A Method and System for Predicting the Mechanical Response of Tower Line Systems Based on Hybrid Expert Networks

Abstract

A method and system for predicting the mechanical response of a transmission tower system based on a hybrid expert network includes: transforming multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand; introducing the high-dimensional feature vectors into a gated network with physical rule masks to determine the activated expert models in the hybrid expert model; selecting a solution strategy based on the scenario of the transmission tower system and solving the activated expert models to obtain the dynamic tension time history curve of the conductor end to the tower suspension point; using a graph neural network to abstract the transmission tower into a topological graph, updating the nodes in the graph neural network based on the dynamic tension time history curve, and using the graph neural network to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure to obtain the state matrix of the transmission tower system. This application achieves decoupled solutions for different physical phenomena, real-time linkage simulation from line to tower, and uses a physical rule mask mechanism to eliminate the physical illusion of AI models predicting galloping phenomena under non-galloping conditions.

Description

Technical Field

[0001] This application relates to the field of predicting the mechanical response of transmission towers and lines, specifically to a method and system for predicting the mechanical response of tower and line systems based on hybrid expert networks. Background Technology

[0002] In the power grid operation and maintenance system, the transmission tower system, or simply the tower-line system, serves as a critical energy channel and faces severe safety challenges. Currently, the closest technical solution is a transmission line condition prediction method based on a single deep neural network. This method typically collects meteorological data (wind speed, wind direction) as input and directly uses an end-to-end neural network, trained to output the conductor galloping amplitude. The steps of existing methods are generally as follows: 1. Data construction phase: Historical meteorological data is collected as input characteristics, and the displacement amplitude of the corresponding conductor galloping monitoring device is collected as a label; 2. Model Building Phase: Construct a general deep neural network model (usually using LSTM, GRU, or Transformer architecture), where LSTM is a Long Short-Term Memory network, GRU is a gated recurrent unit, and Transformer is a transformation architecture. This model typically treats the transmission line as a single "black box," without distinguishing its physical structure such as span, number of splits, and boundary conditions (such as tower stiffness). 3. Model training phase: Supervised learning of the model is performed using historical data, with the goal of minimizing the mean square error between the predicted displacement and the actual displacement; 4. Model inference stage: Input real-time meteorological data, and the model outputs the predicted maximum amplitude of conductor galloping.

[0003] The core characteristic of existing methods is that they focus only on a single index of a single object, and their output does not include dynamic tension information of the conductor, nor does it have a calculation module for the response of the tower structure.

[0004] Existing technologies attempt to fit all operating conditions using the same set of neural network architecture weight parameters. However, galloping is a nonlinear self-excited vibration, light wind vibration is a fluid-structure interaction forced vibration, and de-icing is a transient impact response. The structures and solution methods of the governing equations for these three types differ significantly. Using a single model makes it difficult to find the global optimum on the complex non-convex loss surface during training, causing the model to easily learn incorrect statistical correlations, which seriously affects the reliability of early warnings.

[0005] Existing purely data-driven models rely solely on statistical regularities. Since extreme disasters in the power grid (such as severe icing and galloping) are typically long-tailed data, the sample size is extremely small and the distribution is highly imbalanced. Without the guidance of physical equations, existing models tend to fit frequently occurring normal operating conditions, treating scarce extreme disasters as noise or outliers. This results in the models being unable to make reasonable extrapolations and predictions when facing such low-probability, high-risk events, unlike physical equations, leading to a risk of underreporting.

[0006] Existing technologies are typically limited to predicting the kinematic parameter of "conductor displacement," neglecting to calculate the dynamic parameter of "dynamic tension at the conductor end." In reality, unbalanced tension in the conductor is the primary cause of tower collapse. Furthermore, existing model outputs often lack an interface with tower mechanics models, making it impossible to assess the cascading effects of the tower-conductor system and to provide early warnings of power pole collapse risks.

[0007] Existing simulation and prediction technologies for the mechanical state of transmission lines still face many bottlenecks when meeting the needs of practical applications: 1. Traditional finite element methods are computationally inefficient: To capture dynamic characteristics such as conductor galloping or aerobatic vibrations, it is usually necessary to establish a refined, fully coupled finite element model that includes split conductors, fittings, and towers, with millions of degrees of freedom. Simulation calculations for nonlinear dynamic processes are extremely time-consuming (often reaching hours or even days), severely lagging behind the power grid's minute-level real-time early warning requirements.

[0008] 2. Single data-driven models struggle to adapt to diverse dynamic mechanisms: The mechanical response of transmission lines is extremely complex, encompassing low-frequency, large-amplitude galloping, high-frequency, small-amplitude wind vibrations, and transient de-icing jumps. Existing artificial intelligence methods often employ single-structure neural network models (such as general-purpose LSTM or Transformer) to fit all operating conditions. However, because the physical control equations for different disasters are drastically different, single models struggle to converge and are prone to producing predictions that violate physical laws under extreme conditions.

[0009] 3. Lack of understanding of tower-line coupling effect: Most existing studies isolate the "conductor" and the "tower," or simplify the tower as a rigid support. In reality, under uneven icing, galloping, or de-icing scenarios, the dynamic unbalanced tension generated by the conductor is the root cause of tower crossarm torsion, tower buckling, and even tower collapse. Simulations lacking a tower-line coupling mechanism cannot realistically assess system-level risks, limiting their application value in disaster prevention and mitigation.

[0010] To address the aforementioned issues, existing technologies lack a comprehensive solution that can adaptively identify different dominant dynamic modes, balance computational efficiency and physical realism, and achieve cascaded mechanical transmission in tower-line systems. Summary of the Invention

[0011] To address the shortcomings of existing technologies, such as their inability to simultaneously achieve multi-mechanism adaptability and computational efficiency, and their lack of system-level coupled evaluation, this application proposes a method for predicting the mechanical response of tower-line systems based on a hybrid expert network, comprising: Transform multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand; In a gated network based on high-dimensional feature vectors and by introducing physical rule masks, the activated expert models in the hybrid expert model are determined. Based on the scenario of the transmission tower system, a solution strategy is selected, and the activated expert model is solved using the solution strategy to obtain the dynamic tension time history curve of the conductor end to the tower hanging point. A graph neural network is used to abstract the transmission tower into a topology graph containing tower body component nodes and component connection relationships. The nodes representing conductor suspension points in the graph neural network are updated based on the dynamic tension in the dynamic tension time history curve. The graph neural network is used to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure and output the tower-line system state matrix. Among them, the multi-source heterogeneous monitoring data includes: static parameters of the line and real-time micro-meteorological data; The hybrid expert model is obtained by training a deep learning model based on historical fluid dynamics features and corresponding mechanical responses.

[0012] Preferably, the step of converting multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand includes: The static parameters of the target line are obtained based on the ledger information. These static parameters include: conductor mass per unit length, bending stiffness, number of split conductors, span, and tower height. The static parameters are mapped to static embedding vectors through a fully connected layer; Physical transformations are performed on real-time micro-meteorological data to form physical codes for dynamic environmental characteristics; The static embedding vector and the physical encoding are concatenated to obtain a high-dimensional feature vector; The multi-source heterogeneous monitoring data includes: ledger information and real-time micro-meteorological data.

[0013] Preferably, the determination of the activated expert model in the hybrid expert model within the gating network based on high-dimensional feature vectors through the introduction of physical rule masks includes: The high-dimensional feature vector is input into the gated neural network, which outputs a set of original weights for each expert. The current physical conditions are verified in real time based on a rule base of typical operating conditions of transmission lines, and a physical mask vector is generated. The final activation probability is calculated using the Softmax function based on the original weights and the physical mask. The expert model with the highest final activation probability or the final activation probability close to 1 is selected as the activated expert model. The hybrid expert models include: the dancing expert model, the vibration expert model, and the de-icing expert model.

[0014] Preferably, the final activation probability is calculated using the following formula:

[0015] in, The physical guiding coefficient. The final activation probability, To normalize the weights, For physical mask, This is a commonly used activation function in machine learning.

[0016] Preferably, the step of updating the nodes representing the conductor suspension points in the dynamic tension time history curve based on the dynamic tension update graph neural network, and using the graph neural network to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure, and outputting the tower-conductor system state matrix, includes: Dynamic tension is injected as message data into the graph neural network; Graph neural networks transmit the injected dynamic tension in the transmission tower topology through the message passing mechanism of graph convolution, simulating the propagation process of dynamic tension in the tower components and outputting the real-time stress and deformation values ​​of all components in the entire tower. Align the real-time stress and deformation values ​​of all components of the tower with the complete state of the conductor on the time axis to obtain the tower-conductor system state matrix.

[0017] Preferably, the nodes representing wire attachment points in the graph neural network are updated according to the following formula:

[0018] In the formula, This is the updated stress state vector of the node. This represents the stress state vector of the current layer's neighbor node u. Let v be the stress state vector of the current node v. Let v be the set of neighbors of node v. Let k be the aggregation function, and k be the current iteration level. This is the activation function.

[0019] Preferably, after obtaining the tower line system state matrix, the process further includes: A comprehensive mechanical state report is generated based on the state matrix of the tower line system, dynamic tension time history curves, and real-time stress and deformation values ​​of all components.

[0020] Preferably, the solution strategy selected based on the scenario of the transmission tower system includes: When the transmission tower system is in a low-frequency, large-displacement galloping scenario, the solution strategy is a physical information neural network with embedded aerodynamic equations. When the transmission tower system is in a scenario of high-frequency, low-amplitude wind vibration, the solution strategy is a convolutional neural network that is good at spectrum analysis. When the transmission tower system is in a transient impact de-icing scenario, the solution strategy is a long short-term memory network with temporal memory capabilities.

[0021] Based on the same inventive concept, this application also provides a tower line system mechanical response prediction system based on a hybrid expert network, including: The multidimensional activation feature encoding module is used to transform multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand. The physical perception gating module is used to determine the activated expert model in the hybrid expert model in a gating network based on high-dimensional feature vectors by introducing physical rule masks; and selects a solution strategy based on the scenario of the transmission tower system, and uses the solution strategy to solve the activated expert model to obtain the dynamic tension time history curve of the conductor end to the tower hanging point. The tower-line coupling graph network module is used to abstract the transmission tower into a topology graph containing tower body component nodes and component connection relationships using graph neural networks. It updates the nodes representing conductor suspension points in the graph neural network based on the dynamic tension in the dynamic tension time history curve, and uses the graph neural network to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure, and outputs the tower-line system state matrix. Among them, the multi-source heterogeneous monitoring data includes: static parameters of the line and real-time micro-meteorological data; The hybrid expert model is obtained by training a deep learning model based on historical fluid dynamics features and corresponding mechanical responses.

[0022] Preferably, the specific implementation steps for determining the activated expert model in the hybrid expert model within the gating network based on high-dimensional feature vectors and the introduction of physical rule masks in the physical perception gating module include: The high-dimensional feature vector is input into the gated neural network, which outputs a set of original weights for each expert. The current physical conditions are verified in real time based on a rule base of typical operating conditions of transmission lines, and a physical mask vector is generated. The final activation probability is calculated using the Softmax function based on the original weights and the physical mask. The expert model with the highest final activation probability or the final activation probability close to 1 is selected as the activated expert model. The hybrid expert models include: the dancing expert model, the vibration expert model, and the de-icing expert model.

[0023] Preferably, the tower-line coupling graph network module is specifically used for: A graph neural network is used to abstract the transmission tower into a topological graph containing the tower body component nodes and the connection relationships between the components; Dynamic tension is injected as message data into the graph neural network; Graph neural networks transmit the injected dynamic tension in the transmission tower topology through the message passing mechanism of graph convolution, simulating the propagation process of dynamic tension in the tower components and outputting the real-time stress and deformation values ​​of all components in the entire tower. Align the real-time stress and deformation values ​​of all components of the tower with the complete state of the conductor on the time axis to obtain the tower-conductor system state matrix.

[0024] Preferably, the multidimensional activation feature encoding module is specifically used for: The static parameters of the target line are obtained based on the ledger information. These static parameters include: conductor mass per unit length, bending stiffness, number of split conductors, span, and tower height. The static parameters are mapped to static embedding vectors through a fully connected layer; Physical transformations are performed on real-time micro-meteorological data to form physical codes for dynamic environmental characteristics; The static embedding vector and the physical encoding are concatenated to obtain a high-dimensional feature vector; The multi-source heterogeneous monitoring data includes: ledger information and real-time micro-meteorological data.

[0025] Preferably, the final activation probability is calculated using the following formula:

[0026] in, The physical guiding coefficient. The final activation probability, To normalize the weights, For physical mask, This is a commonly used activation function in machine learning.

[0027] Preferably, the nodes representing wire attachment points in the graph neural network are updated according to the following formula:

[0028] In the formula, This is the updated stress state vector of the node. This represents the stress state vector of the current layer's neighbor node u. Let v be the stress state vector of the current node v. Let v be the set of neighbors of node v. Let k be the aggregation function, and k be the current iteration level. This is the activation function.

[0029] Preferably, it further includes: a report generation module, used for: A comprehensive mechanical state report is generated based on the state matrix of the tower line system, dynamic tension time history curves, and real-time stress and deformation values ​​of all components.

[0030] In another aspect, this application also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a method for predicting the mechanical response of a tower line system based on a hybrid expert network, as described above, is realized.

[0031] In another aspect, this application also provides a computer-readable storage medium having an executable program stored thereon, which, when executed, implements a prediction of the mechanical response of a tower line system based on a hybrid expert network as described above.

[0032] Compared with the prior art, the beneficial effects of this application are as follows: This application provides a method for predicting the mechanical response of a transmission tower system based on a hybrid expert network. The method includes: converting multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand; determining the activated expert models in the hybrid expert model through a gating network with physical rule masks based on the high-dimensional feature vectors; selecting a solution strategy based on the scenario of the transmission tower system and using the solution strategy to solve the activated expert models, obtaining the dynamic tension time history curve of the conductor end to the tower suspension point; using a graph neural network to abstract the transmission tower into a topological graph containing tower component nodes and component connection relationships, updating the nodes representing the conductor suspension points in the graph neural network based on the dynamic tension in the dynamic tension time history curve, and using the graph neural network to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure, outputting the state matrix of the transmission tower system; wherein, the multi-source heterogeneous monitoring data includes: static parameters of the line and real-time micro-meteorological data; the hybrid expert model is obtained by training a deep learning model based on historical hydrodynamic features and corresponding mechanical responses. This application employs different solution strategies for different scenarios, achieving decoupled solutions for different physical phenomena. It uses a physical rule masking mechanism to eliminate the physical illusion of AI models predicting the dancing phenomenon under non-dancing conditions, and utilizes graph neural networks to achieve real-time linkage simulation from line to tower. Attached Figure Description

[0033] Figure 1The flowchart of the method for predicting the mechanical response of a tower line system based on a hybrid expert network is shown in this application. Figure 2 This is a flowchart illustrating the overall process framework of the method for predicting the mechanical response of the transmission tower system in this application. Figure 3 This is a schematic diagram of the physical sensing gating and physical masking mechanism of this application; Figure 4 This is a schematic diagram of an electronic device structure according to this application. Detailed Implementation

[0034] This application primarily addresses the following three problems existing in current mechanical simulation technologies for power transmission tower systems: Computational efficiency issue: This addresses the problem that traditional finite element methods, when dealing with fully coupled models, suffer from excessively long computation times due to high complexity, making it impossible to meet the requirements for real-time disaster early warning at the minute level for power grids.

[0035] Multi-mechanism adaptability problem: Solving the problem that existing single artificial intelligence models cannot simultaneously and accurately fit dynamic phenomena with different physical mechanisms such as dancing, wind vibration and de-icing.

[0036] Tower-to-line coupling accuracy issue: This addresses the problem that traditional methods calculate conductors and towers separately, neglecting the cascading damage to towers caused by the dynamic unbalanced tension of conductors, and failing to simultaneously obtain the mechanical state of the tower-to-line system.

[0037] This application provides an adaptive dynamic response prediction architecture based on a hybrid expert network using physical mechanisms. This architecture automatically matches the most suitable dynamic solver based on multi-dimensional environmental excitations through a physical perception gating mechanism, and realizes the mechanical transmission from the conductor to the tower through a graph neural network, thereby achieving efficient, accurate, and adaptive prediction of the full-element response from environmental excitation to the tower-conductor system.

[0038] To better understand this application, the content of this application will be further described below in conjunction with the accompanying drawings and embodiments.

[0039] Example 1: A method for predicting the mechanical response of a tower-line system based on a hybrid expert network, such as... Figure 1 As shown, it includes: Step 1: Transform multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand; Step 2: Based on high-dimensional feature vectors, determine the activated expert models in the hybrid expert model through a gated network that incorporates physical rule masks; Step 3: Select a solution strategy based on the scenario of the transmission tower system, and use the solution strategy to solve the activated expert model to obtain the dynamic tension time history curve of the conductor end to the tower hanging point; Step 4: Use a graph neural network to abstract the transmission tower into a topology graph containing tower body component nodes and component connection relationships. Update the nodes representing conductor suspension points in the graph neural network based on the dynamic tension in the dynamic tension time history curve. Use the graph neural network to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure and output the tower-line system state matrix. Among them, the multi-source heterogeneous monitoring data includes: static parameters of the line and real-time micro-meteorological data; The hybrid expert model is obtained by training a deep learning model based on historical fluid dynamics features and corresponding mechanical responses.

[0040] like Figure 2 As shown, this application solves the problem that traditional single models cannot take into account the multiple complex dynamic mechanisms of transmission lines by constructing a three-layer architecture of "physical perception gating - heterogeneous expert solving - graph network coupling". Adaptive and accurate prediction is achieved through the following four levels of technical means.

[0041] 1. Physical Feature Perception Layer: This application does not directly use the raw sensor data as input. Instead, it introduces prior knowledge from the fields of fluid mechanics and structural mechanics to "upgrade" the raw meteorological and structural data into dimensionless feature vectors with clear physical meaning. By calculating key factors such as Reynolds number and icing eccentricity, the neural network can perceive potential physical instabilities at the input end.

[0042] 2. Gated Network Decision Layer: This is the core scheduling unit of this system. Unlike traditional hybrid expert networks that rely solely on statistical data patterns for task distribution, this application designs a physical rule masking mechanism. This mechanism transforms classical mechanical criteria into neural network constraints, forcing the gated network to activate the corresponding expert model when specific physical conditions are met, thereby eliminating the physical illusion of AI models predicting dancing phenomena under non-dancing conditions.

[0043] 3. Hybrid Expert Solver Layer: This application employs different solution strategies for different scenarios. For low-frequency, large-displacement galloping, a physical information neural network with embedded aerodynamic equations is used; for high-frequency, small-amplitude wind vibrations, a convolutional neural network adept at spectral analysis is used; and for de-icing due to transient impacts, a long short-term memory network with temporal memory capabilities is used. These expert models work in parallel, achieving decoupled solutions for different physical phenomena.

[0044] 4. Tower-Line Coupling Assessment Layer: To assess the damage caused by conductor movement to the tower, this application utilizes the non-Euclidean spatial data processing capabilities of graph neural networks to model the transmission tower as a topological graph. By injecting the dynamic tension calculated from the conductor as a message into the graph network, and using the message passing mechanism of graph convolution, the propagation process of force in the tower components is simulated, thereby achieving real-time linkage simulation from "line" to "tower".

[0045] The following example illustrates the specific implementation steps of this application using a transmission line experiencing a combined "icing + strong wind" galloping condition during a winter cold wave.

[0046] Step 1: Transform the multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand, including: The static parameters of the target line are obtained based on the ledger information. These static parameters include: conductor mass per unit length, bending stiffness, number of split conductors, span, and tower height. The static parameters are mapped to static embedding vectors through a fully connected layer; Physical transformations are performed on real-time micro-meteorological data to form physical codes for dynamic environmental characteristics; The static embedding vector and the physical encoding are concatenated to obtain a high-dimensional feature vector; The multi-source heterogeneous monitoring data includes: ledger information and real-time micro-meteorological data.

[0047] Step 1 specifically includes: constructing a multidimensional excitation feature space and physical encoding.

[0048] The purpose of this step is to transform multi-source, heterogeneous monitoring data into high-dimensional feature vectors that are understandable to the model and contain physical laws. The system will extract static parameters of the target line based on the log information, specifically including: conductor mass per unit length, bending stiffness, number of split conductors, span, and tower nominal height. These parameters will be mapped into static embedding vectors through a fully connected layer. Simultaneously, the system will also perform physical transformations on the real-time collected micrometeorological data to form dynamic environmental feature physical codes. For fluid characteristics, the system will calculate the Reynolds number based on real-time wind speed and conductor diameter. With Storoha This serves as a key indicator determining fluid dynamics behavior; for icing conditions, image recognition technology is used to extract the profile of the icing cross-section of the conductor and calculate the cross-sectional eccentricity factor. Finally, the static vector will be concatenated with the dynamic physical code to form the system input tensor. .

[0049] Step 2: Based on high-dimensional feature vectors, through the introduction of physical rule masks in the gated network, determine the activated expert models in the hybrid expert model, including: The high-dimensional feature vector is input into the gated neural network, which outputs a set of original weights for each expert. The current physical conditions are verified in real time based on a rule base of typical operating conditions of transmission lines, and a physical mask vector is generated. The final activation probability is calculated using the Softmax function based on the original weights and the physical mask. The expert model with the highest final activation probability or the final activation probability close to 1 is selected as the activated expert model. The hybrid expert models include: the dancing expert model, the vibration expert model, and the de-icing expert model.

[0050] The galloping expert model can predict the galloping trajectory, galloping amplitude, galloping frequency, and the dynamic tension time history of the conductor caused by galloping, as well as the alternating bending moment and stress generated by galloping on the tower material. This application uses the galloping trajectory of the conductor as an example to introduce the model.

[0051] Vibration expert models can predict the amplitude and frequency of conductor vibration in light wind and secondary span vibration, high-frequency dynamic tension changes caused by vibration, vibration stress and fatigue damage at hanging points, tower nodes, and clamps.

[0052] The de-icing expert model can predict the jump height, jump impact velocity, peak impact tension generated at the moment of de-icing, and data such as instantaneous impact load and internal force change on the tower caused by de-icing.

[0053] Furthermore, the final activation probability is calculated using the following formula:

[0054] in, The physical guiding coefficient. The final activation probability, To normalize the weights, For physical mask, This is a commonly used activation function in machine learning.

[0055] Step 2 specifically includes: a gated network based on physical rule masks, such as... Figure 3 As shown.

[0056] This step utilizes a gating network to determine which expert model should be activated under the current condition. To ensure the physical correctness of the decision, a physical rule mask is introduced. Tensor First, a set of original normalized weights is output through the gated neural network. These correspond to different experts in areas such as galloping, vibration, and de-icing. Simultaneously, the system will verify the current physical conditions in real time based on a rule base of typical transmission line operating conditions, determining the state of different operating conditions and generating physical mask vectors based on parameters such as icing eccentricity, wind speed, and wind direction. Finally, the original weights are combined with the physical mask, and the final activation probability is calculated using the Softmax function.

[0057] in, This is the physical guidance coefficient. Through this step, the system will mathematically push the activation weight of the dance expert close to 1.0, thereby shielding the noise interference generated by other models at this time, thus achieving physical-driven attention focus.

[0058] Step 3: Select a solution strategy based on the scenario of the transmission tower system, and use the solution strategy to solve the activated expert model to obtain the dynamic tension time history curve of the conductor end to the tower suspension point, including: When the transmission tower system is in a low-frequency, large-displacement galloping scenario, the solution strategy is a physical information neural network with embedded aerodynamic equations. When the transmission tower system is in a scenario of high-frequency, low-amplitude wind vibration, the solution strategy is a convolutional neural network that is good at spectrum analysis. When the transmission tower system is in a transient impact de-icing scenario, the solution strategy is a long short-term memory network with temporal memory capability. The activated expert model is solved using the aforementioned solution strategy to obtain the dynamic tension time history curve of the conductor end to the tower suspension point.

[0059] Step 3 specifically includes: sparse solution of heterogeneous hybrid expert model.

[0060] Based on the weights determined in step 2, the system activates the corresponding experts to perform high-precision mechanical calculations, while other experts remain in a dormant or low-power state. For example, activating the "dancing" expert will employ a physical information neural network, embedding a physical residual layer, containing nonlinear partial differential equations related to conductor motion. The physical neural network will receive current wind field and icing parameters, deduce the conductor's three-dimensional spatial trajectory and torsion angle over a future period, accurately simulate the conductor's low-frequency, large-amplitude elliptical trajectory motion, and ultimately output the dynamic tension time-history curve of the conductor's end relative to the tower suspension point. The periodic fluctuation amplitude and frequency of tension, including tension, are used as interface data for subsequent tower-line coupling calculations.

[0061] Step 4: Using a graph neural network, the transmission tower is abstracted into a topology graph containing tower component nodes and component connection relationships. The nodes representing conductor suspension points in the graph neural network are updated based on the dynamic tension in the dynamic tension time history curve. A multi-layer message passing algorithm is then used to simulate the force transmission process within the structure, outputting the tower-conductor system state matrix, including: Dynamic tension is injected as message data into the graph neural network; Graph neural networks transmit the injected dynamic tension in the transmission tower topology through the message passing mechanism of graph convolution, simulating the propagation process of dynamic tension in the tower components and outputting the real-time stress and deformation values ​​of all components in the entire tower. Align the real-time stress and deformation values ​​of all components of the tower with the complete state of the conductor on the time axis to obtain the tower-conductor system state matrix.

[0062] Update the nodes representing wire attachment points in the graph neural network according to the following formula:

[0063] In the formula, This is the updated stress state vector of the node. This represents the stress state vector of the current layer's neighbor node u. Let v be the stress state vector of the current node v. Let v be the set of neighbors of node v. Let k be the aggregation function, and k be the current iteration level. This is the activation function.

[0064] Step 4 specifically includes: tower-line coupling calculation based on graph neural network.

[0065] Abstracting the transmission tower into a single diagram ,node Geometric nodes representing the tower structure, each node containing its spatial coordinates, degrees of freedom under forces, and other characteristics. (Edge) The angle steel or steel pipe component representing the connection node includes material properties such as cross-sectional area, Young's modulus, and yield strength. Based on the dynamic tension time history curve of the conductor end to the tower suspension point obtained in S3... The system updates the nodes representing conductor suspension points in the graph neural network as external stimulus messages, adding a time-varying external force to the graph model. Then, a multi-layer message passing algorithm is executed using the graph neural network to simulate the force transmission process within the structure. The stress state of the suspension point node is first transmitted to the crossarm node, causing crossarm deformation; the internal force of the crossarm is further transmitted to the main tower structure, and finally to the tower legs and foundation. Each layer of graph convolution operation aggregates the stress information of neighboring nodes, updating the stress state vector of the current node.

[0066] After multiple iterations, the graphical model will output the real-time stress and deformation values ​​of all components of the tower, align them with the complete state of the conductor on the time axis, and construct the state matrix of the tower-conductor system.

[0067] After obtaining the state matrix of the tower line system, this application also includes: generating a full-element mechanical state report based on the state matrix of the tower line system, the dynamic tension time history curve, and the real-time stress and deformation values ​​of all components.

[0068] The system then outputs a comprehensive mechanical state report, including the conductor galloping trajectory, tension fluctuation curve, and tower stress cloud diagram. The conductor galloping trajectory is output by the galloping expert model.

[0069] This application achieves adaptive decoupling of multiple physical mechanisms, improving the physical consistency of predictions: through physical gating mechanisms, this application can automatically identify the dominant dynamic mode based on physical factors such as icing eccentricity and Reynolds number, automatically adapt to various working conditions such as light wind, galloping, and de-icing, and activate expert models without manual intervention, thus solving the problem of poor generalization ability of a single model in mixed multi-mechanism scenarios.

[0070] This application achieves a significant improvement in real-time performance and computational efficiency: by adopting a sparse activation strategy, only lightweight experts are activated under normal weather conditions, and different high-precision physical experts are activated only under complex coupled scenarios. Compared with fully coupled finite element simulation, the inference time is reduced from hours to milliseconds.

[0071] This application realizes the safety assessment and fault early warning of tower-line cascade: by injecting the dynamic tension time history output by conductor experts as boundary conditions into the tower diagram network, real-time coupling from conductor dynamics to tower structural mechanics is realized, enabling the system to not only warn of conductor flashover, but also to accurately warn of tower torsion, buckling and tower collapse accidents caused by unbalanced tension.

[0072] Example 2 Based on the same inventive concept, this application also provides a tower line system mechanical response prediction system based on a hybrid expert network, including: The multidimensional activation feature encoding module is used to transform multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand. The physical perception gating module is used to determine the activated expert model in the hybrid expert model in a gating network based on high-dimensional feature vectors by introducing physical rule masks; and selects a solution strategy based on the scenario of the transmission tower system, and uses the solution strategy to solve the activated expert model to obtain the dynamic tension time history curve of the conductor end to the tower hanging point. The tower-line coupling graph network module is used to abstract the transmission tower into a topology graph containing tower body component nodes and component connection relationships using graph neural networks. It updates the nodes representing conductor suspension points in the graph neural network based on the dynamic tension in the dynamic tension time history curve, and uses the graph neural network to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure, and outputs the tower-line system state matrix. Among them, the multi-source heterogeneous monitoring data includes: static parameters of the line and real-time micro-meteorological data; The hybrid expert model is obtained by training a deep learning model based on historical fluid dynamics features and corresponding mechanical responses.

[0073] Furthermore, the specific implementation steps for determining the activated expert model in the hybrid expert model within the gating network based on high-dimensional feature vectors and by introducing physical rule masks in the physical perception gating module include: The high-dimensional feature vector is input into the gated neural network, which outputs a set of original weights for each expert. The current physical conditions are verified in real time based on a rule base of typical operating conditions of transmission lines, and a physical mask vector is generated. The final activation probability is calculated using the Softmax function based on the original weights and the physical mask. The expert model with the highest final activation probability or the final activation probability close to 1 is selected as the activated expert model. The hybrid expert models include: the dancing expert model, the vibration expert model, and the de-icing expert model.

[0074] Furthermore, the tower-line coupling graph network module is specifically used for: A graph neural network is used to abstract the transmission tower into a topological graph containing the tower body component nodes and the connection relationships between the components; Dynamic tension is injected as message data into the graph neural network; Graph neural networks transmit the injected dynamic tension in the transmission tower topology through the message passing mechanism of graph convolution, simulating the propagation process of dynamic tension in the tower components and outputting the real-time stress and deformation values ​​of all components in the entire tower. Align the real-time stress and deformation values ​​of all components of the tower with the complete state of the conductor on the time axis to obtain the tower-conductor system state matrix.

[0075] Furthermore, the multidimensional activation feature encoding module is specifically used for: The static parameters of the target line are obtained based on the ledger information. These static parameters include: conductor mass per unit length, bending stiffness, number of split conductors, span, and tower height. The static parameters are mapped to static embedding vectors through a fully connected layer; Physical transformations are performed on real-time micro-meteorological data to form physical codes for dynamic environmental characteristics; The static embedding vector and the physical encoding are concatenated to obtain a high-dimensional feature vector; The multi-source heterogeneous monitoring data includes: ledger information and real-time micro-meteorological data.

[0076] Furthermore, the final activation probability is calculated using the following formula:

[0077] in, The physical guiding coefficient. The final activation probability, To normalize the weights, For physical mask, This is a commonly used activation function in machine learning.

[0078] Furthermore, update the nodes representing wire attachment points in the graph neural network according to the following formula:

[0079] In the formula, This is the updated stress state vector of the node. This represents the stress state vector of the current layer's neighbor node u. Let v be the stress state vector of the current node v. Let v be the set of neighbors of node v. Let k be the aggregation function, and k be the current iteration level. This is the activation function.

[0080] Furthermore, it also includes: a report generation module, used for: A comprehensive mechanical state report is generated based on the state matrix of the tower line system, dynamic tension time history curves, and real-time stress and deformation values ​​of all components.

[0081] In summary, the method for predicting the mechanical response of a tower-line system based on a hybrid expert network proposed in this application includes a multi-dimensional excitation feature encoding module, a physical perception gating module, a heterogeneous expert model pool, and a tower-line coupled graph network module. Through a gating mechanism, specific dynamic expert models are adaptively activated according to environmental and structural characteristics, and the mechanical transmission between tower-line systems is realized through a graph network.

[0082] This application relates to a physical rule-aware gating mechanism: unlike traditional data-driven gating, this mechanism introduces a physical rule mask during weight generation. Based on prior criteria from fluid dynamics and structural dynamics (such as the wind speed range where galloping occurs, the threshold for icing asymmetry, etc.), the weights output by the neural network are modified and constrained to ensure the physical correctness of the model's dynamic pattern recognition.

[0083] This application relates to a method for constructing a heterogeneous hybrid expert model pool: constructing neural network experts with different structures for different disaster mechanisms of transmission lines. These include nonlinear aeroelastic experts based on physical information neural networks, micro-wind vibration experts based on frequency domain convolutional networks, and transient shock experts based on recurrent neural networks, in order to solve the problem that a single model cannot adapt to multiple physical mechanisms.

[0084] Example 3 like Figure 4As shown, this application also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0085] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the tower line system mechanical response prediction method based on a hybrid expert network in the above embodiments.

[0086] Example 4 Based on the same inventive concept, this application also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the tower-line system mechanical response prediction method based on a hybrid expert network in the above embodiments.

[0087] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] The above are merely embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of the claims of this application.

Claims

1. A method for predicting the mechanical response of a tower-line system based on a hybrid expert network, characterized in that, include: Transform multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand; In a gated network based on high-dimensional feature vectors and by introducing physical rule masks, the activated expert models in the hybrid expert model are determined. Based on the scenario of the transmission tower system, a solution strategy is selected, and the activated expert model is solved using the solution strategy to obtain the dynamic tension time history curve of the conductor end to the tower hanging point. A graph neural network is used to abstract the transmission tower into a topology graph containing tower body component nodes and component connection relationships. The nodes representing conductor suspension points in the graph neural network are updated based on the dynamic tension in the dynamic tension time history curve. The graph neural network is used to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure and output the tower-line system state matrix. Among them, the multi-source heterogeneous monitoring data includes: static parameters of the line and real-time micro-meteorological data; The hybrid expert model is obtained by training a deep learning model based on historical fluid dynamics features and corresponding mechanical responses.

2. The method as described in claim 1, characterized in that, The process of converting multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand includes: The static parameters of the target line are obtained based on the ledger information. These static parameters include: conductor mass per unit length, bending stiffness, number of split conductors, span, and tower height. The static parameters are mapped to static embedding vectors through a fully connected layer; Physical transformations are performed on real-time micro-meteorological data to form physical codes for dynamic environmental characteristics; The static embedding vector and the physical encoding are concatenated to obtain a high-dimensional feature vector; The multi-source heterogeneous monitoring data includes: ledger information and real-time micro-meteorological data.

3. The method as described in claim 1, characterized in that, The gating network based on high-dimensional feature vectors and the introduction of physical rule masks determines the activated expert models in the hybrid expert model, including: The high-dimensional feature vector is input into the gated neural network, which outputs a set of original weights for each expert. The current physical conditions are verified in real time based on a rule base of typical operating conditions of transmission lines, and a physical mask vector is generated. The final activation probability is calculated using the Softmax function based on the original weights and the physical mask. The expert model with the highest final activation probability or the final activation probability close to 1 is selected as the activated expert model. The hybrid expert models include: the dancing expert model, the vibration expert model, and the de-icing expert model.

4. The method as described in claim 1, characterized in that, The final activation probability is calculated using the following formula: in, The physical guiding coefficient. The final activation probability, To normalize the weights, For physical mask, This is a commonly used activation function in machine learning.

5. The method as described in claim 1, characterized in that, The method involves updating the nodes representing conductor suspension points in a graph neural network based on the dynamic tension time history curve, and using the graph neural network to execute a multi-layer message passing algorithm to simulate the force transmission process within the structure, outputting a tower-conductor system state matrix, including: Dynamic tension is injected as message data into the graph neural network; Graph neural networks transmit the injected dynamic tension in the transmission tower topology through the message passing mechanism of graph convolution, simulating the propagation process of dynamic tension in the tower components and outputting the real-time stress and deformation values ​​of all components in the entire tower. Align the real-time stress and deformation values ​​of all components of the tower with the complete state of the conductor on the time axis to obtain the tower-conductor system state matrix.

6. The method as described in claim 5, characterized in that, Update the nodes representing wire attachment points in the graph neural network according to the following formula: In the formula, This is the updated stress state vector of the node. This represents the stress state vector of the current layer's neighbor node u. Let v be the stress state vector of the current node v. Let v be the set of neighbors of node v. Let k be the aggregation function, and k be the current iteration level. This is the activation function.

7. The method as described in claim 5, characterized in that, After obtaining the tower line system state matrix, the following is also included: A comprehensive mechanical state report is generated based on the state matrix of the tower line system, dynamic tension time history curves, and real-time stress and deformation values ​​of all components.

8. The method as described in claim 1, characterized in that, The solution strategy based on the scenario in which the transmission tower system is located includes: When the transmission tower system is in a low-frequency, large-displacement galloping scenario, the solution strategy is a physical information neural network with embedded aerodynamic equations. When the transmission tower system is in a scenario of high-frequency, low-amplitude wind vibration, the solution strategy is a convolutional neural network that is good at spectrum analysis. When the transmission tower system is in a transient impact de-icing scenario, the solution strategy is a long short-term memory network with temporal memory capabilities.

9. A system for predicting the mechanical response of a tower-line system based on a hybrid expert network, characterized in that, include: The multidimensional activation feature encoding module is used to transform multi-source heterogeneous monitoring data into high-dimensional feature vectors that the model can understand. The physical perception gating module is used to determine the activated expert model in the hybrid expert model in a gating network based on high-dimensional feature vectors by introducing physical rule masks; and selects a solution strategy based on the scenario of the transmission tower system, and uses the solution strategy to solve the activated expert model to obtain the dynamic tension time history curve of the conductor end to the tower hanging point. The tower-line coupling graph network module is used to abstract the transmission tower into a topology graph containing tower body component nodes and component connection relationships using graph neural networks. It updates the nodes representing conductor suspension points in the graph neural network based on the dynamic tension in the dynamic tension time history curve, and uses the graph neural network to execute a multi-layer message passing algorithm to simulate the force transmission process inside the structure, and outputs the tower-line system state matrix. Among them, the multi-source heterogeneous monitoring data includes: static parameters of the line and real-time micro-meteorological data; The hybrid expert model is obtained by training a deep learning model based on historical fluid dynamics features and corresponding mechanical responses.

10. The system as described in claim 9, characterized in that, The specific implementation steps for determining the activated expert model in the hybrid expert model within the gating network based on high-dimensional feature vectors and the introduction of physical rule masks in the physical perception gating module include: The high-dimensional feature vector is input into the gated neural network, which outputs a set of original weights for each expert. The current physical conditions are verified in real time based on a rule base of typical operating conditions of transmission lines, and a physical mask vector is generated. The final activation probability is calculated using the Softmax function based on the original weights and the physical mask. The expert model with the highest final activation probability or the final activation probability close to 1 is selected as the activated expert model. The hybrid expert models include: the dancing expert model, the vibration expert model, and the de-icing expert model.

11. The system as described in claim 9, characterized in that, The tower-line coupling diagram network module is specifically used for: Dynamic tension is injected as message data into the graph neural network; Graph neural networks transmit the injected dynamic tension in the transmission tower topology through the message passing mechanism of graph convolution, simulating the propagation process of dynamic tension in the tower components and outputting the real-time stress and deformation values ​​of all components in the entire tower. Align the real-time stress and deformation values ​​of all components of the tower with the complete state of the conductor on the time axis to obtain the tower-conductor system state matrix.

12. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a method for predicting the mechanical response of a tower line system based on a hybrid expert network as described in any one of claims 1 to 8 is implemented.

13. A readable storage medium, characterized in that, It contains an execution program, which, when executed, implements a method for predicting the mechanical response of a tower line system based on a hybrid expert network as described in any one of claims 1 to 8.

Images