Ground wire icing phase state classification prediction method based on feature optimization and neural network
By establishing a correlation model of meteorological environmental conditions and ground wire ice-covered phase state based on feature preference and neural network, the problem of experience dependent on judgment of ground wire ice-covered phase state in the prior art is solved, and accurate prediction of ice-covered phase state and stable and safe operation of the power grid is achieved.
Patent Information
- Application Number
- CN202510307090.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-15
- Publication Date
- 2025-06-24
AI Technical Summary
The existing technology lacks effective methods to uniformly judge and predict the ice-covered phase of the ground wire, resulting in relying on experience and wasting manpower, material resources and financial resources.
The classification prediction method for ground wire ice-covered phase state based on feature preference and neural network is adopted. By obtaining the correlation data of meteorological conditions and ground wire ice-covered phase state, the preferred features are selected and modeled using artificial neural networks, the sample imbalance problem is processed, and the correlation model of meteorological environmental conditions and ground wire ice-covered phase state is established.
Accurate classification prediction of ice-covered phase states is realized, which significantly improves the spatial and temporal prediction accuracy of ice-covered thickness, provides real-time and dynamic physical data support for the power grid's anti-ice melting decisions, and ensures the stable and safe operation of the power grid.
Smart Images
Figure CN120197079A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of transmission line icing and snow covering, and particularly relates to a method for classifying and predicting the icing phase state of overhead conductors based on feature optimization and neural network. Background Art
[0002] The icing phenomenon in nature presents in various forms on the surface of objects. This ice crystal condensation process usually does not pose a hazard in the natural environment and can even form amazing ice and snow wonders. However, when this natural phenomenon acts on artificial facilities such as power grids, it shows completely different destructive characteristics, especially posing a significant threat to the metal conductors of overhead transmission lines. The potential safety hazards caused by ice accumulation are mainly reflected in two dimensions: physical mechanics and power operation:
[0003] From the perspective of mechanical mechanics, the continuously accumulating ice layer will significantly increase the vertical load of the conductor system, and at the same time, the increased conductor cross-section will generate a stronger horizontal load under the action of wind. More seriously, when the ice layer distribution shows regional differences or non-synchronous shedding occurs, the conductor system will bear a severe dynamic impact, thereby inducing a severe lateral swing phenomenon (commonly known as conductor galloping).
[0004] In terms of electrical performance, the ice crystals formed on the surface of insulators may cause the breakdown of the insulation between poles and trigger abnormal discharge phenomena. Moreover, the instantaneous reduction of the distance between adjacent conductors during conductor galloping is more likely to cause short-circuit faults due to arc bridging. Such icing disasters not only cause direct economic losses of hundreds of millions of yuan, but may also lead to regional power outages, having a chain impact on social and economic activities and residents' daily lives.
[0005] Currently, the judgment of the icing phase state of conductors is usually based on experience, without a unified standard and method, wasting manpower, material resources and financial resources. Therefore, it has practical value to deeply study the mapping relationship between meteorological environmental conditions and icing phase states and give a unified model.
[0006] It should be noted that the information disclosed in the above background art section is only used for understanding the background of the present application, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention
[0007] The main object of the present invention is to overcome the defects existing in the above background art, and provide a method for classifying and predicting the icing phase state of overhead conductors based on feature optimization and neural network.
[0008] To achieve the above object, the present invention adopts the following technical solutions:
[0009] A method for classifying and predicting the icing phase state of overhead conductors based on feature optimization and neural network, comprising the following steps:
[0010] S1. Obtain the correlation data between a series of meteorological conditions and the ice phase state of the ground wire, and select the optimal features according to the correlation coefficient.
[0011] S2. Judge whether the sample sizes of each category are balanced. If balanced, use an artificial neural network to build a model; if not, go to step S3.
[0012] S3. Select method one or method two to handle the imbalance of sample data.
[0013] S4. Method one: Use the sample variable weight method to assign weights to samples of each category according to the sample sizes of each category.
[0014] S5. Method two: Use the SMOTE algorithm to generate a part of artificial samples to make the sample sizes of each fault category equal.
[0015] S6. Use the backpropagation algorithm to train the artificial neural network.
[0016] S7. Obtain the correlation model between the meteorological environment conditions and the ice phase state of the ground wire.
[0017] Further, in step S1, the selection of the optimal features is carried out by calculating the correlation coefficient between the meteorological environment conditions and the ice phase state of the ground wire, specifically including: first calculating the correlation coefficient between each meteorological parameter and the ice phase state; eliminating the features with correlation coefficients lower than the preset threshold to retain the meteorological parameters with higher correlation with the ice phase state as the optimal features, and finally determining that the optimal features are temperature, humidity, and wind speed.
[0018] Further, in step S2, judging whether the sample sizes of each category are balanced is specifically done by comparing the numbers of samples of each category. If the differences in the numbers of samples of different categories are too large, perform sample data imbalance processing to ensure that the classification effect of the artificial neural network is not affected by sample imbalance.
[0019] Further, in step S4, the sample variable weight method adjusts the weights of samples of each category according to the differences in the numbers of samples of each category during the training process of the artificial neural network, specifically by increasing the weights of the minority category samples to balance the influence of samples of each category on model training.
[0020] Further, in step S5, the SMOTE algorithm generates artificial samples through the following steps:
[0021] Count the sample sizes of each category and determine the maximum sample size of a single category among all categories.
[0022] Perform feature transformation on the original features and calculate the number of artificial samples that need to be generated for each category according to the maximum sample size.
[0023] From samples of the same category, find several nearest neighbors of the current sample, randomly select a neighboring sample, calculate the difference vector between it and the current sample, and generate new samples based on the difference vector to balance the number of samples in each category.
[0024] Further, in step S6, the artificial neural network consists of an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is determined by the dimension of the input feature vector. The output layer is the ice accretion density, and the hidden layer processes the input data through an activation function and transmits it to the output layer.
[0025] Further, in step S6, the training of the artificial neural network adopts the backpropagation algorithm, which specifically includes: first, calculating the outputs of neurons in each layer through forward propagation, then calculating the error of the output layer, and backpropagating the error to the hidden layer. Finally, adjust the connection weights and biases of the neural network according to the errors of each layer to minimize the cost function.
[0026] Further, the error calculation of the backpropagation algorithm includes: first calculating the error of the output layer, then successively calculating the errors of the hidden layers in reverse, and finally using the errors of each layer to calculate the partial derivatives of the cost function with respect to the connection weights and biases, and adjusting the neural network parameters according to the partial derivatives.
[0027] Further, in step S6, during the training process of the artificial neural network, regularization processing is adopted. By adding a regularization term to the cost function, the model complexity is reduced, the generalization ability of the model is improved, and overfitting is prevented.
[0028] Further, in step S6, the adjustment of the neural network parameters is carried out according to the partial derivatives of the cost function. Specifically, the gradient descent method or the momentum gradient descent method is used for optimization. By iteratively adjusting the connection weights and biases, the value of the cost function is gradually reduced until the model converges.
[0029] The present invention has the following beneficial effects:
[0030] In view of the deficiencies of the prior art, the present invention proposes a method for classifying and predicting the icing phase state of ground wires and conductors based on feature optimization and neural networks, and establishes a correlation model between meteorological environmental conditions and the icing phase state of ground wires and conductors based on an artificial neural network. This method takes meteorological environmental parameters as inputs and the icing phase state as the output, and realizes accurate classification and prediction of the icing phase state by analyzing the non-linear coupling relationship between meteorological parameters and the icing density of transmission lines. The model uses multivariate meteorological data such as wind speed, temperature, and humidity as the input layer, and constructs a dynamic correlation model between meteorological elements and icing parameters in combination with the multi-layer perceptron architecture. This technical system significantly improves the spatio-temporal prediction accuracy of icing thickness, provides real-time dynamic physical field data support for the anti-icing and de-icing decision-making of the power grid, further provides a scientific basis for the stable and safe operation of the power grid, and has important engineering application value for the safe operation and maintenance of transmission corridors under extreme meteorological conditions.
[0031] Other beneficial effects in the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 is a flowchart of the method for classifying and predicting the icing phase state of ground wires and conductors based on feature optimization and neural networks of the present invention.
[0033] Figure 2 is a schematic diagram of a three-layer artificial neural network according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0034] The following provides a detailed description of the embodiments of the present invention. It should be emphasized that the following description is merely exemplary and is not intended to limit the scope of the present invention and its applications.
[0035] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present invention, "a plurality" means two or more unless otherwise specifically defined.
[0036] Refer to Figure 1 , an embodiment of the present invention provides a method for classifying and predicting the icing phase state of ground wires and conductors based on feature optimization and neural networks, including the following steps:
[0037] S1. Obtain a series of correlation data between meteorological conditions and the icing phase state of ground wires and conductors, and select preferred features according to the correlation coefficient.
[0038] In a preferred embodiment, the selection of preferred features is carried out by calculating the correlation coefficient between the meteorological environmental conditions and the icing phase of the ground wire, specifically including: first calculating the correlation coefficient between each meteorological parameter and the icing phase; eliminating features with correlation coefficients lower than a preset threshold value, so as to retain meteorological parameters with a higher correlation with the icing phase as preferred features, and finally determining the preferred features as temperature, humidity and wind speed.
[0039] S2. Determine whether the sample size of each category is balanced. If balanced, use artificial neural network to build a model; if unbalanced, proceed to step S3.
[0040] In a preferred embodiment, it is determined whether the sample sizes of each category are balanced, specifically by comparing the number of samples of each category. If the difference in the number of samples of different categories is too large, the sample data imbalance processing is performed to ensure that the classification effect of the artificial neural network is not affected by the sample imbalance.
[0041] S3. Select method 1 or method 2 to handle sample data imbalance.
[0042] S4. Method 1: Using the sample variable weight method, weights are assigned to samples of each category according to the number of samples of each category. In a preferred embodiment, in step S4, the sample variable weight method adjusts the weights of samples of each category according to the difference in the number of samples of each category during the training of the artificial neural network, specifically increasing the weights of samples of the minority category to balance the impact of samples of each category on model training.
[0043] S5. Method 2: Use the SMOTE algorithm to generate a part of artificial samples so that the sample size of each fault category is equal; in a preferred embodiment, in step S5, the SMOTE algorithm generates artificial samples through the following steps: count the sample size of each category and determine the maximum sample size of a single category among all categories; perform feature transformation on the original features and calculate the number of artificial samples to be generated for each category based on the maximum sample size; find several nearest samples of the current sample from samples of the same category, randomly select a neighboring sample, calculate the differential vector between it and the current sample, and generate a new sample based on the differential vector so that the sample size of each category is balanced.
[0044] S6. Training artificial neural networks using back-propagation algorithm.
[0045] In a preferred embodiment, the artificial neural network consists of an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is determined by the dimension of the input feature vector. The output layer is the ice accretion density, and the hidden layer processes the input data through an activation function and transmits it to the output layer. The training of the artificial neural network uses the backpropagation algorithm, which specifically includes: first, calculating the outputs of the neurons in each layer through forward propagation, then calculating the error of the output layer, and backpropagating the error to the hidden layer. Finally, adjusting the connection weights and biases of the neural network according to the errors of each layer to minimize the cost function. The error calculation of the backpropagation algorithm includes: first calculating the error of the output layer, then sequentially calculating the errors of the hidden layers in reverse, and finally using the errors of each layer to calculate the partial derivatives of the cost function with respect to the connection weights and biases, and adjusting the neural network parameters according to the partial derivatives. During the training process of the artificial neural network, regularization processing is adopted. By adding a regularization term to the cost function, the model complexity is reduced, the generalization ability of the model is improved, and overfitting is prevented. In addition, the adjustment of the neural network parameters is carried out according to the partial derivatives of the cost function, and specifically, the gradient descent method or the momentum gradient descent method is used for optimization. By iteratively adjusting the connection weights and biases, the value of the cost function is gradually reduced until the model converges.
[0046] S7. Obtain the correlation model between meteorological environmental conditions and the ice accretion phase state of the conductor and ground wire.
[0047] The present invention proposes a method for classifying and predicting the ice accretion phase state of a conductor and ground wire based on feature optimization and a neural network. By constructing a correlation model between meteorological environmental conditions and the ice accretion phase state of the conductor and ground wire based on an artificial neural network, combined with feature optimization and sample data imbalance processing methods, it effectively solves the deficiencies in the prior art that the judgment of the ice accretion phase state relies on experience and lacks a unified standard and method. This method uses multi-source meteorological data such as wind speed, temperature, and humidity as inputs, and adopts a multi-layer perceptron architecture to analyze the non-linear coupling relationship between meteorological parameters and the ice accretion density of the transmission line, realizing accurate classification and prediction of the ice accretion phase state. By using the sample variable weight method or the SMOTE algorithm to process the sample imbalance problem, the classification effect and generalization ability of the model are further improved. In addition, this technical system significantly improves the spatio-temporal prediction accuracy of the ice accretion thickness, provides real-time dynamic physical field data support for the ice prevention and melting decision-making of the power grid, and can provide a scientific basis for the safe operation and maintenance of the transmission corridor under extreme meteorological conditions. It has important engineering application value, effectively avoids the waste of manpower, material resources, and financial resources caused by empirical judgment in traditional methods, and provides a strong guarantee for the stable and safe operation of the power grid.
[0048] The following further describes specific embodiments of the present invention and their algorithm examples.
[0049] S1. Obtain a series of correlation data between meteorological conditions and the ice accretion phase state of the conductor and ground wire, and select preferred features according to the correlation coefficient.
[0050] S2. Determine whether the sample sizes of all categories are balanced. If they are balanced, use an artificial neural network for modeling; if not, proceed to step S3.
[0051] S3. Select a method for handling unbalanced sample data.
[0052] S4. Method 1: Use the sample variable weighting method to assign weights to samples of each category according to the sample sizes of each category.
[0053] S5. Method 2: Use the SMOTE algorithm to generate a part of artificial samples to make the sample sizes of each fault category equal.
[0054] S6. Use the backpropagation algorithm to train the artificial neural network.
[0055] S7. Obtain the correlation model between meteorological environmental conditions and the ice phase state of the ground wire.
[0056] Furthermore, in step S1, the feature selection is preferably carried out according to the following principle: First, calculate the Pearson correlation coefficient between the meteorological environmental conditions and the ice phase state of the ground wire:
[0057]
[0058] where r is the Pearson correlation coefficient, x i is the i-th meteorological parameter data, is the average value of the meteorological parameters, y i is the i-th ice density data, is the average density of the ice. In the present invention, it is assumed that the ice density is 0.8 - 0.92 g / cm 3 for glaze, the ice density is 0.25 - 0.8 g / cm 3 for mixed glaze, the ice density is 0.1 - 0.25 g / cm 3 for rime, and the ice density is 0 g / cm 3 means no ice.
[0059] Secondly, eliminate the feature quantities with low correlation coefficients (correlation coefficient |r| < 0.2), compare the correlation coefficients between the remaining meteorological parameters and the ice phase state, select the parameters more relevant to the target, and determine that the preferred features are temperature, humidity, and wind speed.
[0060] Furthermore, in step S2, whether the sample sizes of all categories are balanced means whether the sample numbers of all categories are equal. If the difference in the numbers of samples of different categories is too large, it will lead to very poor classification effect of the artificial neural network and cannot achieve the expected effect.
[0061] Further, in the S4 step, the sample variable weight method is carried out according to the following principle: in the training process of the artificial neural network, the weight of the "minority" samples is increased. For a multi-classification problem, the cost function of the artificial neural network with equal-weight samples is as follows.
[0062]
[0063] The number of layers of the artificial neural network is denoted as L, and the number of neurons in each layer is denoted as S. l , f k (x (i) ) represents the output of the i-th sample and the k-th node in the output layer. The above cost function regularizes the parameters of the neural network. λ is called the penalty factor, aiming to reduce the model complexity and improve the generalization ability of the model.
[0064] When the number of samples in each category is unbalanced and different weights need to be assigned to the samples in each category, the cost function of the artificial neural network with unequal weights becomes:
[0065]
[0066] Assume that the number of samples in category k is less than that in category (k + 1). Then, in order to reduce the bias of the artificial neural network caused by sample imbalance, the weights of the two types of samples satisfy the relationship ω k > ω k+1 .
[0067] Further, in the S5 step, the SMOTE algorithm is carried out according to the following principle: First, count the sample sizes of each category and determine the maximum sample size N of a single category among all categories. max ; Secondly, perform feature transformation on the original features and determine the number of samples that need to be artificially generated for each category according to the maximum sample size N. max Again, among the samples of the same category, find the N near nearest samples of the current sample, randomly select one of the nearest samples, and calculate its difference vector Diff with the current sample according to the following formula:
[0068] Diff = X ran -X
[0069] X represents the feature vector of the current sample, and X ran represents the feature vector of the sample randomly selected from the N near nearest samples. Generate a new sample according to the difference vector:
[0070] X new = X + ran.*Diff
[0071] where, X newIt represents the feature vector of the newly generated sample. ran is a random vector with the same dimension as the feature vector, and the numerical range of the random numbers in each dimension is (0, 1).
[0072] Furthermore, in the step S6, the artificial neural network is composed of three parts: an input layer, a hidden layer, and an output layer, as shown in the appendix Figure 2 As shown. Among them, the number of neurons in the input layer is determined by the dimension of the input feature vector. In the present invention, the input layer has three dimensions: temperature, humidity, and wind speed, and the output layer is the ice accretion density.
[0073] Furthermore, the training of the artificial neural network adopts the backpropagation algorithm. The backpropagation algorithm does not actually refer to a specific algorithm, but rather to a class of algorithms. The common feature of this class of algorithms is to first calculate the error of the last layer, and then calculate the error of the previous layer in reverse, until the second layer. Finally, the partial derivatives of the cost function with respect to the connection weights and bias amounts are calculated using the errors of each layer.
[0074] In the present invention, assuming that the total number of training samples is n, each training sample is denoted as:
[0075] (x (1) ,y (1) ),(x (2) ,y (2) ),…,(x (n) ,y (n) )
[0076] Among them, NI is the number of neurons in the input layer of the artificial neural network, which is also equal to the dimension of the feature vector, and N O is the number of neurons in the output layer of the artificial neural network, which is also equal to the number of categories.
[0077] Training the neural network is to minimize the cost function of the artificial neural network by adjusting the connection weights and bias amounts between neurons. Taking a three-layer artificial neural network as an example, given a training sample (x, y), the forward propagation process of the artificial neural network is as follows:
[0078] a (1) =x
[0079] z (2) =ω (1) a (1)
[0080]
[0081] z (3) =ω (2) a (2)
[0082] a (3) =f(z(3) )
[0083] The above formula a (1) 、a (2) and a (3) are the outputs of the neurons in the first, second, and third layers respectively, represents the bias of the hidden layer, and g and f represent the activation functions of the hidden layer and the output layer of the neural network respectively.
[0084] Taking to represent the error of the j-th neuron in the l-th layer, the process of error backpropagation is as follows:
[0085] δ (3) = a (3) - y
[0086] δ (2) = (ω (2) ) T δ (3) .* g'(z (2) )
[0087] Considering the regularization processing of the neural network parameters, and the training set is a feature matrix rather than a feature vector, at this time, it is necessary to calculate the error for the entire training set to obtain an error matrix. Assuming the use of to represent the error caused by the j-th parameter affecting the i-th activation unit in the l-th layer, the calculation of the error matrix is as follows:
[0088]
[0089] It should be noted that the above formula actually calculates the cumulative error of n training samples. After considering the regularization of the neural network parameters, there is:
[0090]
[0091] The partial derivative of the cost function with respect to the neural network parameters is:
[0092]
[0093] After obtaining the partial derivative of the cost function with respect to the neural network, the adjustment of the neural network parameters is based on the partial derivative, and there are many methods that can be used, such as the steepest descent method, the momentum gradient descent method, and so on.
[0094] The embodiment of the present invention also provides a storage medium for storing a computer program, which when executed, at least executes the method described above.
[0095] An embodiment of the present invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein, when the processor executes the computer program, it at least executes the method described above.
[0096] An embodiment of the present invention also provides a processor, which executes a computer program and at least executes the method described above.
[0097] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM, Read Only Memory), a programmable read-only memory (PROM, Programmable Read-Only Memory), an erasable programmable read-only memory (EPROM, Erasable Programmable Read-Only Memory), an electrically erasable programmable read-only memory (EEPROM, Electrically Erasable Programmable Read-Only Memory), a ferromagnetic random access memory (FRAM, Ferromagnetic Random Access Memory), a flash memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM, Compact Disc Read-Only Memory); the magnetic surface memory can be a disk memory or a tape memory. The storage medium described in the embodiments of the present invention is intended to include, but not limited to, these and any other suitable types of memories.
[0098] In several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed with each other can be through some interfaces. The indirect coupling or communication connection of devices or units can be electrical, mechanical, or other forms.
[0099] The units described as separate components above may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0100] In addition, in each embodiment of the present invention, each functional unit may be entirely integrated into one processing unit, or each unit may be separately regarded as one unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware, or in the form of a combination of hardware and software functional units.
[0101] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including the above method embodiments; and the foregoing storage medium includes: removable storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks and other various media that can store program codes.
[0102] Alternatively, if the above-mentioned integrated units of the present invention are implemented in the form of software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. And the foregoing storage medium includes: removable storage devices, ROM, RAM, magnetic disks, or optical disks and other various media that can store program codes.
[0103] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.
[0104] The features disclosed in the several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.
[0105] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.
[0106] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those skilled in the technical field to which the present invention pertains, without departing from the concept of the present invention, several equivalent substitutions or obvious modifications can be made, and as long as the performance or use is the same, they should all be regarded as falling within the protection scope of the present invention.
Claims
1. A method for classifying and predicting ice-covered ground wire phases based on feature optimization and neural network, characterized in that: The following steps are involved: S1. Obtain a series of correlation data between meteorological conditions and the ice phase of the ground wire, and select the optimal features according to the correlation coefficient; S2. Determine whether the sample size of each category is balanced. If balanced, use artificial neural network to build a model; if unbalanced, proceed to step S3; S3. Select method 1 or method 2 to handle sample data imbalance; S4. Method 1: Using the sample variable weight method, weights are assigned to samples of each category according to the sample size of each category; S5. Method 2: Use the SMOTE algorithm to generate a portion of artificial samples so that the sample size of each fault category is equal; S6. Training artificial neural network using back propagation algorithm; S7. Obtain a correlation model between meteorological environmental conditions and the ice-covered ground wire phase.
2. The method according to claim 1, characterized in that In step S1, the selection of preferred features is carried out by calculating the correlation coefficient between the meteorological environmental conditions and the icing phase of the ground wire, specifically including: first calculating the correlation coefficient between each meteorological parameter and the icing phase; eliminating the features with correlation coefficients lower than a preset threshold value, so as to retain the meteorological parameters with a higher correlation with the icing phase as the preferred features, and finally determining the preferred features as temperature, humidity and wind speed.
3. The method according to claim 1, characterized in that: In step S2, it is determined whether the sample sizes of each category are balanced, specifically by comparing the number of samples of each category. If the difference in the number of samples of different categories is too large, the sample data imbalance processing is performed to ensure that the classification effect of the artificial neural network is not affected by the sample imbalance.
4. The method according to claim 1, characterized in that: In step S4, the sample variable weight method adjusts the weight of each category of samples according to the difference in the number of samples of each category during the training phase of the artificial neural network, specifically increasing the weight of the minority category samples to balance the impact of each category of samples on the model training.
5. The method according to claim 1, characterized in that: In step S5, the SMOTE algorithm generates artificial samples through the following steps: Count the sample size of each category and determine the maximum sample size of a single category among all categories; Perform feature transformation on the original features and calculate the number of artificial samples that need to be generated for each category based on the maximum sample size; From samples of the same category, find several nearest neighboring samples of the current sample, randomly select a neighboring sample, calculate the difference vector between it and the current sample, and generate a new sample based on the difference vector so that the sample size of each category is balanced.
6. The method according to claim 1, characterized in that In step S6, the artificial neural network consists of an input layer, a hidden layer and an output layer. The number of neurons in the input layer is determined by the dimension of the input feature vector. The output layer is the ice density. The hidden layer processes the input data through an activation function and passes it to the output layer.
7. The method according to claim 1, characterized in that In step S6, the training of the artificial neural network adopts the back propagation algorithm, which specifically includes: firstly calculating the output of each layer of neurons through forward propagation, then calculating the error of the output layer, and back propagating the error to the hidden layer, and finally adjusting the connection weights and bias of the neural network according to the errors of each layer to minimize the cost function.
8. The method according to claim 7, characterized in that The error calculation of the back propagation algorithm includes: first calculating the error of the output layer, then reversely calculating the error of the hidden layer in turn, and finally using the errors of each layer to calculate the partial derivatives of the cost function with respect to the connection weights and biases, and adjusting the neural network parameters according to the partial derivatives.
9. The method according to claim 1, characterized in that: In step S6, regularization is used in the training process of the artificial neural network. Regularization terms are added to the cost function to reduce the complexity of the model, improve the generalization ability of the model, and prevent overfitting.
10. The method according to claim 1, characterized in that In step S6, the neural network parameters are adjusted according to the partial derivatives of the cost function, and are specifically optimized using the gradient descent method or the momentum gradient descent method. By iteratively adjusting the connection weights and biases, the value of the cost function is gradually reduced until the model converges.