Intelligent monitoring and early warning method, system and equipment for icing thickness of distribution line and medium

Abstract

The invention relates to the technical field of power system safety monitoring equipment, in particular to an intelligent monitoring and early warning method, system and equipment for the icing thickness of a distribution line and a medium, and synchronous acquisition and preprocessing of multi-dimensional parameters are realized by deploying an ice-resistant packaging multi-parameter sensing array; dynamically optimizing a modal number and a penalty factor of variational modal decomposition by using a genetic algorithm to realize multi-scale feature decoupling of the icing signal; a gating circulation unit is adopted to extract time sequence features, the time sequence features are input into a Kolmophilov-Arnod network to execute nonlinear mapping, and dual-path output is fused through residual connection; deploying the trained hybrid model to an industrial-grade edge computing unit, and completing local real-time reasoning by using edge computing power; triggering incremental updating of the sliding window based on a prediction error threshold value, and dynamically adjusting the weight of an output layer to maintain the adaptability of the model; through hardware-algorithm collaborative innovation, reliable operation is guaranteed in an extremely low-temperature environment, the early warning response time is remarkably shortened, and the prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system safety monitoring equipment, and in particular to a method, system, equipment and medium for intelligent monitoring and early warning of ice thickness on distribution lines. Background Art

[0002] In cold climates, ice covering power distribution lines due to factors such as precipitation and wind can significantly increase line loads and may lead to serious accidents such as conductor fluttering, line breakage, and even tower collapse. However, existing monitoring technologies face significant challenges: traditional equipment relies on manual or single sensors, with low data acquisition frequency and a lack of local computing power, resulting in severe response delays; at the same time, these devices have poor reliability in extremely low temperature environments. In terms of predictive models, physical models require high-performance servers and cannot be deployed on-site, while traditional machine learning and deep learning models on edge devices either have slow inference speeds or have high memory requirements that make them difficult to integrate. In addition, the difficulty of fusing multi-source sensor data and the poor compatibility of models with hardware platforms result in large warning delays in existing systems and an inability to meet real-time prevention and control needs.

[0003] To address these core issues, an innovative solution was proposed. This solution enhances data collection capabilities through a multi-sensor fusion design and employs industrial-grade embedded hardware to ensure stable operation in extreme environments. At the algorithmic level, an innovative solidification and optimization of the hybrid prediction model significantly reduces model complexity and memory usage, enabling efficient computation at the edge. In engineering applications, local real-time prediction and early warning are achieved, significantly shortening response time, and supporting dynamic updates through a data caching mechanism. This integrated sensing, computing, and early warning design effectively addresses key challenges such as poor real-time performance, weak environmental adaptability, and insufficient prediction accuracy, providing timely and reliable all-weather protection for the line. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: to address the key problems existing in the current ice thickness prediction technology - complex physical model parameters, insufficient adaptability of traditional machine learning to high-dimensional data, and the difficulty of a single deep learning model to balance the contradiction between time series characteristics and nonlinear mapping, an innovative VMD-GRU-KAN hybrid prediction method is proposed.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: an intelligent monitoring and early warning method for ice thickness of distribution lines, which comprises: Deploy an ice-resistant encapsulated multi-parameter sensor array, sample data at a fixed frequency, synchronously collect multi-dimensional parameters and pre-process the data; The variational mode decomposition parameters of ice cover data are dynamically optimized by genetic algorithm to determine the mode number and penalty factor. A gated recurrent unit is used to extract ice cover temporal features, which are then fed into a KAN to perform nonlinear mapping, and the dual-path outputs are fused through a residual connection. Deploy the trained VMD-GRU-KAN model to an industrial-grade AI edge computing unit, utilizing 2TOPS computing power to perform real-time inference. The sliding window incremental update mechanism is triggered. When the prediction error exceeds the threshold, new data is automatically collected and the output layer weights are adjusted to keep the model dynamically adaptable.

[0007] As a preferred solution of the intelligent monitoring and early warning method for ice thickness of distribution lines described in the present invention, the deployment of the anti-ice package multi-parameter sensor array includes deploying a multi-dimensional meteorological sensor and installing an ice thickness detection unit; Among them, the multi-dimensional meteorological sensors include temperature, relative humidity, wind speed, wind direction, air pressure, precipitation and conductor surface temperature sensors; Among them, the installation of ice thickness detection units includes using the laser ranging principle, installed below the conductor suspension insulator string, covering the key nodes of the entire line; The multidimensional parameters include temperature, relative humidity, wind speed, wind direction, air pressure, precipitation, ice thickness, ice weight and conductor surface temperature.

[0008] As a preferred solution of the intelligent monitoring and early warning method for ice thickness of distribution lines described in the present invention, the method of dynamically optimizing the variational mode decomposition parameters of ice data using a genetic algorithm includes adaptively decomposing the original data into an intrinsic mode function, which is expressed as: , in, Indicates the decomposition A modal function, represents the center frequency of the mode function, Represents a group and The minimum value of the objective function, represents the time derivative, represents the Dirac impulse function, represents the Hilbert transform kernel, represents the imaginary unit, represents the convolution operation, express paradigm, represents the modal number, represents the modal index, represents the time variable, Represents the base of natural logarithms.

[0009] As a preferred solution of the intelligent monitoring and early warning method for ice thickness of distribution lines described in the present invention, wherein: the determination of the modal number and the penalty factor includes the modal number and penalty factor ;

[0010] in, The value determines the IMF components obtained by signal decomposition. The value affects the bandwidth constraint of the decomposition.

[0011] As a preferred solution of the intelligent monitoring and early warning method for ice thickness of distribution lines described in the present invention, the input KAN performs nonlinear mapping, and the dual-path output is fused through residual connection, including the input layer receiving the GRU layer output and constructing a two-layer KAN network, and using neuron output to predict the ice thickness value; KAN is based on the Kolmogorov-Arnold representation theorem through a finite composite representation of a single variable continuous function as follows: , in, express Continuous function on the dimensional input space, represents the inner univariate continuous function, represents the outer univariate continuous function, Indicates the input variables, represents the input variable dimension, Represents the feature vector position index corresponding to the sensor, represents the nonlinear layer neuron index, represents the first eigenvalue of the input, Indicates the input eigenvalues; Among them, KAN includes inner transformation layer, aggregation layer, and outer transformation layer; Among them, the inner transformation layer maps each input dimension to a new feature space, introduces nonlinear transformation, and processes the input independently as follows: , in, represents the number of spline basis functions, represents the trainable parameters, Indicates the B-spline basis functions, represents the activation function of the inner network of KAN, represents the basis function index, represents the output feature index, represents the input feature index, represents the input feature vector; Among them, the aggregation layer represents the combination of features of different input dimensions, and the adaptive adjustment of feature importance is achieved through weights as follows: , in, represents the learnable weight parameters, Indicates the The output of the aggregation unit, represents the input features, Represents the input features The learnable nonlinear transformations performed; Among them, the outer transformation layer includes nonlinear transformation of the aggregation results to enhance the expression ability of the network, and maps it to the target space as follows: , in, Indicates the final output result. Indicates the KAN network The nonlinear transformation output of each neuron.

[0012] As a preferred solution of the intelligent monitoring and early warning method for ice thickness on distribution lines described in the present invention, the trained VMD-GRU-KAN model is deployed to an industrial-grade AI edge computing unit, and real-time reasoning is performed using 2TOPS computing power, including adopting an industrial-grade embedded AI chip, and memory configuration for real-time reasoning, quantizing and compressing the trained VMD-GRU-KAN hybrid model to adapt to the memory limitations of the edge computing unit.

[0013] As a preferred solution of the intelligent monitoring and early warning method for ice thickness of distribution lines described in the present invention, the triggering sliding window incremental update mechanism includes immediately triggering when the real-time prediction error exceeds the threshold for three consecutive times; collecting 8 hours of ice monitoring data, and automatically shortening the update cycle to 4 hours in extreme weather conditions; applying The criterion automatically filters outliers; freezes the pre-trained weights of GRU-KAN and uses sliding window data to adjust the output fully connected layer parameters.

[0014] The present invention provides an intelligent monitoring and early warning system for ice thickness on distribution lines. It uses an ice-resistant packaged multi-sensor array to synchronously collect eight-dimensional parameters, such as temperature, humidity, and wind speed. It combines a genetic algorithm to dynamically optimize the mode number and penalty factor of variational mode decomposition, thereby achieving multi-scale feature decoupling of icing signals. It uses a gated recurrent unit to extract temporal features and a Kolmogorov-Arnold network to perform nonlinear mapping, fusing dual-path outputs through residual connections. The compressed hybrid model is deployed to an industrial-grade edge computing unit (with 2TOPS computing power) to perform high-precision real-time inference locally. An incremental update mechanism is triggered based on a prediction error threshold to dynamically adjust the output layer weights. Ultimately, a hardware-algorithm collaborative "perception-decomposition-prediction-optimization" closed-loop system is constructed to achieve all-weather intelligent prevention and control of icing disasters.

[0015] To solve the above technical problems, the present invention provides the following technical solutions: an intelligent monitoring and early warning system for ice thickness on distribution lines, comprising: The anti-icing sensor array module deploys an anti-icing packaged multi-parameter sensor array, samples data at a fixed frequency, synchronously collects multi-dimensional parameters and pre-processes the data; Decomposition optimization module, which dynamically optimizes the variational mode decomposition parameters of ice cover data through genetic algorithm to determine the number of modes and penalty factors; The residual fusion module uses a gated recurrent unit to extract ice cover temporal features, inputs them into a KAN to perform nonlinear mapping, and fuses the dual-path outputs through a residual connection. The edge curing module deploys the trained VMD-GRU-KAN model to an industrial-grade AI edge computing unit, utilizing 2TOPS computing power to perform real-time inference. The dynamic fine-tuning module triggers the sliding window incremental update mechanism. When the prediction error exceeds the threshold, it automatically collects new data and adjusts the output layer weights to keep the model dynamically adaptable.

[0016] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the intelligent monitoring and early warning method for ice thickness of distribution lines are implemented.

[0017] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the intelligent monitoring and early warning method for ice thickness of distribution lines are implemented.

[0018] Beneficial effects of the present invention: The intelligent monitoring and early warning system for ice thickness on distribution lines provided by the present invention effectively overcomes the reliability defects of traditional monitoring equipment in extreme low temperature environments through the collaborative innovation of ice-resistant packaged sensor arrays and edge computing hardware, and realizes stable operation around the clock in high-altitude and cold areas; relying on hybrid model edge solidification and adaptive parameter optimization technology, it greatly shortens the ice thickness prediction response time and significantly improves the prediction accuracy; combined with the dynamic incremental update mechanism, it continuously maintains the model prediction stability under complex meteorological conditions such as blizzards and freezing rain, and systematically solves the three major technical problems in power grid ice monitoring, namely poor real-time performance, weak environmental adaptability and insufficient long-term accuracy, providing reliable technical support for the safety protection of distribution lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 This is a logical structure diagram of an ice thickness monitoring and early warning device for an intelligent monitoring and early warning method for ice thickness on distribution lines provided by one embodiment of the present invention.

[0021] Figure 2 A VMD flow chart optimized by a genetic algorithm for an intelligent monitoring and early warning method for ice thickness on distribution lines provided in one embodiment of the present invention.

[0022] Figure 3 A GRU structure diagram of an intelligent monitoring and early warning method for ice thickness on distribution lines provided in one embodiment of the present invention.

[0023] Figure 4 A diagram of the GRU-KAN combined structure of an intelligent monitoring and early warning method for ice thickness on distribution lines provided in one embodiment of the present invention.

[0024] Figure 5 A VMD decomposition diagram of a method for intelligent monitoring and early warning of ice thickness on distribution lines provided in one embodiment of the present invention.

[0025] Figure 6 A comparison chart of predictions from various models of the intelligent monitoring and early warning method for ice thickness on distribution lines provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0026] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0027] Example 1, reference Figures 1 to 4 , is an embodiment of the present invention, which provides an intelligent monitoring and early warning method for ice thickness on distribution lines, comprising: S1: Deploy an ice-resistant packaged multi-parameter sensor array, sample data at a fixed frequency, synchronously collect multi-dimensional parameters and pre-process the data.

[0028] It should be noted that the deployment of an ice-resistant packaged multi-parameter sensing array includes the deployment of multi-dimensional meteorological sensors and the installation of an ice thickness detection unit;

[0029] The deployment of multi-dimensional meteorological sensors includes 7-dimensional meteorological sensors, including temperature (±0.1°C accuracy), relative humidity (±1% accuracy), wind speed (0-30m / s range), wind direction (0-360°), air pressure (500-1100hPa), precipitation (0-50mm / h), and conductor surface temperature (-40°C~100°C). The ice thickness detection unit is installed using a laser ranging principle (resolution 0.01mm) and is installed 50cm below the conductor suspension insulator string. The sensor group spacing is ≤200 meters, covering key nodes along the entire line (such as corner towers and wide-span sections). Specifically, such as Figure 1 As shown, the deployment of ice-resistant packaged multi-parameter sensing arrays also includes edge computing unit configuration, environmental adaptability design, and dual-mode communication module deployment; The edge computing unit includes an embedded AI (artificial intelligence) chip with a core processor computing power of 2TOPS (two trillion operations per second); and is equipped with dedicated memory for real-time model inference; it can store 20 days of raw data (at a 1Hz sampling rate); Among them, the environmental adaptability design includes wide temperature operation within the range of -40℃~85℃, equipped with active heat dissipation, heating and protection; Among them, the dual-mode communication module deployment includes support for SA / NSA (standalone networking / non-standalone networking) dual modes, real-time data backhaul, setting up backup channels, and adopting the LoRa (long-range radio) module activation mechanism.

[0030] Furthermore, the synchronous collection of multi-dimensional parameters includes obtaining historical meteorological data from the weather station, including: temperature (°C), relative humidity (%), wind speed (m / s), wind direction (°), air pressure (hPa) and precipitation (mm); and collecting historical data from the distribution line icing monitoring system, including ice thickness (mm), ice weight (kg / m) and conductor surface temperature; establishing a data collection cycle of 1 hour, and continuously collecting data for two complete freezing seasons.

[0031] Furthermore, data preprocessing includes: use The standard marks abnormal values ​​and records the collected temperature, humidity, ice thickness and other data as , for a single observation in any variable (such as temperature series), it is recorded as collected data , and calculate the mean and standard deviation respectively, expressed as: , , in, 、 Respectively represent the collected data The mean and standard deviation of represents the output feature index, represents the input feature index, represents the input variable dimension, Indicates the input characteristic parameters, including temperature, relative humidity, wind speed, wind direction, air pressure, precipitation, ice thickness, ice weight and conductor surface temperature. Represents the core statistics of ice cover data preprocessing, Indicates the collected data, that is, the corresponding features in the time series Single sampling data of Collected data outside the scope Recorded as outlier data and deleted; Use interpolation to fill in the short-term missing and deleted data. The average value of the sample points in the adjacent time periods is padded, which can be expressed as: , in, represents the calculated average value, Representing variables In the missing moment Previous The observation values ​​at consecutive moments, Indicates the current missing time The previous chapter The sensor data value at a time point, Indicates the current missing time The sensor data value at the first time point before (immediately adjacent to the missing point), Indicates the sliding window length.

[0032] S2: Dynamically optimize the variational mode decomposition parameters of ice cover data through genetic algorithm to determine the mode number and penalty factor.

[0033] It should be noted that after data cleaning and interpolation, the ice thickness data is complex. Specifically, it is a typical non-stationary, nonlinear signal that is a mixture of fluctuation components at different time scales, such as high-frequency noise caused by instantaneous weather changes (such as gusts and short-term precipitation), medium-term daily variations related to the day-night temperature difference, and long-term variation trends determined by seasonal changes. VMD (variational mode decomposition) is used to perform multi-scale decomposition of the ice thickness series. The original signal can be adaptively decomposed into a series of relatively stable intrinsic mode functions (IMFs), each of which corresponds to a specific time scale or oscillation mode. VMD transforms the signal decomposition problem into a variational problem, estimating each mode function and its center frequency, which can be expressed as: , in, Indicates the decomposition The number of modes, represents the center frequency of the mode function, Represents a group and The minimum value of the objective function, represents the time derivative, represents the Dirac impulse function, represents the Hilbert transform kernel, represents the imaginary unit, represents the convolution operation, express paradigm, represents the modal number, represents the modal index, represents the time variable, represents the base of natural logarithms; Satisfy the constraints: , in, represents the original ice thickness time series; by introducing Lagrange multipliers and penalty factors, the constrained optimization problem can be transformed into an unconstrained problem, which can be expressed as: , in, represents the penalty factor, represents the Lagrange multiplier, which controls the bandwidth constraint of the decomposition; The value of directly controls the bandwidth and number of modal components; the VMD algorithm iteratively solves the above optimization problem through the alternating direction multiplier method (ADMM) and finally obtains the IMF and the corresponding center frequency.

[0034] Furthermore, the modal number and penalty factor are determined including the modal number and penalty factor ; in, The value determines the IMF components obtained by signal decomposition. The value affects the bandwidth constraint of the decomposition; e.g. Figure 2 As shown, the genetic algorithm is used for automatic optimization, including: Parameter encoding: The two key parameters of VMD are and Encoded as chromosomes, Integer encoding is used and the value range is ; Real number coding is used, and the value range is , each chromosome is represented as a binary ; Initial population generation: Random generation The initial population consists of chromosomes ; For each chromosome, its Value and The values ​​are randomly generated and expressed as: , , in, yes A uniform random number between Indicates rounding down. Represents the randomly generated modal number, represents the randomly generated penalty factor, represents the total number of samples; Fitness evaluation: For each chromosome , use the corresponding parameters to perform VMD decomposition on the historical data of ice thickness; design the fitness function Expressed as: , in, Represents the fitness function, through Item, guides the genetic algorithm to select stable modes with clear physical meaning, Represents the reconstruction error, and the calculation formula is , is the original data, The data reconstructed after VMD decomposition, the smaller the MSE value, the higher the fidelity of the decomposition; represents the modal stationarity penalty, The first The Hurst index of the modal component is used to measure the long-range correlation of the time series; the ideal stationary mode Hurst index should be close to 0.5; by calculating The sum of the absolute values ​​of , penalizing those that still retain strong trends after decomposition ( Close to 1), strong counter-trend ( The smaller the value of this item, the better the overall stability of the decomposed modes. represents the fitness weight coefficient, which is used to balance the two optimization objectives of reconstruction error and modal stability; Selection operation: Roulette wheel selection method is used to select outstanding individuals to enter the next generation; The probability of being selected is proportional to the fitness, expressed as: , in, express The probability of being selected, express fitness value; adopting the elite retention strategy, the two individuals with the highest fitness in the current population are directly retained to the next generation to ensure that excellent genes will not be lost during the evolution process; Crossover operation: Pair the selected individuals to generate offspring individuals; use single-point crossover method, and set the crossover probability to 0.8; for a pair of parent individuals and , generating two offspring individuals and , expressed as: , , , in, 、 represents the modal number of parent chromosomes 1 and 2, 、 represents the penalty factor for parent chromosomes 1 and 2, 、 represents the modal number of offspring chromosomes 1 and 2, 、 represents the penalty factor for offspring chromosomes 1 and 2, represents the cross-weight coefficient, which is For a random number between Value, inherited from both parents with a probability of 50%; Mutation operation: perform mutation operation on offspring individuals, and the mutation probability is set to 0.1; Value, within its range Generate a new value randomly within Value, Gaussian mutation, expressed as: , in, Represents the value after mutation, is a Gaussian random number with mean 0 and standard deviation 1. It is a variable step length, with an initial value of 500 and decreases linearly with the increase of the number of iterations; Population update and termination conditions: elite individuals, offspring individuals generated by crossover and mutation are combined into a new generation of population; fitness evaluation to mutation operation are repeated until the termination conditions are met, including reaching the maximum number of iterations of 100 and no significant decrease in the optimal fitness for 10 consecutive generations (the decrease is less than 0.0001).

[0035] Furthermore, after optimizing the VMD parameters using a genetic algorithm, the historical ice thickness data was decomposed into IMF components. Each IMF component represents a different frequency feature in the ice thickness data: high-frequency IMFs typically reflect noise and short-term fluctuations, medium-frequency IMFs reflect diurnal variations and changes in weather systems, and low-frequency IMFs reflect seasonal trends. These features, combined with the original meteorological features, form the input to the GRU-KAN (Gated Recurrent Unit-Kolmogorov-Arnold Network) model. For real-time predictions, new meteorological data first undergoes the same VMD decomposition process, using optimized parameters to decompose the historical ice thickness data into multiple IMF components. This is then combined with the latest meteorological data and input into the trained GRU-KAN model to generate a predicted value for ice thickness.

[0036] S3: A gated recurrent unit is used to extract the temporal features of ice cover, which are then fed into a Kolmogorov-Arnold network (KAN) for nonlinear mapping, and the dual-path outputs are fused through a residual connection.

[0037] It should be noted that the extraction of ice cover time series features using gated recurrent units includes constructing time series samples, setting the input time window length to T (e.g., 24 hours), setting the prediction time length to P (e.g., 6 hours), and using a sliding window method to construct input-output sample pairs; The input KAN performs nonlinear mapping and fuses the dual-path output through residual connections. The input layer receives the output of the GRU layer and constructs a two-layer KAN network. The neuron output is used to predict the ice thickness value. The KAN is based on the Kolmogorov-Arnold representation theorem and is represented by a finite composite of single-variable continuous functions as follows: , in, express Continuous function on the dimensional input space, represents the inner univariate continuous function, represents the outer univariate continuous function, Indicates the input variables, represents the input variable dimension, Represents the feature vector position index corresponding to the sensor, represents the nonlinear layer neuron index, represents the first eigenvalue of the input, Indicates the input eigenvalues; Among them, KAN includes an inner transformation layer, an aggregation layer, and an outer transformation layer; Among them, the inner transformation layer maps each input dimension to a new feature space, introduces nonlinear transformation, and processes the input independently as follows: , in, represents the number of spline basis functions, represents the trainable parameters, Indicates the B-spline basis functions, represents the activation function of the inner network of KAN, represents the basis function index, represents the output neuron index, represents the input feature index, represents the input feature vector; Among them, the aggregation layer represents the combination of features of different input dimensions, and the adaptive adjustment of feature importance is achieved through weights as follows: , in, represents the learnable weight parameters, Indicates the The output of the aggregation unit, represents the input features, Represents the input features The learnable nonlinear transformations performed; Among them, the outer transformation layer includes nonlinear transformation of the aggregation results to enhance the expression ability of the network, and maps it to the target space as follows: , in, Indicates the final output result. Indicates the KAN network The nonlinear transformation output of each neuron.

[0038] Furthermore, the GRU (Gated Recurrent Unit) layer design includes an input layer size of dimensional, contains 128 hidden layer neurons, and its operating mechanism is as follows Figure 3 As shown; Update Gate and reset gate It determines how much historical information to retain and how much new information to accept. The calculation formulas are expressed as follows: , , in, represents the update gate, Represents the reset gate, 、 are the input weight matrices of the update gate and reset gate respectively, Indicates the input information at the current moment. 、 are the corresponding cyclic weight matrices, Indicates the state at the previous moment. represent activation function, Represents the activation function; through Activation function generates candidate hidden states , expressed as: , in, represents the Hadamard product (element-wise product), is the input weight matrix, is the cyclic weight matrix; get the final state , expressed as: , in, Indicates the final state, Indicates hidden state.

[0039] Furthermore, hybrid architectures connect Figure 4As shown in the figure, the features of the GRU layer output after dimension reduction by the fully connected layer are input into the KAN network, and residual connections are added to improve the model performance; GRU is used to process time series data, extract time series features, and KAN is used for feature conversion and final prediction.

[0040] S4: Deploy the trained VMD-GRU-KAN (Variational Mode Decomposition-Gated Recurrent Unit-Kolmogorov-Arnold Network) model to an industrial-grade AI edge computing unit, utilizing 2TOPS of computing power to perform real-time inference.

[0041] It should be noted that model training includes: Loss function selection: The mean square error (MSE) is used as the main loss function, expressed as: , in, Represents the predicted data, is the length of the time series; and introduce The regularization term is used to control model complexity and prevent overfitting, and is expressed as: , in, is the data loss of the model, which represents the error between the model’s predicted value and the true label; is the regularization parameter, which is used to control the strength of regularization; is the weight vector of The square of the norm, expressed as the sum of the squares of the parameters in the weight vector, represents the regularization term; In model training, the Adam optimizer is used as the optimization algorithm; the initial learning rate is set to 0.001, which can balance the convergence speed and stability in most cases; A learning rate decay strategy is adopted to improve the convergence performance and generalization ability of the model by gradually reducing the learning rate; after each epoch, if the loss of the validation set does not decrease significantly, the learning rate will be multiplied by a decay factor ; The update formula of learning rate is expressed as: , in, Indicates the The learning rate of the round, is the initial learning rate, is the period of learning rate decay, represents the floor function, Indicates the current training round number, Indicates the number of complete decay cycles; The batch size is set to 64 to balance computational efficiency and model stability, and the number of training epochs is set to 200 to ensure that the model can fully learn the features in the data; Model performance evaluation uses cross-validation, which divides the dataset into several folds and performs training and validation on each fold to ensure consistent performance of the model on different data subsets;

[0042] To prevent overfitting, an early stopping mechanism is applied; if the performance indicators on the validation set (such as validation loss or validation accuracy) do not improve in 5 consecutive training rounds, the training is stopped.

[0043] Furthermore, the trained VMD-GRU-KAN model is deployed to an industrial-grade AI edge computing unit, and real-time inference is performed using 2TOPS computing power. This includes using an industrial-grade embedded AI chip and memory configuration for real-time inference. The trained VMD-GRU-KAN hybrid model is quantized and compressed to adapt to the memory limitations of the edge computing unit.

[0044] S5: Trigger the sliding window incremental update mechanism. When the prediction error exceeds the threshold, it automatically collects new data and adjusts the output layer weights to keep the model dynamically adaptable.

[0045] It should be noted that the triggering of the sliding window incremental update mechanism includes immediate triggering when the real-time prediction error exceeds the threshold for three consecutive times; collecting 8 hours of ice monitoring data, and automatically shortening the update cycle to 4 hours in extreme weather conditions; applying The criterion automatically filters outliers; freezes the pre-trained weights of GRU-KAN and uses sliding window data to adjust the output fully connected layer parameters.

[0046] Furthermore, the system is triggered immediately when the real-time prediction error exceeds the threshold (MAE>2.0mm) three times in a row; extracts ice cover monitoring data from the last eight hours, and automatically shortens the update cycle to four hours (when the error threshold is continuously triggered) in extreme weather conditions (approximately 5,760 samples, 1Hz sampling); applies the 3σ criterion to automatically filter outliers (anomaly processing rate ≤3.5%); retains the pre-trained weights of GRU-KAN (the GRU layer and the main body of the KAN network are not updated); uses sliding window data to fine-tune the parameters of the fully connected layer (learning rate 0.0001); incremental learning memory usage is ≤15MB (30% of the total memory 50MB); and a single update takes ≤5 minutes (guaranteed by 2TOPS of edge chip computing power).

[0047] Furthermore, the prediction results are evaluated by calculating the root mean square error (RMS), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (CDR). Completed; the calculation formulas are expressed as: , , , , in, represents the time observation value of the i-th sample point, Represents the model's predicted value for the i-th sample point, is the arithmetic mean of all observations, represents the input variable dimension, represents the output feature index, represents the root mean square error, represents the mean absolute error, represents the mean absolute percentage error, represents the coefficient of determination.

[0048] Example 2, reference Figure 5 and Figure 6 , which is an embodiment of the present invention, provides an intelligent monitoring and early warning method for ice thickness of distribution lines. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0049] The data of the 110kV distribution line of the power grid was selected as a sample for verification. The area is located in the plateau and mountainous areas, with complex winter climate and frequent icing.

[0050] To ensure the real-time operation of the icing prediction model, a cluster of intelligent monitoring terminals (38 devices in total) was deployed at key nodes of the 110kV power grid. These devices include: Sensing layer: Each device is equipped with a 7-dimensional meteorological sensor (temperature / humidity / wind speed / wind direction / air pressure / precipitation / conductor temperature), and an ice thickness detection unit (resolution 0.1mm) is installed; Edge layer: Equipped with NVIDIA (NVIDIA Jetson Orin NX embedded artificial intelligence computing platform) edge computing unit (computing power 2TOPS, memory 50MB), inference latency was measured in a low temperature environment of -25°C Seconds / time, meeting the real-time requirements of the model; Communication layer: The main channel uses a 5G industrial module (uplink rate of 120Mbps), and the backup channel is configured with a LoRa relay network (transmission distance of 12.3km in mountainous areas), with a 100% success rate for dual-channel switching; Raw data is averaged and aggregated over 10-minute windows to generate hourly features for input into the model. A portion of historical data is extracted as the foundation for the model. The data spans eight complete freezing seasons, from December 2012 to February 2020, and is collected over a 24-hour period to capture subtle changes in ice formation. The dataset includes 652 days of initial samples. After rigorous screening, 625 days of valid samples were retained, as shown in Table 1. The data feature dimensions cover seven indicators: temperature, relative humidity, wind speed, wind direction, air pressure, precipitation, and conductor surface temperature. The target variable is ice thickness. Statistics show that the temperature range is between -10°C and 5°C, the relative humidity is between 60% and 95%, the wind speed is between 0 and 15 m / s, and the ice thickness varies from 0 to 25 mm. In the data preprocessing stage, we use The data were cleaned strictly according to the criteria; 37 outliers were processed mainly by interpolating the average values ​​of sample points in adjacent time periods, and the outlier removal rate was 3.46%; the genetic algorithm optimized the VMD parameters and finally obtained the optimal modal number. is 4, the optimal penalty factor is 142.1; The VMD decomposition process is as follows Figure 5 As shown, the multi-scale characteristics of the ice thickness time series are revealed; the high-frequency IMF reflects short-term noise fluctuations, the medium-frequency IMF captures the daily variation characteristics, and the low-frequency IMF shows the seasonal trend; The model is constructed with an innovative architecture design. The input layer uses a 20-day sliding window, the GRU layer is configured with 128 hidden neurons, the KAN network is designed with two layers, containing 64 and 32 neurons respectively, and the output layer is a single neuron, which directly predicts the ice thickness after 6 hours. The data set is based on The proportion is divided into training set, validation set and test set; The Adam (Adaptive Moment Estimation) optimizer was used in the training process, with an initial learning rate of 0.0001 and an exponential decay strategy. The batch size was set to 64, the training rounds were 120, and an early stopping mechanism was set to terminate the training when there was no performance improvement for 5 consecutive rounds. In order to fully verify the performance of the model, LSTM (Long Short-Term Memory Network), GRU, TCN (Temporal Convolutional Network), LSTM-KAN (Long Short-Term Memory Network-Kolmogorov-Arnold Network Hybrid Model), GRU-KAN, and VMD-GRU models were selected for comparison with the model of the present invention. The prediction curves of the test sets of each model are shown in the following figure. Figure 6 The comparison results of the evaluation indicators of the prediction results of each model are shown in Table 2; Table 1 Operational data collected by monitoring equipment during the 625-day ice-cover period

[0051] Table 2 Evaluation indicators of prediction results of each model

[0052] Experimental results show that the VMD-GRU-KAN model outperforms the comparison methods in various indicators; RMSE is reduced to 1.793mm, a 40.33% reduction compared to the baseline model; MAPE is reduced to 5.761; MAE is the lowest, only 1.233mm; The model's goodness of fit reached 0.962, indicating excellent model fit. The model's superior performance stems from the multi-scale feature extraction of VMD decomposition, the long-term temporal dependencies captured by the GRU layer, and the nonlinear feature mapping of the KAN network. To verify the robustness of the model, parameter sensitivity analysis and performance testing under different meteorological conditions were conducted. The model's performance under various parameter configurations was evaluated by adjusting the number of VMD modes and the number of KAN network neurons. Furthermore, the model maintained its prediction accuracy under extreme weather conditions such as low temperature, high humidity, and strong winds. After deployment, the system is maintained in a dynamic maintenance mechanism to ensure long-term stability. Incremental learning is automatically triggered every 8 hours to fine-tune the output layer weights based on the latest 5760 samples. ) immediately initiates an emergency update; triggered 27 times during the freeze period in January 2022, making the model Stable at 0.958±0.003; execute operation and maintenance instructions through channels (412 operations per year), typical scenarios include reading device status and reading early warning data.

[0053] Example 3 is an embodiment of the present invention, which provides an intelligent monitoring and early warning system for ice thickness on distribution lines, including: The anti-icing sensor array module deploys an anti-icing packaged multi-parameter sensor array, samples data at a fixed frequency, synchronously collects multi-dimensional parameters and pre-processes the data; Decomposition optimization module, which dynamically optimizes the variational mode decomposition parameters of ice cover data through genetic algorithm to determine the number of modes and penalty factors; The residual fusion module uses a gated recurrent unit to extract ice cover temporal features, inputs them into a KAN to perform nonlinear mapping, and fuses the dual-path outputs through a residual connection. The edge curing module deploys the trained VMD-GRU-KAN model to an industrial-grade AI edge computing unit, utilizing 2TOPS computing power to perform real-time inference. The dynamic fine-tuning module triggers the sliding window incremental update mechanism. When the prediction error exceeds the threshold, it automatically collects new data and adjusts the output layer weights to keep the model dynamically adaptable.

[0054] This embodiment also provides an electronic device, which is suitable for the case of an intelligent monitoring and early warning method for the thickness of ice covering on distribution lines, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the intelligent monitoring and early warning method for the thickness of ice covering on distribution lines proposed in the above embodiment.

[0055] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for intelligent monitoring and early warning of ice thickness on distribution lines proposed in the above embodiment is implemented.

[0056] The storage medium proposed in this embodiment and the method for realizing intelligent monitoring and early warning of ice thickness on distribution lines proposed in the above embodiment belong to the same inventive concept. For technical details not fully described in this embodiment, please refer to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0057] Through the above description of the embodiments, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware. Of course, it can also be implemented using hardware, but in many cases the former is a more preferred embodiment. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk, or optical disk, and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0058] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An intelligent monitoring and early warning method for ice thickness on distribution lines, characterized by: include, Deploy an ice-resistant encapsulated multi-parameter sensor array, sample data at a fixed frequency, synchronously collect multi-dimensional parameters and pre-process the data; The variational mode decomposition parameters of ice cover data are dynamically optimized by genetic algorithm to determine the mode number and penalty factor. The variational modal decomposition parameters include the modal number and penalty factor , the mode number is used to determine the number of intrinsic mode functions obtained by signal decomposition, and the penalty factor is used to control the bandwidth constraint of each mode; The genetic algorithm is implemented through parameter encoding, fitness function evaluation and crossover mutation operation. and Search and obtain decomposition parameters; A gated recurrent unit is used to extract ice cover temporal features, which are then fed into a KAN to perform nonlinear mapping, and the dual-path outputs are fused through a residual connection. Deploy the trained VMD-GRU-KAN model to an industrial-grade AI edge computing unit, utilizing 2TOPS computing power to perform real-time inference. The sliding window incremental update mechanism is triggered. When the prediction error exceeds the threshold, new data is automatically collected and the output layer weights are adjusted to keep the model dynamically adaptable.

2. The intelligent monitoring and early warning method for ice thickness on distribution lines according to claim 1, characterized in that: The deployment of the ice-resistant packaged multi-parameter sensor array includes deploying a multi-dimensional meteorological sensor and installing an ice thickness detection unit; Among them, the multi-dimensional meteorological sensors include temperature, relative humidity, wind speed, wind direction, air pressure, precipitation and conductor surface temperature sensors; Among them, the installation of ice thickness detection units includes using the laser ranging principle, installed below the conductor suspension insulator string, covering the key nodes of the entire line; The multidimensional parameters include temperature, relative humidity, wind speed, wind direction, air pressure, precipitation, ice thickness, ice weight and conductor surface temperature.

3. The intelligent monitoring and early warning method for ice thickness on distribution lines according to claim 2, characterized in that: The dynamic optimization of the variational mode decomposition parameters of ice cover data by genetic algorithm includes adaptively decomposing the original data into intrinsic mode functions, which are expressed as: ， in, Indicates the decomposition A modal function, represents the center frequency of the mode function, Represents a group and The minimum value of the objective function, represents the time derivative, represents the Dirac impulse function, represents the Hilbert transform kernel, represents the imaginary unit, represents the convolution operation, express paradigm, represents the modal number, represents the modal index, represents the time variable, Represents the base of natural logarithms.

4. The intelligent monitoring and early warning method for ice thickness on distribution lines according to claim 3, characterized in that: The determination of the modal number and the penalty factor includes the modal number and penalty factor ; in, The value determines the IMF components obtained by signal decomposition. The value affects the bandwidth constraint of the decomposition.

5. The intelligent monitoring and early warning method for ice thickness on distribution lines according to claim 4, characterized in that: The input KAN performs nonlinear mapping and fuses the dual-path output through residual connections, including the input layer receiving the GRU layer output and constructing a two-layer KAN network, using neuron output to predict the ice thickness value; KAN is based on the Kolmogorov-Arnold representation theorem through a finite composite representation of a single variable continuous function as follows: ， in, express Continuous function on the dimensional input space, represents the inner univariate continuous function, represents the outer univariate continuous function, Indicates the input variables, represents the input variable dimension, Represents the feature vector position index corresponding to the sensor, represents the nonlinear layer neuron index, represents the first eigenvalue of the input, Indicates the input eigenvalues; Among them, KAN includes an inner transformation layer, an aggregation layer, and an outer transformation layer; Among them, the inner transformation layer maps each input dimension to a new feature space, introduces nonlinear transformation, and processes the input independently as follows: ， in, represents the number of spline basis functions, represents the trainable parameters, Indicates the B-spline basis functions, represents the activation function of the inner network of KAN, represents the basis function index, represents the output feature index, represents the input feature index, represents the input feature vector; Among them, the aggregation layer represents the combination of features of different input dimensions, and the adaptive adjustment of feature importance is achieved through weights as follows: ， in, represents the learnable weight parameters, Indicates the The output of the aggregation unit, represents the input features, Represents the input features The learnable nonlinear transformations performed; Among them, the outer transformation layer includes nonlinear transformation of the aggregation results to enhance the expression ability of the network, and maps it to the target space as follows: ， in, Indicates the final output result. Indicates the KAN network The nonlinear transformation output of each neuron.

6. The intelligent monitoring and early warning method for ice thickness on distribution lines according to claim 5, characterized in that: The method of deploying the trained VMD-GRU-KAN model to an industrial-grade AI edge computing unit and performing real-time inference using 2TOPS computing power includes using an industrial-grade embedded AI chip, configuring memory for real-time inference, quantizing and compressing the trained VMD-GRU-KAN hybrid model, and adapting it to the memory limitations of the edge computing unit.

7. The intelligent monitoring and early warning method for ice thickness on distribution lines according to claim 6, characterized in that: The trigger sliding window incremental update mechanism includes immediate triggering when the real-time prediction error exceeds the threshold for three consecutive times; collecting 8 hours of ice monitoring data, and automatically shortening the update cycle to 4 hours in extreme weather conditions; applying The criterion automatically filters outliers; freezes the pre-trained weights of GRU-KAN and uses sliding window data to adjust the output fully connected layer parameters.

8. An intelligent monitoring and early warning system for ice thickness on distribution lines, applying the intelligent monitoring and early warning method for ice thickness on distribution lines according to any one of claims 1 to 7, characterized in that: include: The anti-icing sensor array module deploys an anti-icing packaged multi-parameter sensor array, samples data at a fixed frequency, synchronously collects multi-dimensional parameters and pre-processes the data; Decomposition optimization module, which dynamically optimizes the variational mode decomposition parameters of ice cover data through genetic algorithm to determine the number of modes and penalty factors; The residual fusion module uses a gated recurrent unit to extract ice cover temporal features, inputs them into a KAN to perform nonlinear mapping, and fuses the dual-path outputs through a residual connection. The edge curing module deploys the trained VMD-GRU-KAN model to an industrial-grade AI edge computing unit, utilizing 2TOPS computing power to perform real-time inference. The dynamic fine-tuning module triggers the sliding window incremental update mechanism. When the prediction error exceeds the threshold, it automatically collects new data and adjusts the output layer weights to keep the model dynamically adaptable.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for intelligent monitoring and early warning of ice thickness on distribution lines according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for intelligent monitoring and early warning of ice thickness on distribution lines according to any one of claims 1 to 7 are implemented.

Images