise sensors were deployed at 300-meter intervals around the migratory bird habitats along the project route. The measurement range was 30-130 dB(A) with an accuracy of ±1 dB(A). The sampling frequency was once per minute. The monitoring period during the critical nighttime resting period for migratory birds was set from 18:00 to 6:00 the next day to collect nighttime noise indicators at the micro-semantic layer.

[0177] An ecological risk knowledge graph is constructed by integrating nodes, and a complete risk chain is generated based on the node relationships and domain rule base of the ecological risk knowledge graph.

[0178] By inputting the characteristics of relevant nodes in the risk chain into the Phi-4 fine-tuned risk identification model, the model outputs suggestions for the location of temporary sheds for rectification and an optimized nighttime construction schedule, with a confidence level of 0.92.

[0179] We compared two existing typical methods, GIS remote sensing static assessment and SVM machine learning, and conducted monitoring for 6 months. A total of 23 significant risks were observed, and the results are as follows:

[0180] The indicators and methods used are as follows: GIS remote sensing static assessment method, SVM machine learning method; number of macro-level rules extracted: 18 items, 0 items, 0 items; accuracy in identifying meso-level conflicts: 99.6%, 67.8%, 93.2%; accuracy in monitoring micro-level exceedances: 98.2%, 75.4%, 85.3%; complete risk chain formation rate: 100%, 0% (no risk association), 0% (single noise warning only); average risk discovery time: 1.7 days, 15.2 days, 3.4 days; rectification suggestion output time: 0.6 days, 3.1 days, 1.3 days; suggestion confidence level: mean 0.96, 0.73, 0.84; suggestion confidence level standard deviation: 0.03, 0.11, 0.06. surface

[0181] As shown in the table above, this method has achieved a breakthrough from 0 to 1 in terms of extracting macro rules and forming a complete risk chain compared with existing methods. Moreover, the average risk discovery time and the time for outputting rectification suggestions are only 1.7 days and 0.6 days, respectively, which reflects the high efficiency of this method in risk discovery and rectification. The average suggestion confidence level of 0.96 also verifies the high reliability of the rectification suggestions of this method.

[0182] As shown in Figure 6, this method is significantly better than the GIS remote sensing static assessment method in terms of accuracy in identifying meso-level conflicts and monitoring micro-level exceedances, and is 6.4% and 12.9% higher than the SVM machine learning method, respectively, verifying its high accuracy advantage in identifying meso-level conflicts and monitoring micro-level exceedances.

[0183] As shown in Figure 7, the present method is shorter than the existing GIS remote sensing static assessment method and SVM machine learning method in terms of both average risk discovery time and rectification suggestion output time, which verifies the efficiency of the present method in risk discovery.

[0184] As shown in Figure 8, the method suggests a mean confidence level of 0.96 and a standard deviation of only 0.03, which verifies the effectiveness of the double review of the method and the high reliability of the rectification suggestions.

[0185] This embodiment details an experiment conducted by the proposed method on a nature reserve along the route of an ultra-high voltage power transmission and transformation project. Data validation demonstrates the method's significant advantages: macroscopically, it can extract 18 rules, solving the problem of ruleless association in existing technologies; mesoscopically, it achieves 100% accuracy in conflict identification, surpassing both GIS and SVM methods, and requires minimal manual annotation; microscopically, it achieves 98.2% accuracy in monitoring exceedances, avoiding lag and misjudgment. This fully proves the method's comprehensive, high-precision, efficient, and highly reliable ecological risk identification capabilities, making it suitable for engineering management in ecologically sensitive areas.

[0186] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for identifying ecological risks in power transmission and transformation projects based on multi-granularity semantic understanding, characterized in that, include: By analyzing relevant data on power transmission and transformation projects, macro-risk categories, constraints, and evaluation criteria are extracted to form top-level rule nodes in the knowledge graph. An improved RT-DETRv2 model is trained based on the ECA attention mechanism. This model periodically processes satellite remote sensing images and UAV aerial videos along the power transmission and transformation project route to identify meso-level risk factors such as land cover change, water disturbance, and vegetation damage, forming geographic entities and event nodes in the knowledge graph. Real-time collection of micro-environmental data from the project site, including noise, water quality, soil moisture, and dust concentration, forms real-time monitoring nodes in the knowledge graph. Based on the top-level rule nodes, geographic entities and event nodes, and real-time monitoring nodes, a multi-granularity semantic understanding module containing macro, meso, and micro semantics is constructed to form an ecological risk knowledge graph. The Phi-4 model learns terminology and knowledge in the field of power transmission and transformation ecological risk through a masked language modeling task, and is then fine-tuned in layers to obtain a large risk identification model. Based on the ecological risk knowledge graph, a large-scale risk identification model is used to determine the ecological risks in power transmission and transformation projects. Based on the topological structure and semantic information of the ecological risk knowledge graph, a dual-review intelligent agent is constructed. The first review is risk self-checking and improvement. Through graph content query and data confidence test, when data is missing or the data confidence is low in the knowledge graph, the knowledge graph is iteratively optimized. The second level of review is causal deduction and decision optimization. Multiple independent risk identification models are used to judge the same risk structure. When the judgment results are inconsistent, the judgment results of each model and the relevant evidence chain are input into the big language model, which performs comprehensive analysis and outputs the final ruling result, and automatically triggers early warning for high-risk events.

2. The method according to claim 1, characterized in that, The improved RT-DETRv2 model specifically involves enhancing feature extraction at the Backbone layer using an ECA attention mechanism, based on the RT-DETRv2 model architecture. The formula is as follows: ,in, For the first Channel weight, For the output features of Stage 4 or Stage 5 in the Backbone of RT-DETRv2, For 1D convolution kernel size, The Sigmoid activation function is used to enhance the distinction between habitat texture features and construction machinery structural features; the fusion is optimized in the Neck layer using a BiFPN bidirectional feature pyramid mechanism, with the following formula: ,in, For the first Path weight, The total number of paths, The scaling factor is used; the loss function in the Head layer includes CIoU regression loss and Focal classification loss; after training, the improved RT-DETRv2 model is used to identify meso-level engineering risk factors and output the risk factor identification results. At intervals, the identification results are converted into geographic entities and event nodes in the knowledge graph to form geographic entities and event nodes in the knowledge graph.

3. The method according to claim 2, characterized in that, The loss function is specifically as follows: CIoU regression loss ensures that the predicted bounding box accurately covers the target region, providing an accurate regional basis for classification. The formula is: ,in, For intersection, union, and comparison, For prediction boxes With real frame The square of the Euclidean distance from the center point, The length of the diagonal of the minimum bounding rectangle that encloses the predicted bounding box and the ground truth bounding box. For balance coefficient, This parameter measures the consistency between the aspect ratio of the predicted bounding box and the ground truth bounding box; Focal classification loss is used to ensure accurate determination of the target category within the precisely selected region, and the formula is as follows: ,in, For category weights, Predict the probability for the target category. For focusing parameters.

4. The method according to claim 1, characterized in that, The construction of the multi-granularity semantic understanding module specifically involves: providing high-quality, multi-dimensional data for knowledge graph construction through the coordinated linkage of three levels of semantics; macro-level rules guiding the key directions of meso-level identification and micro-level monitoring; and the dynamic optimization of macro-level rules through feedback from meso- and micro-level data. The module processes three types of node data through entity alignment and relationship fusion algorithms. Entity alignment is achieved by calculating node name similarity and attribute similarity, using the following formula: ,in, To assess overall similarity, 、 These are the name similarity weight and the attribute similarity weight, respectively. For name similarity, For attribute similarity; when If the nodes are identified as the same entity, they are merged. The preset judgment threshold is used; the relationship fusion algorithm fills in the missing relationships between nodes through the rule base of the substation ecosystem to ensure that the relationship coverage of the knowledge graph is not lower than the preset minimum coverage.

5. The method according to claim 1, characterized in that, The process of fine-tuning the Phi-4 model in a hierarchical manner includes: collecting a corpus of ecological risk domain data for power transmission and transformation, containing ecological protection regulations, historical risk event reports, and environmental impact assessment technical documents; inputting the corpus of ecological risk domain data for power transmission and transformation, fine-tuning the Phi-4 small parameter model through a masked language model task with preset training parameters, and constructing a preliminary model; based on the sample set of labeled risk data and judgment results, optimizing the preliminary model with a cross-entropy loss function, and using an incremental pre-training strategy to freeze the bottom weights on the domain corpus and fine-tune only the top layer to construct a large-scale risk identification model.

6. The method according to claim 1, characterized in that, The process by which the risk identification big data model determines the ecological risks in power transmission and transformation projects is as follows: Through a dual-input mode combining graph features and domain knowledge, the node feature matrix of the target project is extracted from the ecological risk knowledge graph. Adjacency matrix ,in For the number of nodes, The feature dimension is used; node embedding vectors are generated through a graph convolutional neural network (GCN), and the calculation formula is as follows: ,in, for Layer node characteristics, （ (the identity matrix) for The degree matrix, For the first The weight matrix of the layer, The ReLU activation function is used; the DPO algorithm is used to fuse the node embedding vectors with the domain knowledge learned by the large model, outputting the risk level and corresponding judgment criteria, as shown in the formula: ,in For temperature parameters, For engineering command input, High-quality risk recommendations are manually labeled. Recommendations based on low quality and risk For the Phi-4 model, the input... Generate output The reward score; by standardization to eliminate the dimensional differences in node embedding vectors, the node embedding vectors are made consistent with the parameter distribution of the domain knowledge of the large model, as shown in the formula: ,in, After eliminating dimensional differences, the first Layer node feature vectors for The standard deviation matrix is ​​extracted; the association weights of risk factors and impact consequences learned in the pre-training stage of the risk identification large model are extracted and stored in the multi-head attention matrix of the Transformer layer; the mapping weights of index exceeding the standard and risk level are extracted and stored in the fully connected weight matrix of the output layer; fusion is achieved through three-layer progressive knowledge injection: the first layer, based on the domain terminology association parameters, combines the annotation results of power transmission and transformation engineering risk samples during GCN backpropagation to make the transformation direction of node embedding vectors conform to the domain knowledge logic; the second layer highlights the key risk factors in the domain knowledge through attention weight allocation, and transforms the risk judgment threshold parameters learned by the large model into the bias term of attention calculation, with the formula as follows: ,in, 、 、 These are the query vector, key vector, and value vector generated based on node embedding vectors, respectively. 、 They are query vectors Key vector The bias term, for The feature dimensions; through this layer of fusion, the feature proportion of key risk factors in the global feature vector is consistent with the risk priority in the domain knowledge; the third layer completes the fusion by binding the output layer weights with the judgment logic of the large model when mapping the global feature vector to the risk level probability distribution in the fully connected layer.

7. The method according to claim 1, characterized in that, The active query verification of the review agent is specifically a hybrid search strategy combining depth-first search (DFS) and breadth-first search (BFS): DFS is used to traverse the association between top-level rule nodes and geographic entity nodes; BFS is used to traverse the matching relationship between real-time monitoring nodes and geographic entity nodes to ensure the integrity of the logical chain; the frequency of query verification is synchronized with the knowledge graph update frequency, and each query generates a verification report containing passed items, items to be optimized, and abnormal items.

8. The method according to claim 1, characterized in that, The data confidence level is specifically calculated based on a multi-factor evaluation of sensor status and data quality to determine the reliability of data acquisition. The formula is as follows: ,in To ensure the credibility of data collection, Sensor calibration status. For data integrity, The degree of environmental interference is considered; the confidence level of the model is identified by confidence mapping and fusion calculation for different model types, using the following formula: ,in, To enhance the credibility of the model, The output confidence score of the computer vision model RT-DETRv2. The output confidence score of the large-scale risk identification model fine-tuned based on Phi-4 is given; the rule matching confidence score is calculated based on the semantic matching degree evaluation of the domain rule base, using the following formula: ,in, To improve the credibility of rule matching, This represents the number of rules in the domain rule base that are relevant to the current data. For the first The semantic matching degree between the rule and the current data To preset the minimum rule confidence level, the data confidence level is calculated by combining the data collection confidence level, model recognition confidence level, and rule matching confidence level, using the following formula: ,in, To assess the overall data confidence level, These are data collection weights, model recognition weights, and rule matching weights, respectively; when When the preset confidence threshold is reached, the model retraining or data acquisition process will be automatically triggered.

9. The method according to claim 1, characterized in that, The iterative optimization of the knowledge graph specifically involves: setting a preset threshold for missing data. Through iterative optimization using a strategy combining incremental and full updates, when the percentage of missing data is ≤ Incremental updates are used, supplementing only missing nodes and relationships; the percentage of missing data is greater than [a certain percentage]. If the overall confidence level of the data is lower than the preset confidence threshold, a full update is adopted, and the node generation and fusion process of the multi-granularity semantic understanding module is re-executed. After the optimized knowledge graph passes consistency and integrity checks, it is put into a second round of review.

10. The method according to claim 1, characterized in that, The second review specifically involves: deployment Each independent risk identification model has a preset consistency threshold. The consistency of the model's judgment results is calculated using the following formula: ,in To determine the consistency of the results, Don't be the first 、 The judgment results of each model, The semantic similarity between the two results is when < If the result is deemed inconsistent, the model's judgment result and the corresponding evidence chain are input into the pre-fine-tuned large language model, and the final ruling result is output, including the unified risk level, key credible evidence, explanation of the disputed points, and an early warning is triggered according to the risk level.

Images