A power transmission line icing galloping prediction method and system, device, and storage medium based on an improved time sequence large model

Abstract

The application discloses a power transmission line icing galloping prediction method and system based on an improved time sequence large model, equipment and a storage medium, relates to the technical field of power system safety monitoring, and comprises the following steps: acquiring line parameters, meteorological data and power transmission line icing galloping data in an abnormal accumulated time period; constructing a data feature engineering by using the line parameters, the meteorological data and the power transmission line icing galloping data in the abnormal accumulated time period; performing data synthesis on the meteorological data and the power transmission line icing galloping data in the abnormal accumulated time period through time alignment to obtain a comprehensive time sequence; and taking the comprehensive time sequence as input, outputting a power transmission line icing galloping probability and a power transmission line galloping amplitude based on the improved time sequence large model. The application can quickly and accurately output a power transmission line icing galloping risk early warning, thereby providing reliable protection for safe operation of a power grid.

Description

TECHNICAL FIELD

[0001] The application relates to a power transmission line icing galloping prediction method and system based on an improved time sequence large model, equipment and a storage medium, and belongs to the technical field of power system safety monitoring. BACKGROUND

[0002] The icing galloping of a high-voltage power transmission line ground wire is a major threat to the safe operation of a power grid. In low-temperature, high-humidity or high-altitude areas, the icing of a wire causes low-frequency and large-amplitude vibration under the action of wind load, the icing layer significantly increases the load of the wire and changes the aerodynamic characteristics, leading to chain accidents such as wire breakage, inter-phase flashover and tower structure failure, which poses a serious challenge to the stability of the power grid and the safety of the society and economy.

[0003] At present, icing galloping prediction mainly relies on traditional statistical regression methods, empirical formulas and low-dimensional nonlinear models. These methods are usually based on limited historical data and empirical parameters, and are limited by the sparsity of feature dimensions, and it is difficult to effectively capture the complex nonlinear coupling relationship between meteorological conditions, ice accumulation and structural response. At the same time, the icing process is significantly affected by the time-varying nature of external environmental factors, and the traditional methods have obvious shortcomings in reflecting dynamic changes, real-time response and robustness. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a power transmission line icing galloping prediction method and system based on an improved time sequence large model, equipment and a storage medium, which can quickly and accurately output power transmission line icing galloping risk warning and provide reliable protection for the safe operation of the power grid.

[0005] To achieve the above-mentioned purpose, the application adopts the following technical solutions:

[0006] In a first aspect, the application provides a power transmission line icing galloping prediction method based on a time sequence large model, comprising:

[0007] acquiring real-time meteorological data;

[0008] comparing the real-time meteorological data with a preset threshold condition, if the threshold condition is not reached, it is judged that there is no icing galloping risk, if the threshold condition is reached, the meteorological data abnormal accumulation time is monitored, and the meteorological data abnormal accumulation time is compared with a set threshold value, if it is lower than the set threshold value, it is judged that the icing galloping risk is removed, if it is higher than the set threshold value, the power transmission line icing galloping probability and the power transmission line galloping amplitude are predicted according to the improved time sequence large model, the line parameters and the meteorological data and the power transmission line icing galloping data in the abnormal accumulation time period.

[0009] Further, the transmission line icing galloping probability and transmission line galloping amplitude prediction according to the improved time series big model, the line parameters, and the meteorological data and the transmission line icing galloping data in the abnormal accumulation time period comprises:

[0010] Obtaining the line parameters, and the meteorological data and the transmission line icing galloping data in the abnormal accumulation time period;

[0011] Constructing a data feature engineering by using the line parameters, and the meteorological data and the transmission line icing galloping data in the abnormal accumulation time period;

[0012] Based on the data feature engineering, data synthesis is performed on the data features of the meteorological data and the transmission line icing galloping data by time alignment, and a comprehensive time series is obtained by combining the data features of the line parameters;

[0013] Taking the comprehensive time series as input, the transmission line icing galloping probability and transmission line galloping amplitude are output based on the improved time series big model.

[0014] Further, the meteorological data comprises region, time, weather, temperature, precipitation, wind direction, wind power, wind speed, air pressure, humidity, air quality, and visibility.

[0015] The line parameters comprise line name, voltage grade, number of branches, conductor model, hanging point height, span, and conductor direction.

[0016] The transmission line icing galloping data comprises line icing thickness, galloping frequency, and galloping amplitude.

[0017] Further, the constructing a data feature engineering by using the line parameters, and the meteorological data and the transmission line icing galloping data in the abnormal accumulation time period comprises:

[0018] Data preprocessing is performed on the line parameters, and the meteorological data and the transmission line icing galloping data in the abnormal accumulation time period to obtain preprocessed meteorological data, line parameters, and transmission line icing galloping data;

[0019] Each item of data is taken as a data feature, and a correlation coefficient between the data feature and the transmission line icing galloping state is calculated by using a Pearson correlation coefficient;

[0020] Highly correlated data features are selected according to the correlation coefficient, and a data feature engineering is obtained by combining the highly correlated data features after standardization processing.

[0021] Further, the data preprocessing comprises deleting abnormal values and error values in the data, and filling missing data by using mean filling and interpolation method.

[0022] The standardization processing is one of normalization, log transformation normalization, Z-Score standardization, periodic encoding, directional encoding, and Sigmoid transformation.

[0023] Further, the data features of the meteorological data and the icing galloping data of the transmission line are combined by time alignment, and a comprehensive time sequence is obtained by combining the data features of the line parameters, including:

[0024] The data features of the meteorological data and the icing galloping data of the transmission line are matched according to the time stamp to obtain a meteorological sequence.

[0025] The data features of the line parameters are combined with the meteorological sequence to generate a comprehensive time sequence, and the dimension of the comprehensive time sequence is represented as time step x feature dimension.

[0026] Further, the improved time sequence model is improved based on the Transformer time sequence model, specifically including: replacing the output layer of the Transformer time sequence model with two parallel task heads to obtain an improved time sequence model, wherein one task head is a classification task output layer, and the other task head is a regression task output layer.

[0027] The classification task output layer includes an attention pooling layer, a fully connected layer and a Sigmoid activation function arranged in sequence.

[0028] The attention pooling layer is used to extract time step information from the feature vector output by the Transformer decoder, and to calculate attention weights for each time step information, and to perform weighted summation of the features of each time step to obtain a global feature vector.

[0029] The fully connected layer is used for linear transformation of the global feature vector.

[0030] The Sigmoid activation function is used to operate the global feature vector after linear transformation to output a probability value, and the probability value range is [0, 1], when the probability value ≤0.5, it is judged that the transmission line icing gallops, otherwise it is judged that the transmission line does not icing gallops.

[0031] The regression task output layer includes a fully connected layer, which is used to map the feature vector output by the Transformer decoder to a continuous value through the fully connected layer, and the continuous value represents the icing galloping amplitude of the transmission line.

[0032] Further, the data processing expression of the attention pooling layer is:

[0033]

[0034]

[0035] wherein, denotes the attention weight at the th time step, denotes a learnable weight vector, denotes the output feature vector of the Transformer decoder at the th time step, denotes the inner product of and denotes the influence degree of the output feature vector of the Transformer decoder at the th time step on the icing galloping of the transmission line, denotes the sum of the influence degrees of the output feature vectors of the Transformer decoder at all time steps on the icing galloping of the transmission line, denotes the number of time steps, denotes the feature vector output by the attention pooling layer, i.e., the global feature vector;

[0036] The calculation expression of the Sigmoid activation function is:

[0037] ,

[0038] wherein, denotes the icing galloping probability of the transmission line, denotes a classifiable weight matrix, denotes the global feature vector after linear transformation, denotes a bias term, denotes a Sigmoid activation function.

[0039] Further, the improved time series large model is further pre-trained, and the pre-training method comprises:

[0040] obtaining historical meteorological data, line parameters and historical icing galloping data of the transmission line;

[0041] constructing a data feature engineering using the historical meteorological data, the line parameters and the historical icing galloping data of the transmission line;

[0042] based on the data feature engineering, synthesizing data features of the historical meteorological data and the historical icing galloping data of the transmission line through time alignment, and combining data features of the line parameters to obtain a comprehensive time series;

[0043] taking the comprehensive time series as an input, training the improved time series large model, calculating a loss function in the training process until the loss function is less than a preset threshold, and obtaining the pre-trained improved time series large model. ​

[0044] Further, the loss function adopts a joint loss function, wherein a binary cross-entropy loss function is adopted for a classification task output layer, and a mean square error loss function is adopted for a regression task output layer.

[0045] An expression of the joint loss function is as follows:

[0046]

[0047] wherein, represents the joint loss function, represents a binary cross-entropy loss function, represents a mean square error loss function, , respectively represent weights of the binary cross-entropy loss function and the mean square error loss function.

[0048] In a second aspect, the present application further provides a power transmission line icing galloping prediction system based on an improved time series large model, comprising:

[0049] A data acquisition module configured to acquire real-time meteorological data;

[0050] A meteorological mutation judgment module configured to compare the real-time meteorological data with a preset threshold condition, if the threshold condition is not reached, it is judged that there is no icing galloping risk, if the threshold condition is reached, the meteorological data abnormal accumulation time is monitored, and the meteorological data abnormal accumulation time is compared with a set threshold, if it is lower than the set threshold, it is judged that the icing galloping risk is removed, if it is higher than the set threshold, the power transmission line icing galloping probability and the power transmission line galloping amplitude are predicted according to the improved time series large model, the line parameters, and the meteorological data and the power transmission line icing galloping data in the abnormal accumulation time period.

[0051] Further, the meteorological mutation judgment module specifically comprises a data acquisition unit, a data feature engineering construction unit, a time series generation unit, and an icing galloping unit.

[0052] Further, the data acquisition unit is configured to acquire the line parameters and the meteorological data and the power transmission line icing galloping data in the abnormal accumulation time period.

[0053] The data feature engineering construction unit is configured to construct a data feature engineering by using the line parameters and the meteorological data and the power transmission line icing galloping data in the abnormal accumulation time period.

[0054] Further, the line parameters include line name, voltage grade, number of branches, conductor model, hanging point height, span, conductor direction, etc., and the power transmission line icing galloping data includes line icing thickness, galloping frequency and galloping amplitude.

[0055] Further, the data feature engineering is constructed by using the line parameter, the meteorological data and the icing galloping data of the transmission line in the abnormal accumulation time period, and includes:

[0056] The line parameter, the meteorological data and the icing galloping data of the transmission line in the abnormal accumulation time period are preprocessed to obtain preprocessed meteorological data, line parameter and icing galloping data of the transmission line, and the method of data preprocessing is to delete the abnormal value and error value in the data, and to fill the missing data by using the mean filling and interpolation method.

[0057] Each preprocessed data is used as a data feature, and the correlation coefficient between the data feature and the icing galloping state of the transmission line is calculated by using the Pearson correlation coefficient.

[0058] The data features with high correlation are screened out according to the correlation coefficient, and the data feature engineering is obtained by combining the data features after standardization processing, and the standardization processing can be selected from one of normalization, logarithmic transformation normalization, Z-Score standardization, period encoding, direction encoding and Sigmoid transformation.

[0059] Further, the time sequence generation unit is configured to synthesize the data features of the meteorological data and the icing galloping data of the transmission line by time alignment based on the data feature engineering, and obtain a comprehensive time sequence by combining the data features of the line parameter, and the dimension of the comprehensive time sequence is represented as time step x feature dimension.

[0060] The data features of the meteorological data and the icing galloping data of the transmission line are matched according to the time stamp to obtain a meteorological sequence.

[0061] The comprehensive time sequence is generated by combining the data features of the line parameter and the meteorological sequence.

[0062] Further, the icing galloping prediction unit is configured to take the comprehensive time sequence as input, and output the icing galloping probability and the icing galloping amplitude of the transmission line based on the improved time sequence large model.

[0063] The improved time sequence large model is obtained by improving the time sequence large model based on the Transformer, and the output layer of the time sequence large model based on the Transformer is replaced by two parallel task heads to obtain the improved time sequence large model, wherein one task head is a classification task output layer, and the other task head is a regression task output layer. Specifically:

[0064] Further, the classification task output layer includes an attention pooling layer, a full connection layer and a Sigmoid activation function arranged in sequence.

[0065] The attention pooling layer extracts time-step information from the feature vector output by the Transformer decoder, calculates attention weights for each time-step, and then sums the features from each time-step using a weighted method to obtain the global feature vector. The processing expression is as follows:

[0066]

[0067]

[0068] in, Indicates the first Attention weights at each time step This represents a learnable weight vector. Indicates the Transformer decoder at the 1st The output feature vector at each time step express and The inner product, Indicates the Transformer decoder at the 1st The influence of the output feature vector at each time step on the ice-covered galloping of transmission lines. This represents the sum of the influence of the Transformer decoder's output feature vectors at all time steps on the icing galloping of the transmission line. Indicates the number of time steps. This represents the feature vector output by the attention pooling layer, i.e., the global feature vector.

[0069] Fully connected layers are used to perform linear transformations on global feature vectors to obtain linearly transformed global feature vectors.

[0070] The Sigmoid activation function is used to calculate a probability value from the linearly transformed global feature vector. The probability value ranges from [0, 1]. If the probability value is ≤0.5, the transmission line is considered to be ice-covered and galloping; otherwise, it is considered to be non-ice-covered and galloping. Its calculation expression is as follows:

[0071] ,

[0072] in, This indicates the probability of ice-induced galloping on transmission lines. Represents the classifiable weight matrix. This represents the global eigenvectors after the linear transformation. Indicates the bias term. This represents the Sigmoid activation function.

[0073] Further, the regression task output layer comprises a full connection layer, which is used for mapping the feature vector output by the Transformer decoder to a continuous value through the full connection layer, and the continuous value represents the icing galloping amplitude of the power transmission line.

[0074] In a third aspect, the present application further provides an electronic device, comprising:

[0075] a memory for storing a computer program;

[0076] a processor for executing the computer program to implement the steps of the power transmission line icing galloping prediction method based on the improved time series large model according to any one of the first aspect.

[0077] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the power transmission line icing galloping prediction method based on the improved time series large model according to any one of the first aspect.

[0078] Compared with the prior art, the present application has the following beneficial effects:

[0079] The present application combines real-time weather data, line state information and power transmission line icing galloping risk, obtains the galloping risk probability and galloping amplitude based on the improved time series large model, breaks through the limitation of traditional single task modeling, can realize timely alarm of the power transmission line icing galloping risk, and through the time series reconstruction of the improved time series large model and the transformation of the Transformer double task, the complex time dependence and periodic pattern are captured, the prediction accuracy is greatly improved, and the power grid icing disaster risk can be effectively reduced. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 FIG. 1 is a flowchart of the power transmission line icing galloping prediction method based on the improved time series large model according to an embodiment of the present application;

[0081] Figure 2 FIG. 2 is a schematic diagram of the integrated time series in the power transmission line icing galloping prediction method based on the improved time series large model according to an embodiment of the present application;

[0082] Figure 3 FIG. 3 is a schematic diagram of the architecture of the improved time series large model in the power transmission line icing galloping prediction method based on the improved time series large model according to an embodiment of the present application. DETAILED DESCRIPTION

[0083] The present application will be further described below in conjunction with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot be used to limit the protection scope of the present application.

[0084] Embodiment 1:

[0085] As Figure 1 shown, the embodiment of the present application provides a power transmission line icing galloping prediction method based on an improved time sequence large model, which specifically comprises the following steps:

[0086] Real-time meteorological data is obtained, in this embodiment, the real-time meteorological data includes region, time, weather, precipitation, wind direction, wind force, wind speed, air pressure, humidity, air quality, visibility, and is obtained through meteorological monitoring equipment.

[0087] In this embodiment, the preset threshold condition is: temperature ≤ 0℃, humidity ≥ 80%, wind speed is 5~15 m / s (wind force 3~7 levels), weather is continuous rain and snow weather (such as sleet or freezing rain), and the above conditions are met simultaneously for a duration of ≥ 30 minutes. Compare the real-time meteorological data with the preset threshold condition, if the preset threshold condition is not reached, it is judged that there is no icing galloping risk; if the preset threshold condition is reached, the meteorological data abnormal accumulation time is monitored, and the meteorological data abnormal accumulation time is compared with the set threshold value (in this embodiment, the set threshold value is 2 hours), if it is lower than the set threshold value, it is judged that the icing galloping risk is removed, if it is higher than the set threshold value, the pre-trained improved time sequence large model is used to predict the power transmission line icing galloping probability and power transmission line galloping amplitude.

[0088] Obtain the line parameters and the meteorological data and power transmission line icing galloping data in the abnormal accumulation time period.

[0089] Construct data feature engineering using the line parameters and the meteorological data and power transmission line icing galloping data in the abnormal accumulation time period, specifically:

[0090] For example, the micro weather station carried by the tower is taken as the data source, the wind speed (0~30m / s), the wind direction, the temperature (-30℃±15℃), the humidity (0~100%RH) and the precipitation (0~50mm / h) parameters are collected with a sampling interval of 1 minute; the conductor inclination angle (±30°), the tension (0~50kN), the vibration frequency (0~50Hz) and other parameters are obtained with a sampling frequency of 10Hz.

[0091] The above data is preprocessed, the preprocessing includes data cleaning, specifically, the abnormal values and error data in the data are removed, that is, if the data is obviously deviated from the normal range of galloping amplitude value, it is deleted. The missing data is processed by mean filling and interpolation method, if the data is missing at a certain moment, the average value of the adjacent moment data is used to fill the missing data.

[0092] The correlation coefficients between each feature and the icing galloping state of the conductor are calculated using the Pearson correlation coefficient, and features with high correlation are selected. In this embodiment, the features with high correlation are: in the meteorological data, the region, time, weather, temperature, precipitation, wind direction, wind force, wind speed, air pressure, humidity, air quality, and visibility; in the line parameters, the line name, voltage grade, number of branches, conductor model, hanging point height, span, and conductor orientation; and in the icing galloping data of the transmission line, the icing thickness, galloping frequency, and galloping amplitude.

[0093] The features with high correlation that are selected are standardized. The feature data has characteristics such as diverse sources, complex types, different scales, and strong spatiotemporal correlation, so different processing methods are used for different features to unify the data scale and enhance the model's ability to identify key factors such as periodicity and spatial direction. The standardization processing methods include but are not limited to normalization, log transformation normalization, Z-Score standardization, period encoding, direction encoding, and Sigmoid transformation.

[0094] In this embodiment, the standardization processing methods corresponding to each feature are shown in Tables 1 and 2:

[0095] Table 1: Processing methods corresponding to some features with high correlation in meteorological data

[0096]

[0097] Table 2: Processing methods corresponding to some features with high correlation in line parameters and icing galloping data of the transmission line

[0098]

[0099] The features after standardization are combined to obtain data feature engineering.

[0100] Next, based on the data feature engineering, the data features of discrete conductor galloping data and continuous meteorological monitoring data in the data feature engineering are synthesized through time alignment, and then the data features of the line parameters in the data feature engineering are combined to generate comprehensive time series, which specifically includes:

[0101] For each transmission line, the data features of discrete galloping data and continuous meteorological data are timestamp matched, and the meteorological sequence is obtained according to the recorded region. If the sequence is in the duration of galloping, the galloping amplitude, galloping frequency, and icing thickness are filled with the recorded data, otherwise they are filled with 0. At the same time, the data features of the line parameters are added to the sequence, because the line parameters are static indicators that do not change with time. Finally, the comprehensive time series is generated, with one comprehensive time series corresponding to one transmission line.

[0102] The dimension of the integrated time series is [time step x feature dimension], as shown in Figure 2 The feature dimension in this embodiment includes region, time, weather, temperature, precipitation, wind direction, wind power, wind speed, air pressure, humidity, air quality, visibility, line name, voltage level, branch number, conductor model, hanging point height, span, conductor orientation, line icing thickness, dancing frequency, and dancing amplitude.

[0103] Next, an improved time series model is constructed.

[0104] The time series model in this embodiment is improved based on the Transformer time series model, and the output layer of the existing Transformer time series model is replaced by two parallel task heads, one of which is a classification task output layer and the other is a regression task output layer.

[0105] In combination with Figure 3 The basic architecture of the improved time series model includes an input layer for inputting integrated time series data (time step x feature dimension), the output end of the input layer is respectively connected with a feature embedding layer and a position encoding layer, the feature embedding layer is used to project the input data to a high-dimensional space through linear transformation to obtain a feature embedding vector, which functions to convert physical quantities of different dimensions into a unified vector representation while preserving the correlation between features. The position encoding layer is used to inject time sequence information by using sine / cosine position encoding to generate a unique position vector for each time step, so that the model can identify the time sequence relationship.

[0106] The position vector and the feature embedding vector are added element by element and then input into a Transformer encoder and a Transformer decoder in turn, both the Transformer encoder and the Transformer decoder include multiple layers, each layer of the Transformer encoder includes a multi-head attention mechanism unit and a feedforward network arranged in sequence, and each layer of the Transformer decoder includes a multi-head causal self-attention mechanism unit, a cross-attention mechanism unit, and a feedforward network arranged in sequence.

[0107] The classification task output layer comprises an attention pooling layer, a fully connected layer and a Sigmoid activation function arranged in sequence, the attention pooling layer is used to extract time step information from the feature vector output by the Transformer decoder, and an attention weight is calculated for each time step information, the features of each time step are weighted and summed to obtain a global feature vector, then the global feature vector is linearly transformed through the fully connected layer, and finally a probability value is output by operating the linearly transformed global feature vector through the Sigmoid activation function, the probability value ranges from 0 to 1, when the probability value is less than or equal to 0.5, it is judged that the transmission line is icing and dancing, otherwise it is judged that the transmission line is not icing and dancing.

[0108] The regression task output layer comprises a fully connected layer, which is used to map the feature vector output by the Transformer decoder through the fully connected layer to obtain a continuous value, and the continuous value represents the icing and dancing amplitude of the transmission line.

[0109] In the embodiment, the improved time series large model selects the time series large model TimesFM (timesfm-1.0-200m) published by Google as the base model, and the model parameter amount is 200m. The model structure design of TimesFM aims to process zero sample prediction tasks of multi-domain and multi-granularity time series, and the core architecture comprises an input layer, a feature embedding layer, a position encoding layer, a stacked Transformer layer and an output layer, and high-efficiency pre-training and generalization capability are realized through a specific patch strategy and a mask mechanism.

[0110] The input layer comprises:

[0111] (1) Patching: the input comprehensive time series is divided into non-overlapping patches of fixed length as the input tokens of the Transformer.

[0112] (2) Residual Block Processing Unit: each patch is converted into a vector with a dimension of model_dim through a residual block. A binary mask is also provided to the Transformer along with the input. The binary mask is used to indicate whether the corresponding data point should be considered (0) or ignored (1). The residual block is essentially a multi-layer perceptron (MLP) with a hidden layer and a skip connection. In order to let the model capture the time sequence information, the model adopts sine-cosine position encoding.

[0113] (3) Random Masking Unit: used to flexibly handle any context length, all possible context lengths are covered through a random masking strategy: a random offset r is selected during training, and the first r time points are masked, ensuring that the model adapts to inputs from 1 to the maximum context length (such as 512).

[0114] Stacked Transformer includes:

[0115] (1) Decoder-only architecture: uses a pure decoder structure, only uses causal self-attention, ensures that each output token can only focus on the previous input token (including itself).

[0116] (2) Layer structure: multi-head causal self-attention: each Transformer layer uses multi-head causal self-attention mechanism, allowing the model to focus on different parts of the input sequence at the same time.

[0117] The feedforward network FFN is set after the self-attention mechanism, and each layer independently applies a feedforward network to each position in the sequence. This further processes the information of attention and enables the model to learn higher-level representations.

[0118] Output layers include:

[0119] (1) Variable-length patch prediction: the output patch length h can be greater than the input patch length p, supporting the prediction of a longer time window at a time, such as input patch 32 and output patch 128. This design reduces the autoregressive step, such as predicting 256 steps only needs 2 times generation, instead of 8 times, improving the efficiency of long sequence prediction.

[0120] (2) Adapt to multi-granularity prediction: by adjusting h and p, the model can handle different time granularities, such as hours, days, and months, without modifying the architecture, only need to cover multi-granularity data in pre-training.

[0121] TimesFM is a decoder-only architecture that will input time series data through multiple self-attention layers and feedforward networks, and finally output a hidden vector sequence. The present application modifies the model structure of TimesFM for the task of icing dance prediction, replaces the output layer of TimesFM model with a classification task output layer and a regression task output layer to obtain an improved time series large model.

[0122] The improved time series large model is pre-trained, and the pre-training method includes:

[0123] First, collect raw data from meteorological monitoring equipment, line state sensors and historical icing dance records, preprocess and extract features, and build data feature engineering.

[0124] Based on data feature engineering, the data features of historical meteorological data and historical transmission line icing galloping data are synthesized by time alignment, and the comprehensive time series is obtained by combining the data features of the line parameters.

[0125] The comprehensive time series is taken as input to train the improved time sequence large model. In the training process, the loss function is calculated until the loss function is less than the preset threshold, and the pre-trained improved time sequence large model is obtained. In this embodiment, the loss function adopts a joint loss function, wherein the binary cross entropy loss function is adopted for the classification task output layer, and the mean square error loss function is adopted for the regression task output layer, and the expression of the joint loss function is:

[0126]

[0127] wherein, represents the joint loss function, represents the binary cross entropy loss function, represents the mean square error loss function, , respectively represent the weights of the binary cross entropy loss function and the mean square error loss function.

[0128] Based on the pre-trained improved time sequence large model, the model is fine-tuned. The fine-tuning adopts linear probing fine-tuning, low-rank adaptation fine-tuning, and weight decomposition adaptation fine-tuning strategies, and the effects are comprehensively compared.

[0129] Linear probing fine-tuning (Linear Probing) only trains input / output and embedding, and freezes the transformer block. The general features of the pre-trained model are taken as a fixed feature extractor, and only a simple linear layer is trained for a specific task. This method has small parameter overhead and fast update, and is suitable for scenarios where it is desired to quickly adapt to new tasks without disturbing the original knowledge of the model.

[0130] Low-rank adaptation fine-tuning (Low-Rank Adaptation, LoRA) adds a low-rank fine-tuning term to a certain weight matrix in the original network, that is, only the low-rank matrix is trained without changing the original large model weight, so as to achieve the effect of efficient fine-tuning of parameters, and the update formula is usually:

[0131] , ;

[0132] wherein, represents the LoRA fine-tuned weight matrix, is the original weight matrix in the pre-trained model; and These are two low-rank matrices (low-rank factors) Typically much smaller During fine-tuning, these are trainable parameters; updates are performed only on the parameters being fine-tuned. and Perform gradient updates, and By keeping the parameters frozen, the number of parameters that need to be tuned is significantly reduced, while pre-trained knowledge can be reused. This approach can capture task-specific variations while reducing memory and computational overhead.

[0133] Directional LoRA (DoRA) can be seen as a further extension of LoRA, aiming to further distinguish between "direction" and "magnitude" information. DoRA first decomposes the weights of the pre-trained model, breaking down each weight matrix into an magnitude vector and a direction matrix. This decomposition allows the model to better control the learning process of the weights. During fine-tuning, DoRA uses LoRA for directional updates, adjusting only the direction parameters while keeping the magnitude parameters unchanged. This reduces the number of parameters that need to be adjusted, improving the efficiency of fine-tuning. This method simplifies the task compared to traditional fine-tuning methods, which require adjusting both magnitude and direction. Weight decomposition analysis helps DoRA enhance the model's learning ability and training stability. This method aims to simulate the learning ability of full fine-tuning while avoiding any additional inference overhead.

[0134] The DoRA update formula is:

[0135]

[0136] in, This represents the weight matrix after DoRA fine-tuning. This represents the trainable magnitude variable. This represents the input feature vector of a certain layer during the forward propagation of the model. express right The fine-tuned increment is used to control the weight strength. This represents the initial direction matrix for freezing. Represents a low-rank matrix and The adapters that make up the configuration This represents the constraint norm.

[0137] Since DoRA maintains high parameter efficiency while achieving performance closest to full parameter fine-tuning, significantly outperforming LoRA and Linear Probing, this embodiment ultimately selects DoRA as the fine-tuning strategy, which can achieve a balance between performance and efficiency.

[0138] The icing galloping probability of the power transmission line and the galloping amplitude of the power transmission line are obtained based on the improved time sequence large model output.

[0139] Embodiment 2:

[0140] Based on the embodiment 1, the application provides a power transmission line icing galloping prediction system based on an improved time sequence large model, which is realized by deploying the improved time sequence large model on an edge computing terminal, and specifically includes the following steps:

[0141] First, the pre-trained improved time sequence large model is structured pruned, and the improved time sequence large model is preprocessed by using a pruning technique, wherein the pruning technique evaluates the importance of each parameter or network module of the model, screens and eliminates parameters with lower contribution or redundancy to the prediction result, so as to reduce the total amount of model parameters and the overall model size, and reduce the calculation complexity of the model.

[0142] The model after pruning is subjected to numerical precision reduction processing by using a 16-bit quantization technology, the original 32-bit floating point number weight and activation value in the model are converted into 16-bit floating point numbers or fixed point numbers, and the memory occupation and inference delay are reduced.

[0143] The model after pruning and 16-bit quantization processing is converted into an ONNX format, and a corresponding inference framework is selected for deployment according to the hardware characteristics of the edge computing terminal, so as to realize efficient real-time inference.

[0144] The system of the embodiment specifically includes:

[0145] The data acquisition module is configured to acquire real-time meteorological data, in the embodiment, the real-time meteorological data includes region, time, weather, precipitation, wind direction, wind force, wind speed, air pressure, humidity, air quality, visibility, and is acquired by a meteorological monitoring device.

[0146] The meteorological mutation judgment module (edge deployment module) is configured to compare the real-time meteorological data with a preset threshold condition, if the threshold condition is not reached, it is judged that there is no icing galloping risk, if the threshold condition is reached, the meteorological data abnormal accumulation time is monitored, and the meteorological data abnormal accumulation time is compared with a set threshold value, if it is lower than the set threshold value, it is judged that the icing galloping risk is removed, if it is higher than the set threshold value, the icing galloping probability of the power transmission line and the galloping amplitude of the power transmission line are predicted according to the improved time sequence large model, the line parameters, and the meteorological data and the power transmission line icing galloping data in the abnormal accumulation time period.

[0147] The meteorological mutation judgment module specifically includes a data acquisition unit, a data feature engineering construction unit, a time sequence generation unit and an icing galloping unit.

[0148] The data acquisition unit is configured to acquire line parameters and meteorological data and power transmission line icing galloping data in an abnormal accumulation time period.

[0149] The data feature engineering construction unit is configured to construct data feature engineering by using the line parameters and the meteorological data and the power transmission line icing galloping data in the abnormal accumulation time period.

[0150] The line parameters include line name, voltage grade, number of branches, conductor model, hanging point height, span, and conductor orientation, and the power transmission line icing galloping data includes line icing thickness, galloping frequency, and galloping amplitude.

[0151] The line parameters and the meteorological data and the power transmission line icing galloping data in the abnormal accumulation time period are subjected to data preprocessing to obtain preprocessed meteorological data, line parameters, and power transmission line icing galloping data. In this embodiment, the method of data preprocessing is to delete abnormal values and error values in the data, and to fill in missing data by using mean filling and interpolation method.

[0152] Each item of preprocessed data is taken as a data feature, and the correlation coefficient between the data feature and the power transmission line icing galloping state is calculated by using the Pearson correlation coefficient.

[0153] Data features with high correlation are screened out according to the correlation coefficient, and a data feature engineering is obtained after standardization processing and combination of the data features. The standardization processing can be selected from one of normalization, logarithmic transformation normalization, Z-Score standardization, periodic encoding, directional encoding, and Sigmoid transformation.

[0154] The time sequence generation unit is configured to perform data synthesis on the meteorological data and the power transmission line icing galloping data in the abnormal accumulation time period by time alignment to obtain a comprehensive time sequence. The dimension of the comprehensive time sequence is represented as time step x feature dimension.

[0155] For each power transmission line, real-time meteorological data and real-time power transmission line icing galloping data are matched according to time stamp to obtain a meteorological sequence.

[0156] The line parameters of the power transmission line and the meteorological sequence are combined to generate a comprehensive time sequence.

[0157] The icing galloping prediction unit is configured to take the comprehensive time sequence as input and output a power transmission line icing galloping probability and a power transmission line galloping amplitude based on an improved time series large model.

[0158] The improved time series large model is improved based on the Transformer time series large model, replaces the output layer of the Transformer time series large model with two parallel task heads, to obtain the improved time series large model, wherein one task head is a classification task output layer, and the other task head is a regression task output layer. Specifically:

[0159] The classification task output layer comprises an attention pooling layer, a full connection layer and a Sigmoid activation function arranged in sequence.

[0160] The attention pooling layer is used to extract time step information from the feature vector output by the Transformer decoder, and calculate the attention weight for each time step information, and then perform weighted summation on the features of each time step to obtain a global feature vector, and the processing expression is:

[0161]

[0162]

[0163] wherein, the attention weight of the i-th time step is represented by ai, the learnable weight vector is represented by w, the output feature vector of the Transformer decoder at the i-th time step is represented by xi, the inner product of and is represented by ai·xi, the influence degree of the output feature vector of the Transformer decoder at the i-th time step on the icing galloping of the power transmission line is represented by ai·xi, the total influence degree of the output feature vectors of the Transformer decoder at all time steps on the icing galloping of the power transmission line is represented by and the number of time steps is represented by T, the feature vector output by the attention pooling layer, i.e. the global feature vector, is represented by h. The full connection layer is used to perform linear transformation on the global feature vector to obtain a linearly transformed global feature vector. The Sigmoid activation function is used to perform operation on the linearly transformed global feature vector to output a probability value, and the probability value ranges from 0 to 1. When the probability value is less than or equal to 0.5, it is judged that the power transmission line is icing galloping, otherwise it is judged that the power transmission line is not icing galloping, and the calculation expression is:

[0164]

[0165]

[0166] ,

[0167] ​​​​wherein, denotes the icing galloping probability of the transmission line, denotes the classifiable weight matrix, denotes the global feature vector after linear transformation, denotes the bias term, denotes the Sigmoid activation function.

[0168] The regression task output layer includes a full connection layer, which is used to map the feature vector output by the Transformer decoder through the full connection layer to obtain a continuous value, and the continuous value represents the icing galloping amplitude of the transmission line.

[0169] The improved timing large model of the application can realize efficient operation on resource-limited devices while maintaining good prediction performance, and fully meets the demand of the power system for real-time monitoring. After actual deployment, the lightweight model can quickly and accurately output icing galloping risk warning, and provides reliable guarantee for safe operation of the power grid.

[0170] Embodiment 3:

[0171] The embodiment also provides an electronic device, comprising:

[0172] a memory for storing a computer program;

[0173] a processor for executing the computer program to implement the steps of the icing galloping prediction method of the transmission line based on the improved timing large model according to the embodiment 1.

[0174] Embodiment 4:

[0175] The embodiment also provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the steps of the icing galloping prediction method of the transmission line based on the improved timing large model according to the embodiment 1.

[0176] Those skilled in the art should understand that the embodiments of the application can be provided as a method, a system, or a computer program product. Therefore, the application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. The solutions in the embodiments of the application can be implemented in various computer languages, such as object-oriented programming language Java and interpreted scripting language JavaScript.

[0177] The present application is described in reference to the flowchart and / or block diagram of the method, apparatus (system) and computer program product according to an embodiment of the present application. It is understood that each flow and / or block in the flowchart and / or block diagram, and a combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, a special purpose computer, an embedded processor or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks. Figure 1 one or more blocks or blocks.

[0178] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks. Figure 1 one or more blocks or blocks.

[0179] These computer program instructions can also be loaded onto a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer implemented process such that the instructions which execute on the computer or other programmable device provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more flows and / or blocks. Figure 1 one or more blocks or blocks.

[0180] Although preferred embodiments of the application have been described, those skilled in the art will recognize that many modifications and variations of this application are possible. It is therefore intended that the appended claims shall cover any modifications and variations of the application which fall within the true spirit and scope of the application.

[0181] It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover the modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

[0182] The above description is only preferred embodiments of the present application. It is understood that, for those skilled in the art, some improvements and modifications can be made without departing from the technical principles of the present application, and these improvements and modifications should also be considered as within the scope of the present application.

Claims

1. A power transmission line icing galloping prediction method based on an improved timing large model, characterized in that, include: Obtain real-time weather data; The real-time meteorological data is compared with the preset threshold conditions. If the threshold conditions are not met, it is determined that there is no risk of ice accretion and dancing. If the threshold condition is reached, the cumulative time of abnormal meteorological data is monitored and compared with the set threshold. If it is lower than the set threshold, the risk of icing and galloping is determined to be eliminated. If it is higher than the set threshold, the probability of icing and galloping of the transmission line and the amplitude of the transmission line galloping are predicted based on the improved time series model, line parameters, meteorological data and transmission line icing and galloping data within the abnormal cumulative time period. The prediction of transmission line icing and galloping probability and amplitude based on the improved time-series large model, line parameters, and meteorological data and transmission line icing and galloping data within the abnormal cumulative time period includes: Acquire line parameters, meteorological data during the cumulative period of anomalies, and data on icing and galloping of transmission lines; Data feature engineering is constructed using line parameters, meteorological data during the cumulative abnormal period, and transmission line icing and galloping data. Based on the aforementioned data feature engineering, data features of meteorological data and transmission line icing and galloping data are synthesized by time alignment, and combined with the data features of line parameters to obtain a comprehensive time series. Using the integrated time series as input, the probability of transmission line icing and galloping and the amplitude of transmission line galloping are obtained based on the improved time series large model. The improved time series large model is obtained by improving the Transformer-based time series large model. Specifically, it includes replacing the output layer of the Transformer-based time series large model with two parallel task heads to obtain the improved time series large model, wherein one task head is the output layer of the classification task and the other task head is the output layer of the regression task. The output layer of the classification task includes an attention pooling layer, a fully connected layer, and a sigmoid activation function arranged sequentially. The attention pooling layer is used to extract time step information from the feature vector output by the Transformer decoder, calculate attention weights for each time step, and sum the features of each time step by weight to obtain the global feature vector. The fully connected layer is used to perform a linear transformation on the global feature vector; The Sigmoid activation function is used to operate on the global feature vector after linear transformation and output a probability value. The probability value ranges from [0, 1]. When the probability value is ≤0.5, it is judged that the transmission line is iced and galloping; otherwise, it is judged that the transmission line is not iced and galloping. The regression task output layer includes a fully connected layer, which is used to map the feature vector output by the Transformer decoder to obtain continuous values, which represent the amplitude of the ice-covered dancing of the transmission line.

2. The improved timing-based large model-based power transmission line icing galloping prediction method according to claim 1, characterized in that, The meteorological data includes region, time, weather, temperature, precipitation, wind direction, wind force, wind speed, air pressure, humidity, air quality, and visibility; The line parameters include line name, voltage level, number of splits, conductor type, suspension point height, span, and conductor direction; The data on icing and galloping of the transmission lines includes the thickness of the icing, the frequency of the galloping, and the amplitude of the galloping.

3. The improved timing-based large model-based power transmission line icing galloping prediction method according to claim 1, characterized in that, The data feature engineering, which utilizes line parameters and meteorological data and transmission line icing and galloping data within the abnormal cumulative time period, includes: Data preprocessing is performed on line parameters, meteorological data and transmission line icing and galloping data within the abnormal cumulative time period to obtain preprocessed meteorological data, line parameters and transmission line icing and galloping data; Each preprocessed data point is treated as a data feature, and its correlation coefficient with the icing and galloping state of the transmission line is calculated using the Pearson correlation coefficient. Data features with high correlation are selected based on correlation coefficients, and then standardized and combined to obtain data feature engineering.

4. The improved timing-based large model-based power transmission line icing galloping prediction method according to claim 3, characterized in that, The data preprocessing includes: removing outliers and erroneous values ​​from the data, and filling in missing data using mean filling and interpolation methods; The standardization process is one of the following: normalization, logarithmic transform normalization, Z-Score normalization, periodic coding, directional coding, and sigmoid transform.

5. The improved timing-based large model-based power transmission line icing galloping prediction method according to claim 1, characterized in that, The data feature engineering described above synthesizes data features from meteorological data and transmission line icing and galloping data through time alignment, and combines these with data features from line parameters to obtain a comprehensive time series, including: The meteorological data and the data characteristics of transmission line icing and galloping data are matched according to the timestamp to obtain the meteorological sequence; A comprehensive time series is generated by combining the data characteristics of the line parameters with the meteorological sequence. The dimension of the comprehensive time series is represented as time step × feature dimension.

6. The improved timing-based large model-based power transmission line icing galloping prediction method according to claim 1, characterized in that, The data processing expression for the attention pooling layer is: ， ， wherein, denotes the attention weight at the th time step, denotes a learnable weight vector, denotes the output feature vector of the Transformer decoder at the th time step, denotes the inner product of and , denotes the degree of influence of the output feature vector of the Transformer decoder at the th time step on the galloping of the transmission line icing, denotes the sum of the degrees of influence of the output feature vectors of the Transformer decoder at all time steps on the galloping of the transmission line icing, denotes the number of time steps, denotes the global feature vector; The expression for calculating the Sigmoid activation function is: ， wherein, denotes the probability of icing galloping of the power transmission line, denotes the classifiable weight matrix, denotes the global feature vector after linear transformation, denotes the bias term, denotes the Sigmoid activation function.

7. The method for predicting icing and galloping of transmission lines based on an improved time-series large model according to claim 1, characterized in that, It also includes pre-training the improved time-series large model, and the pre-training methods include: Acquire historical meteorological data, line parameters, and historical transmission line icing and galloping data; Data feature engineering was constructed using historical meteorological data, line parameters, and historical transmission line icing and galloping data. Based on data feature engineering, data features of historical meteorological data and historical transmission line icing galloping data are synthesized by time alignment, and combined with the data features of line parameters to obtain a comprehensive time series. The improved time series model is trained by taking the comprehensive time series as input. During the training process, the loss function is calculated until the loss function is less than a preset threshold, thus obtaining the pre-trained improved time series model.

8. The method for predicting icing and galloping of transmission lines based on an improved time-series large model according to claim 7, characterized in that, The loss function adopted is a joint loss function, wherein the binary cross-entropy loss function is adopted for the output layer of the classification task, and the mean squared error loss function is adopted for the output layer of the regression task. The expression for the joint loss function is: ， in, Denotes the joint loss function. This represents the binary cross-entropy loss function. This represents the mean squared error loss function. , These represent the weights of the binary cross-entropy loss function and the mean squared error loss function, respectively.

9. A transmission line icing and galloping prediction system based on an improved time-series large model, characterized in that, The steps for implementing the transmission line icing galloping prediction method based on an improved time-series large model as described in any one of claims 1 to 8 include: The data acquisition module is configured to acquire real-time meteorological data; The meteorological change judgment module is configured to compare real-time meteorological data with preset threshold conditions. If the threshold conditions are not met, it is determined that there is no risk of icing and galloping. If the threshold conditions are met, the module monitors the cumulative time of abnormal meteorological data and compares it with the set threshold. If the cumulative time of abnormal meteorological data is lower than the set threshold, it is determined that the risk of icing and galloping has been eliminated. If the cumulative time of abnormal data is higher than the set threshold, the module predicts the probability of icing and galloping of the transmission line and the amplitude of the transmission line galloping based on the improved time series model, line parameters, meteorological data and transmission line icing and galloping data within the abnormal cumulative time period.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the transmission line icing and galloping prediction method based on an improved time-series large model as described in any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the transmission line icing galloping prediction method based on the improved time-series large model as described in any one of claims 1 to 8.

Images