CN116231603B

CN116231603B - Self-adaptive control method for feeder line of traction power supply wide-area protection measurement and control system

Info

Publication number: CN116231603B
Application number: CN202310504938.6A
Authority: CN
Inventors: 罗杨; 郝海峰; 陈德明; 何顺江; 林伟; 熊列彬
Original assignee: Chengdu Jiaoda Yunda Electrical Co ltd
Current assignee: Chengdu Jiaoda Yunda Electrical Co ltd
Priority date: 2023-05-08
Filing date: 2023-05-08
Publication date: 2023-07-14
Anticipated expiration: 2043-05-08
Also published as: CN116231603A

Abstract

The invention relates to the technical field of power supply wide area protection, in particular to a feeder line self-adaptive control method of a traction power supply wide area protection measurement and control system. Firstly, obtaining a time characteristic vector and an impedance angle corresponding to each traction power supply interval, and constructing a delay time characteristic vector according to the time characteristic vector and the impedance angle; performing abnormality judgment on the delay time length feature vector to obtain an abnormality judgment result; determining a working condition feature code of the traction power supply section according to the delay time length feature vector of the traction power supply section and the correction label of the delay time length feature vector, and determining the abnormality degree of the working condition feature code; and measuring and controlling the feeder line in the traction power supply section according to the abnormality degree and the abnormality judgment result. The invention combines the network communication state context and the self-adaptive control method of the multi-mode feeder monitoring data, thereby improving the reliability of single-point override of the feeder in the prior art.

Description

Self-adaptive control method for feeder line of traction power supply wide-area protection measurement and control system

Technical Field

The invention relates to the technical field of power supply wide area protection, in particular to a feeder line self-adaptive control method of a traction power supply wide area protection measurement and control system.

Background

In the electrified railway traction power supply wide area protection measurement and control system, the feeder line protection measurement and control device is provided with wide area protection, has the characteristics of selective tripping, power failure interval reduction, high action speed and the like, and has great significance in improving the reliability and stability of traction power supply. The wide area protection generally adopts a current or impedance principle, performs data interaction in real time through a wide area network, and realizes quick tripping or quick locking among feeder devices through a jump or locking signal.

In practical application, when the overhead contact system of the electrified railway breaks down, the wide area protection action breaks off the feeder circuit breakers and all parallel circuit breakers on the power supply arm at the fault side, and the non-fault side circuit breakers do not act. If the communication delay of the wide area network is too long, the feeder line of the non-fault side substation cannot timely receive other transmitted blocking signals to block the wide area protection, or the feeder line of the fault side cannot timely receive other transmitted jump signals to jump off the circuit breaker, the feeder line of the other substation is involved, the non-fault side tripping is caused, and the protection of the existing interval is lost.

At present, the existing wide area protection logic is too simple, is sensitive to the communication delay of a wide area network, and generally requires the communication delay to be not more than 0.01s. In the existing data transmission communication network between railways, the communication delay is difficult to be less than 0.01s due to the influence of communication such as routing, message, core network and the like, and the time-out time is usually about 100 ms-1 s for common network communication time Jitter (Jitter). Therefore, in practical application, the wide area network is generally directly connected and networked by adopting a special optical fiber data channel, and special monitoring and maintenance are required for the network. The cost of the distributed node is too high, so a new method is needed to realize self-adaptive distributed control, thereby reducing the difficulty of design, construction and maintenance.

The traditional distributed control of the traction power supply interval can only be realized by analyzing the related trigger signals sent by the wide area protection measurement and control device, and the judgment of circuit breaking can not be carried out according to the field situation. In the existing method, when timely tripping action cannot be performed under the condition of poor network communication, whether the local feeder line is abnormal or not cannot be known, and further measurement and control of the feeder line are difficult to realize.

Disclosure of Invention

In order to solve the technical problem that the control of the feeder line is difficult to realize, the invention aims to provide a traction power supply wide area protection measurement and control system feeder line self-adaptive control method, which adopts the following technical scheme:

obtaining a time characteristic vector and an impedance angle corresponding to each traction power supply interval, and constructing a delay time length characteristic vector according to the time characteristic vector and the impedance angle;

marking corresponding labels on delay time length feature vectors in the delay time length feature vector sequence to obtain corresponding label sequences; correcting the tag sequence to obtain a corresponding corrected tag sequence, and obtaining corrected tags corresponding to the characteristic vectors of each delay time length; performing abnormality judgment on the delay time length feature vector to obtain an abnormality judgment result;

determining working condition feature codes of the traction power supply interval according to the delay time length feature vector of the traction power supply interval, the similarity between the delay time length feature vectors of the traction power supply interval and other traction power supply intervals and the correction label of the delay time length feature vector;

Determining a sample distance according to the difference of the corresponding delay time length characteristic vectors among the traction power supply intervals and the time characteristic vector of the traction power supply interval; determining the abnormality degree of the working condition feature code based on the sample distance;

and (5) performing override on the feeder line in the traction power supply section according to the abnormality degree and the abnormality judgment result.

Preferably, the labeling the delay duration feature vector in the delay duration feature vector sequence to obtain a corresponding label sequence includes:

classifying the delay time length feature vectors by using a K-MEANS algorithm based on the distance between the delay time length feature vectors to obtain at least two categories, and labeling the delay feature vectors in each category; and obtaining labels corresponding to each delay characteristic vector in the delay time length characteristic vector sequence, and constructing a corresponding label sequence.

Preferably, the modifying the tag sequence to obtain a corresponding modified tag sequence includes:

and carrying out smoothing treatment on the label sequence to obtain a corresponding corrected label sequence.

Preferably, the performing anomaly determination on the delay time feature vector to obtain an anomaly determination result includes:

taking part of delay time length feature vectors as a training set; calculating the distance between the delay time length feature vector and the training set to be judged; when the distance to be judged is larger than an abnormal threshold value, taking the delay time length characteristic vector as an abnormal delay time length characteristic vector; and when the distance to be judged is smaller than or equal to the abnormal threshold value, taking the delay time length characteristic vector as a normal delay time length characteristic vector.

Preferably, the constructing a delay time length feature vector according to the time feature vector and the impedance angle includes:

and constructing a delay time length feature vector by the impedance angle acquired at the current moment, the absolute value of the difference value of the impedance angles at the current moment and the previous moment, the L2 distance between the predicted time feature vector and the actual time feature vector, and the maximum value element and the minimum value element of the time feature vector.

Preferably, the determining the working condition feature code of the traction power supply section according to the delay time feature vector of the traction power supply section, the similarity between the delay time feature vectors of the traction power supply section and other traction power supply sections, and the correction label of the delay time feature vector includes:

taking a high-dimensional vector formed by the delay time length characteristic vector of the current time and the delay time length characteristic vectors of the first two times of the traction power supply interval as input to obtain a single-heat code corresponding to a correction label of the delay time length characteristic vector of the current time;

and constructing a working condition feature code by the delay time length feature vector of the traction power supply section, the minimum cosine similarity of the delay time length feature vector of the traction power supply section and the delay time length feature vector of other traction power supply sections, the independent heat code corresponding to the delay time length feature vector of the traction power supply section, and the shortest time feature of the traction power supply section and the other traction power supply sections in the neighborhood.

Preferably, the determining the sample distance according to the difference of the corresponding delay time length feature vectors between the traction power supply intervals and the time feature vector of the traction power supply interval includes:

for any two traction power supply intervals, calculating cosine similarity of delay time length feature vectors corresponding to the two traction power supply intervals, and taking the cosine similarity as a first similarity; taking a difference value between a preset first threshold value and a first similarity as a first factor; respectively acquiring two minimum cosine similarities of delay time length feature vectors of the two traction power supply intervals and other traction power supply intervals, and taking the ratio of a larger value to a smaller value in the two minimum cosine similarities as a second factor; respectively acquiring two shortest time features of the traction power supply intervals in the neighborhood corresponding to the traction power supply intervals, and taking the maximum value of the two shortest time features as a third factor; taking the hamming distances of the independent heat codes corresponding to the two traction power supply intervals as a fourth factor; the first factor, the second factor, the third factor and the fourth factor are all in positive correlation with the sample distance.

Preferably, the determining the abnormality degree of the working condition feature code based on the sample distance includes:

And taking the sample distance as the reachable distance between the working condition feature codes, calculating the local outlier factor of the working condition feature codes based on the sample distance, and taking the local outlier factor of the working condition feature codes as the abnormality degree of the working condition feature codes.

Preferably, the overriding the feeder line in the traction power supply section according to the abnormality degree and the abnormality judgment result includes:

and when the abnormality judgment result is continuously abnormal within the preset timeout period and the maximum value of the abnormality degree within the preset timeout period is greater than the maximum abnormality degree at the historical moment, the feeder is overridden.

Preferably, the shortest time feature is: and the minimum element value in the time characteristic vector corresponding to the traction power supply interval and other traction power supply intervals in the neighborhood.

The embodiment of the invention has at least the following beneficial effects:

the invention relates to the technical field of power supply wide area protection. The method comprises the steps of firstly, carrying out abnormality judgment on the delay time length feature vector to obtain an abnormality judgment result, and realizing abnormality judgment and monitoring of the time length feature vector and the impedance angle. Further, according to the delay time length feature vector of the traction power supply section, the similarity between the delay time length feature vectors of the traction power supply section and other traction power supply sections and the correction label of the delay time length feature vector, the working condition feature code of the traction power supply section is determined, the working condition feature code further analyzes the delay time length feature vector and combines the difference of the time feature vectors to realize abnormal monitoring of the working condition feature code, and finally, the abnormal degree and the abnormal judgment result are combined to measure and control the feeder line of the traction power supply section. The time delay of the traditional control network is limited to 0.01s, and the time length of the tolerant network Jitter can be prolonged to be a preset time-out time length based on the override logic. When other traction power supply intervals cannot be connected within the preset timeout period, if tripping conditions occur in adjacent traction power supply intervals, the traction power supply intervals are guaranteed to be in a power supply state until the timeout tolerance time is reached, so that the work of a main power supply line is protected, and the interval protection after connection recovery still has selectivity.

Drawings

In order to more clearly illustrate the embodiments of the invention or the technical solutions and advantages of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are only some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.

Fig. 1 is a flow chart of a method for adaptively controlling a feeder line of a traction power supply wide area protection measurement and control system according to an embodiment of the present invention.

Detailed Description

In order to further describe the technical means and effects adopted by the invention to achieve the preset aim, the following is a detailed description of specific implementation, structure, characteristics and effects of a feeder adaptive control method of a traction power supply wide area protection measurement and control system according to the invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" means that the embodiments are not necessarily the same. Furthermore, the particular features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

The embodiment of the invention provides a specific implementation method of a feeder self-adaptive control method of a traction power supply wide-area protection measurement and control system, which is suitable for a traction power supply wide-area protection measurement and control scene. The traction power supply section in the scene is composed of a transformer substation, a section of feeder line, networking equipment and a matched power system. In order to solve the technical problem that the control of the feeder line is difficult to realize, the invention combines the network communication state context and the self-adaptive control method of the multi-mode feeder line monitoring data, and improves the reliability of single-point override of the feeder line in the prior art.

The invention provides a specific scheme of a feed line self-adaptive control method of a traction power supply wide-area protection measurement and control system, which is specifically described below with reference to the accompanying drawings.

Referring to fig. 1, a method flowchart of a method for adaptively controlling a feeder line of a traction power supply wide area protection measurement and control system according to an embodiment of the present invention is shown, where the method includes the following steps:

step S100, a time characteristic vector and an impedance angle corresponding to each traction power supply interval are obtained, and a delay time characteristic vector is constructed according to the time characteristic vector and the impedance angle.

Firstly, constructing a time feature vector N according to network fluctuation of information collected by each traction power supply interval, wherein the time feature vector is used for representing RTT features of one traction power supply interval. The network fluctuation is the change of the RTT between the total node of the wide area protection system and the nodes of each traction power supply interval in each traction power supply interval. The RTT (Round-Trip Time) is a Round Trip Time, which refers to a Time required for data to be sent from a source host to a target host and then returned from the target host to the source host, and specifically, traceroute command is one of network diagnostic tools, and may track a path of a data packet from the source host to the target host, so as to obtain an RTT corresponding to each source host. It should be noted that, the method for obtaining RTT is a well-known technology for those skilled in the art, and will not be described herein. In the embodiment of the invention, RTT corresponding to a traction power supply interval is a total time delay from the start of sending data from one traction power supply interval to the time when the traction power supply interval receives confirmation from other traction power supply intervals, wherein the traction power supply interval is taken as a sending end or a source host, and the other traction power supply intervals are taken as receiving ends or target hosts. The RTT value can also be said to be a delay period.

The time feature vector N is used to describe the network quality features of one traction power supply section facing other traction power supply sections and the wide area protection system in combination with the fluctuation sizes of each RTT of the one traction power supply section and the other traction power supply sections.

And transmitting the RTT values acquired by information acquisition to the rest traction power supply intervals and the wide area protection system by one traction power supply interval a, and arranging the RTT values in sequence from large to small to obtain the time feature vector corresponding to the traction power supply interval a.

The time feature vector corresponding to the traction power supply interval is acquired once per second, which can be said to be updated once per second. Since there are more than one traction power supply section, the corresponding temporal feature vector is multidimensional, e.g. there are 10 traction power supply sections, and the dimension of the corresponding temporal feature vector is 10.

And training the corresponding first RNN for each traction power supply interval, and predicting the time feature vector. For the updated time feature vector per second, an observation period, for example, one month, is specified by the practitioner, and the time feature vector N is recorded for a long time. Because the continuously recorded sequence is too long to be directly trained, the segmentation process is performed, and for a sequence with too long a length, the sequence can be changed into a training set by segmenting the sequence into shorter sequences. In the embodiment, the original sequence is segmented according to a fixed length by using a fixed length method, and each sub-sequence has equal length, so that the number of the sub-sequences can be controlled by adjusting the sub-sequence length and the overlapping rate. It should be noted that, the method of splitting a long sequence into a plurality of shorter sequences is a well-known technique for those skilled in the art, and will not be described herein.

And taking the segmented short-time feature vector sequence corresponding to each traction power supply interval as a data set of the first RNN network, and training the first RNN network corresponding to each traction power supply interval to output a predicted time feature vector of the next moment of the traction power supply interval in the first RNN network with trained time feature vectors corresponding to the traction power supply intervals.

The RNN has an internal circulation structure and can transmit the previous state information to the current moment, so that modeling and prediction of sequence data are realized, and the sequence of a data set is only required to be used as an input sequence during training, and the prediction result is obtained through forward propagation of an RNN network. In the forward propagation process, a deep learning framework such as Pytorch needs to be utilized to save the hidden state of each moment so as to calculate the gradient in the backward propagation process. Because the time feature vector N is multidimensional, the corresponding short-time feature vector sequence is multidimensional, the loss difference is calculated by using a mean square error (Mean Squared Error, MSE), the next moment is taken as a predicted real label, and the MSE loss can be obtained after the RNN predicted result is compared with the real label. The deep learning framework may calculate gradients of parameters from the values of the loss function and update the parameters with an optimizer. The practitioner sets a reasonable number of training rounds, e.g., 10000epoch, until the model converges or reaches a preset number of training rounds.

And recording real-time impedance angles for each traction power supply interval, and constructing a characteristic vector C of the difference between the predicted result and the real-time result, the impedance angle and the delay time at the moment.

In the circuit, the alternating voltage and current are both in the form of sine waves, which can be expressed as

And

wherein

And

maximum voltage and current, respectively, ω is angular frequency,

and

the phase angles of the voltage and current, respectively, sin is a sine function and t is time. The impedance angle theta may be expressed as

Representing the difference between the voltage phase angle and the current phase angle. If the voltage in the circuit leads the current, the impedance angle is positive; if the current in the circuit leads the voltage, the impedance angle is negative. In a pure resistor circuit, the voltage and current phases are the same, and the impedance angle is zero; in a pure inductance circuit, the voltage leads the current by 90 degrees, and the impedance angle is positive by 90 degrees; in a purely capacitive circuit, the voltage lags the current by 90 °, the impedance angle being minus 90 °.

The results between the impedance angle and the time feature vector N should be irrelevant, but for a system operating for a long period of time and with multiple potential operating conditions, there will be common integrated operating conditions. Therefore, further, the corresponding delay time length characteristic vector is constructed by combining the impedance angle and the time characteristic vector, and the comprehensive operation condition of the system is reflected by the delay time length characteristic vector.

Because of a random phenomenon when the network fluctuates, but common time Jitter (Jitter) features exist, the delay duration feature vector can avoid non-self-consistency caused by the result of a specific time feature vector N, because the situation of the time feature vector N is changeable, and if reasonable processing is performed in a later period, the influence that the time feature vector N cannot completely cover all possible situations needs to be greatly reduced.

Because each first RNN network can accurately predict the time feature vector N of the next moment under a typical condition, the delay time length feature vector C carries out description of different working conditions of the traction power supply section by means of the difference size and the real-time impedance angle of the predicted time feature vector and the actual time feature vector obtained by the first RNN network, so that the problem of long tail effect is avoided regardless of all the possibility of the time feature vector N, the features of the working conditions of the traction power supply section are accurately described, and later processing and analysis are facilitated.

Constructing a delay duration feature vector according to the time feature vector and the impedance angle, and specifically: and constructing a delay time length feature vector by the impedance angle acquired at the current moment, the absolute value of the difference value of the impedance angles at the current moment and the previous moment, the L2 distance between the predicted time feature vector and the actual time feature vector, and the maximum value element and the minimum value element of the time feature vector, namely the delay time length feature vector is a 5-dimensional vector. Wherein the impedance angle acquired at the current time is used to describe a typical load; the absolute value of the difference value of the impedance angle between the current moment and the previous moment is used for describing the fluctuation condition of the impedance angle; the L2 distance of the predicted temporal feature vector and the actual temporal feature vector is used to describe the difference between the predicted outcome value and the actual outcome value; the maximum value element and the minimum value element of the time feature vector are used for describing the communication performance between the traction power supply interval and other traction power supply intervals.

The collection frequency of the time characteristic vector and the impedance angle in the step S100 is 1Hz, namely, the collection is carried out once per second, and the collection is carried out aiming at different traction power supply intervals.

Step S200, marking corresponding labels on delay time length feature vectors in the delay time length feature vector sequence to obtain corresponding label sequences; correcting the tag sequence to obtain a corresponding corrected tag sequence, and obtaining corrected tags corresponding to the characteristic vectors of each delay time length; and carrying out abnormality judgment on the delay time length feature vector to obtain an abnormality judgment result.

An observation period is set in the wide area system, clustering is carried out based on the similarity according to the data of the delay duration feature vector C in the observation period, and a clustering result with too short time is corrected according to the time continuity of samples in the class, so that a corrected data set is obtained.

Firstly, marking corresponding labels on delay time length feature vectors in the delay time length feature vector sequence to obtain corresponding label sequences; and then correcting the tag sequence to obtain a corresponding corrected tag sequence, and obtaining corrected tags corresponding to the characteristic vectors of each delay time length.

Marking corresponding labels on delay duration feature vectors in the delay duration feature vector sequences to obtain corresponding label sequences, and specifically: classifying the delay time length feature vectors by using a K-MEANS algorithm based on the distance between the delay time length feature vectors to obtain at least two categories, and labeling the delay feature vectors in each category; and obtaining labels corresponding to each delay characteristic vector in the delay time length characteristic vector sequence, and constructing a corresponding label sequence. In the embodiment of the invention, the distance between the delay time feature vectors is L2 distance, and in other embodiments, the practitioner can adjust the distance acquisition method according to the actual situation.

Based on the records of the delay time feature vectors of all traction power supply interval observation periods, according to the L2 distance, the implementer can classify the states of the delay time feature vectors of different situations based on a K-MEANS algorithm, in this embodiment, k=10 is used as the classification target number, and in other embodiments, the implementer can adjust the value according to the actual situation. It should be noted that the number of working conditions that the delay time length feature vector C of the traction power supply section can represent may be less than 10, so that the value of K is set to 10 for covering and classifying as much as possible. It should be noted that, since the K-MEANS algorithm is a well-known technique for those skilled in the art, the clustering principle is not repeated in the present invention. The excessive classification of the delay time length feature vectors is realized through a K-MEANS algorithm, and the delay time length feature vectors of different categories are labeled, so that each category can be assigned with one label and 0-9 total 10 labels because of being divided into 10 categories.

Correcting the tag sequence to obtain a corresponding corrected tag sequence, and obtaining corrected tags corresponding to the characteristic vectors of each delay duration, specifically: and carrying out smoothing treatment on the label sequence to obtain a corresponding corrected label sequence. If the delay time length feature vector sequence { a1, a2, a3, a4}, is a1, a2, a3, a4 is marked with labels b1, b2, b3, b4 respectively, a corresponding label sequence { b1, b2, b3, b4}, that is, the label corresponding to the delay time length feature vector a1 is b1, the label corresponding to the delay time length feature vector a2 is b2, the label corresponding to the delay time length feature vector a3 is b3, and the label corresponding to the delay time length feature vector a4 is b4. And correcting the tag sequences { b1, b2, b3 and b4} to obtain corresponding corrected tag sequences { c1, c2, c3 and c4}, wherein the corrected tag corresponding to the delay duration feature vector a1 is c1, the corrected tag corresponding to the delay duration feature vector a2 is c2, the corrected tag corresponding to the delay duration feature vector a3 is c3 and the corrected tag corresponding to the delay duration feature vector a4 is c4.

Performing smoothing processing on the label sequence, namely performing label smoothing processing on the clustered labels to prevent failure of FCN training, and specifically: the method is characterized in that a time domain smoothing mode is used for processing, and for a traction power supply interval, each moment of a delay duration characteristic vector sequence formed by a delay duration characteristic vector C is provided with a label, namely each element in the delay duration characteristic vector sequence is provided with a label, and when the label of the delay duration characteristic vector C jumps at different label values, the characteristic of oversubsign of the label in the time domain is described. The tag sequence is smoothed using a median filtering algorithm over the tag values of abrupt changes in the tag sequence. The median filtering algorithm is a nonlinear filtering algorithm for removing noise and outliers from the signal. For interrupts or outliers of the tag values in a sequence, smoothing may be performed by a median filtering algorithm. The specific steps of the smoothing process are as follows: firstly, arranging label values corresponding to elements in a delay time length feature vector sequence C into a label sequence to be processed according to a time sequence; setting a window size w, in this embodiment, w=5, that is, the working mode within 5s is regarded as unlikely to be significantly suddenly changed, the window length is approximately equal to the duration of a working condition assumed, and in other embodiments, the practitioner can adjust the window size according to the actual situation. Starting with sample 3, the window w is moved by moving the window one by one and calculating the median of the tag values in the window, which can be said to be the moving step size of 1. It should be noted that, the window is moved from the 3 rd sample, that is, the 3 rd delay duration feature vector in the delay duration feature vector sequence is taken as the center point, and then the window is moved to the left, and the window directly taking the 3 rd delay duration feature vector as the center point is taken as the initial position of the window, so that the initial window can completely contain 5 delay duration feature vectors. For example, for moving a window over a tag sequence when the window size is 5, the 1 st window covers samples 1 through 5, the 2 nd window covers samples 2 through 6, and so on. And taking the median of the five tag values in the window as a corrected tag after the center point of the current window is smoothed. For samples at the last two ends of the tag sequence, a smaller window may be used or truncated to avoid excessive smoothing. For example, when the first element of the tag sequence is taken as the center point of the window, since there is no other element on the left side of the first element, the median of the elements having practical significance in the window can be taken as the corrected tag value of the first element, and the median of the first element, the second element and the third element can be taken as the corrected tag corresponding to the first element; when the second element of the tag sequence is taken as the center point of the window, only one element is left of the second element, so that the median value of the elements with practical significance in the window can be taken as the correction tag value of the second element, and the median values of the first element, the second element, the third element and the fourth element can be taken as the correction tags corresponding to the second element. The label is a label value, and the correction label is a correction label value.

The label correction is carried out on all traction power supply intervals, so that each delay time length characteristic vector has a corresponding correction label.

Further, according to the data set of the correction tag, the delay time length characteristic vector C which belongs to each moment is marked as One-Hot codes of different types. Training a fully-connected neural network (FCN), and training the correction label of the marked delay time length feature vector C so as to identify the correction label corresponding to the delay time length feature vector after the observation period, wherein the correction label can also be said to reflect the type of the delay time length feature vector.

Because the delay time length characteristic vector is a 5-dimensional vector and the context information is too little, the invention uses a high-dimensional vector formed by the first n moments of the delay time length characteristic vector as the input of the FCN network. In the embodiment of the present invention, the value of n is 3, and in other embodiments, the practitioner can adjust the value according to the actual situation.

Therefore, for one training data, the delay time length feature vectors of the current moment and the delay time length feature vectors of the first two moments jointly form a 12-dimensional high-dimensional vector, and the sequence is that

Wherein

The delay time length characteristic vector is the t time;

A delay time length characteristic vector at the t-1 time;

is the characteristic vector of the delay time length at the t-2 time.

And (3) directly performing One-Hot coding of 0-9 on the correction label corresponding to the delay time length feature vector, wherein the delay time length feature vector at each current moment is used as input, the delay time length feature vectors at the first two moments jointly form a 12-dimensional high-dimensional vector, one-Hot coding corresponding to the correction label is used as output, a prediction result is compared with a real correction label, and the value of the cross entropy loss function is calculated. A certain number of training rounds, e.g. 10000epoch, is set by the practitioner until the model converges or reaches a preset number of training rounds. The FCN network can classify the class according to the correction label of the delay time length feature vector, and output the 10 values of 0 or 1 in a one-hot coding form can be obtained through post-processing, wherein only one position is 1, namely the position of the corresponding type.

In addition, the observer also needs to construct an OCSVM identification model based on the delay time length feature vector for each traction power supply interval, judge whether the delay time length feature vector is abnormal or not, and determine a threshold value to be used as a real-time abnormality determiner. The specific training mode is as follows: a portion of the data set of normally distributed delay duration feature vectors is randomly selected for training, and the remaining data is used as test data. Among them, the vector machine model (One-Class SVM, OCSVM) requires selection of kernel functions and parameters for mapping data into a high-dimensional space. Cross-validation was performed using the grid search tool gridsearch cv to select the best kernel and parameters. Parameters kernel and nu represent kernel and penalty factors, respectively. The specific kernel parameter is [ rbf, linear ], and the nu parameter is [0.01,0.05,0.1], wherein rbf is a Gaussian kernel, and linear is a linear kernel function. After training the One-Class SVM model using the selected kernel function and parameters, new data points are input into the trained OCSVM model, and the new data is classified as anomalous based on its distance from the training data. Specifically, if a new data point is more than an anomaly threshold from the center of the training data, it is classified as an anomaly delay duration feature vector. In the embodiment of the present invention, the value of the anomaly threshold is 0.7, and in other embodiments, the practitioner can adjust the value according to the actual situation. Therefore, when the distance to be judged is larger than the abnormal threshold value, the delay time length characteristic vector is taken as an abnormal delay time length characteristic vector; and when the distance to be judged is smaller than or equal to the abnormal threshold value, taking the delay time length characteristic vector as a normal delay time length characteristic vector.

Step S300, working condition feature codes of the traction power supply section are determined according to the delay time length feature vector of the traction power supply section, the similarity between the delay time length feature vector of the traction power supply section and other traction power supply section, and the correction label of the delay time length feature vector.

The method comprises the steps of establishing a real-time adjacent matrix for each traction power supply section in a wide area measurement and control system, carrying out unidirectional representation on delay time length feature vectors in a jurisdiction area of each power supply station, and forming a binary group by the delay time length feature vectors of the two sides, wherein each element in the adjacent matrix is a binary group, for example, an element in a first row and a second column of adjacent matrixes is a binary group formed by delay time length feature vectors of a first traction power supply section and a second traction power supply section, an element in a second row and a third column of adjacent matrixes is a binary group formed by delay time length feature vectors of the second traction power supply section and a third traction power supply section, and an element in the first row and the second column of adjacent matrixes is a binary group formed by delay time length feature vectors of the first traction power supply section and the first traction power supply section, namely 0. The data of the two groups is processed into a working condition characteristic code D, so that the working condition characteristic code D is used as a joint representation of different working condition situations of each power supply station when communication is normal.

For the adjacency matrix, in order to observe the working condition of the traction power supply section of the neighborhood, the working condition feature code D of the current traction power supply section is described in the adjacency matrix in an 8-neighborhood mode.

According to the number of traction power supply intervals, the embodiment of the invention assumes that a wide area protection system needs to manage 10 traction power supply intervals, and then the adjacent matrix comprises 100 binary groups which are compared pairwise, namely each tuple comprises two delay duration feature vectors of the two parties at the moment, and the two delay duration feature vectors are sequentially arranged according to the corresponding object modes of rows and columns. The binary set is pre-information for evaluating the working condition between the two traction power supply sections at this time, and is a format for conveniently representing two necessary information sources because the relevance of the working condition between the two traction power supply sections and the working condition related index are described later. Between any two traction power supply sections, the traction power supply section corresponding to the row is communicated with the traction power supply section corresponding to the column, so that the network quality is further described according to the relation between the uplink and the downlink of the network.

For the eight neighborhoods of any element in the adjacency matrix, the characteristics of communication between one traction power supply interval and a target traction power supply interval and communication between the traction power supply intervals of the geographic position neighborhood of the target traction power supply interval are described. Because the backbone network used in adjacent traction power supply intervals of one traction power supply interval may be common and the ground factor is similar, it may be used for reference representation of adjacent conditions. The eight neighborhoods are areas to which the neighborhoods in 8 directions, which are adjacent in rows and columns and adjacent in diagonal directions, exist at one position in the matrix in each direction. Therefore, in the normal working process, the local 8 neighborhood is subjected to description of the working condition feature code D of the traction power supply section.

Under the condition that the actual condition of the traction power supply section is unknown, the working condition characteristics between the traction power supply section and the neighborhood are described, the integral working condition change and network change conditions can be described, the section potential relation of the adjacent traction power supply sections can be uniformly described according to the prediction error characteristics of the network RTT of each traction power supply section, and the section potential relation can be specifically represented by the characteristics of the RTT and the impedance angle of each traction power supply section.

Characteristic coding of the delay duration characteristic vector is carried out based on a traction power supply interval, and the characteristic vector is used as a working condition characteristic code, and the specific coding mode is as follows: and (5) carrying out numerical value collection of delay time length feature vectors in the traction power supply interval and the neighborhood. Taking a high-dimensional vector formed by the delay time length characteristic vector of the current time and the delay time length characteristic vectors of the first two times of the traction power supply interval as input to obtain a single-heat code corresponding to a correction label of the delay time length characteristic vector of the current time; and constructing a working condition feature code by the delay time length feature vector of the traction power supply section, the minimum cosine similarity of the delay time length feature vector of the traction power supply section and the delay time length feature vector of other traction power supply sections, the independent heat code corresponding to the delay time length feature vector of the traction power supply section, and the shortest time feature of the traction power supply section and the other traction power supply sections in the neighborhood. The shortest time characteristic is the lowest time delay from the traction power supply interval to other traction power supply intervals in the neighborhood, namely the minimum RTT value, namely the shortest time delay duration. In the embodiment of the invention, the traction power supply intervals in the neighborhood of the traction power supply interval are two adjacent traction power supply intervals.

Step S400, determining a sample distance according to the difference of corresponding delay time length feature vectors among traction power supply intervals and the time feature vector of the traction power supply interval; and determining the degree of abnormality of the working condition feature code based on the sample distance.

To prevent cracking of distributed control, the wide area protection system dynamically enables overriding authority for a portion of the traction power supply interval. Based on the adjacency matrix, local abnormal factors of working condition feature codes of each traction power supply interval are counted according to the duration of 1 hour, so that the first n power supply stations with stronger abnormal movements are determined, and the power supply stations can still be actively switched to the standby power supply station under the condition of large network delay. It can also be said that the sample distance is determined according to the difference of the corresponding delay time length feature vectors between the traction power supply intervals and the time feature vector of the traction power supply interval, and the abnormality degree of the working condition feature code is further determined based on the sample distance.

The sample distance acquisition method comprises the following steps: for any two traction power supply intervals, calculating cosine similarity of delay time length feature vectors corresponding to the two traction power supply intervals, and taking the cosine similarity as a first similarity; taking a difference value between a preset first threshold value and a first similarity as a first factor; respectively acquiring two minimum cosine similarities of delay time length feature vectors of the two traction power supply intervals and other traction power supply intervals, and taking the ratio of a larger value to a smaller value in the two minimum cosine similarities as a second factor; respectively acquiring two shortest time features of the traction power supply intervals in the neighborhood corresponding to the traction power supply intervals, and taking the maximum value of the two shortest time features as a third factor; taking the hamming distances of the independent heat codes corresponding to the two traction power supply intervals as a fourth factor; the first factor, the second factor, the third factor and the fourth factor are all in positive correlation with the sample distance. In the embodiment of the present invention, the value of the first threshold is preset to be 1, and in other embodiments, the practitioner can adjust the value according to the actual situation. The sample distance is the sample distance in the acquisition of the local anomaly factor, and the local density can be obtained according to the sample distance, so that the local anomaly factor of the sample can be obtained according to the local density. It should be noted that the steps of obtaining the local abnormality factor (Local Outlier Factor, LOF) of the sample are known to those skilled in the art, and will not be described herein.

First, the LOF value at each time is calculated:

for the working condition feature codes of any traction power supply interval at any time within one hour, each sample traction power supply interval is used as a sample to evaluate local abnormality factors, and the local abnormality factors are used for indicating rareness: the local outlier factor detection method is a method for performing unsupervised outlier detection according to the spatial distribution of samples. The method calculates an outlier LOF for each sample point and determines whether it is an outlier by determining if LOF is close to 1. The greater the LOF is, the more likely it is an outlier factor, the closer to 1, and the more likely it is a normal point. Firstly, a hypothesis space is established, and a spatial distribution relation is established for a sample set through a measurement criterion function Z, namely a sample distance. The distribution relationship between samples D needs to be defined, so as to obtain the distribution relationship of the sample space.

Wherein,,

the sample distance is corresponding between the traction power supply interval a and the traction power supply interval b;

cosine similarity of delay time length feature vectors corresponding to the traction power supply interval a and the traction power supply interval b, namely first similarity corresponding to the traction power supply interval a and the traction power supply interval b; 1 is a preset first threshold value;

A first factor corresponding to a traction power supply interval a and a traction power supply interval b;

the minimum cosine similarity of delay time length feature vectors of the traction power supply interval a and other traction power supply intervals is obtained;

the minimum cosine similarity of delay time length feature vectors of the traction power supply interval b and other traction power supply intervals is obtained;

a second factor corresponding to the traction power supply interval a and the traction power supply interval b;

is a maximum function;

to take the minimum function;

the shortest time characteristic of the traction power supply interval a and other traction power supply intervals is as follows;

the shortest time characteristic of the traction power supply interval b and other traction power supply intervals is as follows;

a third factor corresponding to the traction power supply interval a and the traction power supply interval b;

the single-heat codes corresponding to the traction power supply interval a;

the single heat codes corresponding to the traction power supply interval b;

a fourth factor corresponding to the traction power supply interval a and the traction power supply interval b;

is the hamming distance.

Wherein the sample distance may reflect the distance between the two traction power supply intervals. Any two samples, based on sample distance, can indicate that there are closer distances between samples in the distribution of samples, and there are farther. Therefore, based on the sample distribution, the K-nearest neighbor sub-block samples of any sample in the sample set F in the space can be known by using the K-nearest neighbor algorithm. After the cosine similarity of the two traction power supply sections is calculated, a first factor is obtained according to the cosine similarity, the first factor can also be said to reflect the cosine distance of the two traction power supply sections, the larger the cosine similarity of the two traction power supply sections is, the smaller the corresponding cosine distance is, the smaller the corresponding sample distance between the two traction power supply sections is, otherwise, the smaller the cosine similarity of the two traction power supply sections is, the larger the corresponding cosine distance is, and the larger the corresponding sample distance between the two traction power supply sections is. The ratio of the larger value to the smaller value in the two minimum cosine similarities of the delay time length feature vectors of the two traction power supply intervals and the other traction power supply intervals, namely the second factor, is used as a constraint term of the cosine similarities of the two traction power supply intervals, and if the minimum cosine similarity of the delay time length feature vectors relative to the other traction power supply intervals is smaller at the moment, the fact that the C value has abnormal distribution to a certain degree at the moment is indicated, so that the constraint sample distance is larger. The shortest time feature is the time length of the lowest delay of the two traction power supply intervals to the neighborhood, represents the theoretical performance of the two communication, and is the minimum element value in the time feature vector corresponding to the traction power supply interval and other traction power supply intervals in the neighborhood.

For the number of different bit numbers between the independent thermal codes or H of the delay time length characteristic vectors of the two traction power supply intervals, namely the difference between different types of modes, the larger the fourth factor is, the description of the two different traction power supply interval modes and the corresponding 8 neighborhood traction power supply areasThe greater the pattern difference between.

To this end, a distance relationship between samples can be obtained, thereby calculating the distance reachable density (Local Reachablity density, LRD) and LOF. The sample is the traction power supply section in the step of calculating the local abnormality factor. Based on the calculation of the K-nearest neighbor reachable distance, the local reachable density LRD of each sample point is calculated. It should be noted that, since the calculation of the local reachable density is a well-known technique for those skilled in the art, the details are not repeated here. Then, based on lrd per sample, LOF is calculated: consider a sample point p. Local anomaly factor (LOF) is one method for measuring whether a sample point p is an outlier. The LOF calculates an average of the ratio of the local reachable densities of the K nearest neighbors of the sample point p to the local reachable densities of the sample point p. If the value of LOF is closer to 1, it is indicated that the density of the neighborhood of sample point p is similar to that of sample point p, meaning that sample point p may belong to the same cluster as its neighborhood. Conversely, if the LOF value is less than 1, it indicates that the density of the sample point p is higher than the density of its neighboring points, which means that the sample point p is a dense point. On the other hand, if the LOF value is greater than 1, it is explained that the density of the point p is lower than that of its neighborhood point, which means that the point p may be an abnormal point. Therefore, the LOF size at each time point of each traction power supply section is calculated based on the records of all traction power supply sections within 1 hour. The larger the LOF, the more rare.

If the working condition feature code of one traction power supply section has rare conditions, the load and the communication condition of the traction power supply section are unstable, and if the traction power supply section of the neighborhood is abnormal, tripping is generated, the traction power supply section can possibly generate abnormal tripping protection action, namely false triggering protection action, and LOF is larger. Conversely, the LOF is smaller.

Step S400 may be said to be taking the sample distance as the reachable distance between the working feature codes, calculating the local outlier factor of the working feature code based on the sample distance, and taking the local outlier factor of the working feature code as an index reflecting the degree of abnormality of the working feature code. It should be noted that the outlier factor is a local outlier factor.

And S500, overriding the feeder line in the traction power supply section according to the abnormality degree and the abnormality judgment result.

The method comprises the steps of dynamically endowing the first n power supply stations with temporary override authority when the network is disconnected in real time, and releasing the override authority for the power supply stations which do not belong to the first n power supply stations. Based on the power supply station with authority override, according to the change condition of the abnormality degree and the abnormality judgment result of the output of the abnormality judgment device, the abnormality judgment result is an abnormal delay time length characteristic vector, so that the abnormality judgment result is continuously abnormal within a preset timeout time length, and when the maximum value of the abnormality degree within the preset timeout time length is greater than the maximum abnormality degree of the historical moment, the feeder is overridden. It should be noted that the abnormality determination result is in an abnormality, that is, the delay time length feature vector is an abnormal delay time length feature vector. In the embodiment of the invention, the preset timeout duration is 1s, that is, when the abnormality determination result exceeds 1s and the maximum value of LOF in 1s is greater than the maximum value before 1s, the feeder line of the traction power supply section is overridden.

Since the adjacent traction power supply section may be in a protection state, the traction power supply section only needs to observe the change rate of the impedance angle of the traction power supply section at the moment, if the change rate is exceeded, the load is considered to have a large change in position, and the influence caused by the trip moment due to the protection action of the adjacent traction power supply section is likely to be caused, so that the trip action is not executed in the networking overtime process.

Specifically, the method for overriding the feeder line in the traction power supply section comprises the following steps:

the maximum timeout tolerance time of the timeout time is set by the implementer, and the preset timeout time is 1s and the timeout tolerance time is 5s in the embodiment of the invention.

If the delay time length feature vector is judged as the abnormal delay time length feature vector by the OCSVM in the previous second, no matter whether the trip signal is triggered by the protection circuit or not at the moment, the trip signal is not tripped, because the power supply trip of the adjacent interval possibly causes interference, and the interval cannot be contacted with the wide area control system at the moment. When the time exceeds 5s, the network is not recovered, namely the control signal and the heartbeat packet of the wide area system are not received, and whether tripping occurs is determined according to the tripping signal.

In summary, the present invention relates to the technical field of power supply wide area protection. The method comprises the steps of obtaining a time characteristic vector and an impedance angle corresponding to each traction power supply interval, and constructing a delay time length characteristic vector according to the time characteristic vector and the impedance angle; marking corresponding labels on delay time length feature vectors in the delay time length feature vector sequence to obtain corresponding label sequences; correcting the tag sequence to obtain a corresponding corrected tag sequence, and obtaining corrected tags corresponding to the characteristic vectors of each delay time length; performing abnormality judgment on the delay time length feature vector to obtain an abnormality judgment result; constructing corresponding binary groups of each traction power supply interval according to the delay duration characteristic vectors corresponding to each other among the traction power supply intervals; determining working condition feature codes of the traction power supply interval according to the delay time length feature vector of the traction power supply interval, the similarity between the traction power supply interval and other traction power supply intervals and the correction label of the delay time length feature vector; determining a sample distance according to the difference of the corresponding delay time length characteristic vectors among the traction power supply intervals and the time characteristic vector of the traction power supply interval; determining the abnormality degree of the working condition feature code based on the sample distance; and measuring and controlling the feeder line in the traction power supply section according to the abnormality degree and the abnormality judgment result. The delay of the traditional control network is limited to 0.01s, and the time length of the tolerant network Jitter can be prolonged to be 1s based on the override logic. When other traction power supply intervals cannot be connected within 1s, if tripping conditions occur in adjacent traction power supply intervals, the traction power supply intervals are guaranteed to be in a power supply state until the longest timeout time, so that the work of a main power supply line is protected, and the interval protection after connection recovery still has selectivity.

It should be noted that: the sequence of the embodiments of the present invention is only for description, and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

In this specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments.

Claims

1. A method for adaptively controlling a feed line of a traction power supply wide area protection measurement and control system is characterized by comprising the following steps:

2. The method for adaptively controlling a feeder line of a traction power supply wide area protection measurement and control system according to claim 1, wherein the labeling the delay duration feature vector in the delay duration feature vector sequence to obtain a corresponding label sequence comprises:

3. The method for adaptively controlling a feeder line of a traction power supply wide area protection measurement and control system according to claim 1, wherein the modifying the tag sequence to obtain a corresponding modified tag sequence comprises:

4. The method for adaptively controlling a feeder line of a traction power supply wide area protection measurement and control system according to claim 1, wherein the performing anomaly determination on the delay time length feature vector to obtain an anomaly determination result comprises:

5. The method for adaptively controlling the feeder line of the traction power supply wide area protection measurement and control system according to claim 1, wherein the constructing the delay time length feature vector according to the time feature vector and the impedance angle comprises the following steps:

6. The method for adaptively controlling the feeder line of the traction power supply wide area protection measurement and control system according to claim 1, wherein the determining the working condition feature code of the traction power supply section according to the delay time feature vector of the traction power supply section, the similarity between the delay time feature vectors of the traction power supply section and other traction power supply sections, and the correction label of the delay time feature vector comprises the following steps:

7. The method for adaptively controlling a feeder line of a wide area protection and measurement and control system for traction power supply according to claim 6, wherein determining the sample distance according to the difference of the corresponding delay time length feature vectors between traction power supply intervals and the time feature vector of the traction power supply interval comprises:

8. The method for adaptively controlling the feeder line of the traction power supply wide area protection measurement and control system according to claim 1, wherein the determining the degree of abnormality of the working condition feature code based on the sample distance comprises:

9. The method for adaptively controlling the feeder line of the traction power supply wide area protection measurement and control system according to claim 1, wherein the step of overriding the feeder line of the traction power supply section according to the abnormality degree and the abnormality judgment result comprises the steps of:

10. The method for adaptively controlling a feeder line of a traction power supply wide area protection and measurement and control system according to claim 6, wherein the shortest time is characterized in that: and the minimum element value in the time characteristic vector corresponding to the traction power supply interval and other traction power supply intervals in the neighborhood.