Intelligent Analysis and Early Warning Methods for Visual Monitoring Data of Power Transmission Lines

Abstract

This invention discloses an intelligent analysis and early warning method for visualized monitoring data of transmission lines, specifically relating to the field of power system condition monitoring technology. The method includes: S1, constructing a dynamically updated spatiotemporal heterogeneous map of transmission lines; S2, performing dynamic risk simulation based on a spatiotemporal map attention network; S3, generating a comprehensive risk assessment conclusion by integrating data prediction and knowledge rules through a disagreement game mechanism; and S4, triggering tiered early warnings based on risk levels and using feedback data from early warning responses to drive model and strategy optimization. This invention solves the problems of delayed early warnings and rigid decision-making in existing technologies, achieving dynamic prediction and extrapolation of line risks, intelligent integrated decision-making, and continuous self-optimization of the system, significantly improving the foresight, accuracy, and operational efficiency of early warnings.

Description

Technical Field

[0001] This invention relates to the field of power system condition monitoring technology, and more specifically, to a method for intelligent analysis and early warning of visualized monitoring data of transmission lines. Background Technology

[0002] As the core artery of the power system, the safe and stable operation of power transmission lines is of paramount importance. Traditional inspection and monitoring mainly rely on manual labor, which has problems such as low efficiency, high risk and difficulty in real-time coverage. With the development of sensor and visualization technology, an integrated air-ground intelligent monitoring system based on drones, fixed cameras and various online monitoring devices has been gradually established, which can collect visible light, infrared and ultraviolet images of the lines in real time, as well as multimodal data such as temperature, tilt angle and vibration.

[0003] Existing intelligent analysis and early warning solutions mostly focus on a single technical approach:

[0004] Firstly, computer vision-based defect identification can effectively detect apparent defects such as insulator damage and foreign objects in the channel, but it lacks the ability to assess the evolution trend of defects and systemic risks.

[0005] Secondly, threshold-based early warning based on sensor time-series data can alarm for exceeding limits such as temperature and sag, but the early warning rules are rigid, the false alarm rate is high, and the spatiotemporal propagation of risks cannot be predicted.

[0006] Third, some advanced solutions attempt to introduce graph neural networks to perform cluster analysis on monitoring points, but their models are mostly static or shallow dynamic, making it difficult to characterize the complex spatiotemporal coupling relationship between line equipment and the dynamic evolution of risks.

[0007] In response to the above situation, the present invention provides a method for intelligent analysis and early warning of power transmission line visualization monitoring data. Summary of the Invention

[0008] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a method for intelligent analysis and early warning of power transmission line visualization monitoring data, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent analysis and early warning of power transmission line visualization monitoring data, comprising the following steps:

[0010] S1. Construct a dynamically updated spatiotemporal heterogeneous graph of transmission lines: Based on the collected multimodal monitoring data, construct a spatiotemporal heterogeneous graph with towers, insulator strings and conductor segments as physical nodes and meteorological and topographical elements as environmental nodes. The attributes of nodes and edges are feature vectors that change over time.

[0011] S2. Risk dynamic simulation based on spatiotemporal graph attention network: The spatiotemporal heterogeneous graph is input into the pre-trained spatiotemporal graph attention network model. Through the spatial attention mechanism and temporal attention mechanism coupled by the model, the spatiotemporal dependency of node states is learned, and the risk prediction results for a specified future time period are output by forward iterative reasoning.

[0012] S3, Dissent Game and Adaptive Decision Making: The risk prediction results are compared and game-based with the rule-based reasoning results based on the pre-set power equipment anomaly knowledge graph to generate a comprehensive risk assessment conclusion.

[0013] S4. Cascaded Early Warning Triggering and Feedback Learning: Trigger tiered early warnings based on comprehensive risk assessment conclusions, and drive the optimization of models and strategies based on feedback data from early warning handling.

[0014] Preferably, in step S1, the nodes of the spatiotemporal heterogeneous graph include tower nodes, insulator string nodes, and conductor segment nodes that represent the physical entities of the transmission line, as well as meteorological grid nodes and terrain grid nodes that represent virtual risk units.

[0015] The edges of a spatiotemporal heterogeneous graph include topological connection edges representing electrical connections and power flow directions, spatial proximity edges representing physical distance and geographical proximity, and correlation edges representing the transmission relationship of environmental factors.

[0016] Preferably, the multimodal monitoring data includes: image and video data from visible light, infrared and ultraviolet imaging devices; time-series data from conductor tilt, temperature, vibration and current sensors; temperature, humidity, wind speed and direction data from micro-weather stations; and topology and power flow data from the power grid dispatching system.

[0017] Preferably, in step S2, the processing procedure of the spatiotemporal graph attention network model includes:

[0018] Spatial attention calculation: For a spatiotemporal heterogeneous graph, at each time slice, calculate the attention coefficient between any two nodes connected by an edge. This coefficient dynamically represents the weight of the immediate influence of the state of one node on the state of another node.

[0019] Temporal attention computation: For each node, attention weights are computed on the time series of its state features to capture the long-short-term dependency patterns of its state evolution.

[0020] Spatiotemporal feature aggregation and updating: Based on spatial attention coefficients and temporal attention weights, the neighbor information and its own historical information of a node are aggregated to update the risk hidden state representation of the node.

[0021] Preferably, in step S2, the dynamic evolution of risk is simulated by forward iterative reasoning. Specifically, the state of the spatiotemporal heterogeneous graph at the current moment is used as the initial graph, and the predicted future environmental sequence data is input. The spatiotemporal graph attention network model starts from the initial graph and performs multi-step forward calculation. Each step outputs the graph state prediction for the next moment, which is used as the input for the next step. The process is iterated sequentially to generate a series of risk state graph sequences for a series of consecutive time points in the future.

[0022] Preferably, the objection game and adaptive decision-making in step S3 specifically include:

[0023] Based on the risk prediction results, high-risk nodes and their risk evolution trajectories are extracted;

[0024] From the knowledge graph of power equipment anomalies, we match expert rules and historical cases related to high-risk nodes and current environmental conditions, and perform logical reasoning to obtain rule-driven risk assessment conclusions.

[0025] When data-driven risk prediction conclusions and rule-driven risk assessment conclusions are inconsistent in terms of risk level or key nodes, game theory decision-making is triggered. The game theory decision-making objective function is to minimize the comprehensive decision loss. The loss function integrates the mean square error of the prediction model, the degree of rule conflict, and the costs of false positives and false negatives. The final risk level determination and the set of key risk nodes are obtained by solving through optimization algorithms.

[0026] Preferably, in step S4, the specific rules for triggering the graded early warning instruction are as follows:

[0027] When the overall risk level is low and the scope of impact is limited, a Level 1 warning is triggered, and a visual prompt is only displayed on the monitoring interface.

[0028] When the overall risk level is medium or high, or when the predicted impact involves critical line segments, a level two assessment and warning is triggered, and warning information and a risk simulation visualization report are pushed to the mobile terminals of designated maintenance personnel.

[0029] When the overall risk level is emergency, or when the forecast indicates that a chain of failures may occur, a level 3 emergency response warning is triggered, simultaneously triggering audible and visual alarms, automatically generating and pushing a response plan containing operational suggestions, and linking relevant systems to initiate preventive checks.

[0030] Preferably, the feedback learning in step S4 specifically includes two closed loops:

[0031] Inner loop online fine-tuning: Using the prediction-verification feedback data pairs obtained after a single early warning response, a small sample training set is constructed to quickly fine-tune some parameters in the spatiotemporal graph attention network model online.

[0032] Outer-loop strategy optimization: Periodically collect all early warning handling cases within a certain period of time to form a reinforcement learning environment. With the optimization goal of reducing overall operation and maintenance costs and reducing power outage time, train a strategy optimization agent to automatically adjust the early warning classification threshold and game decision weight parameters.

[0033] Preferably, before step S2, a pre-training step of the spatiotemporal graph attention network model is also included: using historical, labeled multi-source monitoring data sequences of transmission lines and corresponding fault or early warning records, with future state prediction and event classification as joint training objectives, the spatiotemporal graph attention network model is subjected to end-to-end supervised training.

[0034] Preferably, the pre-training step of the model adopts a course learning strategy, which first uses data with a longer time granularity for initial training, and then gradually introduces high-frequency, noisy real-time data sequences for refined training, so as to improve the convergence speed and generalization ability of the model.

[0035] The technical effects and advantages of this invention are as follows:

[0036] 1. This invention constructs a dynamic spatiotemporal heterogeneous graph that integrates multi-source data and uses a spatiotemporal graph attention network model to learn the spatiotemporal dependencies of node states. It can simulate the dynamic evolution of risks in time and space dimensions, output the risk prediction results and evolution trajectory for a specified future period, thereby transforming the early warning mode from post-event alarm to pre-event prediction, providing a key basis for proactive operation and maintenance decision-making.

[0037] 2. This invention compares and plays a game between the data-driven risk prediction output by the spatiotemporal graph attention network model and the rule-based reasoning results based on the knowledge graph of power equipment anomalies. With the goal of minimizing the comprehensive decision loss, it makes adaptive decisions, which overcomes the limitations of the black box decision-making of a single AI model and the rigidity of traditional rule systems. This makes the final comprehensive risk assessment conclusion both forward-looking and interpretable, and respects the experience of domain experts, significantly reducing the false positive and false negative rates.

[0038] 3. This invention establishes a three-level cascaded early warning triggering rule that matches the dynamic risk level, and designs a dual-loop feedback learning mechanism that includes inner-loop online fine-tuning and outer-loop strategy optimization. This enables the system to not only respond accurately to the current risk, but also to use the feedback data from each early warning response to drive the automated and continuous optimization of the analysis model parameters and early warning strategy parameters. This achieves autonomous optimization of the system's intelligence level and effectively reduces operation and maintenance costs and failure losses in the long run. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0040] Figure 2This is a schematic diagram of the construction of the spatiotemporal heterogeneity diagram of the transmission line according to the present invention.

[0041] Figure 3 This is a flowchart illustrating the structure and risk simulation of the spatiotemporal graph attention network model of the present invention.

[0042] Figure 4 This is a closed-loop diagram illustrating the cascaded early warning triggering and feedback learning of the present invention. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Example 1

[0044] This embodiment provides a method for intelligent analysis and early warning of power transmission line visualization monitoring data, as shown in the attached figure. Figure 1 As shown, the specific steps include:

[0045] S1. Construct a dynamically updated spatiotemporal heterogeneity map of transmission lines.

[0046] Specifically, in the implementation of this step, firstly, the system integrates and collects multi-source monitoring data, which includes:

[0047] Image and video data: High-definition visible light images, infrared thermal images, and ultraviolet corona detection videos from drone inspections, helicopter inspections, and ground-based fixed cameras;

[0048] Sensor timing data: Data from miniature sensors installed on conductors and towers, including conductor tilt (sag), temperature, vibration acceleration, and current values, with a sampling frequency of 1-10Hz;

[0049] Environmental data: Data from micro-weather stations along the route, including temperature, humidity, wind speed, wind direction, and rainfall;

[0050] Power grid operation data: Real-time power grid topology and line power flow data obtained from energy management systems or wide-area measurement systems;

[0051] Based on a unified spatiotemporal reference (such as the BeiDou Navigation Satellite System and Coordinated Universal Time UTC), the aforementioned multimodal data are time-synchronized and spatially registered. Subsequently, a dynamically updated spatiotemporal heterogeneity map of transmission lines is constructed, as shown in the attached figure. Figure 2 As shown, the process is as follows:

[0052] When defining nodes:

[0053] Physical nodes: Each tower, each insulator string, and each conductor segment between two towers is modeled as an independent node;

[0054] Virtual environment nodes: The area where the route corridor is located is divided into regular geographical grids (such as 500m×500m), and each grid serves as a meteorological and terrain node;

[0055] When defining edges:

[0056] Topology connection edge: Based on the power grid electrical wiring diagram, an edge is established between the connected tower nodes, with the direction consistent with the power flow direction, and the weight can be initialized as the line impedance;

[0057] Spatial proximity edge: An undirected edge is established between any two physical nodes within a physical distance (e.g., 1 kilometer), with the weight inversely proportional to the distance;

[0058] Environmental association edge: A directed edge is established between a meteorological grid node and all physical nodes within its influence range. The direction represents the environmental impact (such as wind blowing from a meteorological node to a conductor node), and the weight is determined by wind speed, wind direction and the line angle, etc.

[0059] Dynamic attribute updates: Each node and edge is attached with a feature vector. For example, the feature vector of a tower node may include its tilt angle and vibration frequency, the feature vector of a conductor segment node may include current, temperature and sag value, and the feature vector of a meteorological grid node may include wind speed, wind direction and temperature. These feature vectors are updated in real time or near real time (e.g. every minute) as new data flows in, making the whole map a dynamic digital mirror of the line status.

[0060] S2. Risk dynamic simulation based on spatiotemporal graph attention network

[0061] Specifically, in practice, this step involves inputting the constructed spatiotemporal heterogeneous graph into a pre-trained spatiotemporal graph attention network model for dynamic risk simulation, as shown in the attached figure. Figure 3 As shown, the model's pre-training and inference processes are as follows:

[0062] When performing model pre-training, historical data should be used to supervise the pre-training of the spatiotemporal graph-based attention network model, as follows:

[0063] Training data preparation: Collect multi-source monitoring data sequences (such as hourly snapshots) from the same sources as the current step S1 within the past 1-3 years, as well as real fault reports and early warning work orders recorded within the corresponding time period;

[0064] Training Objectives: Employing a multi-task learning approach, the model is jointly trained to achieve two objectives: First, future state prediction: Based on the graphs from the previous T historical time points, the model is required to predict the state of key nodes (such as conductor temperature and sag) at time t in the future; Second, event classification: The model is required to determine whether a specific risk event (such as conductor galloping or insulator flashover) will occur within a certain time period in the future.

[0065] Course learning strategy: To improve training efficiency, a course learning approach is adopted. In the first stage, data with coarse time granularity (such as one sample per day) but high data quality and complete labeling are used for initial training to allow the model to master basic spatiotemporal correlation patterns. In the second stage, high-frequency (such as per minute) real-time monitoring data sequences containing more noise are gradually introduced for refined training to improve the model's ability to capture real-time dynamics and its robustness.

[0066] When performing online risk simulation reasoning, the model operates using a forward iterative reasoning method, as follows:

[0067] Spatial attention calculation: For each time slice, the model calculates the attention coefficient between any two nodes connected by an edge in the graph. For example, it calculates the attention weight of a strong wind weather node to a conductor segment node. This weight is not fixed, but is dynamically calculated by the neural network and depends on the characteristics of the two nodes at the moment (such as wind speed, conductor type, and historical galloping). This allows for precise quantification of the intensity of the environment's impact on the equipment.

[0068] Temporal attention calculation: For each node, the model calculates attention weights on its historical state feature sequence, enabling the model to determine, for example, whether the sustained high temperature over the past 2 hours or the temperature rise rate over the past 10 minutes contributes more to the current overheating risk of the conductor.

[0069] Iterative Risk Prediction: Taking the current spatiotemporal heterogeneous graph as the initial state, and inputting the forecast data sequence (wind speed, precipitation probability, etc.) for the next 6-24 hours obtained from the meteorological department, the model performs multi-step forward calculations starting from the current graph. In the first step, combining the current graph state and the environmental forecast for the first time slice, the graph state for the first time slice is predicted (i.e., the hidden risk state of each node). In the second step, this predicted state is used as the new current state, and combined with the environmental forecast for the second time slice, the state for the second time slice is predicted. This process is iterated sequentially, and finally, a risk state graph sequence for a series of future time points (e.g., one point per hour in the next 6 hours) is generated. This sequence intuitively shows the distribution and evolution trend of risks (such as high temperature, galloping index) in the future spatiotemporal space.

[0070] S3, Dissent Game and Adaptive Decision Making

[0071] Specifically, in the implementation of this step, the data-driven risk prediction output by the spatiotemporal graph attention network model will be fused with the results of knowledge-based rule reasoning. The specific process is as follows:

[0072] Knowledge graph rule reasoning: The system maintains a power equipment anomaly knowledge graph, which stores rules summarized by domain experts (such as when the insulator pollution degree is >0.5 and the humidity is >90%, the pollution flashover risk level is high) and typical historical cases. When the spatiotemporal graph attention network model outputs high-risk nodes, the system matches relevant rules and similar cases from the graph, performs logical reasoning, and obtains a risk assessment conclusion based on rules and experience.

[0073] Triggering game theory decision-making: Compare data-driven predictions and rule-driven conclusions. If there is a significant discrepancy between the two in terms of risk level (e.g., AI predicts medium, rule-based judgment is high) or key risk points, then the game theory decision-making mechanism is triggered.

[0074] Game Theory Solution: The core of the game is a loss function: L = α * MSE (prediction error) + β * Conflict (rule conflict degree) + γ * Cost (false positive / false negative cost), where α, β, and γ are weighting coefficients, MSE measures the historical uncertainty of the prediction model, Conflict measures the degree of deviation from expert rules, and Cost is determined by the operational economic model. Through optimization algorithms such as gradient descent, the final risk level determination and key risk node set that minimize the total loss L are solved, ensuring that the decision-making respects data patterns, does not violate key domain knowledge, and seeks the optimal economic outcome.

[0075] S4, Cascaded Early Warning Triggering and Feedback Learning

[0076] Specifically, in its implementation, this step involves the system issuing early warnings and providing feedback based on the comprehensive risk assessment conclusions following the game theory decision, forming a closed loop, as shown in the attached figure. Figure 4 As shown, the details are as follows:

[0077] When conducting a three-level cascaded early warning:

[0078] Level 1 Attention Warning: When the risk level is low and the impact is limited to a single tower or line segment, the area will be highlighted on the digital twin screen in the monitoring center without disturbing human staff.

[0079] Level 2 Assessment and Early Warning: When the risk level is medium or high, or when the predicted impact may affect important load lines, the system automatically generates an early warning report and pushes it to the relevant operation and maintenance team leader via mobile app. The report includes: risk type, predicted evolution animation, impact range map, suggested inspection route and key inspection items.

[0080] Level 3 Response Warning: When the risk level is emergency, or when the prediction indicates that a chain trip may be triggered (such as a tree obstacle simultaneously touching multiple lines), the system immediately triggers an audible and visual alarm in the control room. At the same time, it automatically generates a response plan (such as: recommending to reduce the load of a certain line by 30% and notifying relevant personnel to carry out live-line troubleshooting in the corresponding section within 1 hour), and pushes it to the dispatch and operation and maintenance system with one click to activate the emergency response plan.

[0081] When conducting two-loop feedback learning:

[0082] Inner Ring Online Fine-tuning: After each early warning response, on-site personnel provide feedback on the verification results via the App (such as confirming that the insulator is damaged, it is a false alarm, or it is due to the bird's nest blocking the view). This set of prediction-verification data immediately constitutes a small sample training set, which is used to quickly fine-tune a small number of parameters at the top layer of the spatiotemporal graph attention network model, so that it can quickly adapt to the latest local conditions.

[0083] Outer-loop strategy optimization: Every quarter or every six months, the system collects all early warning cases and their handling costs, failure losses, and other data to build a reinforcement learning environment. A strategy optimization agent is trained with the goal of maximizing long-term rewards (i.e. reducing total maintenance costs and failure losses). This agent automatically outputs optimization suggestions for strategy parameters such as early warning classification thresholds and game loss function weights (α, β, γ). After confirmation by engineers, the system is updated to achieve continuous strategy-level optimization. Example 2

[0084] To better understand this invention, this embodiment uses the example of a power transmission line in a mountainous area facing the risk of wildfires to illustrate the specific application of this method:

[0085] Data and Mapping: Satellite remote sensing detects fire points, micro-weather stations report wind direction from the fire site to the power line, and the system constructs a spatiotemporal map: the grid where the fire point is located is a high-risk virtual node, and the towers and conductors of the downwind line are physical nodes. Environmental association edges representing the propagation of smoke and heat radiation are established.

[0086] Risk simulation: The spatiotemporal graph attention network model, combined with the fire spread model and wind speed and direction, simulates the diffusion path of smoke particles and high-temperature air masses in the spatiotemporal graph within the next 2 hours through iterative reasoning, and predicts which conductor segments and insulator strings will have their pollution and temperature characteristics rise rapidly, and the time point when they reach the danger threshold.

[0087] Game theory decision: The model predicts that the insulators in a certain section will reach a high risk in 1.5 hours. The knowledge graph rules suggest that the open flame is within 5 kilometers of the line and the wind speed is greater than level 3, which is directly judged as an emergency risk. Both of them point to high risk, and there is no significant disagreement. The comprehensive conclusion is that it is an emergency.

[0088] Early warning and feedback: The system immediately triggered a Level 3 emergency response warning, recommending that the dispatching department urgently adjust the power grid operation mode in the area, reduce the load on the affected lines, and notify the fire and maintenance teams to go to the specific coordinates to prepare for fire fighting and maintenance. Afterwards, feedback data such as the wildfire spread trajectory and the actual outage status of the lines were recorded and used to optimize the model's prediction accuracy for similar scenarios.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent analysis and early warning of visualized monitoring data of power transmission lines, characterized by: Includes the following steps: S1. Construct a dynamically updated spatiotemporal heterogeneous graph of transmission lines: Based on the collected multimodal monitoring data, construct a spatiotemporal heterogeneous graph with towers, insulator strings and conductor segments as physical nodes and meteorological and topographical elements as environmental nodes. The attributes of nodes and edges are feature vectors that change over time. S2. Risk dynamic simulation based on spatiotemporal graph attention network: The spatiotemporal heterogeneous graph is input into the pre-trained spatiotemporal graph attention network model. Through the spatial attention mechanism and temporal attention mechanism coupled by the model, the spatiotemporal dependency of node states is learned, and the risk prediction results for a specified future time period are output by forward iterative reasoning. S3, Dissent Game and Adaptive Decision Making: The risk prediction results are compared and game-based with the rule-based reasoning results based on the pre-set power equipment anomaly knowledge graph to generate a comprehensive risk assessment conclusion. The objection game and adaptive decision-making in step S3 specifically include: Based on the risk prediction results, high-risk nodes and their risk evolution trajectories are extracted; From the knowledge graph of power equipment anomalies, we match expert rules and historical cases related to high-risk nodes and current environmental conditions, and perform logical reasoning to obtain rule-driven risk assessment conclusions. When data-driven risk prediction conclusions and rule-driven risk assessment conclusions are inconsistent in terms of risk level or key nodes, game-theoretic decision-making is triggered. The game-theoretic decision-making takes minimizing the comprehensive decision loss as the objective function. The loss function integrates the mean square error of the prediction model, the degree of rule conflict, and the costs of false positives and false negatives. The final risk level determination and the set of key risk nodes are obtained by solving through optimization algorithms. S4, Cascaded Early Warning Triggering and Feedback Learning: Trigger tiered early warnings based on comprehensive risk assessment conclusions, and drive the optimization of models and strategies based on feedback data from early warning handling; The feedback learning in step S4 specifically includes two closed loops: Inner loop online fine-tuning: Using the prediction-verification feedback data pairs obtained after a single early warning response, a small sample training set is constructed to quickly fine-tune some parameters in the spatiotemporal graph attention network model online. Outer-loop strategy optimization: Periodically collect all early warning handling cases within a certain period of time to form a reinforcement learning environment. With the optimization goal of reducing overall operation and maintenance costs and reducing power outage time, train a strategy optimization agent to automatically adjust the early warning classification threshold and game decision weight parameters.

2. The intelligent analysis and early warning method for visualized monitoring data of power transmission lines according to claim 1, characterized in that: In step S1, the nodes of the spatiotemporal heterogeneous graph include tower nodes, insulator string nodes, and conductor segment nodes that represent the physical entities of the transmission line, as well as meteorological grid nodes and terrain grid nodes that represent virtual risk units. The edges of a spatiotemporal heterogeneous graph include topological connection edges representing electrical connections and power flow directions, spatial proximity edges representing physical distance and geographical proximity, and correlation edges representing the transmission relationship of environmental factors.

3. The intelligent analysis and early warning method for visualized monitoring data of power transmission lines according to claim 1, characterized in that: Multimodal monitoring data includes: image and video data from visible light, infrared and ultraviolet imaging devices; time-series data from conductor tilt, temperature, vibration and current sensors; temperature, humidity, wind speed and direction data from micro-weather stations; and topology and power flow data from the power grid dispatching system.

4. The intelligent analysis and early warning method for visualized monitoring data of power transmission lines according to claim 1, characterized in that: In step S2, the processing procedure of the spatiotemporal graph attention network model includes: Spatial attention calculation: For a spatiotemporal heterogeneous graph, at each time slice, calculate the attention coefficient between any two nodes connected by an edge. This coefficient dynamically represents the weight of the immediate influence of the state of one node on the state of another node. Temporal attention computation: For each node, attention weights are computed on the time series of its state features to capture the long-short-term dependency patterns of its state evolution. Spatiotemporal feature aggregation and updating: Based on spatial attention coefficients and temporal attention weights, the neighbor information and its own historical information of a node are aggregated to update the risk hidden state representation of the node.

5. The intelligent analysis and early warning method for visualized monitoring data of power transmission lines according to claim 1, characterized in that: In step S2, the dynamic evolution of risk is simulated by forward iterative reasoning. Specifically, the current state of the spatiotemporal heterogeneous graph is used as the initial graph, and the predicted future environmental sequence data is input. The spatiotemporal graph attention network model starts from the initial graph and performs multi-step forward calculations. Each step outputs the graph state prediction for the next time moment, which is used as the input for the next step. This process is iterated sequentially to generate a series of risk state graph sequences for a series of consecutive time points in the future.

6. The intelligent analysis and early warning method for visualized monitoring data of power transmission lines according to claim 1, characterized in that: In step S4, the specific rules for triggering the graded early warning instruction are as follows: When the overall risk level is low and the scope of impact is limited, a Level 1 warning is triggered, and a visual prompt is only displayed on the monitoring interface. When the overall risk level is medium or high, or when the predicted impact involves critical line segments, a level two assessment and warning is triggered, and warning information and a risk simulation visualization report are pushed to the mobile terminals of designated maintenance personnel. When the overall risk level is emergency, or when the forecast indicates that a chain of failures may occur, a level 3 emergency response warning is triggered, simultaneously triggering audible and visual alarms, automatically generating and pushing a response plan containing operational suggestions, and linking relevant systems to initiate preventive checks.

7. The intelligent analysis and early warning method for visualized monitoring data of power transmission lines according to claim 1, characterized in that: Before step S2, a pre-training step of the spatiotemporal graph attention network model is also included: using historical, labeled multi-source monitoring data sequences of transmission lines and corresponding fault or early warning records, with future state prediction and event classification as joint training objectives, the spatiotemporal graph attention network model is trained end-to-end under supervision.

8. The intelligent analysis and early warning method for visualized monitoring data of power transmission lines according to claim 7, characterized in that: The model's pre-training steps employ a course-based learning strategy, initially using data with a longer time granularity for preliminary training, and then gradually introducing high-frequency, noisy real-time data sequences for refined training, in order to improve the model's convergence speed and generalization ability.

Images