Power transmission line icing probability prediction method based on icing identification and quantile regression
By combining k-means clustering, CNN and LSTM neural networks, a probability prediction model for ice-covered thickness in transmission lines is constructed, which solves the problem of low ice-covered prediction accuracy in the existing technology, and realizes more accurate ice-covered pattern recognition and probability prediction, providing more comprehensive early warning and maintenance information for the safe and stable operation of the power grid.
Patent Information
- Application Number
- CN202411722604.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-03-04
AI Technical Summary
The existing transmission line ice-covered prediction model is difficult to provide accurate probability prediction under different ice-covered modes, resulting in low prediction accuracy and cannot provide decision makers with comprehensive early warning information and maintenance solutions.
The method based on ice-covering pattern recognition and quantile regression is adopted, and the ice-covering pattern is identified through the k-means clustering algorithm and the CNN model. The ice-covering thickness probability prediction model is constructed based on the LSTM neural network and quantile regression, providing the probability distribution interval of ice-covering thickness under different ice-covering patterns.
It improves the accuracy and reliability of ice-covering prediction, provides more comprehensive prediction information, helps decision makers formulate effective early warning and maintenance plans, and ensures the safe and stable operation of the power grid.
Smart Images
Figure CN119557851A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of probability prediction of ice thickness of power transmission lines, and relates to a method for predicting the probability of ice coating on power transmission lines based on ice coating pattern recognition and quantile regression. Background Art
[0002] Since the 21st century, due to the rapid development of the national economy and the continuous increase in electricity demand, my country's power grid construction has entered a period of rapid development. However, due to the significant differences in the natural environment in different regions, transmission lines face a higher risk of meteorological disasters, which poses a challenge to the safe and stable operation of the lines. For transmission lines, icing hazards are one of the most common and serious hazards. In severe cases, they may even cause tripping, line breaking, tower collapse, insulator ice flash, communication interruption and other problems, posing a huge threat to the safety and stability of the lines and causing huge economic losses. By predicting the thickness of line ice, we can respond to the damage caused by ice in a timely manner and ensure the safe and stable operation of the power grid.
[0003] In recent years, domestic and foreign scholars have conducted a lot of research on the prediction of icing on transmission lines and have made relatively significant progress. At present, icing prediction models can be roughly divided into three categories, namely physical models, statistical models, and artificial intelligence models. Although the existing model methods have achieved relatively superior results in their respective experimental scenarios, most transmission line icing predictions are deterministic modeling for the entire icing cycle. The prediction accuracy is often not high enough, or the model is too complex. Few studies have made probabilistic predictions of different characteristics under different icing modes, and it is impossible to provide decision makers with more comprehensive prediction information to formulate and make decisions on early warning information and maintenance plans. Summary of the invention
[0004] The technical problem to be solved by the present invention is to provide a method for predicting the probability of icing on transmission lines based on icing recognition and quantile regression. By improving the algorithm, the method can provide accurate icing thickness probability prediction results, which is of great significance to the anti-icing work of power grid transmission lines, as well as the safe operation, maintenance and inspection of the power grid.
[0005] The technical solution adopted by the present invention is: a transmission line icing probability prediction method based on icing recognition and quantile regression, and the transmission line icing probability prediction method comprises: Step S10, obtaining historical ice thickness data, meteorological data and altitude data of the transmission lines in the forecast area, and obtaining key features that affect the icing mode of the transmission lines by enhancing the above data; Step S20, based on the key features extracted in step S10, the transmission line icing patterns are classified into N types of labels using a k-means clustering algorithm, where N∈[2,9]; Then the silhouette coefficient is selected to evaluate the N types of labels, the best value of N is evaluated, and the sub-datasets under different icing mode labels are obtained; Step S30, constructing a transmission line icing pattern recognition model with the help of a CNN model, the model takes the historical icing data and meteorological data two moments before the predicted time T as input, and outputs the label of the icing pattern at the predicted time T; Step S40, construct an ice thickness probability prediction model by combining quantile regression with LSTM neural network. The model takes the historical ice thickness data and meteorological data of the transmission line at two moments before the predicted moment T as input, outputs the ice thickness probability distribution interval of the transmission line at the predicted moment T, and obtains the ice thickness probability prediction results under different ice modes.
[0006] Further, the data enhancement processing in step S10 is: according to the historical ice thickness data, meteorological data, and altitude data, by analyzing the mechanism of the icing process of the transmission line, the cumulative change of ice thickness F1, the gradient of ice thickness F2, the third-order derivative of ice thickness F3, the relative ambient temperature F4, the ambient humidity F5, and the wind speed F6 are extracted as key features affecting the icing mode; The cumulative change in ice thickness F1, ; The gradient of ice thickness F2, ; The third-order derivative of ice thickness F3, ; Temperature F4, F4∈(-15-0℃); Relative humidity F5, F5>80%; Wind speed F6, F6∈ (0-10m / s).
[0007] Where h represents the ice thickness; h(t2) and h(t1) represent the ice thickness at time t2 and t1 respectively; t2 and t1 are two adjacent moments, and ≥1.
[0008] Further, the silhouette coefficient formula is ; in, represents the average Euclidean distance between the ith ice thickness data and the ice thickness data in its category, Represents the average Euclidean distance between ice thickness data and all ice thickness data in other categories.
[0009] Furthermore, the training and evaluation method of the transmission line icing pattern recognition model in step S30 is: The sub-datasets under different icing pattern labels are divided into training sets and test sets in a ratio of 8:2. With the help of the constructed transmission line icing pattern identification model, the icing patterns of all test sets are obtained. Based on the CNN-based ice pattern recognition model, the user accuracy PA and global accuracy OA are used to evaluate the results of ice pattern recognition. User precision PA evaluates the correct probability of the model in predicting a certain type of icing pattern label, and global precision OA evaluates the overall prediction accuracy; The calculation formula of user accuracy PA is as follows: , The global accuracy OA calculation formula is as follows: , Where m represents the number of samples of ice cover data classified into a certain type of ice cover mode, It represents the total number of ice-covered data of category e that are classified into ice-covered data of category f, and n is the total number of samples in the test set of all original ice-covered thickness data.
[0010] Furthermore, the training and evaluation method of the ice thickness probability prediction model in step S40 is: The sub-datasets under different icing mode labels are classified according to the label category, and the sub-datasets of each category are divided into training sets and test sets in a ratio of 8:2; The model is trained separately for each category of sub-datasets to obtain the probability prediction model of ice thickness with different ice mode labels, thereby obtaining the upper and lower bounds of ice thickness probability prediction under different ice mode labels; The probability of ice thickness in different icing mode labels is comprehensively evaluated using the interval coverage probability (PICP) and the interval normalized average width (PINAW). The probability prediction results under all icing mode labels are sorted in time series to obtain the probability prediction results of icing thickness for all test sets.
[0011] The present invention also provides a storage medium for storing a computer program, wherein when the computer program is executed by a processor, a method for predicting the probability of icing of a power transmission line based on icing pattern recognition and quantile regression is implemented.
[0012] The beneficial effects produced by the use of the present invention are as follows: 1) The icing prediction model based on icing pattern recognition of the present method fully considers the characteristics of ice thickness changes under different icing modes, and improves the prediction accuracy by using different data sets for targeted training for different icing modes, thereby avoiding the loss of accuracy caused by using a single model to uniformly predict multiple icing modes; 2) The icing probability prediction method for transmission lines based on icing pattern recognition and quantile regression of the present method can not only effectively judge the icing mode at the time to be predicted, but also use the probability prediction method combining LSTM with quantile regression to obtain the probability prediction intervals under different icing modes, and obtain the probability distribution of the icing thickness of the transmission lines at the time to be predicted, so as to provide decision makers with more comprehensive prediction information. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0014] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.
[0015] See attached Figure 1 , a transmission line icing probability prediction method based on icing recognition and quantile regression, the transmission line icing probability prediction method comprising: Step S10, obtaining historical ice thickness data, meteorological data and altitude data of the transmission lines in the forecast area, and obtaining key features that affect the icing mode of the transmission lines by enhancing the above data; The data enhancement process in step S10 is as follows: according to the historical ice thickness data, meteorological data, and altitude data, by analyzing the mechanism of the icing process of the transmission line, the cumulative change of ice thickness F1, the gradient of ice thickness F2, the third-order derivative of ice thickness F3, the relative ambient temperature F4, the ambient humidity F5, and the wind speed F6 are extracted as key features affecting the icing mode; The cumulative change in ice thickness F1, ; The gradient of ice thickness F2, ; The third-order derivative of ice thickness F3, ; Temperature F4, F4∈(-15-0℃); Relative humidity F5, F5>80%; Wind speed F6, F6∈ (0-10m / s).
[0016] Where h represents the ice thickness; h(t2) and h(t1) represent the ice thickness at time t2 and t1 respectively; t2 and t1 are two adjacent moments, and ≥1.
[0017] Step S20, based on the key features extracted in step S10, the transmission line icing patterns are classified into N types of labels using a k-means clustering algorithm, where N∈[2,9]; Then the silhouette coefficient is selected to evaluate the N types of labels, the best value of N is evaluated, and the sub-datasets under different icing mode labels are obtained; According to the cumulative change of ice thickness F1, the gradient of ice thickness F2, the third-order derivative of ice thickness F3, the ambient relative temperature F4, the ambient humidity F5, and the wind speed F6, the k-means clustering algorithm is used to divide the ice pattern into N categories (N∈[2,9]). The silhouette coefficient (SC) is selected to evaluate the k-means classification result N, and the index coefficients of the cluster number N from 2 to 9 are evaluated to finally confirm the specific value of the classification number N.
[0018] The silhouette coefficient formula is ; in, represents the average Euclidean distance between the ith ice thickness data and the ice thickness data in its category, Represents the average Euclidean distance between ice thickness data and all ice thickness data in other categories.
[0019] , Assuming that the i-th data point belongs to class A, then a(i) can be expressed as the distance from the i-th data point to other data points x in class A. j is the average of the Euclidean distances. |A| is the total number of original ice thickness data points in class A, d(x i， x j ) is the data point x i and x j The Euclidean distance between .
[0020] , b ( i ) refers to the average distance from the ith data point to the nearest non-self class. First, we need to find the non-self class B closest to the ith data point, and then calculate the distance from the ith data point to the data point x in class B. k The average value of the Euclidean distance of . Where |B| is the total number of original ice thickness data points in class B.
[0021] Step S30, constructing a transmission line icing pattern recognition model with the help of a CNN model, the model takes the historical icing data and meteorological data two moments before the predicted time T as input, and outputs the label of the icing pattern at the predicted time T; The training and evaluation method of the transmission line icing pattern identification model in step S30 is: The sub-datasets under different icing pattern labels are divided into training sets and test sets in a ratio of 8:2. With the help of the constructed transmission line icing pattern identification model, the icing patterns of all test sets are obtained. Based on the CNN-based ice pattern recognition model, the user accuracy PA and global accuracy OA are used to evaluate the results of ice pattern recognition. User precision PA evaluates the correct probability of the model in predicting a certain type of icing pattern label, and global precision OA evaluates the overall prediction accuracy; The calculation formula of user accuracy PA is as follows: , The global accuracy OA calculation formula is as follows: , Where m represents the number of samples of ice cover data classified into a certain type of ice cover mode, It represents the total number of ice-covered data of category e that are classified into ice-covered data of category f, and n is the total number of samples in the test set of all original ice-covered thickness data.
[0022] Step S40, construct an ice thickness probability prediction model by combining quantile regression with LSTM neural network. The model takes the historical ice thickness data and meteorological data of the transmission line at two moments before the predicted moment T as input, outputs the ice thickness probability distribution interval of the transmission line at the predicted moment T, and obtains the ice thickness probability prediction results under different ice modes.
[0023] According to the training set icing pattern classification results and the test set icing pattern identification results obtained in step S30, the QR-LSTM model is trained and predicted using the training set and test set in different modes. The historical ice thickness data and meteorological data two moments before the time to be predicted are used as model input, and the ice thickness prediction interval at the time to be predicted is output to obtain the ice thickness probability prediction results under different icing modes.
[0024] Compared with the traditional linear regression method to calculate the expected value of the dependent variable, Quantile Regression (QR) fits the conditional quantile relationship between the independent variable and the dependent variable, providing more statistical information. It more accurately reflects the impact of the independent variable on the range of variation of the dependent variable. As a typical non-parametric probability prediction method, QR does not need to assume the distribution of the predicted variable, but directly predicts the quantile at each quantile level. It can accurately describe the range and distribution shape of the dependent variable and effectively reflect the impact of outliers.
[0025] For conditional quantiles, it is defined as the conditional cumulative distribution function of a random variable (X, Y) where Y is given X=x. Then the τ-order quantile of the random variable Y under the condition X=x is defined as: , in is the τth quantile. The above formula is also equivalent to the τth quantile of the random variable Y satisfy: , It is also equivalent to: , , in, is the Pinball loss function. Based on the Pinball loss function, for a certain τ∈(0,1), find the minimum expectation of the loss function. The expectation of the loss function is , Differentiate the above formula and set it to 0, and we get , because It is monotonous and satisfying of The value can minimize the expectation of the loss function. That is Therefore, the solution obtained is consistent with the quantile The LSTM quantile regression method specifically uses the expected sum of the loss functions of each quantile as the loss function of the neural network for training, thereby obtaining the quantile of each quantile. During the network training process, the quantile of the corresponding quantile can be obtained by minimizing the loss function. The specific loss function calculation formula is: , in, is the true value, is the predicted value, and x is the input number.
[0026] The training and evaluation method of the ice thickness probability prediction model in step S40 is: The model is trained separately for each category of sub-datasets to obtain the probability prediction model of ice thickness with different ice mode labels, thereby obtaining the upper and lower bounds of ice thickness probability prediction under different ice mode labels; The probability of ice thickness in different icing mode labels is comprehensively evaluated using the interval coverage probability (PICP) and the interval normalized average width (PINAW). Then, the interval coverage probability and interval normalized average width method are used to comprehensively evaluate the ice thickness probability results within different ice pattern labels; When evaluating the ice thickness probability results within different icing mode labels, interval prediction performance indicators are used, including the interval coverage probability (PICP), which is used to measure the proportion of actual observations contained in the prediction interval; the normalized average width of the interval (PINAW), which is used to evaluate the economy and accuracy of the prediction interval width; and the comprehensive evaluation index (CWC), which combines the coverage probability and interval width to provide a balanced evaluation standard for the overall performance of the prediction interval.
[0027] The calculation method of interval coverage probability PICP is as follows: , , Where n is the number of samples of all ice thickness data in the test set; is a binary value, , are the lower and upper bounds of the ice thickness prediction interval, respectively. When the kth actual measurement value falls within the prediction interval, the value is 1; otherwise, it is 0. PICP is the probability that the actual value falls within the prediction interval and is an important indicator for evaluating the reliability of interval prediction. The larger its value, the greater the probability that the actual value falls within the prediction interval and the better the model prediction performance.
[0028] The calculation method of interval normalized average width PINAW is as follows: , in, Indicates the range of actual measured values (used for normalization). If the interval is too wide, the uncertainty of the prediction results will increase, the redundant information will increase, and the information validity of the prediction results will be reduced. PINAW is used to evaluate the width of the prediction interval. The smaller its value, the narrower the prediction interval and the better the prediction performance.
[0029] Generally speaking, we hope that the larger the PICP is, the better, and the smaller the PINAW is, the better, but at the same time, the two are contradictory. Therefore, a comprehensive index CWC is set, which is defined as the difference between PICP and PINAW. The larger the CWC is, the higher the comprehensive score is, and the better the model performance is to a certain extent. Its calculation method is as follows: , For N types of icing modes, N icing thickness probability prediction models are trained. Each model is trained 10 times, and the model with the highest CWC score is selected as the model with the best performance in icing thickness probability prediction for each type of icing mode.
[0030] Finally, the probability prediction results under all icing mode labels are sorted in time series to obtain the probability prediction results of icing thickness for all test sets.
[0031] Since the time of the sub-datasets under different icing modes is not completely arranged in chronological order, it is necessary to sort them in chronological order according to the time of each sub-dataset, and finally obtain the icing probability prediction results arranged in chronological order at all times.
[0032] The present invention also provides a storage medium for storing a computer program, wherein when the computer program is executed by a processor, a method for predicting the probability of icing of a power transmission line based on icing pattern recognition and quantile regression is implemented.
[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting the probability of icing on power transmission lines based on icing identification and quantile regression, characterized in that: The method for predicting the probability of icing on a transmission line comprises: Step S10, obtaining historical ice thickness data, meteorological data and altitude data of the transmission lines in the forecast area, and obtaining key features that affect the icing mode of the transmission lines by enhancing the above data; Step S20, based on the key features extracted in step S10, the transmission line icing patterns are classified into N types of labels using a k-means clustering algorithm, where N∈[2,9]; Then the silhouette coefficient is selected to evaluate the N types of labels, the best value of N is evaluated, and the sub-datasets under different icing mode labels are obtained; Step S30, constructing a transmission line icing pattern recognition model with the help of a CNN model, the model takes the historical icing data and meteorological data two moments before the predicted time T as input, and outputs the label of the icing pattern at the predicted time T; Step S40, construct an ice thickness probability prediction model by combining quantile regression with LSTM neural network. The model takes the historical ice thickness data and meteorological data of the transmission line at two moments before the predicted moment T as input, outputs the ice thickness probability distribution interval of the transmission line at the predicted moment T, and obtains the ice thickness probability prediction results under different ice modes.
2. The method for predicting the probability of icing on a power transmission line according to claim 1, characterized in that: The data enhancement process in step S10 is as follows: according to the historical ice thickness data, meteorological data, and altitude data, by analyzing the mechanism of the icing process of the transmission line, the cumulative change of ice thickness F1, the gradient of ice thickness F2, the third-order derivative of ice thickness F3, the relative ambient temperature F4, the ambient humidity F5, and the wind speed F6 are extracted as key features affecting the icing mode; The cumulative change in ice thickness F1, ； The gradient of ice thickness F2, ； The third-order derivative of ice thickness F3, ； Temperature F4, F4∈(-15-0℃); Relative humidity F5, F5>80%; Wind speed F6, F6∈ (0-10m / s); Where h represents the ice thickness; h(t2) and h(t1) represent the ice thickness at time t2 and t1 respectively; t2 and t1 are two adjacent moments, and ≥1.
3. The method for predicting the probability of icing on a power transmission line according to claim 1, characterized in that: The silhouette coefficient formula is ； in, represents the average Euclidean distance between the ith ice thickness data and the ice thickness data in its category, Represents the average Euclidean distance between ice thickness data and all ice thickness data in other categories.
4. The method for predicting the probability of icing on a power transmission line according to claim 1, characterized in that: The training and evaluation method of the transmission line icing pattern identification model in step S30 is: The sub-datasets under different icing pattern labels are divided into training sets and test sets in a ratio of 8:
2. With the help of the constructed transmission line icing pattern identification model, the icing patterns of all test sets are obtained. Based on the CNN-based ice pattern recognition model, the user accuracy PA and global accuracy OA are used to evaluate the results of ice pattern recognition; User precision PA evaluates the correct probability of the model in predicting a certain type of icing pattern label, and global precision OA evaluates the overall prediction accuracy; The calculation formula of user accuracy PA is as follows: ， The global accuracy OA calculation formula is as follows: ， Where m represents the number of samples of ice cover data classified into a certain type of ice cover mode. It represents the total number of the e-th type of ice data that are classified into the f-th type of ice data, and n is the total number of samples in the test set of all original ice thickness data.
5. The method for predicting the probability of icing on a power transmission line according to claim 1, characterized in that: The training and evaluation method of the ice thickness probability prediction model in step S40 is: The sub-datasets under different icing mode labels are classified according to the label category, and the sub-datasets of each category are divided into training sets and test sets in a ratio of 8:2; The model is trained separately for each category of sub-datasets to obtain the probability prediction model of ice thickness with different ice mode labels, thereby obtaining the upper and lower bounds of ice thickness probability prediction under different ice mode labels; The interval coverage probability and the interval normalized average width are used to comprehensively evaluate the ice thickness probability results in different ice pattern labels. The probability prediction results under all icing mode labels are sorted in time series to obtain the probability prediction results of icing thickness for all test sets.
6. A storage medium for storing a computer program, characterized in that: When the computer program is executed by a processor, the method for predicting the probability of icing on a power transmission line based on icing pattern recognition and quantile regression as described in any one of claims 1 to 5 is implemented.