Power transmission line icing risk early warning method fusing induced electricity characteristics and meteorological data

Abstract

The invention relates to the technical field of power system power transmission line on-line monitoring, and discloses a power transmission line icing risk early warning method fusing induced electricity characteristics and meteorological data, and the method comprises the steps: collecting induced electricity and corridor micrometeorological multi-source heterogeneous data, and executing deviation standardization processing; constructing a full-link parameter collaborative optimization strategy, carrying out joint coding on variational mode decomposition parameters and least square support vector machine hyper-parameters, and carrying out global iterative optimization by using an improved lizard optimization algorithm and taking verification set error minimization as a target; performing signal decomposition and sample entropy extraction based on the optimal parameter combination, and constructing a multi-dimensional complexity index; and inputting the reconstructed model to carry out regression calculation, and outputting an icing early warning value through reverse normalization reduction. According to the method, through full-link parameter cooperation and multi-source feature deep fusion, adaptive matching of a signal processing layer and a prediction layer is realized, the defect of independent parameter optimization is overcome, and icing prediction precision and response speed are improved.

Description

Technical Field

[0001] This invention relates to the field of online monitoring technology for power system transmission lines, specifically a method for early warning of icing risk on transmission lines that integrates induced voltage characteristics and meteorological data. Background Technology

[0002] Transmission line icing refers to the physical phenomenon where supercooled water droplets in the air collide with and freeze on the surface of the conductors during low-temperature rain, snow, and freezing weather, or where wet snow accumulates on the conductors. Icing can significantly increase the mechanical load on transmission conductors and towers, and in severe cases, it can lead to serious accidents such as conductor galloping, line breakage, insulator flashover, or even tower collapse and line breakage, seriously threatening the safe and stable operation of the power system. Therefore, conducting early warning of transmission line icing risks, especially accurate prediction of icing thickness, is an important technical foundation for power grid disaster prevention and mitigation and guiding de-icing decisions.

[0003] Existing prediction technologies for icing thickness on transmission lines mainly fall into three categories: physical mechanism models, statistical regression models, and data-driven machine learning models. Physical mechanism models are based on thermodynamics and fluid mechanics principles, establishing equations by analyzing water droplet collisions, freezing, and thermal equilibrium processes. Statistical regression models focus on analyzing the correlation between historical meteorological data and icing thickness, establishing prediction formulas through multiple linear regression or time series analysis. With the development of intelligent sensing technology and big data, data-driven models based on machine learning algorithms such as support vector machines and artificial neural networks are gradually becoming mainstream. These methods improve the accuracy and real-time performance of predictions to some extent by exploring the nonlinear mapping relationship between meteorological parameters and icing thickness.

[0004] However, despite the progress made in existing technologies, several problems remain in practical applications, making it difficult to meet the requirements for high-reliability early warning in terms of prediction accuracy and stability. First, existing methods often rely on single meteorological monitoring data, neglecting the induced electrical characteristics associated with DC transmission lines during operation. These induced electrical signals are extremely sensitive to changes in the conductor surface medium and ground parameters, directly reflecting icing conditions. Relying solely on micro-meteorological data is insufficient to fully characterize the physical process of icing, leading to significant prediction biases during abrupt changes in micro-topography and climate. Second, the collected induced electrical and meteorological data typically exhibit strong nonlinear and non-stationary characteristics, accompanied by substantial noise interference. Traditional methods lack effective signal decomposition and feature extraction mechanisms, directly inputting raw data into the model, making it difficult to capture the multi-scale dynamic characteristics implicit in the signal, thus affecting the model's ability to analyze icing growth trends. Finally, the construction process of existing prediction models often suffers from fragmented parameter optimization. The control parameters of signal decomposition algorithms (such as variational mode decomposition) and the hyperparameters of regression prediction models (such as support vector machines) are usually optimized independently step by step or set based on experience. This local optimization strategy ignores the coupling relationship between the signal processing layer and the model prediction layer, which can easily lead the overall system to get stuck in a local optimum. As a result, the model has weak generalization ability and the prediction accuracy cannot be guaranteed when facing complex and ever-changing icing conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for early warning of icing risk on transmission lines that integrates induced electrical characteristics and meteorological data. This method aims to solve the problems of low accuracy and weak generalization ability in predicting icing on transmission lines caused by neglecting the fusion of induced electrical characteristics, difficulties in extracting features of non-stationary signals, and the fragmented optimization of signal decomposition and prediction model parameters in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a method for early warning of icing risk on transmission lines that integrates induced voltage characteristics and meteorological data. This method addresses the nonlinear and non-stationary mapping relationship between icing thickness on DC transmission lines and induced voltage characteristics and meteorological factors, proposing a prediction scheme based on collaborative optimization of all-link parameters.

[0008] This method solves the local optimum problem caused by traditional step-by-step independent optimization by jointly optimizing the control parameters of the signal decomposition algorithm and the hyperparameters of the prediction model, and achieves deep coupling between feature extraction and regression prediction.

[0009] Specifically, the method includes the following technical processes:

[0010] First, a multi-source data sensing system is established to simultaneously acquire induced electrical characteristic data of DC transmission lines, micro-meteorological data of the line corridor, and measured icing thickness data at corresponding times.

[0011] Considering the significant differences in physical dimensions between different source data, deviation standardization is performed on all acquired data. This process first calculates the maximum and minimum values ​​of each feature data dimension in the historical training dataset, then calculates the difference between the current feature value and the minimum value, and divides the difference by the difference between the maximum and minimum values, thereby linearly mapping the original data to a dimensionless interval and constructing a standardized preprocessed dataset.

[0012] Secondly, based on the preprocessed dataset, a collaborative optimization strategy for end-to-end parameters is constructed. This strategy aims to achieve joint optimization of parameters between the signal processing layer and the model prediction layer by jointly encoding the control parameters (number of modes and quadratic penalty factor) of the variational mode decomposition algorithm and the hyperparameters (regularization parameter and kernel width parameter) of the least squares support vector machine model into a decision vector.

[0013] By iteratively searching the decision vector using a global optimization algorithm, and using the prediction error on the validation set as the evaluation criterion, the globally optimal parameter combination that optimizes the overall prediction performance is obtained.

[0014] In the signal decomposition stage, the variational mode decomposition algorithm is reconstructed using the number of modes and the quadratic penalty factor in the globally optimal parameter combination. This algorithm constructs a constrained variational model to find eigenmode functions with specific center frequencies and finite bandwidths, minimizing the sum of the estimated bandwidths of each component.

[0015] In the solution process, a quadratic penalty factor and Lagrange multipliers are introduced to transform the constrained variational model into an unconstrained augmented Lagrange function. The alternating direction multiplier method is used to alternately update the eigenmode function components, center frequency, and Lagrange multipliers in the frequency domain until the convergence condition is met, thereby decomposing the non-stationary signal into multiple eigenmode function components.

[0016] In the feature extraction stage, the sample entropy algorithm is used to quantify the complexity features of the decomposed components. This process first reconstructs the phase space of the intrinsic mode function components, generates a sequence of reconstructed vectors, and calculates the Chebyshev distance between the reconstructed vectors.

[0017] By statistically analyzing the number of vector pairs with a distance less than a preset similarity tolerance under different embedding dimensions, and calculating the negative natural logarithm of the ratio of vector pair counts as the embedding dimension increases, the sample entropy value is used as a measure of signal complexity. The sample entropy values ​​of each component are then concatenated to generate a multidimensional complexity comprehensive index vector.

[0018] In the regression prediction stage, the least squares support vector machine model is reconstructed using the regularization parameter and kernel width parameter from the globally optimal parameter combination. This model employs a radial basis function (RBF) kernel to map the input multidimensional complexity index vector to a high-dimensional feature space, and the kernel width parameter controls the scope of the RDF. Based on the principle of minimizing structural risk, a regularization parameter is introduced to balance model complexity and training error. By solving a system of linear equations containing Lagrange multipliers, a nonlinear mapping relationship is established between induced electrical characteristics and icing thickness.

[0019] In the specific implementation of parameter optimization, an improved lizard optimization algorithm is adopted to implement a whole-link parameter collaborative optimization strategy. During the initialization phase, the algorithm utilizes the Tent chaotic mapping to generate an initial population to enhance its ergodicity and performs a rounding operation on the dimension representing the number of modes to satisfy the algorithm constraints. The algorithm performs local adjustments through a local search mechanism and global updates through a global exploration mechanism to update the population position.

[0020] The algorithm constructs a fitness evaluation function, transmits the current decision vector parameters to the signal decomposition and model building stages respectively, calculates the root mean square error between the predicted values ​​on the validation set and the actual ice thickness data, and uses this to guide the population iteration, finally outputting the globally optimal parameter combination.

[0021] Finally, the normalized predicted values ​​output by the model are denormalized. Using the maximum and minimum values ​​obtained from the historical training dataset, the normalized predicted values ​​are multiplied by the difference between the extreme values ​​and the minimum value is added to reconstruct the transmission line icing thickness warning value with physical dimensions.

[0022] A second aspect of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the transmission line icing risk early warning method that integrates induced electrical characteristics and meteorological data as described in the first aspect.

[0023] By executing the above methods, the device can automatically complete the entire process from data acquisition, parameter co-optimization, signal processing to ice thickness inversion, and achieve accurate early warning of icing risks for transmission lines.

[0024] This invention provides a method for early warning of icing risk on power transmission lines that integrates induced electrical characteristics and meteorological data. It has the following beneficial effects:

[0025] 1. This invention constructs a full-link parameter collaborative optimization strategy, jointly encoding the decomposition control parameters of the variational mode decomposition algorithm and the model hyperparameters of the least squares support vector machine model into a unified decision vector, and using a global optimization algorithm to iteratively optimize the prediction error on the validation set. This helps overcome the local optimum problem caused by the independent step-by-step optimization of signal decomposition and model prediction in traditional methods, ensuring that the granularity of signal decomposition and the effect of feature extraction directly serve the final prediction target, thereby improving the overall accuracy of icing thickness prediction.

[0026] 2. This invention utilizes variational mode decomposition (VMD) to process the non-stationary characteristics of induced electrical features and meteorological data, and combines this with sample entropy algorithm to quantify the complexity of intrinsic mode function components. This technical approach effectively separates noise interference from effective trend information in the original multi-source data, and mines the implicit nonlinear dynamic characteristics within the data through sample entropy. This provides high-quality input variables for the regression model that better reflect the icing physical process, enhancing the system's ability to capture icing trend changes under complex meteorological environments.

[0027] 3. This invention employs least squares support vector machine as the core model for regression prediction. It balances model complexity and training error by minimizing structural risk, and replaces the quadratic programming solution process of traditional support vector machines by solving a system of linear equations. This model architecture, while ensuring nonlinear mapping capabilities, reduces computational complexity and improves model training efficiency and convergence speed, thus meeting the real-time data processing and response speed requirements of online monitoring scenarios for power transmission lines.

[0028] 4. This invention establishes a deviation standardization and denormalization processing logic based on historical extreme value statistics. In the preprocessing stage, it eliminates the significant difference in physical dimensions between induced electrical characteristics and micrometeorological data. In the prediction output stage, it uses fixed historical extreme values ​​to restore physical quantities. This processing method not only avoids the interference of numerical magnitude differences on the weights of the optimization algorithm, but also ensures that real-time prediction data and historical training data are transformed under the same benchmark, guaranteeing the physical authenticity and numerical stability of the ice thickness inversion results. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0030] Figure 2 This is a comparison chart of icing thickness predictions in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figure 1 and Figure 2 This invention provides a method for predicting the icing thickness of DC transmission lines by fusing induced electrical characteristics and meteorological data. Addressing the technical problems of rigid multi-source feature fusion, fragmented optimization of signal decomposition parameters and prediction model parameters, and the inability of the model kernel structure to dynamically adjust with meteorological conditions in existing technologies, this method proposes a deep prediction architecture based on multi-parameter collaborative optimization and feature adaptive driving, including the following steps:

[0033] S1. Establish a multi-dimensional sensing and data standardization system, synchronously collect induced electrical characteristic data of DC transmission lines, micro-meteorological data of line corridors, and measured ice thickness data at corresponding times, and use deviation standardization method to uniformly process the original multi-source heterogeneous data, eliminate dimensional differences, and construct a standardized basic dataset.

[0034] S2. Based on the strategy of collaborative optimization, adaptive signal decomposition is performed. The key control parameters in the variational mode decomposition algorithm are regarded as dynamic variables to be optimized. The parameter combination generated by the feedback of the global optimization algorithm is received, and the preprocessed induced electricity and meteorological time series signals are decomposed to obtain the intrinsic mode function components that are most conducive to subsequent feature expression.

[0035] S3. Extract modal nonlinear features and construct a comprehensive complexity index. Calculate and fuse the energy entropy and sample entropy of each intrinsic mode function component. Simultaneously, construct a comprehensive entropy index based on the extracted entropy features to reflect the complexity of the current meteorological and line conditions, providing a control basis for the dynamic adjustment of the internal structure of the subsequent prediction model.

[0036] S4. Construct an entropy-driven dynamic deep hybrid kernel extreme learning machine model, build a prediction architecture that includes a deep feature extraction layer and a hybrid kernel function regression layer, establish a nonlinear mapping relationship between the weights of the hybrid kernel function and the comprehensive entropy index in the regression layer, so that the model can adaptively adjust the weight ratio of Gaussian kernel and polynomial kernel according to the complexity of the input signal.

[0037] S5. The improved lizard optimization algorithm is used to perform full-link parameter collaborative optimization. A unified coding space including signal decomposition parameters, dynamic weight control parameters and model hyperparameters is constructed. By introducing specific population initialization, environment simulation and mutation perturbation strategies, the global optimal parameter combination that minimizes the prediction error is found in the multi-dimensional space, so as to achieve the best match between the signal processing end and the model prediction end.

[0038] S6. Based on the global optimal parameters, solidify the model and perform real-time prediction. Use the optimal decomposition parameters to process the real-time collected data and input the extracted features into the solidified prediction model. The model automatically adjusts the kernel function structure according to the feature state of the real-time data and outputs a high-precision predicted value of the icing thickness of DC transmission lines.

[0039] The technical implementation details of each of the above steps will be described in detail below with reference to specific embodiments of the present invention.

[0040] Step S1 involves establishing a multi-dimensional sensing and data standardization system by simultaneously collecting induced electrical characteristic data of DC transmission lines, micro-meteorological data of the line corridor, and measured icing thickness data at corresponding times. The deviation standardization method is used to uniformly process the original multi-source heterogeneous data, eliminate dimensional differences, and construct a standardized basic dataset.

[0041] The specific implementation of step S1 includes sub-steps S1.1 and S1.2.

[0042] S1.1 Construct a multi-source heterogeneous data synchronous acquisition network.

[0043] In order to obtain electrical operating status data of DC transmission lines and meteorological data of the environment in which the lines are located, this step first deploys multi-dimensional sensing equipment at the towers in the icing-prone sections of the DC transmission lines.

[0044] Specifically, non-contact voltage and current sensors and integrated monitoring terminals installed on insulator strings or conductors are used to collect high-frequency induced electrical characteristic data reflecting changes in line load and surface condition.

[0045] Meanwhile, using micro-meteorological stations erected on the same tower or deployed nearby, local micro-meteorological data of the line corridor are collected, including environmental parameters such as ambient temperature, relative humidity, wind speed, wind direction and precipitation. In addition, the measured ice thickness of the conductor at the corresponding time is obtained through tension sensors or online image monitoring terminals as a benchmark label.

[0046] To address the issue of inconsistent sampling frequencies among multiple sensor sources, the system uses GPS or BeiDou timing modules to synchronize the clocks of all data acquisition terminals. Taking the sampling time of the ice thickness data with the lowest sampling frequency as a unified reference, a linear interpolation algorithm is used to map the high-frequency sampled induced electrical characteristic data and meteorological data to this unified reference time point, forming a time-synchronized original multi-source dataset.

[0047] The original dataset consisting of the above data is denoted as . Original dataset The set expression is as follows:

[0048] ;

[0049] In the formula, This is the original multi-source dataset after time synchronization; for The induced current collected at all times; for The induced voltage collected at all times; for The ambient temperature at that moment; for Relative humidity at any given time; for Wind speed at any given moment; for The wind direction at any given moment; for Rainfall at any given time; for Real-time measured ice thickness; This is the index for the sampling time point.

[0050] S1.2 Data normalization preprocessing.

[0051] Obtaining the original dataset Subsequently, considering that the dataset contains multiple physical dimensions and the numerical ranges of each feature value vary greatly (for example, the induced voltage can be as high as several thousand volts, while the ice thickness is only a few millimeters), directly using the original dataset would be problematic. Inputting large numerical features into subsequent prediction models can cause the gradient update direction of the model to be dominated by the numerical features, thereby reducing the model's sensitivity to numerical features and affecting the convergence speed of model training.

[0052] To address the aforementioned issues, this step employs the Min-Max Normalization algorithm to linearly transform each dimension of the feature data. The Min-Max Normalization algorithm calculates the extreme values ​​of the feature columns, linearly mapping all original data to the dimensionless interval [0,1], thus preserving the distribution pattern of the original data while eliminating dimensional differences. The normalization calculation formula is as follows:

[0053] ;

[0054] In the formula, These are the normalized eigenvalues; These are the original eigenvalues; This is the minimum value of the feature in the historical training dataset; This is the maximum value of this feature in the historical training dataset.

[0055] After the above normalization process, the preprocessed dataset is output. Preprocessing datasets The expression is as follows:

[0056] ;

[0057] In the formula, This is the normalized preprocessed dataset; This is the induced current after preprocessing; This is the induced voltage after preprocessing; The ambient temperature after pretreatment; The humidity after pretreatment; The wind speed after pretreatment; The wind direction after pretreatment; This refers to the pre-treated precipitation. The thickness of the ice layer after pretreatment.

[0058] In this embodiment, by constructing a multi-source heterogeneous data synchronous acquisition network and implementing strict deviation standardization processing, the alignment problem between induced electrical characteristics and meteorological data in the spatiotemporal dimension and the dimensional conflict problem in the numerical dimension are effectively solved. A standardized data foundation is constructed, which helps to improve the ability of subsequent variational mode decomposition algorithms to capture signal fluctuations and accelerates the training and convergence speed of the prediction model.

[0059] Step S2 involves implementing adaptive signal decomposition based on a collaborative optimization strategy. Key control parameters in the variational mode decomposition algorithm are treated as dynamic variables to be optimized. The algorithm receives parameter combinations generated by the global optimization algorithm and decomposes the preprocessed induced voltage and meteorological time-series signals to obtain the intrinsic mode function components most favorable for subsequent feature representation. The specific implementation of step S2 includes sub-steps S2.1 and S2.2.

[0060] S2.1 Variational Mode Decomposition Algorithm Model Construction and Solution.

[0061] In response to the nonlinear and nonstationary characteristics of DC transmission induced electrical signals and meteorological data, step S2 employs variational mode decomposition algorithm for time-frequency analysis.

[0062] The core idea of ​​variational mode decomposition (VMD) is to construct and solve a constrained variational problem, decomposing the input signal into several eigenmode function components with specific center frequencies and finite bandwidths. First, the constrained variational model of the VMD algorithm is constructed. The objective function of the VMD algorithm is as follows:

[0063] ;

[0064] The constraints of the variational mode decomposition algorithm are as follows:

[0065] ;

[0066] In the formula, For decomposition A set of intrinsic mode function components; This is the set of center frequencies corresponding to each eigenmode function component; This represents the total number of intrinsic mode function components; It is a time variable; For time The partial derivatives; The Dirac distribution function; The imaginary unit; This is the convolution operator; The input is the timing signal to be decomposed; It is an L2 norm.

[0067] To solve the constrained variational model described above, the variational mode decomposition algorithm introduces a quadratic penalty factor and Lagrange multipliers, transforming the constrained variational problem into an unconstrained augmented Lagrangian function. The formula for the augmented Lagrangian function is as follows:

[0068] ;

[0069] In the formula, To augment the Lagrange function; It is the set of intrinsic mode function components; For the set of center frequencies; For Lagrange multipliers; It is a secondary penalty factor; This represents the total number of intrinsic mode function components; For time The partial derivatives; The Dirac distribution function; The imaginary unit; This is the convolution operator; Pi; For the first One eigenmode function component; is the base of the natural logarithm; For the first The center frequencies of the intrinsic mode function components.

[0070] Step S2 uses the alternating direction multiplier method to iteratively solve the augmented Lagrange function, alternately updating each eigenmode function component, center frequency, and Lagrange multiplier in the frequency domain. The frequency domain update formula for the eigenmode function components is as follows:

[0071] ;

[0072] In the formula, For the first After the first iteration update The frequency domain values ​​of each intrinsic mode function component; For frequency variables; The Fourier transform of the input signal; Indicates except All except Perform summation; For the first The first iteration The frequency domain values ​​of each intrinsic mode function component; For the first The frequency domain values ​​of the Lagrange multipliers at the next iteration; It is a secondary penalty factor; For frequency variables; For the first The first iteration The center frequencies of the intrinsic mode function components.

[0073] The frequency domain update formula for the center frequency is as follows:

[0074] ;

[0075] In the formula, For the first After the first iteration update The center frequencies of the eigenmode function components; For the first After the first iteration update The frequency domain values ​​of the eigenmode function components.

[0076] The frequency domain update formula for the Lagrange multipliers is as follows:

[0077] ;

[0078] In the formula, For the first The frequency domain values ​​of the Lagrange multipliers after the next iteration; In this embodiment, the update step size for the Lagrange multipliers is... Set to 0.

[0079] The variational mode decomposition algorithm terminates its iteration by determining a precision threshold. The convergence criterion formula is as follows:

[0080] ;

[0081] In the formula, Indicates index From 1 to Summation operation is performed on all terms; This represents the total number of intrinsic mode function components; The index of the intrinsic mode function component; For the first After the first iteration update The frequency domain values ​​of each intrinsic mode function component; For the first The first iteration The frequency domain values ​​of each intrinsic mode function component; This represents the current iteration number of the variational mode decomposition algorithm. This is the preset convergence tolerance value.

[0082] S2.2 Cooperative parameter-driven signal decomposition execution.

[0083] Step S2 establishes the parameter transfer mechanism between the variational mode decomposition algorithm and the improved lizard optimization algorithm in subsequent step S5. Step S2 reads the dynamic parameter combinations generated by the improved lizard optimization algorithm during the optimization iteration process. and combine dynamic parameters This serves as the current configuration parameter for the variational mode decomposition algorithm.

[0084] Based on the number of modes read With secondary penalty factor Step S2 processes the preprocessed dataset output from step S1. Variational mode decomposition is performed on the induced current, induced voltage, ambient temperature, humidity and wind speed.

[0085] Taking induced current as an example, the variational mode decomposition algorithm executes the iterative update process in step S2.1 until the calculation result meets the requirements of the convergence criterion formula, and finally outputs... The eigenmode function components of the induced current are denoted as... .

[0086] Similarly, the ambient temperature, humidity, and wind speed data are decomposed to obtain the corresponding intrinsic mode function (EMF) component sets. All the decomposed EMF components will be used as the input source data for feature extraction in step S3.

[0087] In this embodiment, by constructing an adaptive signal decomposition mechanism based on collaborative optimization, the traditional method of fixed parameter decomposition is changed, thereby increasing the number of modes in the variational mode decomposition algorithm. With secondary penalty factor By incorporating a full-link optimization closed loop, the granularity and bandwidth of signal decomposition can be adaptively adjusted based on the feedback from the final prediction model. This enables parameter correlation between the data processing layer and the model prediction layer, which helps to extract more representative signal features, thereby improving the adaptability of the overall prediction system.

[0088] Step S3 involves extracting modal nonlinear features and constructing a comprehensive complexity index. Step S3 introduces a sample entropy algorithm to calculate the complexity of the intrinsic mode function components. The intrinsic mode function components of the meteorological and induced electrical data output in step S2 are calculated to construct a multidimensional feature vector matrix. The specific implementation of step S3 includes sub-steps S3.1 and S3.2.

[0089] S3.1 Construction of the Sample Entropy Algorithm Model.

[0090] To address the nonlinear and non-stationary characteristics of induced current meteorological data, step S3 uses sample entropy to quantify the complexity of the time-series signal. A lower sample entropy value indicates higher self-similarity of the time-series signal; a higher sample entropy value indicates higher complexity. For a length of... For time-series signals, the calculation of sample entropy first involves reconstructing the phase space of the time-series signal to build a sequence of reconstructed vectors. The expression for the reconstructed vectors is as follows:

[0091] ;

[0092] In the formula, For the first One reconstructed vector; For the timing signal at the 1st The value at each moment; For the embedding dimension; This is a time index, with values ​​ranging from 1 to... ; This represents the total length of the timing signal.

[0093] The distance between two reconstructed vectors is defined as the maximum absolute value of the difference between corresponding elements of the two reconstructed vectors, and the distance formula is as follows:

[0094] ;

[0095] In the formula, For two reconstructed vectors and Chebyshev distance between them; For from the first Starting from a point 3D reconstructed vector; For from the first Starting from a point 3D reconstructed vector; For the embedding dimension; This is for calculating the maximum value. This is the offset index within the vector; For the timing signal at the 1st The value at each moment; For the timing signal at the 1st The value at each moment.

[0096] Based on the above distance formula, statistics are calculated for embedding dimensions of... and At that time, the number of vector pairs whose distance between reconstructed vectors is less than the preset similarity tolerance r is counted, and the sample entropy is calculated. The formula for calculating the sample entropy is as follows:

[0097] ;

[0098] In the formula, The calculated sample entropy value; For natural logarithm operations; The total length of the timing signal; For the embedding dimension; For similarity tolerance; Indicates index From 1 to Summing all terms; For embedding dimension is When, with vector The distance between them is less than the similarity tolerance. And satisfy The number of vectors; For embedding dimension is When, with vector The distance between them is less than the similarity tolerance. And satisfy The number of vectors.

[0099] S3.2 Modal feature extraction and comprehensive index construction.

[0100] Step S3 reads the set of all intrinsic mode function components output in step S2. The set of intrinsic mode function components includes induced current components, induced voltage components, and meteorological data components. The embedding dimension of the sample entropy algorithm is then set. The value is usually set to 2, which represents the similarity tolerance. The value is taken as 0.1 to 0.25 times the standard deviation of the intrinsic mode function component sequence.

[0101] Step S3 calculates the sample entropy value of each intrinsic mode function component sequentially. Taking induced current as an example, the calculated sample entropy values ​​are used to form an induced current feature vector. The sample entropy values ​​of the intrinsic mode function components of all types of data are concatenated to construct a multidimensional complexity feature vector. This multidimensional complexity feature vector characterizes the nonlinear dynamic characteristics of the signal in different frequency bands and will serve as the input feature set for establishing the prediction model in the subsequent step S4.

[0102] In this embodiment, step S3 introduces sample entropy to perform secondary feature extraction on the components after variational mode decomposition, which helps overcome the problem that it is difficult to quantify the nonlinearity of the signal using only time-domain or frequency-domain features. Sample entropy can measure the complexity changes of induced electricity and meteorological data at multiple scales. By constructing a multi-dimensional complexity feature vector, it can more effectively capture abrupt changes in the data and provide input features containing dynamic characteristics for subsequent prediction models.

[0103] Step S4 involves establishing a prediction model based on least squares support vector machines. Step S4 receives the multidimensional complexity comprehensive index vector constructed in step S3 as input, and uses least squares support vector machines to map the nonlinear features to a high-dimensional space for regression analysis, thus constructing a mapping relationship for icing thickness prediction. The specific implementation of step S4 includes sub-steps S4.1 and S4.2.

[0104] S4.1 Construction of the least squares support vector machine regression model.

[0105] For the nonlinear regression problem of icing thickness, step S4 constructs a least squares support vector machine model. Given the training dataset output from step S3, which contains multiple samples, the least squares support vector machine maps the input feature vectors to a high-dimensional feature space through a nonlinear mapping function, and constructs the optimal linear regression function in this high-dimensional feature space based on the principle of minimizing structural risk.

[0106] In the model building process, step S4 introduces regularization parameters to balance the complexity of the model and the training error, and uses a least squares linear system as the loss function.

[0107] By introducing Lagrange multipliers, the constrained optimization problem is transformed into a dual problem. Furthermore, the Karl-Kuhn-Tucker conditions (KKT conditions) are used to further transform the optimization problem into solving a system of linear equations, thereby obtaining the optimal weight vector and bias. The regularization parameter determines the model's fit to the training samples and its generalization ability, and is one of the key parameters that need to be optimized in the subsequent step S5.

[0108] S4.2 Kernel Function Mapping and Predictive Execution.

[0109] In solving the linear equation system, step S4 introduces a kernel function to replace the inner product operation in the high-dimensional feature space to avoid errors. This embodiment uses the radial basis function (RBF) kernel as the core mapping tool. The calculation formula for the RBF kernel is as follows:

[0110] ;

[0111] In the formula, Let be the kernel function, representing the th . training samples With the training samples Inner product in a high-dimensional feature space; Using natural constants An exponential function with base 0; The square of the Euclidean distance between the two eigenvectors; The width parameter (kernel width) of the kernel function is used to control the range of the radial basis functions. For the first The input feature vector of each training sample; For the first The input feature vector of each training sample.

[0112] Based on the obtained Lagrange multipliers and biases, step S4 outputs the final least squares support vector machine prediction model expression:

[0113] ;

[0114] In the formula, Samples to be predicted The predicted value of ice thickness; This represents the total number of training samples; Indicates index From 1 to Summation operation is performed on all terms; For sample index; To obtain the first One Lagrange multiplier; Samples to be predicted With the training samples Kernel function values ​​between; This is the bias obtained by solving.

[0115] In this embodiment, step S4 establishes a nonlinear regression prediction architecture based on least squares support vector machines. The performance of the prediction model is highly dependent on the selection of regularization parameters and kernel width parameters, which will be used as optimization targets in the subsequent step S5 to improve the lizard optimization algorithm.

[0116] Step S5 involves optimizing the end-to-end parameters based on the improved lizard optimization algorithm. Step S5 constructs the improved lizard optimization algorithm, which includes the number of modes defined in step S2. With secondary penalty factor The regularization parameter and kernel width parameter defined in step S4 are combined to form a four-dimensional decision vector. Through iterative optimization, the optimal parameter combination that minimizes the prediction model error is obtained. The specific implementation of step S5 includes sub-steps S5.1, S5.2, and S5.3.

[0117] S5.1 Population Initialization and Decision Vector Encoding.

[0118] Step S5 first determines the search space for the decision vector. The formula for constructing the four-dimensional decision vector is as follows:

[0119] ;

[0120] In the formula, The number of modalities defined for step S2 is an integer variable; The quadratic penalty factor defined in step S2, The regularization parameters defined in step S4, The kernel width parameters defined in step S4 are all continuous variables.

[0121] Step S5 sets the population size for the improved lizard optimization algorithm to be... The maximum number of iterations is To enhance the ergodicity of the initial population, a Tent chaotic map is used to generate initial positions. For the ... For each individual lizard, an initial value belonging to the interval (0,1) and not equal to 0.5 is randomly generated as the chaos seed. Then, the Tent chaos mapping formula is used to generate a chaotic sequence of values, which are then mapped to the decision vector space. The initialization formula is as follows:

[0122] ;

[0123] In the formula, Indicates the first The first lizard individual in the... In the decision space of the first dimension The sequence values ​​generated by the next chaotic iteration; Indicates the first The first lizard individual in the... In the 3D decision space The chaotic sequence value at the next iteration; This serves as an index for the individual lizards within the population. Number the dimensions of the decision vector; This represents the current iteration number of the chaotic mapping.

[0124] ;

[0125] In the formula, For pre-set The chaotic value selected after the next iteration; For the first The first lizard individual in the search space The initial position value of the dimension; For the first The lower bound of the values ​​that a decision variable can take; For the first The upper limit of the values ​​that a decision variable can take; Population size.

[0126] To address the constraint that the number of modes in the decision vector must be an integer, step S5 performs a rounding operation on the first dimension of the position vector during the initialization phase and after each subsequent position update operation.

[0127] ;

[0128] In the formula, For the first The modality count parameter value corresponding to each individual lizard; This is the value of the first dimension of the position vector.

[0129] S5.2 Construct the full-link fitness evaluation function.

[0130] Step S5 establishes a fitness function connecting steps S2, S3, and S4 to evaluate the quality of the current parameter combination. For the... The parameter combination represented by each individual lizard The fitness value is calculated as follows:

[0131] Candidate parameters and The data is transmitted to step S2 as configuration parameters for the variational mode decomposition algorithm to decompose the preprocessed induced electrical and meteorological data and obtain the intrinsic mode function components.

[0132] The intrinsic mode function components obtained from the decomposition are transferred to step S3 to calculate the sample entropy of each component and construct a multidimensional complexity comprehensive index vector.

[0133] The multidimensional complexity comprehensive index vector is divided into a training set and a validation set, and candidate parameters are used. and Construct a least-squares support vector machine model based on the training set;

[0134] The constructed model is used to predict the validation set, and the root mean square error is calculated as the fitness value.

[0135] The formula for calculating the fitness function is as follows:

[0136] ;

[0137] In the formula, For the first Fitness value of an individual lizard; The number of samples in the validation set; To verify the first set The true value of the ice thickness for each sample; To utilize parameter combinations The trained model is used on the validation set. Predicted ice thickness for each sample.

[0138] S5.3 Population position iterative update and optimal parameter output.

[0139] An improved lizard optimization algorithm simulates the lizard's predation behavior, including a refuge-assisted phase and a natural predation phase. In each iteration, the algorithm updates the position of the globally optimal lizard based on its fitness value.

[0140] During the shelter support phase (partial development), individual lizards move based on the current best individual's position, and the position update formula is as follows:

[0141] ;

[0142] In the formula, For the first During the nth iteration Updated location for each individual lizard; For the first The globally optimal lizard position at the next iteration; For the first During the nth iteration The current location of each individual lizard; is a random scalar distributed in the interval [0,1].

[0143] During the natural predation phase (global exploration), when encountering a predator, individual lizards will either flee or reposition themselves. The position update formula is as follows:

[0144] ;

[0145] In the formula, For the first The update position in the next iteration; For the first The current position at the next iteration; , These are random numbers distributed in the interval [0,1]. It is the upper bound vector of the decision vector; This is the lower bound vector of the decision vector; It is a constant for small perturbations; To follow a standard normal distribution A random vector.

[0146] Step S5 repeats the above evaluation and update process until the maximum number of iterations is reached. After the algorithm terminates, it outputs the optimal parameter combination corresponding to the minimum global fitness value. Step S5 will and Configure to step S2, and Configure to step S4 to complete the final deployment of the prediction model.

[0147] In this embodiment, the joint optimization of parameters in the data decomposition layer and the prediction model layer is achieved by improving the lizard optimization algorithm. The Tent chaotic mapping enhances the ergodicity of the initial population, and the full-link fitness function ensures that the optimization direction directly points to the final prediction accuracy. This helps to solve the local optimum problem caused by traditional step-by-step independent parameter tuning and guarantees the best performance of the prediction system when processing nonlinear and non-stationary induced electrical data.

[0148] Step S6: Execution of end-to-end optimal prediction and real-time icing early warning

[0149] Step S6, as the online application stage of the technical solution of this invention, mainly involves loading the globally optimal parameter combination determined in step S5 into the signal processing and prediction model to construct a real-time prediction system capable of adapting to complex meteorological environments. Step S6 receives real-time induced voltage characteristics and meteorological monitoring data transmitted from the DC transmission line, performs online calculations using the optimal model, and outputs the current predicted value of the line icing thickness, thereby achieving real-time early warning of the icing risk of the transmission line. The specific implementation of step S6 includes sub-steps S6.1 and S6.2.

[0150] S6.1 Optimal Parameter Deployment and Real-time Feature Engineering

[0151] Step S6 first reads the optimal number of modes output in step S5. With the optimal quadratic penalty factor This is then solidified into the operating parameters of the variational mode decomposition algorithm. When the monitoring terminal transmits data in real time... When the system analyzes the induced electrical characteristic sequence and meteorological data sequence at a given time, it uses a pre-configured variational mode decomposition algorithm to decompose various time-series signals online.

[0152] For the real-time intrinsic mode function components obtained from the decomposition, step S6 calculates the complexity value of each component according to the sample entropy algorithm in step S3. During this process, the system maintains the same embedding dimension and similarity tolerance settings as in the training phase, and concatenates the calculated real-time sample entropy values ​​in a predetermined order to construct a multi-dimensional complexity comprehensive index vector that can characterize the current line operating state. This vector is then input into the prediction model as a decision-making basis.

[0153] S6.2 Online Inversion and Numerical Output of Icing Thickness

[0154] Step S6 takes the optimal regularization parameters output from step S5. With the optimal kernel width parameter Load the data into the least squares support vector machine model, use all historical data to complete the final training and weight locking of the model, and generate an online prediction model.

[0155] Step S6 inputs the real-time multidimensional complexity comprehensive index vector generated in S6.1 into the online prediction model. The model, through kernel function mapping and weighted calculation, outputs the normalized predicted icing thickness value for the current moment. To obtain physically meaningful engineering values, step S6 uses the maximum and minimum values ​​of historical icing thickness data collected in step S1 to perform inverse normalization on the normalized predicted value. The inverse normalization calculation formula is as follows:

[0156] ;

[0157] In the formula, for Predicted true value of icing thickness after inverse normalization at any time; Output of online prediction model Time-normalized predicted values; The maximum ice thickness found in the historical training dataset in step S1; The minimum ice thickness in the historical training dataset statistically analyzed in step S1.

[0158] The system will calculate The system compares the predicted value with the preset icing alarm threshold. If the predicted value exceeds the threshold, the system immediately triggers an early warning signal, notifying maintenance personnel to take de-icing or ice removal measures.

[0159] In this embodiment, by deploying the optimal parameter combination obtained from end-to-end optimization to the online monitoring system, the best match between signal decomposition and model prediction parameters is achieved. The variational mode decomposition and least squares support vector machine models are driven by real-time collected induced current and meteorological data, and high-precision icing thickness prediction values ​​are directly output without manual intervention. This helps the operation and maintenance department to grasp the icing status of the lines in real time and improve the response speed of power grid disaster prevention and mitigation under extreme weather conditions.

Claims

1. A method for early warning of icing risk on transmission lines that integrates induced electrical characteristics and meteorological data, characterized in that, Includes the following steps: Acquire induced current characteristic data of DC transmission lines, micro-meteorological data of the line corridor, and measured ice thickness data at corresponding times. Perform deviation standardization on all acquired data to construct a standardized preprocessed dataset. Based on the preprocessed dataset, a full-link parameter collaborative optimization strategy is constructed. The control parameters of the variational mode decomposition algorithm and the hyperparameters of the least squares support vector machine model are jointly encoded into decision vectors. The globally optimal parameter combination that minimizes the prediction error of the validation set is obtained through iterative optimization. The variational mode decomposition algorithm is reconstructed using the control parameters in the global optimal parameter combination. The induced electrical characteristic data and micro-meteorological data at the time to be predicted are decomposed into signals. The complexity features of the decomposed components are extracted by the sample entropy algorithm to generate a multi-dimensional complexity comprehensive index vector. The least squares support vector machine model is reconstructed using the hyperparameters in the global optimal parameter combination. The multidimensional complexity comprehensive index vector is then input into the reconstructed model for regression calculation. The calculation results are then inversely normalized to output the icing thickness warning value for the transmission line.

2. The method for early warning of icing risk of transmission lines by integrating induced electrical characteristics and meteorological data according to claim 1, characterized in that, The deviation standardization process specifically includes: Find the maximum and minimum values ​​of each feature data dimension in the historical training dataset. Calculate the difference between the current feature value and the minimum value, and divide the difference by the difference between the maximum value and the minimum value to obtain the normalized feature value mapped to the dimensionless interval; In the inverse normalization process, the maximum and minimum values ​​obtained from the historical training dataset are used to reverse the calculation results.

3. The method for early warning of icing risk of transmission lines by integrating induced electrical characteristics and meteorological data according to claim 1, characterized in that, The control parameters of the variational mode decomposition algorithm include the number of modes and the quadratic penalty factor; The specific process of signal decomposition includes: A constrained variational model for variational mode decomposition is constructed, and the eigenmode functions with specific center frequencies and finite bandwidths are found to minimize the sum of the estimated bandwidths of each component. By introducing the quadratic penalty factor and the Lagrange multiplier, the constrained variational model is transformed into an unconstrained augmented Lagrange function; The intrinsic mode function components, center frequency, and Lagrange multipliers are alternately updated in the frequency domain according to the number of modes until the convergence condition is met, and then the intrinsic mode function components are output.

4. The method for early warning of icing risk of transmission lines by integrating induced electrical characteristics and meteorological data according to claim 1, characterized in that, The extraction of complexity features of decomposed components using the sample entropy algorithm specifically includes: The decomposed components, i.e., the eigenmode function components, are reconstructed in phase space to generate a reconstructed vector sequence; Calculate the Chebyshev distance between the reconstructed vectors and count the number of vector pairs whose distance is less than the preset similarity tolerance under different embedding dimensions; The negative of the natural logarithm of the vector pair scalar ratio is calculated as the embedding dimension increases, and used as the sample entropy value to measure the complexity of the intrinsic mode function components.

5. The method for early warning of icing risk of transmission lines by integrating induced electrical characteristics and meteorological data according to claim 1, characterized in that, The hyperparameters of the least squares support vector machine model include regularization parameters and kernel width parameters; the specific process of regression calculation includes: The radial basis function is used to map the input multidimensional complexity index vector to a high-dimensional feature space, and the kernel width parameter is used to control the range of the radial basis function. An optimization objective is constructed based on the principle of minimizing structural risk, and the regularization parameter is introduced to balance the complexity and training error of the model. By solving a system of linear equations containing Lagrange multipliers, the weight vector and bias of the model are determined, and a nonlinear mapping relationship from induced electrical characteristics to ice thickness is established.

6. The method for early warning of icing risk of transmission lines by integrating induced electrical characteristics and meteorological data according to claim 1, characterized in that, The end-to-end parameter collaborative optimization strategy employs an improved lizard optimization algorithm, the initialization process of which includes: Determine the upper and lower limits of the decision vector and construct the search space; A chaotic sequence is generated using the Tent chaotic mapping, and the chaotic sequence is mapped to the search space to generate the position vector of the initial population; The dimension representing the number of modes in the position vector is rounded to meet the parameter constraints of the variational mode decomposition algorithm.

7. The method for early warning of icing risk of transmission lines by integrating induced electrical characteristics and meteorological data according to claim 6, characterized in that, The improved lizard optimization algorithm guides the optimization process by constructing a fitness evaluation function, the calculation process of which includes: The parameter combination represented by the current individual is analyzed, the control parameters are transmitted to the signal decomposition stage, and the hyperparameters are transmitted to the model construction stage. Based on the current parameters, variational mode decomposition and sample entropy feature extraction are performed on the training set data to generate a training feature set. The least squares support vector machine model is trained using the training feature set, and the trained model is used to predict the validation set data. The root mean square error between the predicted values ​​on the validation set and the actual icing thickness data is calculated, and the root mean square error is used as the fitness value of the current individual.

8. The method for early warning of icing risk of transmission lines by integrating induced electrical characteristics and meteorological data according to claim 6, characterized in that, The position update process of the improved lizard optimization algorithm includes: During the shelter support phase, the positions of individuals in the population are locally adjusted based on the current position of the globally optimal individual and the random distribution factor. During the natural predation phase, an update strategy is selected based on random probability, and the position of individuals in the population is updated globally by simulating escape behavior or position adjustment after being disturbed. After each location update, the fitness value of the individual is recalculated, and the global optimal parameter combination is updated.

9. The method for early warning of icing risk of transmission lines by integrating induced electrical characteristics and meteorological data according to claim 1, characterized in that, The specific calculation logic for the inverse normalization process is as follows: Multiply the normalized predicted value obtained from the regression calculation by the difference between the maximum and minimum historical icing thickness to obtain the restored change amplitude; The change amplitude is added to the historical minimum icing thickness to obtain an icing thickness warning value with physical dimensions.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the transmission line icing risk early warning method that integrates induced electrical characteristics and meteorological data as described in any one of claims 1 to 9.

Images