A method and device for predicting galloping of a power transmission line and a storage medium

Abstract

The present application relates to the technical field of power transmission line, and discloses a galloping prediction method and device of power transmission line and a storage medium, the method comprising: acquiring sample labels and sample data of prediction variables of a sample power transmission line, constructing a time series for each prediction variable, and the prediction variables including line parameter variables and meteorological variables; performing Fourier transform with an adaptive window length based on the time series of the prediction variables to obtain a frequency spectrum; generating a sample multi-channel image based on the frequency spectrum corresponding to each prediction variable, training an initial galloping prediction model based on the sample multi-channel image and the sample labels, and obtaining a target galloping prediction model; and performing prediction based on the target multi-channel image by using the target galloping prediction model to obtain a galloping prediction result of a power transmission line to be predicted. The present application improves the ability to describe the causes of galloping by considering multiple prediction variables and introducing Fourier transform, so that the model can still make reliable prediction under the condition of sample scarcity or extreme weather, and has strong generalization ability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission lines, in particular to a galloping prediction method and device for power transmission lines and a storage medium. BACKGROUND

[0002] In the operation process of a power transmission line, galloping is a low-frequency large-amplitude vibration phenomenon caused by wind, which mainly occurs in the case of ice-coated or other asymmetric conductor sections. The galloping phenomenon not only causes conductor jumping and interline short circuit, but also may cause broken wires, tower damage and other serious consequences, which poses a significant threat to the safe and stable operation of the power system. Therefore, how to timely and accurately predict galloping has become an important technical problem in the operation and maintenance of power transmission.

[0003] Most of the existing galloping prediction methods directly model and predict time series using support vector machines, random forests, recurrent neural networks or long short-term memory networks. Although these methods capture certain time trends and nonlinear characteristics, they ignore other galloping causes and are difficult to capture the correlation between features, resulting in poor prediction accuracy and weak generalization ability. SUMMARY

[0004] Therefore, the present application provides a galloping prediction method and device for power transmission lines and a storage medium to solve the problem of poor prediction accuracy and weak generalization ability of existing galloping prediction methods.

[0005] In a first aspect, the present application provides a galloping prediction method for a power transmission line, which comprises:

[0006] For each sample power transmission line, obtain the sample label and sample data of a plurality of prediction variables of the sample power transmission line, and construct a time series for each prediction variable based on the sample data of the plurality of prediction variables, wherein the prediction variables include line parameter variables and meteorological variables;

[0007] Perform Fourier transform with an adaptive window length based on the time series of each prediction variable to obtain a frequency spectrum corresponding to each prediction variable;

[0008] Generate a sample multi-channel image based on the frequency spectrum corresponding to each prediction variable, and train an initial galloping prediction model based on the sample multi-channel image and the sample label of each sample power transmission line to obtain a target galloping prediction model;

[0009] Collect data of a plurality of prediction variables of a to-be-predicted power transmission line, and generate a target multi-channel image based on the data of the plurality of prediction variables;

[0010] Use the target galloping prediction model to make a prediction based on the target multi-channel image to obtain a galloping prediction result for the to-be-predicted power transmission line.

[0011] The present application covers the structural and environmental causes of galloping by simultaneously considering line parameter variables and meteorological variables, solves the problem of incomplete cause description caused by single variable in traditional methods, introduces Fourier transform into galloping prediction, converts the one-dimensional time series constructed by the above data into a two-dimensional frequency spectrum, retains the time-frequency characteristics of meteorological variables, makes up for the defects of traditional time series methods that ignore frequency domain information, and solves the inherent limitations of fixed window length in time-frequency resolution through an adaptive window length mechanism, more effectively captures the internal relationship between meteorological changes and galloping occurrence, improves the description ability of galloping risk under complex working conditions, and improves the pertinence and robustness of the frequency spectrum characteristics. By generating a multi-channel frequency spectrum based on all frequency spectra, the model is trained, the cooperative relationship between variables is retained, local overfitting caused by traditional variable-by-variable modeling is avoided, and the model can learn the complex patterns of galloping causes from a holistic perspective. In practical applications, the trained target model is used to predict the transmission line to be predicted, which not only quantifies the risk level, but also provides accurate quantitative basis for operation and maintenance decisions. By considering multiple prediction variables and introducing Fourier transform, the description ability of galloping causes is improved, the model can still reliably predict under sample scarcity or extreme weather conditions, and can be extended to other wind-induced failure prediction fields such as icing and jumper, with wide application scope and strong generalization ability.

[0012] In an optional implementation, Fourier transform with adaptive window length is performed based on the time series of each prediction variable to obtain a frequency spectrum corresponding to each prediction variable, including:

[0013] For each prediction variable, determine the local change rate of the time series of the prediction variable at each time;

[0014] Determine the maximum local change rate from all time local change rates;

[0015] Based on the local change rate at each time, the maximum local change rate, the preset maximum window function length, and the preset minimum window function length, determine the window function length corresponding to each time;

[0016] For the time series of the prediction variable, perform short-time Fourier transform on the time series using the window function length corresponding to each time to obtain a complex matrix corresponding to each time;

[0017] The complex modulus of the complex matrix corresponding to each time is taken as the pixel value corresponding to each time;

[0018] Based on the pixel value corresponding to each time in the time series of the prediction variable, generate a frequency spectrum corresponding to the prediction variable.

[0019] This invention, through local rate of change quantification and adaptive window length adjustment, enables time-frequency analysis to maintain high resolution in both abrupt and gradual change regions, more effectively capturing the intrinsic relationship between meteorological change trends and galloping events, enhancing the ability to characterize galloping risks under complex operating conditions. Furthermore, the conversion process from time series to spectrogram fully preserves the time-frequency coupling relationship of variables, making up for the shortcomings of traditional time series methods that ignore frequency domain features, and providing more comprehensive feature inputs for galloping prediction.

[0020] In one alternative implementation, the local rate of change is calculated using the following formula:

[0021]

[0022] In the formula, Indicating the time series of predictor variables in The local rate of change at time t; This represents the number of all time steps in the time series. Indicates the index of the time step; and These represent the predictor variables in Time and Sample data at any given time; Indicates that the predictor variable is continuous Time and The absolute value of the change in sample data between time points.

[0023] This invention characterizes the intensity of change of the predictor variable at different times by determining the local rate of change at each moment in the time series, thus providing data support for dynamically adjusting the window length.

[0024] In one alternative implementation, the window function length is calculated using the following formula:

[0025]

[0026] In the formula, Indicates the predictor variable in The length of the window function at time t; Indicates the preset maximum window function length; Indicates the preset minimum window function length; express The local rate of change at time t; This represents the maximum local rate of change.

[0027] This invention takes into account the inconsistent rates of change of the predictor variable at different times and adaptively adjusts the length of the window function used in the Fourier transform, making the window shorter for drastic changes and longer for stable changes, thereby enhancing the response capability to abrupt events and improving the resolution and robustness of the spectral features.

[0028] In an optional embodiment, generating the sample multi-channel image based on the spectrum corresponding to each prediction variable comprises:

[0029] Adjusting the size of the spectrum corresponding to each prediction variable to a preset size;

[0030] Preprocessing each spectrum after adjusting the size to obtain a preprocessed spectrum;

[0031] Stacking all preprocessed spectra in the channel dimension to obtain a sample multi-channel image.

[0032] The present application retains the unique time-frequency characteristics of each variable by adjusting the size and preprocessing, and strengthens the correlation between variables by channel stacking, providing high-quality input for subsequent model training, which helps to improve the accuracy, robustness and generalization ability of the dance prediction model.

[0033] In an optional embodiment, the initial dance prediction model is trained based on the sample multi-channel image and the sample label of each sample transmission line to obtain a target dance prediction model, comprising:

[0034] Dividing the sample multi-channel images of all sample transmission lines into a training set and a validation set;

[0035] Obtaining an expert weight of each prediction variable, and taking the expert weight of each prediction variable as an initial weight of each prediction variable;

[0036] Using the initial dance prediction model, weighting each sample multi-channel image in the training set based on the initial weight of each prediction variable to obtain a weighted feature map;

[0037] Based on the weighted feature map, a sample dance prediction probability is obtained;

[0038] Based on the sample dance prediction probability, the sample label corresponding to the sample multi-channel image, the expert weight of each prediction variable and the initial weight, a training loss is calculated;

[0039] Based on the training loss, the initial weight of each prediction variable and the model parameters of the initial dance prediction model are updated;

[0040] Based on the validation set, the performance of the initial dance prediction model is evaluated, and the step of using the initial dance prediction model and weighting each sample multi-channel image in the training set based on the initial weight of each prediction variable to obtain a weighted feature map is returned until the performance of the initial dance prediction model on the validation set reaches a preset stopping condition. The initial dance prediction model obtained by the last update is taken as the target dance prediction model.

[0041] This invention, through a systematic training process involving variable weighting, loss design, and iterative optimization, enables the target dancing prediction model to possess high precision, strong generalization, and high reliability, thereby improving the accuracy and generalization ability of dancing prediction.

[0042] In one alternative implementation, the training loss is calculated using the following formula:

[0043]

[0044] In the formula, Indicates training loss; This represents the loss between the predicted probability of sample dancing and the sample label corresponding to the multi-channel image of the sample; This indicates the preset weight adjustment parameters; Indicates the number of predictor variables; Indicates the index of the predictor variable; Indicates the first The current initial weights of the predictor variables; Indicates the first Expert weights for each predictor variable.

[0045] This invention introduces expert weights when calculating training loss, combining the advantages of data-driven and experience-driven approaches, which can effectively alleviate the problem of uneven sample distribution and improve the ability to recognize the boundary scenes where dancing occurs.

[0046] In one alternative implementation, the initial weights of each predictor variable are updated using the following formula:

[0047]

[0048] In the formula, Indicates the first Initial weights of each predictor variable after update; Indicates the first The current initial weights of the predictor variables, i.e., the initial weights before the update; This indicates the preset learning rate.

[0049] This invention updates the weights so that the model gradually learns the feature patterns related to dancing, while dynamically adjusting the variable weights to optimize the feature contribution.

[0050] In one alternative implementation, a time series is constructed for each predictor variable based on sample data of multiple predictor variables, including:

[0051] A sliding time window mechanism is used to construct a sample matrix based on sample data from multiple predictor variables;

[0052] The sample matrix is ​​processed to remove missing values ​​and then normalized to obtain the target matrix;

[0053] Each column in the target matrix is taken as a time series of a corresponding prediction variable.

[0054] The present application captures the cumulative effect and hysteresis of meteorological changes through a sliding time window mechanism, and missing value processing and normalization ensure data continuity, consistency and comparability, reducing the interference of noise and scale difference on model training, and the extracted time series provides a data basis for subsequent model training.

[0055] In a second aspect, the present application provides a galloping prediction device for a power transmission line, the device comprising:

[0056] A construction module is configured to, for each sample power transmission line, obtain a sample label and sample data of a plurality of prediction variables of the sample power transmission line, and construct a time series for each prediction variable based on the sample data of the plurality of prediction variables, the prediction variables including line parameter variables and meteorological variables;

[0057] A transformation module is configured to perform Fourier transform with an adaptive window length based on the time series of each prediction variable to obtain a frequency spectrum corresponding to each prediction variable;

[0058] A training module is configured to generate a sample multi-channel image based on the frequency spectrum corresponding to each prediction variable, and train an initial galloping prediction model based on the sample multi-channel image and the sample label of each sample power transmission line to obtain a target galloping prediction model;

[0059] An acquisition module is configured to acquire data of a plurality of prediction variables of a to-be-predicted power transmission line, and generate a target multi-channel image based on the data of the plurality of prediction variables;

[0060] A prediction module is configured to use the target galloping prediction model to perform prediction based on the target multi-channel image to obtain a galloping prediction result of the to-be-predicted power transmission line.

[0061] In an optional implementation, the transformation module comprises:

[0062] A first determination unit is configured to, for each prediction variable, determine a local change rate of the time series of the prediction variable at each time point;

[0063] A second determination unit is configured to determine a maximum local change rate from the local change rates at all time points;

[0064] A third determination unit is configured to determine a window function length corresponding to each time point based on the local change rate at each time point, the maximum local change rate, a preset maximum window function length and a preset minimum window function length;

[0065] The transformation unit is used to perform a short-time Fourier transform on the time series of the predicted variable using the window function length corresponding to each time step, so as to obtain the complex matrix corresponding to each time step.

[0066] The fourth determining unit is used to take the complex modulus of the complex matrix at each time step as the pixel value at each time step.

[0067] The generation unit is used to generate a spectrogram of the predictor variable based on the pixel value at each time step in the time series of the predictor variable.

[0068] In one alternative implementation, the local rate of change is calculated using the following formula:

[0069]

[0070] In the formula, The time series of predictor variables are represented in The local rate of change at time t; This represents the number of all time steps in the time series. Indicates the index of the time step; and These represent the predictor variables in Time and Sample data at any given time; Indicates that the predictor variable is continuous Time and The absolute value of the change in sample data between time points.

[0071] In one alternative implementation, the window function length is calculated using the following formula:

[0072]

[0073] In the formula, Indicates the predictor variable in The length of the window function at time t; Indicates the preset maximum window function length; Indicates the preset minimum window function length; express The local rate of change at time t; This represents the maximum local rate of change.

[0074] In one alternative implementation, the training module includes:

[0075] The adjustment unit is used to adjust the size of the spectrogram corresponding to each predictor variable to a preset size.

[0076] The preprocessing unit is used to preprocess each resized spectrogram to obtain a preprocessed spectrogram.

[0077] stacking unit, configured to stack all pre-processed spectrograms in a channel dimension to obtain a sample multi-channel image.

[0078] In an optional implementation, the training module comprises:

[0079] a dividing unit, configured to divide the sample multi-channel images of all sample transmission lines into a training set and a validation set;

[0080] an obtaining unit, configured to obtain an expert weight of each prediction variable and take the expert weight of each prediction variable as an initial weight of each prediction variable;

[0081] a weighting unit, configured to take an initial galloping prediction model, weight each sample multi-channel image in the training set based on the initial weight of each prediction variable, and obtain a weighted feature map;

[0082] a prediction unit, configured to perform prediction based on the weighted feature map to obtain a sample galloping prediction probability;

[0083] a calculation unit, configured to calculate a training loss based on the sample galloping prediction probability, a sample label corresponding to the sample multi-channel image, the expert weight of each prediction variable, and the initial weight;

[0084] an updating unit, configured to update the initial weight of each prediction variable and a model parameter of the initial galloping prediction model based on the training loss;

[0085] a training unit, configured to evaluate the performance of the initial galloping prediction model based on the validation set, return to the step of taking the initial galloping prediction model, weighting each sample multi-channel image in the training set based on the initial weight of each prediction variable, and obtaining a weighted feature map, until the performance of the initial galloping prediction model on the validation set reaches a preset stop condition, and take the initial galloping prediction model obtained through the last update as a target galloping prediction model.

[0086] In an optional implementation, the training loss is calculated by the following formula:

[0087]

[0088] In the formula, denotes the training loss; denotes a loss between the sample galloping prediction probability and a sample label corresponding to the sample multi-channel image; denotes a preset weight adjustment parameter; denotes the number of prediction variables; denotes the index of the prediction variable; denotes the current initial weight of the th prediction variable; Indicates the first Expert weights for each predictor variable.

[0089] In one alternative implementation, the initial weights of each predictor variable are updated using the following formula:

[0090]

[0091] In the formula, Indicates the first Initial weights of each predictor variable after update; Indicates the first The current initial weights of the predictor variables, i.e., the initial weights before the update; This indicates the preset learning rate.

[0092] In one alternative implementation, the building module includes:

[0093] The building unit is used to construct a sample matrix based on sample data from multiple predictor variables using a sliding time window mechanism.

[0094] The processing unit is used to handle missing values ​​in the sample matrix and normalize it to obtain the target matrix;

[0095] The fifth determining unit is used to take each column in the target matrix as the time series of the corresponding predictor variable.

[0096] Thirdly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the transmission line galloping prediction method of the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0097] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0098] Figure 1 This is a flowchart illustrating a method for predicting the galloping of transmission lines according to an embodiment of the present invention.

[0099] Figure 2 This is a structural block diagram of a transmission line galloping prediction device according to an embodiment of the present invention;

[0100] Figure 3 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0101] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application. The present application improves the ability to describe the galloping inducement by considering multiple prediction variables and introducing Fourier transform, so that the model can still make reliable prediction under the condition of sample scarcity or extreme meteorological scene, and the generalization ability is improved.

[0102] Most of the existing galloping prediction methods directly model and predict the time series by using a model, which captures certain time trends and nonlinear characteristics, but ignores other galloping inducements and is difficult to capture the correlation between characteristics, resulting in poor prediction accuracy and weak generalization ability.

[0103] According to the embodiments of the present application, a galloping prediction method for a power transmission line is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0104] In the present embodiment, a galloping prediction method for a power transmission line is provided, which can be used in terminals such as computers and the like, Figure 1 The galloping prediction method for a power transmission line according to the embodiments of the present application is shown in the flowchart as Figure 1 The flowchart includes the following steps:

[0105] In step S101, for each sample power transmission line, the sample label and the sample data of multiple prediction variables of the sample power transmission line are obtained, and a time series is constructed for each prediction variable based on the sample data of the multiple prediction variables, the prediction variables including line parameter variables and meteorological variables.

[0106] Specifically, if only a single type of variable is relied on when performing the dance prediction, the dance inducement is not fully characterized, resulting in poor prediction accuracy. Therefore, in the embodiments of the present application, both line parameter variables and meteorological variables are used as prediction variables. The line parameter variables include line direction and span, which directly reflect the influence of the structure characteristics of the line itself on the dance sensitivity. The meteorological variables include temperature, relative humidity, precipitation, wind speed, wind direction, ice thickness, and ice duration, which reflect the influence of environmental factors on the dance. According to the above-mentioned multiple prediction variables, corresponding sample data is obtained. When obtaining the sample data corresponding to the meteorological variables, the sample data can be queried in the ice dance database according to the latitude and longitude of the sample transmission line. Considering the lag correlation between the dance and the weather evolution process, such as the need for a certain time accumulation for ice formation and the need for continuous influence of wind force, the sample input sequence of each prediction variable is constructed according to the sample data of the above-mentioned prediction variables, so as to capture the influence of the previous meteorological changes on the dance event.

[0107] In step S102, Fourier transform is performed on the time sequence of each prediction variable based on an adaptive window length, to obtain a frequency spectrum corresponding to each prediction variable.

[0108] Specifically, the prediction variables, especially the meteorological variables, usually exhibit complex dynamic change characteristics. It is difficult to capture the change trend and energy pattern of the meteorological conditions by directly using the time sequence for prediction. Therefore, by performing short-time Fourier transform on the time sequence, the time sequence is converted into a frequency spectrum in the time-frequency domain, which can reflect the energy distribution of the prediction variables at different frequencies. Moreover, the influence of the prediction variables such as wind speed and wind direction on the dance has a certain time lag effect. Therefore, the frequency spectrum image extracted by the short-time Fourier transform can better capture these time lag effects, thereby improving the accuracy of the dance prediction. In addition, by using the short-time Fourier transform with an adaptive window length, unlike the traditional short-time Fourier transform with a fixed window length, the internal relationship between the meteorological change trend and the dance occurrence can be more effectively captured, and the ability to characterize the dance risk under complex working conditions can be improved.

[0109] In step S103, a sample multi-channel image is generated based on the frequency spectrum corresponding to each prediction variable, and an initial dance prediction model is trained based on the sample multi-channel image and the sample label of each sample transmission line, to obtain a target dance prediction model.

[0110] Specifically, when making galloping prediction, the influence of multiple prediction variables on the galloping of the power transmission line needs to be considered at the same time. The spectrogram of each prediction variable represents the time-frequency distribution characteristics of the variable, and these characteristics have different influences on galloping prediction. Therefore, the spectrogram of each prediction variable is taken as an independent channel, and all image channels are fused to obtain a multi-channel image. The model is trained using the multi-channel image, so that the model can simultaneously focus on the changes of each prediction variable at different frequencies and time periods, thereby better identifying the time-frequency characteristics related to the galloping state, and helping the model to learn which frequency bands have a strong correlation with the galloping state. For example, some frequency components may be closely related to changes in wind speed, while other frequency components may be closely related to changes in temperature or humidity. Moreover, the model can better learn the complex relationship between meteorological factors and the galloping state, avoid the local overfitting problem caused by traditional variable-by-variable modeling, and have stronger generalization ability compared to traditional galloping prediction with scarce samples.

[0111] In step S104, data of multiple prediction variables of the power transmission line to be predicted are collected, and a target multi-channel image is generated based on the data of the multiple prediction variables.

[0112] Specifically, in the actual galloping prediction scenario, for the power transmission line to be predicted, the corresponding data is extracted according to the multiple prediction vectors in step S101, and the processing process of the above data is referred to the processing process of the above data in steps S102 and S103, and a target multi-channel image suitable for inputting the model is generated.

[0113] In step S105, a target galloping prediction model is used to make prediction based on the target multi-channel image, and a galloping prediction result of the power transmission line to be predicted is obtained.

[0114] Specifically, the target multi-channel image is input into the target galloping prediction model which has been trained. The model has learned the internal correlation between multivariate time-frequency characteristics and galloping through pre-training, can automatically extract key features from the target multi-channel image, and perform inference calculation based on these features. A value between 0 and 1 is output, which is the probability of galloping of the power transmission line to be predicted, as the galloping prediction result. The galloping probability quantifies the risk degree of galloping of the power transmission line to be predicted, which can be used to evaluate the line operation state, such as the higher the galloping probability, the higher the risk level, and can also be used to judge "galloping occurs" or "galloping does not occur" through the size relationship between the preset threshold value, thereby providing accurate data support for the operation and maintenance decision of the power transmission line.

[0115] This invention, by simultaneously considering line parameter variables and meteorological variables, covers both structural and environmental causes of galloping, solving the problem of incomplete characterization of causes due to single variables in traditional methods. It introduces Fourier transform into galloping prediction, converting the one-dimensional time series constructed from the aforementioned data into a two-dimensional spectrogram, preserving the time-frequency characteristics of meteorological variables and overcoming the deficiency of traditional time series methods in ignoring frequency domain information. Simultaneously, an adaptive window length mechanism overcomes the inherent limitations of fixed window lengths in time-frequency resolution, more effectively capturing the intrinsic relationship between meteorological trends and galloping occurrence, improving the ability to characterize galloping risks under complex operating conditions, and enhancing the specificity and robustness of spectral features. By generating multi-channel spectrograms based on all spectrograms and training the model accordingly, the synergistic relationships between variables are preserved, avoiding local overfitting caused by traditional variable-by-variable modeling, enabling the model to learn the complex patterns of galloping causes from a holistic perspective. In practical applications, the trained target model is used to predict the transmission line under test, quantifying the degree of risk and providing accurate quantitative basis for operation and maintenance decisions. By considering multiple predictor variables and introducing Fourier transform, the ability to characterize the causes of galloping is improved, enabling the model to make reliable predictions even in scenarios with scarce samples or extreme weather conditions. It can also be extended to other wind-induced fault prediction fields such as icing and wire skipping, making it widely applicable and highly generalizable.

[0116] This embodiment provides a method for predicting the galloping of transmission lines, which specifically includes the following steps:

[0117] Step S201: For each sample transmission line, obtain the sample label of the sample transmission line and sample data of multiple predictor variables, and construct a time series for each predictor variable based on the sample data of multiple predictor variables. The predictor variables include line parameter variables and meteorological variables.

[0118] In some optional implementations, step S201 above constructs a time series for each predictor variable based on sample data of multiple predictor variables, including:

[0119] Step S2011: Using a sliding time window mechanism, a sample matrix is ​​constructed based on the sample data of multiple predictor variables. Specifically, assuming that the data sampling frequency of the predictor variables is once per hour, and the collection time window is set to the past T hours, the sample matrix shown in equation (1) can be constructed.

[0120] (1)

[0121] In the formula, Represents the sample matrix; Indicates the first The predictor variables are in Sample data at any given time; Indicates the data collection time window; The number of prediction variables.

[0122] In step S2012, missing value processing is performed on the sample matrix, and normalization is performed to obtain a target matrix. Specifically, for the above-mentioned sample matrix, part of the data may be missing due to sensor failure, transmission interruption, etc. during data acquisition. If the data containing missing values is directly used for subsequent processing, it will cause prediction deviation. Therefore, by using interpolation method to fill in the missing values, the continuity of the data can be maintained, and the interference of the missing values can be avoided. Then, since the numerical ranges of different prediction variables are quite different, such as temperature may be-10℃~30℃, wind speed may be 0~20m / s, and span may be 100m~500m, if these data are directly used for model training, the prediction variables with large numerical range may cover up the influence of the prediction variables with small numerical range, resulting in unbalanced learning of the model to different prediction variables. Therefore, Min-Max normalization is adopted, and each sample data of all prediction variables is mapped to a unified numerical interval by the following formula (2).

[0123] (2)

[0124] In the formula, denotes the normalized sample data; denotes the sample data; denotes the minimum value in the numerical range of the prediction variable corresponding to the sample data; denotes the maximum value in the numerical range of the prediction variable corresponding to the sample data.

[0125] In step S2013, each column in the target matrix is taken as a time series of the corresponding prediction variable. Specifically, by performing missing value processing and normalization on the sample matrix, each column data in the target matrix is taken as a time series of the corresponding prediction variable, which solves the problem of incomplete data, ensures the continuity of time series, eliminates the influence of dimension, balances the role of each prediction variable, provides high-quality and standardized input data for subsequent model training, and improves the stability and prediction accuracy of the model.

[0126] In step S202, Fourier transform with adaptive window length is performed based on the time series of each prediction variable to obtain a frequency spectrum corresponding to each prediction variable.

[0127] Specifically, the above-mentioned step S202 includes:

[0128] In step S2021, for each prediction variable, the local change rate of the time series of the prediction variable at each time is determined. Specifically, for the time series of each prediction variable, the local change rate at each time is determined by the following formula (3) to characterize the change intensity of the prediction variable at different times.

[0129] (3)

[0130] wherein, denotes the local change rate of the time series of the prediction variable at the time instant denotes the number of all time steps of the time series; denotes the index of the time step; and denote the sample data of the prediction variable at the time instant and the time instant denotes the absolute value of the change amount of the sample data of the prediction variable between the consecutive time instants and

[0131] Step S2022, determining the local change rate maximum value from the local change rates of all time instants. Specifically, the local change rates of all time instants in the time series are traversed to determine the maximum value therein.

[0132] Step S2023, determining the window function length corresponding to each time instant based on the local change rate of each time instant, the local change rate maximum value, the preset maximum window function length and the preset minimum window function length. Specifically, if a fixed window length is used for short-time Fourier transform, the abrupt change points and slowly varying segments in the data cannot be effectively captured, therefore the window function length corresponding to each time instant is dynamically determined by the following formula (4) to improve the accuracy and robustness of time-frequency analysis. Optionally, the embodiment of the present application adopts a Hamming window function, which can be selected by the actual demand, and the present application does not limit the window function.

[0133] (4)

[0134] wherein, denotes the window function length at the time instant denotes the preset maximum window function length; denotes the preset minimum window function length; denotes the local change rate at the time instant denotes the local change rate maximum value.

[0135] Step S2024, for the time series of the prediction variable, performing short-time Fourier transform on the time series by using the window function length corresponding to each time instant to obtain the complex matrix corresponding to each time instant. Specifically, since it is difficult to capture the change trend and energy pattern of the prediction variable by only considering the time sequence characteristics, the time-frequency characteristics of each time instant are extracted by using the window function length corresponding to each time instant in the time series by the short-time Fourier transform shown in the following formula (5), which is helpful to improve the accuracy of the flutter prediction.​​​​​

[0136] (5)

[0137] wherein, denotes the time instant, a complex matrix corresponding to the frequency; denotes the time index; denotes the sample data at time; denotes the window function length at time instant; denotes the window function with length . denotes the complex exponential term, which is the basic kernel function of the short-time Fourier transform, used to extract the component with frequency .

[0138] At step S2025, the complex modulus of the complex matrix corresponding to each time instant is taken as the pixel value corresponding to each time instant. Specifically, for the complex matrix obtained at each time instant, the complex modulus is calculated and taken as the pixel value corresponding to the time instant.

[0139] At step S2026, a frequency spectrum graph corresponding to the prediction variable is generated based on the pixel value corresponding to each time instant in the time series of the prediction variable. Specifically, the frequency spectrum graph is a two-dimensional image, which can be effectively recognized by the model, and converts the time-frequency feature into an image feature, facilitating the model to capture the frequency band and time pattern related to the dance. By converting the one-dimensional time series into a frequency spectrum graph with both time and frequency information, the dynamic change rule of the prediction variable is retained, and the resolution of the key feature is improved through an adaptive mechanism, providing high-quality input features for subsequent model training, which helps to improve the accuracy and robustness of dance prediction.

[0140] At step S203, a sample multi-channel image corresponding to each prediction variable is generated based on the frequency spectrum graph, and an initial dance prediction model is trained based on the sample multi-channel image and the sample label of each sample transmission line, to obtain a target dance prediction model.

[0141] Specifically, the above step S203 includes:

[0142] At step S2031, the size of the frequency spectrum graph corresponding to each prediction variable is adjusted to a preset size. Specifically, since the time series of each prediction variable is adaptively set with a window function length for each time instant when performing short-time Fourier transform, the generated frequency spectrum graphs have inconsistent sizes. Therefore, all the frequency spectrum graphs are adjusted to the preset size by using the bilinear interpolation shown in the following formula (6) or by using other methods, so that the final multi-channel image has a uniform size, facilitating subsequent processing and model training.

[0143] (6)

[0144] wherein, represents a preset size of a spectrogram; represents an interpolation function; represents a spectrogram before adjustment; represents a preset size, represents a number of frames in a time axis direction, represents a number of frequency points in a frequency direction.

[0145] In step S2032, each spectrogram after adjustment is preprocessed to obtain a preprocessed spectrogram. Specifically, a logarithmic transformation (Log Transformation) is used to enhance the contrast of the low value area in the spectrogram, highlighting the weak change characteristics. Subsequently, Gaussian filtering (Gaussian Filtering) is applied to smooth the spectrogram, suppress the noise interference caused by local sharp abnormal points, and enhance the continuity of the image, which helps to improve the robustness and generalization ability of the model.

[0146] In step S2033, all preprocessed spectrograms are stacked in the channel dimension to obtain a sample multi-channel image. Specifically, when predicting the dance, multiple prediction variables need to be considered simultaneously, and the spectrogram of each prediction variable represents the time-frequency distribution characteristics of the variable, which has different effects on the dance state prediction. Therefore, the spectrograms of all prediction variables are stacked in the channel dimension to form a sample multi-channel image with a size of as the input of the model.

[0147] In step S2034, the sample multi-channel images of all sample transmission lines are divided into a training set and a validation set. Specifically, all sample multi-channel images are randomly divided according to a preset ratio, the training set is used for model parameter updating, and the validation set is used to monitor the performance change in the training process to prevent overfitting of the model.

[0148] In step S2035, the expert weight of each prediction variable is obtained, and the expert weight of each prediction variable is used as the initial weight of each prediction variable. Specifically, the expert weight is assigned to each prediction variable according to the expert experience to reflect the theoretical influence degree of different prediction variables in the occurrence of dance. At the same time, the expert weight is used as the initial weight, which is updated through model training.

[0149] ​Step S2036, using the initial galloping prediction model, weighting each sample multi-channel image in the training set based on the initial weight of each prediction variable to obtain a weighted feature map. Specifically, the initial galloping prediction model uses the structure of an existing deep learning model, which requires it to be able to process multi-channel image input and adapt to different spectral distributions, such as convolutional neural networks and time convolutional networks. Input any sample multi-channel image in the training set into the initial galloping prediction model. The model weights the spectral graph corresponding to each prediction variable in the sample multi-channel image according to the initial weight of each prediction variable by formula (7) below, and takes all the weighted spectral graphs as the weighted feature map to adjust the feature importance of each prediction variable and make the model focus more on the time-frequency features of the key variables.

[0150] (7)

[0151] In the formula, represents the weighted spectral graph of the spectral graph in the i-th prediction variable; represents the initial weight, which is the expert weight when step S2036 is performed for the first time, i.e., the first round of training; represents the spectral graph.

[0152] Step S2037, based on the weighted feature map, the sample galloping prediction probability is obtained. Specifically, the high-level features are extracted from the weighted feature map using the multiple convolutional layers and pooling layers of the initial galloping prediction model, and finally the probability value between 0 and 1, i.e., the sample galloping prediction probability, is output through the fully connected layer and the sigmoid activation function.

[0153] Step S2038, based on the sample galloping prediction probability, the sample label corresponding to the sample multi-channel image, the expert weight and the initial weight of each prediction variable, the training loss is calculated. Specifically, the sample label is used to identify whether the sample transmission line has galloping, if it has galloping, the sample label is 1, if it does not have galloping, the sample label is 0. The cross-entropy loss between the sample galloping prediction probability and the sample label is calculated as the prediction loss of the model. At the same time, in order to prevent the model from deviating too far from the expert experience setting during optimization, and at the same time improve the stability of the model in small sample and complex weather scenarios, an expert weight constraint term is introduced in the calculation of the training loss, and the training loss shown in formula (8) is obtained.

[0154] (8)

[0155] In the formula, represents the training loss; represents the loss between the sample galloping prediction probability and the sample label corresponding to the sample multi-channel image; represents a preset weight adjustment parameter;​ represents the number of prediction variables; represents the index of the prediction variable; represents the initial weight of the i-th prediction variable at present; represents the initial weight of the i-th prediction variable at present; represents the expert weight of the i-th prediction variable. Step S2039, updating the initial weight of each prediction variable and the model parameter of the initial dance prediction model based on the training loss. Specifically, using a suitable optimizer such as the Adam optimizer, the model parameter and the initial weight are updated by gradient descent method, so that the model gradually learns the feature pattern related to the dance, and the variable weight is dynamically adjusted to optimize the feature contribution. Wherein, the initial weight of each prediction variable is updated by the following formula (9).

[0156]

[0157] (9) In the formula,

[0158] represents the updated initial weight of the i-th prediction variable; represents the initial weight of the i-th prediction variable at present, i.e. the initial weight before updating; represents the initial weight of the i-th prediction variable at present, i.e. the initial weight before updating; represents the preset learning rate. Step S20310, evaluating the performance of the initial dance prediction model based on the validation set, returning to the step of weighting the sample multi-channel image in the training set based on the initial weight of each prediction variable using the initial dance prediction model to obtain the weighted feature map, until the performance of the initial dance prediction model on the validation set reaches the preset stopping condition, and taking the initial dance prediction model obtained by the last update as the target dance prediction model. Specifically, the preset stopping condition is set based on the early stopping mechanism, for example, the training is stopped when the model performance does not improve for a preset number of consecutive times. After each round of training, the sample multi-channel image of the validation set is input into the current initial dance prediction model to evaluate the performance of the model obtained by this training. If the stopping condition is not met, return to step S2036 to weight, predict, calculate the loss and update the parameter based on the current updated initial weight and model parameter for the training set sample, and enter the next round of training. If the stopping condition is met, the training is stopped, and the model obtained by the last update is taken as the target dance prediction model. By training the model, it is ensured that the target dance prediction model can better adapt to the dance prediction scene of the actual power transmission line and still maintain high-precision prediction under complex weather conditions, providing reliable basis for the operation and maintenance decision of the power transmission line.

[0159] Step S20310, evaluating the performance of the initial dance prediction model based on the validation set, returning to the step of weighting the sample multi-channel image in the training set based on the initial weight of each prediction variable using the initial dance prediction model to obtain the weighted feature map, until the performance of the initial dance prediction model on the validation set reaches the preset stopping condition, and taking the initial dance prediction model obtained by the last update as the target dance prediction model. Specifically, the preset stopping condition is set based on the early stopping mechanism, for example, the training is stopped when the model performance does not improve for a preset number of consecutive times. After each round of training, the sample multi-channel image of the validation set is input into the current initial dance prediction model to evaluate the performance of the model obtained by this training. If the stopping condition is not met, return to step S2036 to weight, predict, calculate the loss and update the parameter based on the current updated initial weight and model parameter for the training set sample, and enter the next round of training. If the stopping condition is met, the training is stopped, and the model obtained by the last update is taken as the target dance prediction model. By training the model, it is ensured that the target dance prediction model can better adapt to the dance prediction scene of the actual power transmission line and still maintain high-precision prediction under complex weather conditions, providing reliable basis for the operation and maintenance decision of the power transmission line.

[0160] ​In some optional embodiments, when evaluating the performance of the model by using the validation set, the sample multichannel image in the validation set is input into the model to obtain a probability value. The probability value is compared with a preset threshold to determine whether the validation result is "galloping occurs" or "galloping does not occur", so as to calculate a plurality of validation indexes such as accuracy, precision, recall and F1 score based on the validation result and the sample label. Optionally, the accuracy is calculated by formula (10) below, which measures the overall prediction correctness of the validation set; the precision is calculated by formula (11) below, which measures the proportion of true galloping in the samples predicted to gallop; the recall is calculated by formula (12) below, which measures the proportion of the samples actually galloping that are correctly predicted; and the F1 score is calculated by formula (13) below, which is the harmonic mean of the precision and the recall.

[0161] (10)

[0162] In the formula, denotes the accuracy; denotes the true positive, that is, the number of samples whose sample label is 1 and whose validation result is galloping occurs; denotes the true negative, that is, the number of samples whose sample label is 0 and whose validation result is galloping does not occur; denotes the false positive, that is, the number of samples whose sample label is 0 but whose validation result is galloping occurs; denotes the false negative, that is, the number of samples whose sample label is 1 but whose validation result is galloping does not occur.

[0163] (11)

[0164] In the formula, denotes the precision; denotes the true positive; denotes the false positive.

[0165] (12)

[0166] In the formula, denotes the recall; denotes the true positive; denotes the false negative.

[0167] (13)

[0168] In the formula, denotes the F1 score; denotes the precision; denotes the recall.

[0169] In step S204, data of a plurality of prediction variables of the power transmission line to be predicted is collected, and a target multichannel image is generated based on the data of the plurality of prediction variables. For details, please refer toFigure 1 Step S104 of the embodiment shown will not be repeated here.

[0170] Step S205, using the target galloping prediction model, based on the target multi-channel image, the prediction result of the galloping prediction of the to-be-predicted transmission line is obtained. For details, please refer to Figure 1 Step S105 of the embodiment shown will not be repeated here.

[0171] The present application covers the structural and environmental causes of galloping by considering line parameter variables and meteorological variables simultaneously, solves the problem of incomplete cause description caused by single variable in traditional methods, introduces Fourier transform into galloping prediction, converts the one-dimensional time series constructed by the above data into a two-dimensional frequency spectrum, retains the time-frequency characteristics of meteorological variables, makes up for the defects of traditional time series methods ignoring frequency domain information, and solves the inherent limitations of fixed window length in time-frequency resolution through the adaptive window length mechanism, more effectively captures the internal relationship between meteorological change trend and galloping occurrence, and improves the description ability of galloping risk under complex working conditions. By generating a multi-channel frequency spectrum based on all frequency spectra, the model is trained, the cooperative relationship between variables is retained, local overfitting caused by traditional variable-by-variable modeling is avoided, and the model can learn the complex pattern of galloping causes from a holistic perspective. In practical application, the trained target model is used to predict the to-be-predicted transmission line, which not only quantifies the risk level, but also provides accurate quantitative basis for operation and maintenance decision-making. By considering multiple prediction variables and introducing Fourier transform, the description ability of galloping causes is improved, the model can still reliably predict under the condition of sample scarcity or extreme weather, and can be extended to other wind-induced failure prediction fields such as icing and jumper, with wide application range and strong generalization ability.

[0172] In the embodiment, a galloping prediction device for a transmission line is also provided, which is used to implement the above embodiments and preferred embodiments, and will not be repeated here. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware or a combination of software and hardware is also possible and contemplated.

[0173] The embodiment provides a galloping prediction device for a transmission line, which comprises Figure 2 as shown, comprising:

[0174] The construction module 201 is configured to, for each sample transmission line, acquire a sample label and sample data of a plurality of prediction variables of the sample transmission line, and construct a time series for each prediction variable based on the sample data of the plurality of prediction variables, the prediction variables including line parameter variables and meteorological variables.

[0175] The transformation module 202 is configured to perform a Fourier transform with an adaptive window length based on the time sequence of each prediction variable to obtain a frequency spectrum corresponding to each prediction variable.

[0176] The training module 203 is configured to generate a sample multi-channel image based on the frequency spectrum corresponding to each prediction variable, and train an initial galloping prediction model based on the sample multi-channel image and the sample label of each sample transmission line to obtain a target galloping prediction model.

[0177] The collection module 204 is configured to collect data of a plurality of prediction variables of a transmission line to be predicted, and generate a target multi-channel image based on the data of the plurality of prediction variables.

[0178] The prediction module 205 is configured to perform prediction based on the target multi-channel image by using the target galloping prediction model to obtain a galloping prediction result of the transmission line to be predicted.

[0179] In some optional embodiments, the transformation module 202 includes:

[0180] The first determination unit is configured to determine, for each prediction variable, a local change rate of the time sequence of the prediction variable at each time.

[0181] The second determination unit is configured to determine a maximum local change rate from the local change rates at all times.

[0182] The third determination unit is configured to determine, based on the local change rate at each time, the maximum local change rate, a preset maximum window function length, and a preset minimum window function length, a window function length corresponding to each time.

[0183] The transformation unit is configured to perform a short-time Fourier transform on the time sequence of the prediction variable by using the window function length corresponding to each time to obtain a complex matrix corresponding to each time.

[0184] The fourth determination unit is configured to take a complex modulus of the complex matrix corresponding to each time as a pixel value corresponding to each time.

[0185] The generation unit is configured to generate a frequency spectrum corresponding to the prediction variable based on the pixel value corresponding to each time in the time sequence of the prediction variable.

[0186] In some optional embodiments, the local change rate is calculated by the following formula:

[0187]

[0188] In the formula, represents the local change rate of the time sequence of the prediction variable at time t; and represents a first derivative of the time sequence of the prediction variable at time t. ​​a number of all time steps of the time series; an index of a time step; and respectively represent sample data of the prediction variable at and respectively represent sample data of the prediction variable at represent an absolute value of a change amount of sample data of the prediction variable between consecutive and respectively.

[0189] In some optional embodiments, the window function length is calculated by the following formula:

[0190]

[0191] wherein, represents a window function length of the prediction variable at represents a preset maximum window function length; represents a preset minimum window function length; represents a local change rate at represents a maximum value of the local change rate.

[0192] In some optional embodiments, the training module 203 comprises:

[0193] an adjusting unit, configured to adjust a size of the spectrum diagram corresponding to each prediction variable to a preset size.

[0194] a preprocessing unit, configured to pre-process each spectrum diagram after the size adjustment to obtain a pre-processed spectrum diagram.

[0195] a stacking unit, configured to stack all pre-processed spectrum diagrams in a channel dimension to obtain a sample multi-channel image.

[0196] In some optional embodiments, the training module 203 comprises:

[0197] a dividing unit, configured to divide the sample multi-channel images of all sample transmission lines into a training set and a verification set.

[0198] an obtaining unit, configured to obtain an expert weight of each prediction variable, and take the expert weight of each prediction variable as an initial weight of each prediction variable.

[0199] a weighting unit, configured to use an initial galloping prediction model to weight each sample multi-channel image in the training set based on the initial weight of each prediction variable, to obtain a weighted feature map.

[0200] ​​The prediction unit is used to make predictions based on the weighted feature map to obtain the predicted probability of sample dancing.

[0201] The computation unit is used to calculate the training loss based on the sample dance prediction probability, the sample label corresponding to the multi-channel image of the sample, the expert weight and the initial weight of each prediction variable.

[0202] The update unit is used to update the initial weights of each predictor variable and the model parameters of the initial dance prediction model based on the training loss.

[0203] The training unit is used to evaluate the performance of the initial dance prediction model based on the validation set. It returns to the step of using the initial dance prediction model, weighting any sample multi-channel image in the training set based on the initial weight of each predictor variable, and obtaining a weighted feature map, until the performance of the initial dance prediction model on the validation set reaches the preset stopping condition. The initial dance prediction model obtained from the last update is used as the target dance prediction model.

[0204] In some alternative implementations, the training loss is calculated using the following formula:

[0205]

[0206] In the formula, Indicates training loss; This represents the loss between the predicted probability of sample dancing and the sample label corresponding to the multi-channel image of the sample; This indicates the preset weight adjustment parameters; Indicates the number of predictor variables; Indicates the index of the predictor variable; Indicates the first The current initial weights of the predictor variables; Indicates the first Expert weights for each predictor variable.

[0207] In some optional implementations, the initial weights of each predictor variable are updated using the following formula:

[0208]

[0209] In the formula, Indicates the first Initial weights of each predictor variable after update; Indicates the first The current initial weights of the predictor variables, i.e., the initial weights before the update; This indicates the preset learning rate.

[0210] In some alternative implementations, the construction module 201 includes:

[0211] The building unit is used to construct a sample matrix based on sample data from multiple predictor variables using a sliding time window mechanism.

[0212] The processing unit is used to handle missing values ​​in the sample matrix and normalize it to obtain the target matrix.

[0213] The fifth determining unit is used to take each column in the target matrix as the time series of the corresponding predictor variable.

[0214] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0215] In this embodiment, the transmission line galloping prediction device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0216] This invention also provides a computer device having the above-described features. Figure 2 The device shown is for predicting the galloping of power transmission lines.

[0217] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 3 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 3 Take a processor 10 as an example.

[0218] The processor 10 can be a central processing unit, a network processing unit, or a combination thereof. The processor 10 can further include a hardware chip. The hardware chip can be an application specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device can be a complex programmable logic device, a field programmable logic device, a general array logic, or any combination thereof.

[0219] The memory 20 stores instructions executable by the at least one processor 10 to cause the at least one processor 10 to perform the methods illustrated in the above embodiments.

[0220] The memory 20 can include a program storage area and a data storage area. The program storage area can store an operating system and application programs required by at least one function. The data storage area can store data created according to the use of the computer device, and the like. In addition, the memory 20 can include a high-speed random access memory, and can further include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some alternative embodiments, the memory 20 can optionally include a memory disposed remotely from the processor 10, which can be connected to the computer device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0221] The memory 20 can include a volatile memory, such as a random access memory, and can also include a non-volatile memory, such as a flash memory, a hard disk, or a solid state disk. The memory 20 can further include a combination of the above-mentioned types of memories.

[0222] The computer device further includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30, and the output device 40 can be connected by a bus or other means, Figure 3 For example, by a bus.

[0223] The input device 30 can receive input digital or character information, and generate key signal inputs related to the user settings and function controls of the computer device, such as a touch screen, a keypad, a mouse, a trackpad, a touchpad, a pointing stick, one or more mouse buttons, a trackball, a joystick, and the like. The output device 40 can include a display device, an auxiliary lighting device (e.g., an LED), a tactile feedback device (e.g., a vibration motor), and the like. The display device includes, but is not limited to, a liquid crystal display, a light emitting diode, a display, and a plasma display. In some alternative embodiments, the display device can be a touch screen.

[0224] The embodiments of the present application further provide a computer readable storage medium, and the method according to the embodiments of the present application can be implemented in hardware, firmware, or recorded in a storage medium, or stored in a remote storage medium or a non-transitory machine readable storage medium and downloaded to a local storage medium through network, so that the method described herein can be processed by such software on a storage medium using a general purpose computer, a special purpose processor, or programmable or special hardware. The storage medium can be a magnetic disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid state disk, etc. Further, the storage medium can also include a combination of the above-mentioned memories. It can be understood that the computer, the processor, the microprocessor controller, or the programmable hardware includes a storage component that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the method shown in the above embodiments is implemented.

[0225] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, the operation of the computer can invoke or provide the method and / or technical solutions according to the present application. Those skilled in the art should understand that the form of computer program instructions in computer readable medium includes but is not limited to source file, executable file, installation package file, etc. Correspondingly, the way of computer program instructions executed by computer includes but is not limited to: the computer directly executes the instructions, or the computer compiles the instructions and then executes the corresponding compiled program, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer readable medium can be any available computer readable storage medium or communication medium accessible to the computer.

[0226] Although the embodiments of the present application are described in conjunction with the accompanying drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A galloping prediction method of a power transmission line, characterized by, The method comprises: For each sample transmission line, sample labels and sample data of a plurality of prediction variables of the sample transmission line are obtained, and a time series is constructed for each prediction variable based on the sample data of the plurality of prediction variables, the prediction variables including line parameter variables and meteorological variables; Performing Fourier transform with an adaptive window length based on the time series of each prediction variable to obtain a frequency spectrum corresponding to each prediction variable; Generating a sample multi-channel image based on the frequency spectrum corresponding to each prediction variable, and training an initial galloping prediction model based on the sample multi-channel image and the sample labels of each sample transmission line to obtain a target galloping prediction model; Collecting data of a plurality of prediction variables of a to-be-predicted transmission line, and generating a target multi-channel image based on the data of the plurality of prediction variables; Using the target galloping prediction model to perform prediction based on the target multi-channel image to obtain a galloping prediction result of the to-be-predicted transmission line; The Fourier transform with the adaptive window length based on the time series of each prediction variable to obtain the frequency spectrum corresponding to each prediction variable comprises: For each prediction variable, determining a local change rate of the time series of the prediction variable at each time; Determining a maximum local change rate from all local change rates; Determining a window function length corresponding to each time based on the local change rate at each time, the maximum local change rate, a preset maximum window function length, and a preset minimum window function length; For the time series of the prediction variable, performing short-time Fourier transform on the time series using the window function length corresponding to each time to obtain a complex matrix corresponding to each time; Taking a complex modulus of the complex matrix corresponding to each time as a pixel value corresponding to each time; Generating the frequency spectrum corresponding to the prediction variable based on the pixel value corresponding to each time in the time series of the prediction variable; The window function length is calculated by the following formula: In the formula, represents a window function length at a time point of represents a preset maximum window function length; represents a preset minimum window function length; represents a local change rate at a time point of represents a local change rate maximum value.​​ 2. The method of claim 1, wherein, The local change rate is calculated by the following formula: wherein denotes the local rate of change of the time series of the prediction variable at time ; denotes the number of all time steps of the time series; denotes the index of a time step; and denote the sample data of the prediction variable at time and time , respectively; denotes the absolute value of the amount of change of the sample data of the prediction variable between consecutive time and time .

3. The method of claim 1, wherein, The generating of the sample multi-channel image based on the frequency spectrum corresponding to each prediction variable comprises: Adjusting the size of the frequency spectrum corresponding to each prediction variable to a preset size; Preprocessing each frequency spectrum after the size is adjusted to obtain a preprocessed frequency spectrum; Stacking all preprocessed frequency spectrums in the channel dimension to obtain the sample multi-channel image.

4. The method of claim 1, wherein, The training of the initial galloping prediction model based on the sample multi-channel image and the sample labels of each sample transmission line to obtain the target galloping prediction model comprises: Dividing the sample multi-channel images of all sample transmission lines into a training set and a validation set; Obtaining an expert weight of each prediction variable, and taking the expert weight of each prediction variable as an initial weight of each prediction variable; Using the initial galloping prediction model to weight each sample multi-channel image in the training set based on the initial weight of each prediction variable to obtain a weighted feature map; Performing prediction based on the weighted feature map to obtain a sample galloping prediction probability; calculate a training loss based on the sample flutter prediction probability, the sample label corresponding to the sample multi-channel image, the expert weight of each prediction variable, and the initial weight; update the initial weight of each prediction variable and the model parameter of the initial flutter prediction model based on the training loss; evaluate the performance of the initial flutter prediction model based on the validation set, return to the step of weighting the sample multi-channel image in the training set based on the initial weight of each prediction variable by using the initial flutter prediction model to obtain a weighted feature map, and repeat the process until the performance of the initial flutter prediction model on the validation set reaches a preset stop condition, and take the initial flutter prediction model obtained after the last update as the target flutter prediction model.

5. The method of claim 4, wherein, The training loss is calculated by the following formula: In the formula, Indicates training loss; This represents the loss between the predicted probability of sample dancing and the sample label corresponding to the multi-channel image of the sample; This indicates the preset weight adjustment parameters; Indicates the number of predictor variables; Indicates the index of the predictor variable; Indicates the first The current initial weights of the predictor variables; Indicates the first Expert weights for each predictor variable.

6. The method of claim 4, wherein, The initial weight of each prediction variable is updated by the following formula: In the formula, represents the initial weight of the first prediction variable after updating; represents the initial weight of the first prediction variable after updating; represents the initial weight of the first prediction variable before updating; represents the initial weight of the first prediction variable before updating; represents a preset learning rate.

7. The method of claim 1, wherein, The sample data of the plurality of prediction variables is used to construct a time series for each prediction variable, including: a sliding time window mechanism is used to construct a sample matrix based on the sample data of the plurality of prediction variables; the sample matrix is processed for missing values and normalized to obtain a target matrix; each column in the target matrix is taken as a time series of the corresponding prediction variable.

8. A galloping prediction device of a power transmission line, characterized by, The device comprises: a construction module configured to, for each sample transmission line, acquire a sample label and sample data of a plurality of prediction variables of the sample transmission line, and construct a time series for each prediction variable based on the sample data of the plurality of prediction variables, the prediction variables including line parameter variables and meteorological variables; a transformation module configured to perform Fourier transform with an adaptive window length based on the time series of each prediction variable to obtain a frequency spectrum corresponding to each prediction variable; a training module configured to generate a sample multi-channel image based on the frequency spectrum corresponding to each prediction variable, and train an initial flutter prediction model based on the sample multi-channel image and the sample label of each sample transmission line to obtain a target flutter prediction model; an acquisition module configured to acquire data of a plurality of prediction variables of a to-be-predicted transmission line, and generate a target multi-channel image based on the data of the plurality of prediction variables; a prediction module configured to perform prediction based on the target multi-channel image by using the target flutter prediction model to obtain a flutter prediction result of the to-be-predicted transmission line; The transformation module comprises: a first determination unit configured to, for each prediction variable, determine a local change rate of the time series of the prediction variable at each time; a second determination unit configured to determine a maximum local change rate from the local change rates at all times; a third determination unit configured to determine a window function length corresponding to each time based on the local change rate at each time, the maximum local change rate, a preset maximum window function length, and a preset minimum window function length; a transformation unit configured to, for the time series of the prediction variable, perform short-time Fourier transform on the time series by using the window function length corresponding to each time to obtain a complex matrix corresponding to each time; a fourth determination unit configured to take a complex modulus of the complex matrix corresponding to each time as a pixel value corresponding to each time. The generating unit is configured to generate a spectrum corresponding to the prediction variable based on pixel values corresponding to each time point in a time sequence of the prediction variable. The window function length is calculated by the following formula: In the formula, represents the window function length at the time point of ; represents the preset maximum window function length; represents the preset minimum window function length; represents the local change rate at the time point of ; represents the maximum value of the local change rate.

9. The apparatus of claim 8, wherein, The local change rate is calculated by the following formula: wherein denotes the local rate of change of the time series of the prediction variable at time ; denotes the number of all time steps of the time series; denotes the index of a time step; and denote the sample data of the prediction variable at time and time , respectively; denotes the absolute value of the amount of change of the sample data of the prediction variable between consecutive time and time .

10. The apparatus of claim 8, wherein, The training module comprises: The adjusting unit is configured to adjust the size of the spectrum corresponding to each prediction variable to a preset size. The preprocessing unit is configured to preprocess each spectrum after size adjustment to obtain a preprocessed spectrum. The stacking unit is configured to stack all preprocessed spectra in a channel dimension to obtain the sample multi-channel image.

11. The apparatus of claim 8, wherein, The training module comprises: The dividing unit is configured to divide the sample multi-channel images of all sample transmission lines into a training set and a validation set. The obtaining unit is configured to obtain an expert weight of each prediction variable and use the expert weight of each prediction variable as an initial weight of each prediction variable. The weighting unit is configured to use the initial galloping prediction model to weight any sample multi-channel image in the training set based on the initial weight of each prediction variable to obtain a weighted feature map. The prediction unit is configured to perform prediction based on the weighted feature map to obtain a sample galloping prediction probability. The calculation unit is configured to calculate a training loss based on the sample galloping prediction probability, a sample label corresponding to the sample multi-channel image, the expert weight and the initial weight of each prediction variable. The updating unit is configured to update the initial weight of each prediction variable and a model parameter of the initial galloping prediction model based on the training loss. The training unit is configured to evaluate the performance of the initial galloping prediction model based on the validation set, return to the step of using the initial galloping prediction model to weight any sample multi-channel image in the training set based on the initial weight of each prediction variable to obtain a weighted feature map, and repeat until the performance of the initial galloping prediction model on the validation set reaches a preset stop condition. The initial galloping prediction model obtained after the last update is used as the target galloping prediction model.

12. The apparatus of claim 11, wherein, The training loss is calculated by the following formula: In the formula, Indicates training loss; This represents the loss between the predicted probability of sample dancing and the sample label corresponding to the multi-channel image of the sample; This indicates the preset weight adjustment parameters; Indicates the number of predictor variables; Indicates the index of the predictor variable; Indicates the first The current initial weights of the predictor variables; Indicates the first Expert weights for each predictor variable.

13. The apparatus of claim 11, wherein, The initial weight of each prediction variable is updated by the following formula: In the formula, represents the initial weight of the first prediction variable after updating; represents the initial weight of the first prediction variable after updating; represents the initial weight of the first prediction variable before updating; represents the initial weight of the first prediction variable before updating; represents a preset learning rate.

14. The apparatus of claim 8, wherein, The construction module comprises: The construction unit is configured to use a sliding time window mechanism to construct a sample matrix based on sample data of the plurality of prediction variables. The processing unit is configured to perform missing value processing on the sample matrix and perform normalization to obtain a target matrix. The fifth determination unit is configured to use each column in the target matrix as a time sequence of a corresponding prediction variable.

15. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions are used to make a computer execute the galloping prediction method of the transmission line in any one of claims 1 to 7.

Images