Power transmission line spacer anti-vibration scheme generation method and system

Abstract

The invention discloses a power transmission line spacer anti-vibration scheme generation method and system, which are used for solving the technical problem that the quality of a finally generated scheme is relatively poor due to the fact that an existing power transmission line spacer anti-vibration scheme generation method is difficult to accurately predict oscillation response under a complex working condition. The method comprises the following steps: acquiring field detection data of a power transmission line, constructing an input parameter space, and generating a target training data set through Latin hypercube sampling; constructing a physical information neural network as an initial amplitude prediction model and an initial stability classification model, and training based on the data set to obtain a target amplitude prediction model and a target stability classification model; and constructing a multi-objective optimization function and setting constraint conditions, solving the multi-objective optimization function according to the reference state of the power transmission line under the constraint by adopting a deep Q network or a non-dominated sorting genetic algorithm based on the model, generating a Pareto optimal solution set, and finally outputting a power transmission line spacer anti-vibration scheme.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission line engineering and anti-vibration safety, and particularly relates to a power transmission line spacer rod anti-vibration scheme generation method and system. BACKGROUND

[0002] The subspan oscillation is a unique self-excited vibration phenomenon of the multi-bundled conductor under the action of wind excitation, and is closely related to the wake effect. The wake field formed by the disturbed airflow of the upwind conductor can periodically aerodynamically excite the downwind conductor, thereby causing the sustained vibration of the conductor in the vertical direction or the torsional direction. The frequency of this vibration is usually between 0.1-10 Hz, and the amplitude is between several centimeters and several tens of centimeters. Although this vibration does not have the high-frequency characteristics of the breeze vibration and the large-amplitude characteristics of the galloping, it can still act on the conductor, the spacer rod and the line hardware for a long time, leading to a series of faults such as conductor fatigue fracture, spacer rod loosening and falling off, and hardware wear and cracking. In severe cases, it can even cause conductor interphase flashover, tower tilting and other major safety accidents, directly threatening the long-term stable operation of the power transmission line.

[0003] With the development of power transmission lines towards the direction of extra-high voltage and ultra-high voltage, the number of conductor bundles increases, the span distance increases, and the conductor cross section is optimized. The inducing conditions of the subspan oscillation become more complex, and small changes in parameters such as wind speed, wind direction angle, conductor tension and bundle spacing can significantly increase the oscillation intensity, further increasing the difficulty of protection design.

[0004] Most of the existing power transmission line spacer rod anti-vibration scheme generation methods are based on the calculation method of a simplified physical model. By ignoring complex factors such as wake interference and nonlinear aerodynamic effects, a simplified mechanical equation is established for solving. However, the model applied in this way has limited accuracy, and it is difficult to accurately predict the oscillation response under complex working conditions, resulting in poor quality of the final generated scheme. SUMMARY

[0005] The present application provides a power transmission line spacer rod anti-vibration scheme generation method and system, which is used to solve the technical problem that the existing power transmission line spacer rod anti-vibration scheme generation method is difficult to accurately predict the oscillation response under complex working conditions, resulting in poor quality of the final generated scheme.

[0006] The first aspect of the present application provides a power transmission line spacer rod anti-vibration scheme generation method, comprising:

[0007] Obtaining the field detection data of the power transmission line, and constructing an input parameter space containing the power transmission line spacer rod anti-vibration related data;

[0008] Generating a target training data set according to the field detection data and the input parameter space by using a Latin hypercube sampling method;

[0009] constructing a physical information neural network as an initial amplitude prediction model and an initial stability classification model, and training the initial amplitude prediction model and the initial stability classification model based on the target training data set to output a target amplitude prediction model and a target stability classification model;

[0010] constructing a multi-objective optimization function and setting a constraint condition;

[0011] solving the multi-objective optimization function according to the benchmark state of the power transmission line under the constraint condition based on the target amplitude prediction model and the target stability classification model by using a deep Q network or a non-dominated sorting genetic algorithm to generate a Pareto optimal solution set;

[0012] outputting a power transmission line spacer anti-vibration scheme based on the Pareto optimal solution set.

[0013] Optionally, the Latin hypercube sampling method is used to generate a target training data set according to the field detection data and the input parameter space, including:

[0014] generating a plurality of sample points in the input parameter space based on the Latin hypercube sampling method;

[0015] performing parameterized computational fluid dynamics simulation on each sample point to obtain a corresponding sub-span oscillation response output of each sample point;

[0016] constructing an initial training data set according to the sub-span oscillation response output corresponding to each sample point;

[0017] supplementing and verifying the initial training data set by using the field detection data to output a target training data set.

[0018] Optionally, the training of the initial amplitude prediction model and the initial stability classification model based on the target training data set to output a target amplitude prediction model and a target stability classification model includes:

[0019] extracting a plurality of initial power transmission line features in the target training data set;

[0020] screening each initial feature by using a random forest and a mutual information method to obtain a plurality of target power transmission line features;

[0021] training the initial amplitude prediction model according to a plurality of target power transmission line features by using a preset loss function to determine a target amplitude prediction model;

[0022] training the initial stability classification model according to a plurality of target power transmission line features by using a cross-validation method to determine a target stability classification model.

[0023] Optionally, based on the target amplitude prediction model and the target stability classification model, a deep Q network is used to solve the multi-objective optimization function under the constraint condition according to the reference state of the power transmission line, and a Pareto optimal solution set is generated, comprising:

[0024] The deep Q network is used to solve the multi-objective optimization function under the constraint condition according to the reference state of the power transmission line, and a plurality of candidate scheme actions and a Q value corresponding to each candidate scheme action are output.

[0025] In each of the candidate scheme actions, the power transmission line features corresponding to each of the candidate scheme actions are extracted and input into the target amplitude prediction model, and the predicted maximum amplitudes corresponding to each of the candidate scheme actions are output.

[0026] The power transmission line features corresponding to each of the candidate scheme actions are input into the target stability classification model, and the oscillation stability state results corresponding to each of the candidate scheme actions are output.

[0027] According to the number of spacer bars in each of the candidate scheme actions, and the corresponding oscillation stability state results and predicted maximum amplitudes, the total reward values corresponding to each of the candidate scheme actions are calculated, and the iteration times are counted in real time.

[0028] It is judged whether the iteration times reach a preset number threshold.

[0029] If yes, the total reward values are sorted in descending order, and the candidate scheme actions corresponding to the top preset number of total reward values constitute a Pareto optimal solution set.

[0030] Optionally, it further comprises:

[0031] If the iteration times do not reach the preset number threshold, the deep Q network is updated according to the Q value corresponding to each of the candidate scheme actions, and a new deep Q network is determined.

[0032] The candidate scheme action corresponding to the maximum total reward value is selected and executed, the new reference state of the power transmission line is output, and the step of using the deep Q network to solve the multi-objective optimization function under the constraint condition according to the reference state of the power transmission line is executed again until the iteration times reach the preset number threshold.

[0033] The plurality of total reward values determined when the iteration times reach the preset number threshold are sorted in descending order, and the candidate scheme actions corresponding to the top preset number of total reward values constitute a Pareto optimal solution set.

[0034] Optionally, based on the target amplitude prediction model and the target stability classification model, a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function under the constraint condition according to the reference state of the power transmission line, and a Pareto optimal solution set is generated, including:

[0035] The non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function under the constraint condition according to the reference state of the power transmission line, and an initial population is output, the initial population including a plurality of population individuals, each population individual corresponding to a candidate scheme;

[0036] In each population individual, the power transmission line characteristics corresponding to each population individual are extracted and input into the target amplitude prediction model respectively, and the predicted maximum amplitude corresponding to each population individual is output;

[0037] The power transmission line characteristics corresponding to each population individual are input into the target stability classification model, and the oscillation stability state result corresponding to each population individual is output;

[0038] According to the predicted maximum amplitude and the number of spacer bars corresponding to each population individual, the target function value corresponding to each population individual is calculated;

[0039] Based on the target function value corresponding to each population individual, the Pareto dominance sorting of each population individual is performed, and an initial non-dominated solution set is obtained;

[0040] Based on the oscillation stability state result corresponding to each population individual, the initial non-dominated solution set is filtered to obtain a target non-dominated solution set, and the number of iterations is counted in real time;

[0041] The target non-dominated solution set is simulated by binary crossover and polynomial mutation, and is combined with the initial population to output a combined population;

[0042] Based on the combined population, a plurality of new population individuals are determined, and it is judged whether the number of iterations reaches a preset number threshold;

[0043] If yes, the plurality of new population individuals form a Pareto optimal solution set.

[0044] If no, the step of extracting the power transmission line characteristics corresponding to each population individual in each population individual and inputting into the target amplitude prediction model respectively to output the predicted maximum amplitude corresponding to each population individual is executed until the number of iterations reaches the preset number threshold.

[0045] The second aspect of the application provides a power transmission line spacer anti-vibration scheme generation system, including:

[0046] The acquisition module is used to acquire on-site inspection data of transmission lines and construct an input parameter space containing vibration-damping data of transmission line spacers;

[0047] A module is used to generate a target training dataset based on the field detection data and the input parameter space using the Latin hypercube sampling method;

[0048] The training module is used to construct a physical information neural network as an initial amplitude prediction model and an initial stability classification model, and to train the initial amplitude prediction model and the initial stability classification model based on the target training dataset, and output the target amplitude prediction model and the target stability classification model.

[0049] The building module is used to construct multi-objective optimization functions and set constraints.

[0050] The generation module is used to solve the multi-objective optimization function based on the target amplitude prediction model and the target stability classification model, using a deep Q-network or a non-dominated sorting genetic algorithm under the constraints, according to the reference state of the transmission line, to generate a Pareto optimal solution set.

[0051] The output module is used to output a vibration prevention scheme for transmission line spacers based on the Pareto optimal solution set.

[0052] A computer device provided by a third aspect of the present invention includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the method for generating a vibration damping scheme for transmission line spacers as described in any of the preceding claims.

[0053] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the steps of the method for generating a vibration damping scheme for transmission line spacers as described in any of the preceding claims.

[0054] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the steps of the method for generating a vibration damping scheme for transmission line spacers as described in any of the preceding claims.

[0055] As can be seen from the above technical solutions, the present invention has the following advantages:

[0056] The above-mentioned technical solution of the present invention provides a method for generating a vibration damping scheme for transmission line spacers. This method involves acquiring field inspection data of the transmission line and constructing an input parameter space containing relevant data on vibration damping of transmission line spacers; using the Latin hypercube sampling method to generate a target training dataset based on the field inspection data and the input parameter space; constructing a physical information neural network as an initial amplitude prediction model and an initial stability classification model; and training the initial amplitude prediction model and the initial stability classification model based on the target training dataset to output a target amplitude prediction model and a target stability classification model; constructing a multi-objective optimization function and setting approximately... The invention employs a constraint condition-based approach. Based on the target amplitude prediction model and the target stability classification model, a deep Q-network or a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function under the baseline state of the transmission line, generating a Pareto optimal solution set. Based on the Pareto optimal solution set, a vibration prevention scheme for the transmission line spacer is output. Based on this scheme, the invention deeply integrates field measured data, efficient sampling technology, and a physical information neural network, enabling the model to fully learn the oscillation physical mechanism and data patterns under complex working conditions during training. This significantly improves the accuracy of oscillation response prediction for complex working conditions with multiple coupled factors, thereby enhancing the quality of the final generated scheme. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating the steps of a method for generating a vibration damping scheme for transmission line spacers according to Embodiment 1 of the present invention.

[0059] Figure 2 This is a schematic diagram of Latin hypercube sampling (two-dimensional example) provided in Embodiment 1 of the present invention;

[0060] Figure 3 This is a schematic diagram of the physical information neural network (PINN) structure provided in Embodiment 1 of the present invention;

[0061] Figure 4 This is a schematic diagram of the training process of the Physical Information Neural Network (PINN) provided in Embodiment 1 of the present invention;

[0062] Figure 5 This is a schematic diagram of the Pareto front generated by multi-objective optimization according to Embodiment 1 of the present invention;

[0063] Figure 6This is an overall framework diagram of a method for generating a vibration damping scheme for transmission line spacers provided in Embodiment 1 of the present invention;

[0064] Figure 7 A schematic diagram of the Pareto front provided in Embodiment 1 of the present invention;

[0065] Figure 8 This is a structural block diagram of a vibration damping scheme generation system for transmission line spacer bars provided in Embodiment 2 of the present invention; Detailed Implementation

[0066] This invention provides a method and system for generating vibration damping schemes for transmission line spacers, which solves the technical problem that existing methods for generating vibration damping schemes for transmission line spacers are unable to accurately predict oscillation responses under complex operating conditions, resulting in poor quality of the final generated schemes.

[0067] Terminology Explanation:

[0068] Latin Hypercube Sampling (LHS) is a hierarchical random sampling technique and a crucial Design of Experiment (DoE) method in computer experiments and simulation modeling. Its core objective is to efficiently cover a multidimensional parameter space with as few sample points as possible, while exhibiting good space-filling and projection properties.

[0069] The sampling approach is as follows: Assume the system has m factors, and each factor has n levels. First, the design space for each factor will be divided into n subspaces. A value is randomly selected within each subspace, resulting in a sampling matrix with n samples for each design space. The sampling process must adhere to two principles: first, the sample points within each sub-design space must be randomly selected; second, exactly one value is selected within each sub-design space.

[0070] Suppose there are k input parameters that need to be sampled (e.g., wind speed, tension, splitting distance), and the value range of each parameter is normalized to the interval [0,1]. This invention aims to generate n sample points.

[0071] Algorithm flow:

[0072] 1) Layering: Divide the range [0,1] of each dimension (each parameter) into n non-overlapping intervals with equal probability. The length of each interval is 1 / n, i.e.: [0,1 / n), [1 / n,2 / n),...,[(n-1) / n,1].

[0073] 2) Random Sampling: For each dimension, randomly and without repetition, draw one value from each of the n intervals. Ensure that there is exactly one sample point in each interval. This process is performed independently in all k dimensions.

[0074] 3) Random pairing: At this point, a k×n matrix is ​​obtained. Each row represents a parameter, containing n random values ​​from different intervals. The n values ​​in each parameter row are randomly arranged (shuffled). The shuffled columns 1, 2...n are combined to form the final n k-dimensional sample points Pi=(x1,x2,...,xk).

[0075] The X and Y axes are each divided into 5 intervals. There is exactly one sample point in each row and column (i.e., each interval of each parameter), which ensures the uniformity of the projection and the space filling.

[0076] Latin hypercube sampling is an advanced experimental design method that ensures each dimension of the input parameter is explored uniformly through hierarchical and randomized permutations. In machine learning-based sub-segment oscillation studies, it is a crucial first step in generating high-quality, low-bias training datasets, enabling the use of a limited number of CFD simulations to lay a solid data foundation for subsequent accurate predictions and optimizations.

[0077] Latin hypercube sampling (LHS) has significant advantages over simple Monte Carlo random sampling:

[0078] 1) The space filling property is a forced, layered uniformity that ensures that there are no “blank” areas in the parameter space.

[0079] 2) Its projection characteristics ensure that the sample points are perfectly uniformly distributed in any single-parameter dimension.

[0080] 3) High convergence efficiency. It obtains a more accurate estimate of the output response with fewer samples and lower variance. Low convergence efficiency. More samples are needed to achieve the same accuracy.

[0081] 4) Excellent computational cost. Particularly suitable for simulations with high computational costs (such as CFD).

[0082] 5) Applicable scenarios include building proxy models, sensitivity analysis, and uncertainty quantification.

[0083] Compared to full factorial design, LHS has an overwhelming advantage in sample size. The sample size of a full factorial design grows exponentially with the number of parameters k (nk), while the sample size n of LHS can be directly specified by the user and is linearly related to the dimension k, making it very suitable for medium- to high-dimensional problems.

[0084] Specific applications in sub-segment oscillation analysis: In the machine learning-based sub-segment oscillation analysis process, LHS is the core step in generating high-quality training data.

[0085] Application process:

[0086] 1) Define the parameter space:

[0087] a) Geometric parameters: Split spacing S (10D-20D), sub-span length LL (30m-80m).

[0088] b) Physical parameters: wind speed U (3-15m / s), conductor tension TT (15%-25%RTS).

[0089] c) Material parameters: Damping ratio ζ (0.01-0.05), etc.

[0090] 2) Perform LHS sampling:

[0091] a) Determine the number of simulations n based on computing resources (e.g., n=1000).

[0092] b) Normalize the above parameter range and use the LHS algorithm to generate n unique combinations of input parameters.

[0093] 3) Driving simulation and building datasets:

[0094] a) Input n sets of parameters into the parameterized CFD script for automated simulation.

[0095] b) Extract target output variables from each simulation, such as amplitude A, dominant frequency f, aerodynamic coefficients CL, CD, etc.

[0096] c) Finally, a complete dataset containing n samples is obtained, which is used to train the machine learning model.

[0097] The necessity of applying LHS: CFD simulations are computationally extremely costly, with a single simulation potentially taking hours to days. LHS ensures that every computational resource is used to explore the design space to the maximum extent, avoiding resource waste. The accuracy of machine learning models heavily depends on the quality and representativeness of the training data. The dataset provided by LHS can unbiasedly reflect the complex nonlinear relationship between input and output, preventing cognitive biases in the model.

[0098] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0099] Please see Figure 1 , Figure 1 This is a flowchart illustrating the steps of a method for generating a vibration damping scheme for transmission line spacers, as provided in Embodiment 1 of the present invention.

[0100] This invention provides a method for generating a vibration damping scheme for transmission line spacers, comprising:

[0101] Step 101: Obtain on-site inspection data of the transmission line and construct an input parameter space containing vibration-damping data of the transmission line spacer bars.

[0102] It should be noted that the definition includes the input parameter space containing data related to vibration damping of transmission line spacers (including parameters such as spacing between multiple spacers, length of secondary span, conductor diameter, tension, wind speed, wind direction angle, and damping characteristics).

[0103] Step 102: Using the Latin hypercube sampling method, generate the target training dataset based on the field detection data and the input parameter space.

[0104] It should be noted that the Latin Hypercube Sampling (LHS) method is used to generate 1000-5000 representative sample points (i.e., multiple sample points) within the input parameter space to ensure spatial filling and projection uniformity of the samples. Then, parametric CFD (Computational Fluid Dynamics) simulations are used to calculate the oscillation amplitude, frequency, and other response outputs corresponding to each sample point, constructing the initial training dataset. Simultaneously, acceleration, meteorological, video, and strain data from field monitoring are integrated to supplement and validate the simulation dataset (i.e., the initial training dataset), resulting in the target training dataset and ensuring data quality.

[0105] It is worth mentioning that you should refer to Figure 2Latin hypercube sampling (LHS) is a hierarchical random sampling technique and a crucial Design of Experiment (DoE) method in computer experiments and simulation modeling. Its core objective is to efficiently cover a multidimensional parameter space with as few sample points as possible, while exhibiting good space-filling and projection properties. LHS is an advanced experimental design method that ensures each input parameter dimension is explored uniformly through hierarchical and random permutations. In machine learning-based sub-segment oscillation studies, it is a key first step in generating high-quality, low-biased training datasets, enabling the use of a limited number of CFD simulations to lay a solid data foundation for subsequent accurate predictions and optimizations. The specific steps of the LHS method include: 1) Normalizing the range of values ​​for each k-dimensional input parameter and dividing it into n non-overlapping intervals; 2) Randomly and non-repeatingly sampling one value from each of the n intervals for each parameter; 3) Randomly permuting the n values ​​for each parameter and combining the permuted parameter values ​​column-wise to form n k-dimensional sample points.

[0106] Furthermore, step 102 may include the following sub-steps:

[0107] S21. Based on the Latin hypercube sampling method, multiple sample points are generated in the input parameter space;

[0108] S22. Perform parametric computational fluid dynamics simulation on each sample point to obtain the sub-span oscillation response output corresponding to each sample point;

[0109] S23. Construct the initial training dataset based on the sub-segment oscillation response output corresponding to each sample point;

[0110] S24. Use on-site detection data to supplement and validate the initial training dataset, and output the target training dataset.

[0111] It should be noted that, based on the Latin hypercube sampling method, multiple sample points are generated in the input parameter space. During the sampling process, the four dimensions of the input parameter space, including the core conductor parameters and meteorological characteristic parameters, are divided into layers according to the range of each parameter value to ensure that the parameter range of each dimension is uniformly covered. At the same time, the sampling positions within the layers are randomly shuffled to avoid sample clustering. The final generated sample points cover both single-parameter extreme conditions (such as maximum design wind speed and minimum secondary span length) and typical conditions with multi-parameter coupling (such as medium-speed wind and medium tension combination conditions), ensuring the representativeness and dispersion of the samples. When performing parametric computational fluid dynamics simulations on each sample point, the parameter data of the sample points are first automatically mapped to a preset multi-split conductor simulation model template. Parameters such as conductor diameter and split spacing drive the generation of the geometric model, while parameters such as wind speed and wind direction angle define the flow field boundary conditions, and conductor tension parameters set structural constraints. Subsequently, the RNG k-ε turbulence model is used to construct the flow field control equations, which are then solved discretizedly using the finite volume method. During the simulation, key data such as conductor vibration acceleration, displacement time history curves, maximum amplitude, and oscillation frequency are monitored and recorded in real time, serving as the secondary span oscillation response output for each sample point. Based on the secondary span oscillation response output for each sample point, the input parameters and response outputs of the sample points are integrated according to a one-to-one correspondence principle to construct an initial training dataset containing an input feature matrix and an output label vector. When supplementing and validating the initial training dataset with field detection data, vibration data matching the working conditions of the sample points are first extracted from the field detection data and added to the initial training dataset to fill in special working conditions not covered by the simulation (such as instantaneous strong gusts). Then, by calculating the root mean square error between the simulation output and the field measured data at the same time, abnormal sample points with errors exceeding a preset threshold (such as 5%) are removed. At the same time, the simulation model parameters are corrected to reduce the deviation. Finally, the target training dataset with both the comprehensiveness of simulation data and the accuracy of measured data is output, providing high-quality data support for subsequent model training.

[0112] Among them, a sample point refers to a single multi-dimensional data unit generated in the input parameter space through Latin hypercube sampling. Each sample point contains a complete set of input parameters (such as a set of parameters for a specific diameter, wind speed, and secondary span length), and is the core carrier connecting the parameter space and the simulation output.

[0113] Parametric computational fluid dynamics simulation refers to a computational fluid dynamics simulation technology that automatically drives the construction of simulation models, the setting and solving of boundary conditions based on preset parameter variables. By mapping sample point parameters to the simulation system, it realizes automated flow field and vibration response calculation under multiple working conditions.

[0114] Secondary span oscillation response output: refers to the set of key data reflecting the oscillation characteristics of the secondary span of the conductor, which is monitored and recorded during the parametric CFD simulation process. The core data includes vibration acceleration, displacement time history curve, maximum amplitude, oscillation frequency, etc., and is the direct basis for evaluating the vibration prevention effect.

[0115] Initial training dataset: refers to the original dataset formed by integrating the input parameters of sample points with the corresponding secondary oscillation response outputs in a one-to-one correspondence. It has a complete input-output mapping relationship, but has not been verified and supplemented by field test data.

[0116] Root mean square error (RMSE): A statistical index used to quantify the deviation between simulation output and actual field measurement data. It is obtained by taking the square root of the mean of the sum of the squares of the deviations of the two sets of data. The smaller the value, the better the data consistency. It is the core criterion for selecting effective sample points.

[0117] RNGk-ε turbulence model: A numerical turbulence calculation model based on renormalization group theory, suitable for simulating complex flow field characteristics around multi-split conductors. It can accurately capture unsteady flow caused by wake interference and provide reliable flow field basis data for subspan oscillation response calculation.

[0118] Finite volume method: A numerical discretization method for solving fluid dynamics control equations. By dividing the computational domain into multiple control volumes and performing integral discretization on the conservation equations, it has the characteristics of good computational stability and high accuracy, and is the core solution technology for parametric CFD simulation.

[0119] Input feature matrix: The input part of the target training dataset, which organizes the multi-dimensional parameter data of all sample points in matrix form. Each row represents a sample point and each column represents an input parameter dimension, providing structured input information for model training.

[0120] Output label vector: The output part of the target training dataset, corresponding to the input feature matrix in vector or matrix form, stores the sub-segment oscillation response output data of each sample point, and is the target variable used for fitting in model training.

[0121] Step 103: Construct a physical information neural network as the initial amplitude prediction model and an initial stability classification model. Based on the target training dataset, train the initial amplitude prediction model and the initial stability classification model to output the target amplitude prediction model and the target stability classification model.

[0122] It should be noted that, based on physical mechanisms, key features (i.e., transmission line features) are extracted from the target training dataset, including Reynolds number, Strauhal number (aerodynamic features), dimensionless tension, dimensionless frequency (structural dynamic features), velocity deficit rate, and turbulence intensity (wake interference features). Random forest and mutual information methods are used for feature selection to obtain multiple target transmission line features. For example... Figure 3 As shown, a Physics-Informed Neural Network (PINN) is trained as an amplitude prediction model (i.e., a target amplitude prediction model), with a preset loss function of: Among them, the physical residual term The model is forced to satisfy the energy balance equation, thereby embedding physical laws into the learning process and improving the model's accuracy and extrapolation ability. Simultaneously, a stability classification model is trained, and nested cross-validation is used to evaluate its performance, ultimately yielding the target stability classification model. Figure 3 In this context, I1 represents the mass inertia term, I2 represents the damping dissipation term, I3 represents the stiffness recovery term, I4 represents the aerodynamic excitation term, I5 represents the physical constraint term, F1 represents the tower spacing, F2 represents the span, and F3 represents the conductor amplitude.

[0123] Among these, feature engineering based on physical mechanisms can significantly improve model performance. Key features include:

[0124] 1) Aerodynamic characteristics:

[0125] (The Reynolds number (Re) is a dimensionless number in fluid mechanics used to measure the ratio between inertial forces and viscous forces in a fluid.)

[0126] The Strouhal number is a dimensionless number in fluid mechanics used to describe periodic unsteady flows.

[0127] (Relative position parameter)

[0128] (Lift / Drag Coefficient)

[0129] in, For fluid density, it describes the mass per unit volume of the fluid. The inflow velocity is the velocity of the fluid relative to the object. Characteristic length (such as wire diameter) is used to characterize the size of a flowing object; This is the dynamic viscosity of the fluid, reflecting the magnitude of the fluid's viscosity; The frequency of vibration or vortex shedding describes the frequency characteristics of periodic flow. These are the absolute position coordinates along a certain direction; A reference length (such as wire diameter) is used to make the absolute position dimensionless; These are the absolute position coordinates along the other direction; The lift coefficient is a dimensionless lift force that reflects the characteristic of the force exerted by the fluid on an object perpendicular to the direction of the incoming flow. It is the drag coefficient, a dimensionless drag force that reflects the characteristics of the force exerted by the fluid on an object, parallel to the direction of the incoming flow.

[0130] 2) Structural dynamic characteristics:

[0131] (Dimensionless tension)

[0132] (Dimensionless frequency)

[0133] (Logarithmic decay rate)

[0134] (Bending stiffness parameter)

[0135] in, The tension of the conductor is the pulling force on the conductor, reflecting the stress state of the conductor; Mass per unit length of the conductor; describes the magnitude of the mass per unit length of the conductor. It is the acceleration due to gravity; The frequency of the conductor's vibration describes the periodicity of the conductor's vibration. This refers to the span length of the conductor, which is the horizontal distance between the conductors of two towers. The logarithmic decay rate is used to characterize the rate at which the amplitude decays during vibration, reflecting the damping characteristics of the system. The elastic modulus of a conductor material describes the material's ability to resist elastic deformation. The moment of inertia is the cross-section of the conductor, reflecting the cross-section's ability to resist bending deformation.

[0136] 3) Wake interference characteristics:

[0137] Feature quantities of the wake region are extracted through CFD post-processing:

[0138] • Speed ​​loss rate:

[0139] • Turbulence intensity distribution:

[0140] • Eddy distribution:

[0141] in, It is the velocity deficit in the wake region, that is, the difference between the incoming flow velocity and the actual velocity at a certain point in the wake region; The incoming flow velocity is the free flow velocity undisturbed by the conductor. This is the root mean square value of the velocity pulsation, reflecting the severity of the velocity fluctuation; The time-averaged velocity at a point within the wake region; The vorticity in the z-direction is used to characterize the rotational intensity of fluid micro-elements; The velocity component is in the x-direction; The coordinates are along the direction of the incoming flow; The velocity component in the y-direction; These are coordinates perpendicular to the direction of the incoming flow, used to describe the spatial location of the wake region.

[0142] Specifically, step 103 may include the following sub-steps:

[0143] S31. Extract multiple initial transmission line features from the target training dataset;

[0144] S32. Random forest and mutual information methods are used to filter the initial features to obtain multiple target transmission line features;

[0145] S33. Using a preset loss function, train the initial amplitude prediction model based on the characteristics of multiple target transmission lines to determine the target amplitude prediction model;

[0146] S34. The initial stability classification model is trained using cross-validation based on the characteristics of multiple target transmission lines to determine the target stability classification model.

[0147] It should be noted that multiple initial transmission line features are extracted from the target training dataset. The extraction process closely follows the core influence mechanism of sub-span oscillations and comprehensively covers three core categories: aerodynamic features, structural dynamic features, and wake interference features. Among them, aerodynamic features include Reynolds number, Strouhal number, lift coefficient, and drag coefficient; structural dynamic features include dimensionless tension, dimensionless frequency, logarithmic decay rate, and bending stiffness parameters; and wake interference features include velocity deficit rate, turbulence intensity, and vortex distribution, ensuring that the initial features accurately correspond to the key physical quantities related to oscillations. When using random forest and mutual information methods to screen initial features, the oscillation response output (such as maximum amplitude) in the target training dataset is first used as the dependent variable, and the three core initial features are used as independent variables. Multiple decision tree models are constructed using the random forest algorithm to calculate the predictive contribution of each feature (i.e., feature importance score). Features with scores higher than a preset threshold are retained to initially eliminate redundancy. Then, the statistical dependency between the remaining features and the oscillation response output (i.e., mutual information value) is quantified using the mutual information method, and strongly correlated features with mutual information values ​​higher than the critical value are screened out. Finally, the screening results of the two methods are combined to remove features with conflicting correlations, resulting in multiple target transmission line features focusing on core influencing factors. This provides concise and efficient input feature support for the subsequent training of physical information neural networks and stability classification models.

[0148] Further, please refer to Figure 4 The initial amplitude prediction model is trained based on the characteristics of multiple target transmission lines using a preset loss function to determine the target amplitude prediction model. In this embodiment, the initial amplitude prediction model is a Physical Information Neural Network (PINN). During training, the selected target transmission line characteristics (including Reynolds number and Strouhal number in aerodynamic features, dimensionless tension and dimensionless frequency in structural dynamic features, and velocity deficit rate and turbulence intensity in wake interference features) are used as model inputs. The neural network layers then map and output the predicted value of the maximum amplitude of the second span. The preset loss function includes two parts: data loss and physical constraint loss. The data loss is calculated by comparing the predicted amplitude of the model with the corresponding true amplitude in the target training dataset. The mean squared error (MSE) is obtained, reflecting the degree of agreement between the prediction results and the measured / simulated data. The physical constraint loss is obtained by quantifying the degree to which the model output satisfies the physical laws of the second span oscillation (such as the differential equation of conductor vibration and the fluid-structure interaction equilibrium condition) (i.e., calculating the sum of squares of the residuals of the physical equations), ensuring that the model prediction conforms to basic mechanical principles. During training, the total loss (the weighted sum of data loss and physical constraint loss) is used as the optimization objective. The network weights and biases are iteratively updated through the backpropagation algorithm. After each iteration, the decreasing trend of the total loss is monitored. When the loss value tends to stabilize and is lower than the preset convergence threshold, training stops. The model at this point is the target amplitude prediction model that can accurately predict the maximum amplitude of the second span oscillation under complex working conditions. Among them, the data loss in the preset loss function... Data loss is the mean squared error (MSE) between the predicted and actual values. It ensures that the output of the neural network matches the observed values ​​on known data points. Assume there are N training data points, each including input features (such as wind speed, time, location, etc.) and the corresponding true displacement value. Let y i Let i be the actual displacement value of the i-th data point. Let be the predicted displacement value of the neural network for the i-th data point. Defined as:

[0149] (1 / N)*Σ_{i=1}^{N}| |^2;

[0150] in, This represents the total number of training data points, i.e., the total number of samples used in model training. is the true amplitude value of the i-th data point, which is the maximum amplitude result of the second span obtained by actual measurement or simulation in the target training dataset; The predicted amplitude value for the i-th data point is the maximum amplitude prediction result for the second span output by the model based on the characteristics of the target transmission line.

[0151] For physical loss It is a residual based on physical equations (such as the energy balance equation). It ensures that the neural network's predictions conform to physical laws. The energy balance equation can be used to construct the physical residual. For sub-slot oscillations, the energy balance equation can be written as:

[0152] ;

[0153] Rewrite it in residual form:

[0154] ;

[0155] in, The residual of the dynamic equation is used to quantify the degree to which the model output satisfies the dynamic equation of conductor vibration. The smaller the residual value, the more the model conforms to the physical laws. Mass per unit length of the conductor describes the magnitude of the mass per unit length of the conductor; The vibration displacement of the conductor is the amount of displacement along the vibration direction, i.e., the predicted amplitude value, which reflects the vibration amplitude of the conductor. Time is the time variable in the vibration process, used to describe the changes in displacement and velocity over time. The damping coefficient reflects the magnitude of damping during conductor vibration and affects the rate of amplitude decay. It is the equivalent stiffness coefficient of the conductor, which is related to the conductor material and structure, and reflects the conductor's ability to resist deformation; It is an aerodynamic function; The inflow velocity is the velocity of the fluid relative to the conductor, such as wind speed.

[0156] Physical loss requires that the residual R be as small as possible. This invention can select M points (which can be data points or randomly sampled points, called residual points) within the computational domain to calculate the residual.

[0157] Let R j Let the residual value be at the j-th residual point, then the physical loss is defined as:

[0158] = (1 / M) *Σ_{j=1}^{M} |R j |^2;

[0159] in, Physical loss is used to quantify the degree to which the model output satisfies physical laws (such as the dynamic equation of conductor vibration), and is a core component of physical constraint loss; M is the total number of residual points, that is, the number of points selected in the computational domain for calculating the residuals, which can come from data points or random sampling points; R j The residual value at the j-th residual point is the specific numerical value of the residual of the dynamic equation at that point, reflecting the degree to which the model conforms to the physical equation at that point.

[0160] It is worth noting that when calculating the residual, the neural network output y is required, and then the first derivative (velocity) and second derivative (acceleration) are calculated through automatic differentiation. Aerodynamic force F aero It is a function dependent on displacement, velocity, and wind speed, and may be given by empirical formulas or neural network models. Hyperparameters This is used to balance data loss and physical loss. If the data volume is large enough, it can be reduced. If data is scarce but the physical laws are important, the amount can be increased. .

[0161] Therefore, the preset loss function is:

[0162] L=(1 / N)*Σ_{i=1}^{N}| |^2+ *(1 / M)*Σ_{j=1}^{M}|R j |^2

[0163] During training, the weights of the neural network can be updated by minimizing the total loss L, so that the neural network fits the data and satisfies the physical laws, and finally obtains the target amplitude prediction model.

[0164] Furthermore, regarding the training process of the initial stability classification model, this invention will construct a stability classification model in the secondary span oscillation analysis to determine whether secondary span oscillations will occur under a given operating condition (i.e., whether it is stable). This is a binary classification problem, with stable and unstable corresponding to two classes respectively. Specifically, the training process is roughly divided into five steps: 1) Data preparation: A labeled dataset is required, where each sample includes input features (such as wind speed, split distance, secondary span length, etc.) and a label (stable or unstable). The label can come from CFD simulation results (e.g., by calculating whether the amplitude exceeds a threshold or through stability analysis) or field observations, i.e., the initial training dataset. 2) Feature selection: Select features related to stability, which may include aerodynamic features, structural dynamic features, and wake interference features, i.e., the target training dataset. 3) Model selection: Various machine learning classification models can be selected, such as logistic regression, support vector machine, random forest, gradient boosting tree, or neural network. This invention will train multiple models and select the one with the best performance. 4) Model Training and Validation: Use cross-validation to tune hyperparameters and evaluate model performance (accuracy, precision, recall, F1 score, etc.). 5) Model Ensemble: Consider using ensembles of multiple models (e.g., voting) to improve prediction stability and accuracy.

[0165] This invention uses gradient boosting trees (such as XGBoost) and neural networks as base classifiers, and then uses ensemble learning.

[0166] Furthermore, regarding the data preparation process, (1) Input features: Features extracted from CFD simulation and field monitoring data, including but not limited to: 1) Aerodynamic features: converted wind speed (U / fD), relative spacing (S / D), etc. 2) Structural dynamic features: tension parameters (T / mgL), damping ratio (ζ), etc. 3) Wake features: velocity deficit (ΔU / U), turbulence intensity (I_u), etc. 4) Geometric features: sub-span ratio (L_i / L_avg), etc.

[0167] (2) Labels: stable (0) and unstable (1). They can be defined based on whether the amplitude exceeds a certain threshold (e.g., 5% of the wire diameter).

[0168] For the feature selection process, (1) Feature standardization: Standardize the features to have a mean of 0 and a variance of 1.

[0169] (2) Feature selection: Use random forest or XGBoost to calculate feature importance and select the most important features. Mutual information method can also be used.

[0170] For the model training process, (1) the dataset is divided into training set and test set (e.g., 70% training, 30% test).

[0171] (2) Use cross-validation (e.g., 5-fold cross-validation) to tune hyperparameters.

[0172] For model evaluation: Accuracy, precision, recall, F1 score, and AUC are used to evaluate model performance. Soft voting ensemble is used, which combines the prediction probabilities of multiple models and takes the average probability as the final prediction. Finally, the trained model is integrated into a GUI (Graphical User Interface) for real-time prediction.

[0173] Step 104: Construct a multi-objective optimization function and set constraints.

[0174] It should be noted that the multi-objective optimization function established in this invention is specifically as follows:

[0175] ;

[0176] in, It is a multi-objective optimization function, consisting of three objective functions: maximum amplitude, number of spacers, and uniformity of arrangement. These three objectives need to be minimized simultaneously. The objective function is the maximum amplitude, and its value is the maximum value of the oscillation amplitude at each monitoring point. , Let be the amplitude value of the second span oscillation at the i-th monitoring point; The objective function is the number of spacers (cost), where the value represents the number of spacers. , The number of spacers represents a cost indicator for vibration damping measures. For uniformity of arrangement, i.e., the objective function for uniformity of arrangement, , The actual spacing of the i-th spacer bar. The average spacing between the spacers is used to quantify the uniformity of the spacer arrangement.

[0177] Furthermore, the constraints include:

[0178] Secondary gap length constraint: L min ≤L i ≤L max ;

[0179] Spacer quantity constraint: N min ≤N≤N max ;

[0180] Amplitude constraint: A i ≤A allow ;

[0181] Among them, Lmin L represents the minimum length of the second gear interval, which is the minimum allowable length limit for the second gear interval. max N represents the maximum length of the second gear, which is the maximum allowable length limit for the second gear; min N represents the minimum number of spacers allowed; max The maximum number of spacers is the maximum number of spacers allowed to be arranged; A allow The maximum allowable amplitude is the upper limit threshold of the oscillation amplitude of the secondary span, used to ensure the safe operation of the conductor.

[0182] Step 105: Based on the target amplitude prediction model and the target stability classification model, a deep Q-network or a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function under constraints according to the baseline state of the transmission line, and a Pareto optimal solution set is generated.

[0183] It should be noted that the objectives are to minimize the maximum amplitude, the number of spacers (cost), and the unevenness of their arrangement. Constraints include the range of sub-span lengths, the range of the number of spacers, and amplitude limits. A Deep Q-Network (DQN) and / or a Non-dominated Sorting Genetic Algorithm II (NSGA-II) are used for solving the problem. DQN performs adaptive optimization by defining the state space, action space, and a reward function that integrates multiple objectives; NSGA-II performs a global search through real-number encoding, simulated binary crossover, polynomial mutation, and constraint dominance principles, ultimately outputting a Pareto optimal solution set, such as... Figure 5 As shown, 1) Pareto front concept: The curve in the figure represents the Pareto optimal solution set, where any improvement at any point on the curve must come at the cost of at least one other objective deteriorating. This reflects the trade-offs between objectives in multi-objective optimization.

[0184] Optionally, the process of using a deep Q-network to solve the multi-objective optimization function under constraints based on the target amplitude prediction model and the target stability classification model, and generating a Pareto optimal solution set, may include the following sub-steps:

[0185] S511. Using a deep Q-network, the multi-objective optimization function is solved under constraints based on the baseline state of the transmission line, and multiple candidate scheme actions and the Q value corresponding to each candidate scheme action are output.

[0186] S512. In each candidate scheme action, extract the transmission line features corresponding to each candidate scheme action, and input them into the target amplitude prediction model respectively, and output the predicted maximum amplitude corresponding to each candidate scheme action.

[0187] S513. Input the transmission line features corresponding to the actions of each candidate scheme into the target stability classification model, and output the oscillation stability state results corresponding to the actions of each candidate scheme.

[0188] S514. Based on the number of spacers in each candidate scheme action, as well as the corresponding oscillation stability state result and predicted maximum amplitude, calculate the total reward value corresponding to each candidate scheme action, and count the number of iterations in real time.

[0189] S515. Determine whether the number of iterations has reached the preset threshold.

[0190] S516. If so, sort the total reward values ​​in descending order and select the candidate action corresponding to the total reward value of the first preset number of positions to form the Pareto optimal solution set.

[0191] S517. If the number of iterations does not reach the preset threshold, the deep Q-network is updated according to the Q-value corresponding to the action of each candidate scheme to determine a new deep Q-network.

[0192] S518. Select the candidate scheme action corresponding to the largest total reward value and execute it. Output the new baseline state of the transmission line and jump to execute the step of using a deep Q network to solve the multi-objective optimization function under constraints based on the baseline state of the transmission line, output multiple candidate scheme actions and the Q value corresponding to each candidate scheme action, until the number of iterations reaches the preset number threshold.

[0193] S519. Sort the multiple total reward values ​​determined when the number of iterations reaches the preset threshold in descending order, and select the candidate scheme actions corresponding to the total reward values ​​of the first preset number of positions to form the Pareto optimal solution set.

[0194] Candidate action schemes: Possible optimization operations output by the deep Q-network, such as adjusting the position of spacers or increasing or decreasing their number, are specific measures to achieve the optimization goal.

[0195] Q-value: Quantifies the expected cumulative reward of a candidate action in the current state. A higher value indicates that the action is more likely to bring better results in long-term optimization.

[0196] Baseline state: The current parameters and state of the transmission line (such as spacer arrangement, conductor characteristics, etc.) are the initial inputs for solving the optimization problem using a deep Q-network.

[0197] It should be noted that a deep Q-network is used to solve a multi-objective optimization function under constraints based on the baseline state of the transmission line. This outputs multiple candidate actions (such as adjusting the position of spacers, increasing or decreasing their quantity, etc.) and the Q-value corresponding to each candidate action (quantifying the expected cumulative reward of the action in the current state). From each candidate action, transmission line features corresponding to that action are extracted (covering aerodynamic features such as Reynolds number and Strouhal number, structural dynamic features such as dimensionless tension and dimensionless frequency, and wake interference features such as velocity deficit rate and turbulence intensity). These features are then input into the target amplitude prediction model, outputting the predicted maximum amplitude (reflecting the severity of the sub-span oscillation) for each candidate action. Finally, the transmission line features corresponding to each candidate action are input into the target stability classification model, outputting the oscillation stability state for each candidate action. The results are analyzed (e.g., "stable" or "unstable," indicating whether there is a risk of dangerous oscillation in the conductor); based on the number of spacers in each candidate scheme action (reflecting cost indicators, fewer spacers mean lower cost), the corresponding oscillation stability state results (stable state grants higher basic rewards), and the predicted maximum amplitude (rewards increase when amplitude is less than the allowable value, and decrease otherwise), the total reward value corresponding to each candidate scheme action is calculated using preset weights (a quantitative indicator for comprehensively evaluating the quality of actions), and the number of iterations is counted in real time (recording the number of loops in the current optimization process); it is determined whether the number of iterations has reached the preset threshold (a pre-set upper limit for the optimization termination loop); if so, the total reward values ​​are sorted in descending order, and the candidate scheme actions corresponding to the total reward values ​​of the first preset number of positions are selected to form a Pareto optimal solution set (containing multiple non-dominated optimal schemes, taking into account both oscillation control effect and cost). If the number of iterations does not reach the preset threshold, the parameters of the deep Q-network are updated using a temporal difference algorithm based on the Q-value corresponding to each candidate action and the actual total reward value (to make the network more accurate in predicting action value), and a new deep Q-network is determined. The candidate action corresponding to the largest total reward value is selected and executed (the scheme is applied to the transmission line), and the new baseline state of the transmission line (updated line parameters and state) is output. Then, the process jumps to step S511 until the number of iterations reaches the preset threshold. The candidate action corresponding to the total reward value of the first preset number of iterations determined when the number of iterations reaches the preset threshold constitutes the Pareto optimal solution set.

[0198] The calculation process for the total reward value is as follows:

[0199] ;

[0200] in, Total reward value; , , These are the weighting coefficients; The reward is based on the maximum amplitude. ; For rewards based on the number of interval bars, ; For stability-based rewards, the reward can be set by judging the steady-state result of the oscillation. For example, if the oscillation state of the transmission line is "stable", then... =1; if it is "unstable", then =-1 (The specific value can be adjusted according to the needs, so as to guide the optimization towards a stable state).

[0201] Optionally, the process of generating a Pareto optimal solution set by using a non-dominated sorting genetic algorithm to solve the multi-objective optimization function under constraints based on the target amplitude prediction model and the target stability classification model, according to the baseline state of the transmission line, may include the following sub-steps:

[0202] S521. Using a non-dominated sorting genetic algorithm, the multi-objective optimization function is solved under constraints based on the baseline state of the transmission line, and an initial population is output. The initial population includes multiple population individuals, and each population individual corresponds to a candidate solution.

[0203] S522. Extract the transmission line features corresponding to each population individual, and input them into the target amplitude prediction model respectively, and output the predicted maximum amplitude corresponding to each population individual.

[0204] S523. Input the transmission line characteristics corresponding to each group of individuals into the target stability classification model, and output the oscillation stability state results corresponding to each group of individuals.

[0205] S524. Calculate the objective function value for each population based on the predicted maximum amplitude and the number of spacers for each individual.

[0206] S525. Based on the objective function values ​​corresponding to individuals in various groups, Pareto dominance sort is performed on individuals in various groups to obtain the initial non-dominated solution set.

[0207] S526. Based on the oscillation stability results corresponding to individuals in various groups, the initial non-dominated solution set is filtered to obtain the target non-dominated solution set, and the number of iterations is counted in real time.

[0208] S527. Perform simulated binary crossover and polynomial mutation on the target non-dominated solution set, and merge it with the initial population, outputting the merged population.

[0209] S528. Based on the merged population, identify multiple new population individuals and determine whether the number of iterations has reached the preset threshold.

[0210] S529. If so, then multiple new population individuals will form a Pareto optimal solution set.

[0211] S5210. If not, then proceed to the step of extracting the transmission line features corresponding to each population individual and inputting them into the target amplitude prediction model, and outputting the predicted maximum amplitude corresponding to each population individual, until the number of iterations reaches the preset threshold.

[0212] It should be noted that the non-dominated sorting genetic algorithm uses real-number encoding to represent the sub-range length, employs simulated binary crossover and polynomial mutation as genetic operators, and uses the constraint dominance principle to handle constraints. Combined with NSGA-II (non-dominated sorting genetic algorithm) for global optimization, the algorithm flow is as follows:

[0213] 1) Initialization: Generate an initial population, with individuals representing the spacing bar arrangement scheme;

[0214] 2) Non-dominated ranking: Ranking the population based on Pareto dominance;

[0215] 3) Crowding degree calculation: Maintaining the diversity of solution sets;

[0216] 4) Selection, crossover, and mutation: generating offspring populations;

[0217] 5) Environmental selection: Selecting the next generation from merged populations;

[0218] 6) Convergence judgment: Repeat steps 2-5 until the termination condition is met.

[0219] Key operator design: Encoding scheme: Real number encoding, directly representing the length of each interval; Crossover operator: Simulated Binary Crossover (SBX); Mutation operator: Polynomial mutation; Constraint handling: Constraint dominance principle.

[0220] Specifically, a non-dominated sorting genetic algorithm is used under the constraint condition (secondary gap length range L). min To L max The range of the number of spacers N min To N max Maximum allowable amplitude A allowUnder the condition of the transmission line's baseline state (current spacer arrangement, conductor parameters, and other initial states), a multi-objective optimization function (minimizing maximum amplitude, number of spacers, and uniformity of arrangement) is solved. An initial population satisfying the constraints is randomly generated, consisting of multiple individuals, each corresponding to a candidate scheme (including specific spacer positions, quantities, and other parameters). Within each individual, transmission line characteristics (such as Reynolds number and Strouhal number for aerodynamic features, dimensionless tension and frequency for structural dynamic features, and velocity deficit rate and turbulence intensity for wake interference features) are extracted and input into the target amplitude prediction model. The predicted maximum amplitude for each individual is then output as the target amplitude prediction model. The calculation basis); input the transmission line characteristics corresponding to each group of individuals into the target stability classification model, and output the oscillation stability state results ("stable" or "unstable") corresponding to each group of individuals; according to the predicted maximum amplitude corresponding to each group of individuals (calculated) ), number of spacers (directly used as) ) and spacing between spacers (calculated) The objective function values ​​for individuals in each population group are obtained. Based on these objective function values, Pareto dominance ranking is performed on individuals in each population group according to the non-dominated ranking rule (if individual a is not inferior to individual b on all objectives and is better on at least one objective, then a dominates b). Individuals not dominated by other individuals are selected to form the initial non-dominated solution set. Based on the oscillatory stable state results for individuals in each population group, the initial non-dominated solution set is filtered, removing individuals in the "unstable" state and retaining individuals in the "stable" state as the target non-dominated solution set. The current iteration number is counted in real time (recording the number of cycles in the population evolution). Simulation 2 is performed on the individuals in the target non-dominated solution set. The process employs binary crossover (SBX) and polynomial mutation (simulating the crossover and mutation of continuous variables to generate offspring individuals, maintaining population diversity), and merges these offspring individuals with the initial population, outputting a merged population containing both parents and offspring. Based on the merged population, superior individuals are selected through non-dominated sorting and crowding calculation (quantifying the density of individuals in the solution set, retaining individuals in sparse regions to maintain diversity), determining multiple new population individuals (maintaining the same size as the initial population), and determining whether the number of iterations has reached a preset threshold (a pre-set upper limit for evolution termination). If yes, the multiple new population individuals form a Pareto optimal solution set (containing the optimal solution that takes into account multiple objectives). If not, the process jumps to step S522 until the number of iterations reaches the preset threshold.

[0221] The initial population is a set of feasible solutions randomly generated at the start of the genetic algorithm. Each solution corresponds to a candidate scheme, and its parameters must meet the constraints. It is the starting point for the evolution of the population.

[0222] Population Individual: The basic unit that makes up the population. Each individual corresponds to a specific candidate solution, which includes key parameters for achieving the optimization goal (such as the position and number of spacers).

[0223] Multinomial mutation: A mutation operator used in genetic algorithms to generate offspring individuals. It achieves mutation by adding perturbations based on a multinomial distribution to the parameters of the parent individuals, thereby enhancing population diversity.

[0224] Pareto dominance ranking: The process of ranking individuals in a population based on Pareto dominance relationships. The core is to select non-dominant individuals that are not dominated by other individuals, which is used to distinguish the quality of individuals.

[0225] Initial non-dominated solution set: The set of individuals obtained after Pareto domination sorting that are not dominated by other individuals in the population, and is a potential better solution.

[0226] Target non-dominated solution set: The set obtained after stability filtering of the initial non-dominated solution set, retaining only individuals with "stable" oscillation states to ensure the feasibility of the solution.

[0227] Merged population: A population formed by merging parent individuals (the initial population or the previous generation population) with offspring individuals (generated through crossover and mutation), used for subsequent screening of new populations.

[0228] New population individuals: The better individuals selected from the merged population through non-dominated sorting and crowding calculation constitute the next generation of the population, driving the population to evolve towards a better solution.

[0229] Step 106: Based on the Pareto optimal solution set, output the vibration prevention scheme for the spacer bars of the transmission line.

[0230] Vibration suppression scheme for transmission line spacers: Based on the specific spacer arrangement strategy output by the Pareto optimal solution set, including parameters such as the number of spacers and their arrangement position (secondary span length), it is used to suppress the oscillation of the secondary span of the transmission line, taking into account the oscillation control effect, cost and arrangement uniformity.

[0231] It should be noted that the final solution (i.e., the vibration isolation scheme for transmission line spacers) is selected from the Pareto optimal solution set based on the decision-maker's preferences for safety and economy. A graphical user interface (GUI) is developed, integrating pre-trained machine learning models and optimization algorithms to provide a one-stop solution, including parameter input, real-time prediction, optimization design, and visualization of vibration patterns, power spectra, and Pareto fronts.

[0232] Specifically, based on the Pareto optimal solution set, the spacer arrangement parameters corresponding to each individual in the population are extracted, including the number of spacers, the arrangement position of each spacer (corresponding to the secondary span length), and other specific information. These are then compiled into multiple transmission line spacer vibration prevention schemes that take into account oscillation control effect, cost, and uniformity of arrangement. The predicted maximum amplitude, number of spacers, and oscillation stability state of each scheme are clearly defined, allowing engineers to select and implement them according to actual needs (such as cost budget, safety level, etc.) to achieve effective control of secondary span oscillations in transmission lines.

[0233] For example, the secondary span oscillation control of a 500kV transmission line (four-split conductor, span 450m) is taken as an example.

[0234] 1. Data Preparation (S1): Define 7-dimensional input parameters (split spacing, sub-span length, wind speed, wind direction angle, tension, conductor density, damping ratio). Use LHS to generate 2400 sample points for parametric CFD simulation to obtain amplitude and stability labels. Simultaneously, collect 61 sets of field monitoring data for the line for validation.

[0235] 2. Model Training (S2): 28 physical features were extracted to train the PINN amplitude prediction model. The final model achieved an amplitude prediction RMSE of 1.98 mm on the test set, a stability classification accuracy of 95.7%, and a single-sample prediction time of approximately 1.2 ms.

[0236] 3. Optimization Design (S3): Set the objective function and constraints (e.g., single span length 30-70m, number of spacers 5-10). Run the NSGA-II optimization algorithm to obtain Pareto optimal solutions, including economic, safe, and balanced solutions. The final balanced solution selected is: using 7 spacers, with an estimated maximum amplitude of 10.3cm, which is a 17.6% reduction in amplitude compared to the traditional 8-spacer solution (amplitude 12.5cm). The evaluation results are shown in Table 1. Three typical solutions are explained below:

[0237] •Safety Option (A): This option has the most spacers (9), the smallest amplitude (8.7cm), and the highest safety but the worst economy.

[0238] • Balanced scheme (B): The number of spacers is moderate (7), the amplitude is moderate (10.3cm), and the best balance is achieved between safety and economy (recommended scheme).

[0239] • Economic Option (C): Minimum number of spacers (6), maximum amplitude (14.2cm), best economy but lowest safety.

[0240] Execution efficiency (S4): Outputs the final detailed layout of the unequal spacing of the spacers through the GUI interface and provides a vibration response prediction report. The entire optimization design cycle is shortened to 25-30% of the original.

[0241] Table 1 Evaluation Results

[0242]

[0243] After comparison, the "balanced" solution was ultimately chosen, which controls costs while ensuring safety.

[0244] For comparison of technical effects, existing technologies can be used as a reference. Secondary span oscillation is a medium-frequency, medium-amplitude self-excited vibration of multi-span conductors under wind excitation, mainly caused by the wake effect, which seriously threatens the safe operation of transmission lines. Traditional protection design methods mainly rely on empirical formulas, simplified physical models, and engineers' experience. These methods have obvious limitations: accurate simulation methods such as computational fluid dynamics are computationally expensive and difficult to use for comparing a large number of schemes; empirical formulas have limited accuracy and cannot fully consider complex multi-objective optimization needs, such as simultaneously taking into account vibration suppression effect, economic cost, and layout rationality.

[0245] With the development of artificial intelligence technology, machine learning has provided new approaches to solving complex engineering problems. However, effectively applying machine learning to subspan oscillation analysis and protection still faces many challenges, such as how to train high-precision models with limited and expensive simulation data, how to embed physical laws into data-driven models to improve their extrapolation capabilities and interpretability, and how to efficiently solve optimization problems involving multiple objectives such as safety and economy. Therefore, there is an urgent need for a new design method that integrates physical mechanisms and data intelligence.

[0246] To address the aforementioned problems, this invention proposes a method for generating vibration damping schemes for transmission line spacers. This method integrates high-fidelity simulation, machine learning, and multi-objective optimization to achieve accurate prediction of secondary span oscillations and efficient, optimized design of spacer arrangement schemes. The method includes the following steps:

[0247] S1. Multi-source data acquisition and dataset construction: Define an input parameter space containing geometric, physical, and material parameters, and generate a preset number of sample points in this space using the Latin hypercube sampling method; use parametric computational fluid dynamics simulation to calculate the sub-span oscillation response output corresponding to each sample point to construct an initial training dataset; integrate field monitoring data to supplement and validate the initial training dataset;

[0248] S2. Feature Engineering and Machine Learning Model Training: Extract aerodynamic features, structural dynamic features, and wake interference features from the dataset; construct an amplitude prediction model based on a physical information neural network, whose loss function includes a data fitting term and a physical constraint term based on the energy balance equation; train a stability classification model; and use nested cross-validation to train and evaluate the model's performance.

[0249] S3. Multi-objective optimization of spacer bar arrangement: Establish a multi-objective optimization function with the objectives of minimizing the maximum amplitude, minimizing the number of spacers, and maximizing the uniformity of arrangement, and set constraints on the secondary span length, the number of spacers, and the amplitude; use a deep Q-network and / or a non-dominated sorting genetic algorithm to solve the multi-objective optimization function under the constraints, and output the Pareto optimal solution set;

[0250] S4. Scheme Generation and Deployment: Select the final spacer arrangement scheme from the Pareto optimal solution set according to engineering preferences; develop a graphical user interface, integrate pre-trained machine learning models and optimization algorithms, and provide functions for parameter input, real-time prediction, optimization design and result visualization.

[0251] Further, please refer to Figure 6 1) Data layer (step S1): Identified with a blue background, it shows the complete data preparation process from parameter space definition, Latin hypercube sampling, CFD simulation to data integration, providing a high-quality training data foundation for subsequent analysis.

[0252] 2) Algorithm layer (step S2): Identified with a pink background, it shows the machine learning model construction process from feature extraction and feature selection to physical information neural network training and model verification, reflecting the core idea of ​​integrating data-driven and physical laws.

[0253] 3) Optimization layer (step S3): Identified with a green background, this layer shows the definition of the multi-objective optimization problem, the selection of optimization algorithms (deep Q-network and NSGA-II parallel), and the generation process of the Pareto optimal solution set, achieving a balance between safety and economy in optimization.

[0254] 4) Application layer (step S4): Identified with a light orange background, it shows the entire process of engineering application from scheme selection and GUI integration to the final scheme output, reflecting the complete transformation path of this invention from theoretical research to engineering practice.

[0255] Among them, such as Figure 7As shown, the Pareto front refers to an ideal state of resource allocation (in this algorithm, it refers to the resource state of the computer server). Assuming a fixed group of people and available resources, the change from one allocation state to another should improve the situation of at least one person without making anyone worse off. This is called Pareto improvement or Pareto optimization.

[0256] Compared with the prior art, the present invention has the following significant advantages:

[0257] 1) High efficiency and high accuracy coexist: LHS is used to build a high-quality dataset with the fewest number of fluid dynamics (CFD) simulations. The machine learning model trained is several orders of magnitude faster than CFD. At the same time, physical laws are incorporated through PINN to ensure prediction accuracy and physical consistency.

[0258] 2) Achieve true multi-objective optimization: Through advanced algorithms such as NSGA-II and DQN, simultaneously optimize vibration suppression effect, economic cost and layout rationality, providing users with a series of Pareto optimal solutions that balance advantages and disadvantages, supporting scientific decision-making.

[0259] 3) Adaptability and Knowledge Discovery: The model can be updated online based on new field monitoring data, exhibiting stronger adaptive capabilities. Feature importance analysis can deepen our understanding of the sub-span oscillation mechanism.

[0260] 4) High engineering practicality: By developing an operating system GUI interface, complex algorithms are encapsulated into easy-to-use tools, which greatly reduces the threshold for promoting and applying this method in engineering practice.

[0261] In this embodiment of the invention, a method for generating vibration damping schemes for spacer bars in transmission lines is provided. First, representative samples are generated through Latin hypercube sampling, driving computational fluid dynamics simulation to construct a training dataset. Next, key features are extracted based on physical mechanisms, and machine learning models such as physical information neural networks are trained to achieve high-precision and rapid prediction of the amplitude and stability of secondary span oscillations. Finally, combining deep Q-networks and multi-objective evolutionary algorithms, with the objectives of minimizing amplitude, number of spacers, and uneven arrangement, a Pareto-optimal spacer arrangement scheme is output. This invention deeply integrates data-driven approaches with physical models, overcoming the shortcomings of traditional methods such as reliance on experience, high computational costs, and difficulty in multi-objective optimization, thus achieving a safe, economical, and efficient design for secondary span oscillation protection in transmission lines.

[0262] Please see Figure 8 , Figure 8 This is a structural block diagram of a vibration damping scheme generation system for transmission line spacers provided in Embodiment 2 of the present invention.

[0263] This invention provides a system for generating vibration damping schemes for transmission line spacers, comprising:

[0264] The acquisition module 801 is used to acquire on-site inspection data of transmission lines and construct an input parameter space containing vibration-damping data of transmission line spacers;

[0265] Module 802 is used to generate the target training dataset based on the field detection data and the input parameter space using the Latin hypercube sampling method;

[0266] Training module 803 is used to construct a physical information neural network as an initial amplitude prediction model and an initial stability classification model, and to train the initial amplitude prediction model and the initial stability classification model based on the target training dataset, and output the target amplitude prediction model and the target stability classification model.

[0267] Module 804 is used to construct multi-objective optimization functions and set constraints.

[0268] The generation module 805 is used to solve the multi-objective optimization function based on the target amplitude prediction model and the target stability classification model, using a deep Q-network or a non-dominated sorting genetic algorithm under constraints according to the baseline state of the transmission line, and generate a Pareto optimal solution set.

[0269] Output module 806 is used to output vibration prevention schemes for transmission line spacers based on Pareto optimal solution sets.

[0270] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and modules described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0271] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the method for generating a vibration damping scheme for transmission line spacers as described in Embodiment 1 above.

[0272] This invention also provides a computer-readable storage medium storing a computer program / instruction thereon, which, when executed by a processor, implements the steps of the method for generating a vibration damping scheme for transmission line spacers as described in Embodiment 1 above.

[0273] This invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the method for generating a vibration damping scheme for transmission line spacers as described in Embodiment 1 above.

[0274] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating a vibration damping scheme for transmission line spacers, characterized in that, include: Acquire field inspection data of transmission lines and construct an input parameter space containing vibration-damping data of transmission line spacers; The Latin hypercube sampling method is used to generate a target training dataset based on the field detection data and the input parameter space; A physical information neural network is constructed as an initial amplitude prediction model and an initial stability classification model is constructed. Based on the target training dataset, the initial amplitude prediction model and the initial stability classification model are trained to output a target amplitude prediction model and a target stability classification model. Construct a multi-objective optimization function and set constraints; Based on the target amplitude prediction model and the target stability classification model, a deep Q-network or a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function under the constraints according to the baseline state of the transmission line, generating a Pareto optimal solution set. Based on the Pareto optimal solution set, a vibration prevention scheme for transmission line spacers is output.

2. The method for generating a vibration damping scheme for transmission line spacers according to claim 1, characterized in that, The method employing Latin hypercube sampling generates a target training dataset based on the on-site detection data and the input parameter space, including: Based on the Latin hypercube sampling method, multiple sample points are generated in the input parameter space; Parametric computational fluid dynamics simulations were performed on each of the sample points to obtain the sub-segment oscillation response output corresponding to each sample point; Based on the sub-segment oscillation response output corresponding to each of the sample points, an initial training dataset is constructed. The initial training dataset is supplemented and validated using the on-site detection data, and the target training dataset is output.

3. The method for generating a vibration damping scheme for transmission line spacers according to claim 1, characterized in that, The process of training the initial amplitude prediction model and the initial stability classification model based on the target training dataset, and outputting the target amplitude prediction model and the target stability classification model, includes: Multiple initial transmission line features are extracted from the target training dataset; Random forest and mutual information methods are used to filter the initial features to obtain multiple target transmission line features; The initial amplitude prediction model is trained using a preset loss function based on the characteristics of multiple target transmission lines to determine the target amplitude prediction model; The initial stability classification model is trained using cross-validation based on the characteristics of multiple target transmission lines to determine the target stability classification model.

4. The method for generating a vibration damping scheme for transmission line spacers according to claim 1, characterized in that, Based on the target amplitude prediction model and the target stability classification model, a deep Q-network is used to solve the multi-objective optimization function under the constraints according to the baseline state of the transmission line, generating a Pareto optimal solution set, including: The deep Q-network is used to solve the multi-objective optimization function based on the baseline state of the transmission line under the constraints, and outputs multiple candidate scheme actions and the Q value corresponding to each candidate scheme action; In each of the candidate scheme actions, the transmission line features corresponding to each candidate scheme action are extracted and input into the target amplitude prediction model, and the predicted maximum amplitude corresponding to each candidate scheme action is output. The transmission line features corresponding to each of the candidate scheme actions are input into the target stability classification model, and the oscillation stability state results corresponding to each of the candidate scheme actions are output. Based on the number of spacers in each candidate scheme action, the corresponding oscillation stability state result and the predicted maximum amplitude, calculate the total reward value corresponding to each candidate scheme action, and count the number of iterations in real time; Determine whether the number of iterations has reached a preset threshold. If so, sort the total reward values ​​in descending order and select the candidate action corresponding to the total reward value of the first preset number of positions to form a Pareto optimal solution set.

5. The method for generating a vibration damping scheme for transmission line spacers according to claim 4, characterized in that, Also includes: If the number of iterations does not reach the preset threshold, the deep Q-network is updated according to the Q-value corresponding to each candidate scheme action to determine a new deep Q-network. The process involves selecting and executing the candidate action corresponding to the largest total reward value, outputting a new baseline state for the transmission line, and then jumping to execute the step of using a deep Q-network to solve the multi-objective optimization function based on the baseline state of the transmission line under the constraints, outputting multiple candidate action actions and the Q value corresponding to each candidate action, until the number of iterations reaches the preset threshold. The total reward values ​​determined when the number of iterations reaches the preset threshold are sorted in descending order, and the candidate scheme actions corresponding to the total reward values ​​of the first preset number of positions are selected to form a Pareto optimal solution set.

6. The method for generating a vibration damping scheme for transmission line spacers according to claim 1, characterized in that, Based on the target amplitude prediction model and the target stability classification model, a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function under the constraints according to the baseline state of the transmission line, generating a Pareto optimal solution set, including: The non-dominated sorting genetic algorithm is used to solve the multi-objective optimization function based on the baseline state of the transmission line under the constraints, and outputs an initial population, which includes multiple population individuals, each of which corresponds to a candidate solution. In each of the population individuals, the transmission line features corresponding to each population individual are extracted and input into the target amplitude prediction model, and the predicted maximum amplitude corresponding to each population individual is output. The transmission line features corresponding to each individual in the population are input into the target stability classification model, and the oscillation stability state results corresponding to each individual in the population are output. Calculate the objective function value for each individual population based on the predicted maximum amplitude and the number of spacers for each individual population. Based on the objective function value corresponding to each individual in the population, Pareto dominance sorting is performed on each individual in the population to obtain an initial non-dominated solution set. Based on the oscillation stability results corresponding to each individual in the population, the initial non-dominated solution set is filtered to obtain the target non-dominated solution set, and the number of iterations is counted in real time. Simulated binary crossover and polynomial mutation are performed on the target non-dominated solution set, and then merged with the initial population to output the merged population. Based on the merged population, multiple new population individuals are identified, and it is determined whether the number of iterations has reached a preset threshold. If so, then the multiple new population individuals will form a Pareto optimal solution set; If not, then proceed to the step of extracting the transmission line features corresponding to each individual in the population, inputting them into the target amplitude prediction model, and outputting the predicted maximum amplitude corresponding to each individual in the population, until the number of iterations reaches the preset threshold.

7. A system for generating vibration damping schemes for transmission line spacers, characterized in that, include: The acquisition module is used to acquire on-site inspection data of transmission lines and construct an input parameter space containing vibration-damping data of transmission line spacers; A module is used to generate a target training dataset based on the field detection data and the input parameter space using the Latin hypercube sampling method; The training module is used to construct a physical information neural network as an initial amplitude prediction model and an initial stability classification model, and to train the initial amplitude prediction model and the initial stability classification model based on the target training dataset, and output the target amplitude prediction model and the target stability classification model. The building module is used to construct multi-objective optimization functions and set constraints. The generation module is used to solve the multi-objective optimization function based on the target amplitude prediction model and the target stability classification model, using a deep Q-network or a non-dominated sorting genetic algorithm under the constraints, according to the reference state of the transmission line, to generate a Pareto optimal solution set. The output module is used to output a vibration prevention scheme for transmission line spacers based on the Pareto optimal solution set.

8. A computer device, characterized in that, The system includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for generating a vibration damping scheme for transmission line spacers as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the method for generating a vibration damping scheme for transmission line spacers as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the method for generating a vibration damping scheme for transmission line spacers as described in any one of claims 1-6.

Images