Relay protection system risk assessment method and system based on semi-supervised MD algorithm

Abstract

This invention relates to a risk assessment method and system for relay protection systems based on a semi-supervised MD algorithm. Belonging to the field of power system relay protection technology, it addresses the problems of poor adaptability and self-correction capabilities, as well as strong subjectivity, in existing assessment methods. By establishing an operational status assessment model of the relay protection system based on the analytic hierarchy process (AHP), which serves as the training set for the semi-supervised MD algorithm, weights are calculated, and the operational status of the relay protection device is subjected to ridge function fuzzy processing. Machine learning is then applied to the established training set to eliminate the subjective defects resulting from the fuzzy processing. Compared with other machine learning algorithms, the proposed semi-supervised MD algorithm in this invention describes the correlation between variables without being affected by dimensionality, has significant advantages in handling distance metric learning between multivariate high-dimensional samples, requires the smallest training set size, and provides good and accurate prediction results.

Description

Technical Field

[0001] This invention belongs to the field of power system relay protection technology, and relates to a risk assessment method and system for relay protection systems based on semi-supervised MD algorithm. Background Technology

[0002] In recent years, an increasing number of intelligent power electronic devices and new energy devices have been connected to the power system. Improving the maintenance efficiency of power system equipment is of great significance for the stable operation of the power system. Correct, rapid, and reliable operation of relay protection devices will effectively prevent the expansion of the fault range after a power system failure, thereby effectively improving the stability of the power system and reducing economic losses.

[0003] With the development of big data technology, machine learning is widely used in power systems. For example, the paper "A Method for Status Assessment of Relay Protection Devices Based on GA-Optimized SVM Parameters and Cloud Model" (Chen Haitao et al., Smart Power, 2020, 48(07):88-92+117.) proposes an improved classification method for optimizing support vector machine parameters using a genetic algorithm. It utilizes the randomness and stable tendency of the cloud model to realize the uncertainty mapping between health status and evaluation domain, making the assessment results more consistent with the actual situation. However, existing methods have significant limitations: 1) Advanced algorithms are needed to improve the accuracy of pattern recognition; the assessment methods have poor adaptability and self-correction capabilities; 2) The schemes are highly subjective and lack sufficient data support; 3) The simulation results of some models may not be very helpful for actual operation and maintenance.

[0004] To address the aforementioned shortcomings, this invention proposes a risk assessment method for relay protection systems based on a semi-supervised MD (Mahalanobis distance) machine learning algorithm. Summary of the Invention

[0005] The purpose of this invention is to design a risk assessment method and system for relay protection systems based on a semi-supervised MD algorithm, so as to solve the problems of poor adaptability and self-correction ability and strong subjectivity of existing assessment methods.

[0006] The present invention solves the above-mentioned technical problems through the following technical solutions:

[0007] The risk assessment method for relay protection systems based on the semi-supervised MD algorithm includes the following steps:

[0008] S1. Select the static and real-time parameters of the MC device, merging unit, protection device, and smart terminal as evaluation indicators, calculate the weights of the evaluation indicators, and establish a state evaluation model based on the AHP fuzzy comprehensive algorithm; the specific method for calculating the weights of the evaluation indicators is as follows: use a nine-level judgment table to obtain an importance comparison matrix through pairwise comparison, and process the importance comparison matrix to obtain a judgment matrix; use a security random indicator to check the consistency of the judgment matrix, and then obtain the judgment matrix of the MC device, merging unit, protection device, and smart terminal through transformation matrix, and calculate the evaluation indicator weights of the MC device, merging unit, protection device, and smart terminal;

[0009] S2. Establish a risk assessment model for relay protection systems based on the semi-supervised MD algorithm to overcome the subjective defects of the fuzzy synthesis algorithm.

[0010] The technical solution of this invention establishes an operational status evaluation model for a relay protection system based on the analytic hierarchy process (AHP) as the training set for a semi-supervised machine learning (MD) algorithm. Weights are calculated on this model, and the operational status of the relay protection device is fuzzy-processed using a ridge function. The established training set is then subjected to machine learning to eliminate the subjective defects resulting from the fuzzy processing. Compared to other machine learning algorithms, the semi-supervised MD algorithm proposed in this invention describes the correlation between variables without being affected by dimensionality, has significant advantages in handling distance metric learning between multivariate high-dimensional samples, requires the smallest training set size, and provides good and accurate prediction results.

[0011] Furthermore, the formula for the judgment matrix is ​​as follows:

[0012]

[0013] The largest eigenvalue of the judgment matrix is:

[0014]

[0015] The eigenvectors of the judgment matrix are:

[0016] W = [ω'1, ω'2, ... ω'] n ]

[0017] The comparison coefficient is defined as a. ij =x i / x j , i=1,2,3,...,m; j=1,2,3,...,m.

[0018] Furthermore, the method for performing consistency checks on the judgment matrix using security random indicators is as follows: Consistency is defined as: CR = CI / RI, where RI is a random index, CI is the consistency check standard, CI = (λ-N) / (N-1), where λ is the largest eigenvalue of the judgment matrix, and N is the non-zero eigenvalue of the n-order consistent matrix;

[0019] Furthermore, the method for establishing a state evaluation model based on the AHP fuzzy synthesis algorithm is as follows:

[0020] S11. Considering the fuzzy relationship between the various evaluation index variables, when evaluating the operating status of the relay protection system, the fuzzy relationship of the continuous interval [0, 1] is used to represent the fuzzy feature domain of the relay protection system. The operating status of the relay protection system includes: normal, warning, error and hidden danger; the fuzzy feature domain of the relay protection system includes: index set, weight set and evaluation set.

[0021] S12. Standardize the data in the fuzzy feature domain of the relay protection system, convert the actual values ​​to the interval [0, 1], and classify the state evaluation index into: benefit type, cost type, and interval type according to the exchange method.

[0022] S13. Calculate the membership matrix of each state evaluation index using the ridge distribution;

[0023] S14. Based on statistical data, set up an indicator set, a weight set, and an evaluation set, and use these as a training set to calculate the evaluation results of the protection devices in this training set.

[0024] Furthermore, the method for establishing a risk assessment model for relay protection systems based on the semi-supervised MD algorithm in step S2 to overcome the subjectivity defects of the fuzzy synthesis algorithm is as follows:

[0025] The cost function of the semi-supervised MD algorithm is:

[0026]

[0027] The calculation yields the result that satisfies the following formula: Value:

[0028]

[0029] The population sample G-index is m = 24, and the mean vector of the high-dimensional data sample is μ = (μ1, μ2, ..., μ...). 24 The population sample G and the individual sample X = (x1, x2, ..., x) 24 The MD between ')' is:

[0030] d 2 (X,G)=(X-μ)'∑(X-μ)-1

[0031] The covariance between the evaluation indicators is:

[0032]

[0033] The solution formula is expressed as:

[0034]

[0035] In the formula, N0 is the number of labeled samples; It represents the number of implicit layer nodes; is the labeled sample vector; n is the dimension of the sample vector; yes The class-labeled sample vector; C is the number of output nodes of the network; w i =[w i1 ,…,w in ] T It is the input weight vector between the input node and the implicit layer node; b i β is the offset of node i; g is the activation function of the implicit layer node; β i =[β i1 ,β i2 ,…,β iC ] T N represents the output weight vector between implicit layer nodes and output nodes. e It is the number of expanded samples; It is an extended sample vector; yes The class-labeled sample vectors; H0 is the hidden layer output matrix of the labeled sample set; β is the weight output matrix; T0 is the class label matrix of the labeled samples; H e The hidden layer output matrix for expanding the sample; T e It is the class label matrix of the expanded samples.

[0036] A risk assessment system for relay protection systems based on a semi-supervised MD algorithm includes: an operation status assessment module and a risk assessment module. The operation status assessment module selects static and real-time parameters of the MC device, merging unit, protection device, and intelligent terminal as assessment indicators, calculates the weights of these indicators, and establishes a status assessment model based on the AHP fuzzy comprehensive algorithm. The method for calculating the weights of the assessment indicators is as follows: a nine-level judgment table is used to obtain an importance comparison matrix through pairwise comparisons, and the importance comparison matrix is ​​processed to obtain a judgment matrix. A safety random index is used to check the consistency of the judgment matrix, and then a transformation matrix is ​​used to obtain the judgment matrices for the MC device, merging unit, protection device, and intelligent terminal. The weights of the assessment indicators for the MC device, merging unit, protection device, and intelligent terminal are calculated. The risk assessment module is used to establish a risk assessment model for relay protection systems based on a semi-supervised MD algorithm to overcome the subjective defects of the fuzzy comprehensive algorithm.

[0037] Furthermore, the formula for the judgment matrix is ​​as follows:

[0038]

[0039] The largest eigenvalue of the judgment matrix is:

[0040]

[0041] The eigenvectors of the judgment matrix are:

[0042] W = [ω'1,ω'2,…ω'] n ]

[0043] The comparison coefficient is defined as a. ij =x i / x j , i=1,2,3,...,m; j=1,2,3,...,m;

[0044] Furthermore, the method for performing consistency checks on the judgment matrix using security random indicators is as follows: Consistency is defined as: CR = CI / RI, where RI is a random index, CI is the consistency check standard, CI = (λ-N) / (N-1), where λ is the largest eigenvalue of the judgment matrix, and N is the non-zero eigenvalue of the n-order consistent matrix.

[0045] Furthermore, the method for establishing a state evaluation model based on the AHP fuzzy synthesis algorithm is as follows:

[0046] (1) Considering the fuzzy relationship between the various evaluation index variables, when evaluating the operating status of the relay protection system, the fuzzy relationship of the continuous interval [0, 1] is used to represent the fuzzy feature domain of the relay protection system. The operating status of the relay protection system includes: normal, warning, error and hidden danger; the fuzzy feature domain of the relay protection system includes: index set, weight set and evaluation set.

[0047] (2) Standardize the data of the fuzzy feature domain of the relay protection system, convert the actual values ​​to the interval [0, 1], and classify the state evaluation index into benefit type, cost type and interval type according to the exchange method;

[0048] (3) The membership matrix of each state evaluation index is calculated using the ridge distribution;

[0049] (4) Based on statistical data, set up an index set, a weight set, and an evaluation set, and use these as a training set to calculate the evaluation results of the protection device in this training set.

[0050] Furthermore, the method for establishing a risk assessment model for relay protection systems based on a semi-supervised MD algorithm to overcome the subjectivity defects of fuzzy synthesis algorithms is as follows:

[0051] The cost function of the semi-supervised MD algorithm is:

[0052]

[0053] The calculation yields the result that satisfies the following formula: Value:

[0054]

[0055] The population sample G-index is m = 24, and the mean vector of the high-dimensional data sample is μ = (μ1, μ2, ..., μ...). 24 The population sample G and the individual sample X = (x1, x2, ..., x) 24 The MD between ')' is:

[0056] d 2 (X,G)=(X-μ)'∑(X-μ) -1

[0057] The covariance between the evaluation indicators is:

[0058]

[0059] The solution formula is expressed as:

[0060]

[0061] In the formula, N0 is the number of labeled samples; It represents the number of implicit layer nodes; is the labeled sample vector; n is the dimension of the sample vector; yes The class-labeled sample vector; C is the number of output nodes of the network; w i =[w i1 ,…,w in ] T It is the input weight vector between the input node and the implicit layer node; b i β is the offset of node i; g is the activation function of the implicit layer node; β i =[β i1 ,β i2 ,…,β iC ] T N represents the output weight vector between implicit layer nodes and output nodes. e It is the number of expanded samples; It is an extended sample vector; yes The class-labeled sample vectors; H0 is the hidden layer output matrix of the labeled sample set; β is the weight output matrix; T0 is the class label matrix of the labeled samples; H e The hidden layer output matrix for expanding the sample; T e It is the class label matrix of the expanded samples.

[0062] The advantages of this invention are:

[0063] The technical solution of this invention establishes an operational status evaluation model for a relay protection system based on the analytic hierarchy process (AHP) as the training set for a semi-supervised machine learning (MD) algorithm. Weights are calculated on this model, and the operational status of the relay protection device is fuzzy-processed using a ridge function. The established training set is then subjected to machine learning to eliminate the subjective defects resulting from the fuzzy processing. Compared to other machine learning algorithms, the semi-supervised MD algorithm proposed in this invention describes the correlation between variables without being affected by dimensionality, has significant advantages in handling distance metric learning between multivariate high-dimensional samples, requires the smallest training set size, and provides good and accurate prediction results. Attached Figure Description

[0064] Figure 1 This is a network topology diagram of a relay protection system based on a semi-supervised MD algorithm for risk assessment of relay protection systems according to an embodiment of the present invention.

[0065] Figure 2 This is a triangular distribution membership function diagram of the risk assessment method for relay protection systems based on the semi-supervised MD algorithm according to an embodiment of the present invention;

[0066] Figure 3This is a ridge distribution membership function graph of the risk assessment method for relay protection systems based on the semi-supervised MD algorithm according to an embodiment of the present invention.

[0067] Figure 4 This is an evaluation diagram of the operation status of a 110kV circuit I transformer protection in a certain area, based on the risk assessment method of the relay protection system based on the semi-supervised MD algorithm according to an embodiment of the present invention. (a) is the index diagram without considering weights, and (b) is the index diagram with weights.

[0068] Figure 5 This is a residual diagram of the risk assessment method for relay protection systems based on the semi-supervised MD algorithm according to an embodiment of the present invention;

[0069] Figure 6 This is a modified residual diagram of the risk assessment method for relay protection systems based on the semi-supervised MD algorithm according to an embodiment of the present invention.

[0070] Figure 7 This is a graph of the ROC evaluation curve of the risk assessment method for relay protection system based on semi-supervised MD algorithm according to an embodiment of the present invention.

[0071] Figure 8 This is a comparative analysis diagram of machine learning algorithms for the risk assessment method of relay protection system based on semi-supervised MD algorithm according to an embodiment of the present invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0074] Example 1

[0075] 1. Establish an operational status assessment model for relay protection systems based on the analytic hierarchy process (AHP).

[0076] 1.1 Selection of Evaluation Indicators

[0077] According to the DL / T860 engineering implementation standard, the network topology of a smart substation for relay protection systems is as follows: Figure 1 As shown, it is divided into three layers: station control layer, interval layer and process layer.

[0078] Station control layer: The station control layer includes background monitors, remote controllers, clock systems, communication power systems, etc., and mainly realizes the monitoring functions of operating status, alarm information, CPU, memory utilization, hard disk utilization, network interface, etc.

[0079] Bay layer: Bay layer equipment includes protection devices, measurement-control (MC) devices, fault recorders, network analyzers, etc. Its main functions are monitoring A / D converters, General Object-Oriented Substation Events (GOOSE), sampled values ​​(SV), and I / O.

[0080] Process layer: Composed of merging units, intelligent terminals, mutual inductors, etc., its main task is to collect information such as clock matching, chip status, and optical power.

[0081] After extensive research, and with reference to the "Guidelines for Secondary Equipment Status Assessment and Risk Assessment" and the "Information Standard for Relay Protection of Smart Substations" of China Southern Power Grid, the static and real-time parameters of MC devices, merging units, protection devices and smart terminals that best reflect the operating status of the relay protection system were selected as the modeling objects.

[0082] The MC device is of great significance for dispatchers to grasp substation information and control it. Its evaluation indicators include 28 indicators: ROM fault, RAM fault, output fault, input fault, parameter setting fault, protection pressure plate fault, EEPROM fault, serial communication interruption, system configuration error, CPU communication interruption, configuration table error, logic table error, inconsistent indication, indicator light fault, system operation fault, circuit breaker tripping without cause, communication abnormality, optical power abnormality, crystal oscillator frequency deviation, GOOSE alarm information (abnormal), GOOSE alarm information (critical), GOOSE alarm information (error), SV alarm information (abnormal), SV alarm information (critical), SV alarm information (error), MMS alarm information (abnormal), MMS alarm information (critical), MMS alarm information (error).

[0083] The merging unit can acquire and transmit voltage and current data in a time sequence. It includes 18 evaluation indicators: ROM fault, RAM fault, output fault, input fault, parameter setting fault, protection pressure plate fault, EEPROM fault, serial communication interruption, system configuration error, CPU communication interruption, configuration table error, logic table error, inconsistent indication, indicator light fault, system operation fault, circuit breaker tripping without cause, communication abnormality, and optical power abnormality.

[0084] The function of the integrated protection device for high, medium and low voltage lines is to disconnect faulty equipment and prevent the fault from escalating. It includes 24 evaluation indicators: ROM fault, RAM fault, output fault, input fault, parameter setting fault, protection pressure plate fault, EEPROM fault, serial communication interruption, system configuration error, CPU communication interruption, configuration table error, logic table error, inconsistent indication, indicator light fault, system operation fault, circuit breaker tripping without cause, communication abnormality, optical power abnormality, crystal oscillator frequency deviation, GOOSE alarm information (abnormal), GOOSE alarm information (critical), GOOSE alarm information (error), SV alarm information (abnormal), SV alarm information (critical), SV alarm information (error).

[0085] The intelligent terminal is responsible for circuit breaker commands and device location, and includes 23 evaluation indicators: ROM fault, RAM fault, output fault, input fault, parameter setting fault, protection pressure plate fault, EEPROM fault, serial communication interruption, system configuration error, CPU communication interruption, configuration table error, logic table error, inconsistent indication, indicator light fault, system operation fault, circuit breaker tripping without cause, communication abnormality, optical power abnormality, crystal oscillator frequency deviation, GOOSE alarm information (abnormal), GOOSE alarm information (critical), GOOSE alarm information (error), SV alarm information (abnormal), SV alarm information (critical).

[0086] 1.2 Calculation of Indicator Weight Values

[0087] 1.2.1 Evaluation Indicators for Relay Protection Systems

[0088] The performance evaluation indicators for relay protection systems consist of dynamic and static indicators. The dynamic indicators for intelligent relay protection devices are derived from self-test information, primarily including memory, setting values, signal inputs and outputs, relay protection boards, communication interfaces, and CPUs. Static indicators are derived from historical operation records, mainly including alarm information records from GOOSE web, SV web, and MMS web. Alarm information is categorized as normal, alarm, fault, and critical.

[0089] 1.2.2 Judgment Matrix

[0090] The importance of each indicator was calculated using a nine-level judgment table as shown in Table 1, where "1" represents x. i and x j Equally important; "3" represents x i Its importance is slightly higher than x j "5" represents x i Compared to x j Important; "7" represents x i Compared to x jMore importantly; "9" represents x i Its importance far exceeds x j .

[0091] Table 1. Level 9 Judgment Table

[0092]

[0093] To obtain the degree of influence of indicator X, which contains m indicators, on indicator Y, an importance comparison matrix is ​​established by pairwise comparisons.

[0094] The comparison coefficient is defined as a ij =x i / x j ,go through This comparison process yields the judgment matrix A:

[0095]

[0096] The largest eigenvalue λ and the corresponding eigenvector W of the judgment matrix A are obtained according to formulas (2)-(6):

[0097]

[0098] M i The nth root is:

[0099]

[0100] Will Normalization yields:

[0101]

[0102] The largest eigenvalue can be expressed as:

[0103]

[0104] The feature vector is represented as:

[0105] W = [ω'1, ω'2, ... ω'] n (6)

[0106] 1.2.3 Weight Calculation and Consistency Check

[0107] The weight calculation result is considered correct only if the judgment matrix satisfies the consistency check criterion. The consistency check criterion is defined as follows:

[0108] CI=(λ-N) / (N-1) (7)

[0109] Where λ is the largest eigenvalue of the n-order positive reciprocal matrix A; N is the non-zero eigenvalue of the n-order uniform matrix.

[0110] Consistency radio (CR) is defined as:

[0111] CR = CI / RI (8)

[0112] RI is a stochastic index, and the random indicators for security are shown in Table 2.

[0113] Table 2 Random Index Table

[0114] N RI N RI N RI N RI 1 0 8 1.41 15 1.59 22 1.64 2 0 9 1.46 16 1.60 23 1.65 3 0.58 10 1.49 17 1.61 24 1.65 4 0.90 11 1.52 18 1.61 25 1.66 5 1.12 12 1.54 19 1.62 26 1.66 6 1.24 13 1.56 20 1.63 27 1.66 7 1.32 14 1.58 21 1.64 28 1.67

[0115] If CR < 0.10, the consistency check of the judgment matrix is ​​considered acceptable. The CR calculation results for the equipment within the station are as follows: CR m-c =0.0069, CR m =0.0123, CR p =0.0118, CR i =0.0122, which meets the consistency check standard.

[0116] The judgment matrices for different devices can be obtained through transformation matrices:

[0117]

[0118]

[0119]

[0120]

[0121] The index weights of different devices are calculated using formulas (2)-(6), as shown in Table 3.

[0122] Table 3 Calculation of Index Weights for Different Equipment

[0123]

[0124]

[0125] 1.3 Fuzzy Comprehensive Evaluation

[0126] Considering the fuzzy relationships between various indicator variables, the actual operation of relay protection cannot be simply described using binary logic of 0-1 true / false. Instead, it should be represented by fuzzy relationships within the continuous interval [0,1]. The fuzzy feature domain of relay protection includes an indicator set, an indicator weight set, and an evaluation set. To standardize the data, the actual values ​​are converted to the interval [0,1]. Based on the exchange method, the status evaluation indicators are divided into three types: benefit-type, cost-type, and interval-type. These three types correspond to three evaluation methods: 1. Benefit-type: the larger the indicator value, the better. 2. Cost-type: the smaller the value, the better. 3. Interval-type: the value should be within the ideal interval.

[0127] In relay protection systems, operating states are generally classified into four levels: progress, warning, error, and critical state. Comparing trigonometric function distributions and ridge function distributions, we find that the ridge distribution provides more membership information when the index is at a high membership degree position. Conversely, at lower membership degrees, the ridge distribution eliminates less membership information. The fuzzy membership degree of the trigonometric distribution tends to yield conservative results; therefore, the ridge distribution is chosen to calculate the membership matrix for each state.

[0128] Finally, this paper takes 80 devices from the statistical data of China Southern Power Grid Company from 2017 to 2020 as examples for analysis. An evaluation index set with 24 indicators and 4 evaluation level types was set; an evaluation set divided into four levels: normal, alarm, fault, and potential hazard; and a weight set for the 80 devices. This was used as the training set to calculate the evaluation results of the protection devices in this training set.

[0129] Given the fuzzy relationships between variables, the actual operation of a relay protection system cannot be simply described using 0-1 logic when evaluating its operational status. Fuzzy relationships should be represented by the continuous interval [0,1], rather than by binary logic of 0 (false) and 1 (true).

[0130] 1.3.1 Fuzzy Comprehensive Evaluation Feature Domain

[0131] The fuzzy feature domain of a relay protection system includes an index set, a weight set, and an evaluation set, denoted by I = {i1, i2, ..., i...}. n} represents the set of indicators for the condition assessment of the relay protection system; the weight values ​​of the indicators are calculated using the analytic hierarchy process (AHP), and are represented as W = {w1, w2, ..., w...} n Assume the evaluation set consists of... m The evaluation results are composed of several elements, which can be represented as E = {e1, e2, ..., e}. n}

[0132] 1.3.2 Data Standardization Processing

[0133] To standardize the data, the actual values ​​are converted to the range [0,1]. Based on the exchange method, the state evaluation indicators are divided into three categories: benefit type, cost type, and range type.

[0134] Benefit-oriented: Benefit-oriented means the higher the value, the better. Its transformation function is g. b (x) can be expressed as

[0135]

[0136] Where x is the actual running value, [x min ,x max [] represents the corresponding lower and upper limits.

[0137] Cost type: For cost types, the smaller the value, the better. Cost type g c The transformation function of (x) is

[0138]

[0139] Where x is the actual running value, [x min ,x max [] represents the corresponding lower and upper limits.

[0140] Interval type: Interval type indicates that the value should be within an ideal interval. Interval type g i The transformation function of (x) is

[0141]

[0142] Where x is the actual running value, [x i_min ,x i_max ] is the ideal running interval, [x min ,x max [ ] is the required running interval.

[0143] 1.3.3 Fuzzy Comprehensive Membership Function

[0144] Triangular and ridge membership functions have been widely used in engineering practice. The operating states of relay protection systems are divided into four levels: normal, warning, error, and potential hazard. The triangular membership functions for different states are as follows:

[0145]

[0146]

[0147]

[0148]

[0149] Combining equations (16)-(19), the triangular distribution membership function calculation model is as follows: Figure 2 As shown.

[0150] The membership functions of the ridge distribution in different states are:

[0151]

[0152]

[0153]

[0154]

[0155] Combining formulas (20)-(23), the ridge distribution membership function calculation model is as follows: Figure 3 As shown.

[0156] from Figure 2 and Figure 3 It can be seen that, compared with the membership function of the triangular distribution, the ridge distribution provides higher membership information when the index is at a high membership position. However, at lower membership positions, the ridge distribution eliminates less membership information. The fuzzy membership function of the triangular distribution tends to yield conservative results. Considering the above analysis, this invention uses the ridge distribution to calculate the membership matrix for each state.

[0157] 1.3.4 Obtaining the training set

[0158] The data used in this invention are taken from the statistical data of China Southern Power Grid Company from 2017 to 2020, including mobile communication equipment, combined generating units, protection devices, and smart terminals. Eighty devices, such as the 110kV Qingshan Substation and the 110kV Ziguang Substation in Guizhou Province, are analyzed as examples.

[0159] The number of indicators used to evaluate protection devices is n. p =24; Assessment level type m p =4; Index setting I p ={Device failure, ROM failure, parameter setting failure, ...}; Evaluation set is E p = {Normal, Alarm, Fault, Potential Problem}; Weight set is W p ={0.1049,0.1049,0.1049,0.0057,…,0.1049}.

[0160] The data standardization processing standards for protection devices are shown in Table 4. The index database shown in Table 4 comes from the station control layer, bay layer, and process layer: remote control personnel perceive operating status and warnings, and data interaction between station levels is achieved through DL / 860; MC devices and protection devices collect data from the MMS network, and switches acquire data through DL / T860; self-monitoring information from intelligent terminals and merging units acquired by device chips is sent to MC devices via GOOSE. Data collected by MC devices is uploaded to the station-level online status monitoring module via MMS. The network analyzer acquires and analyzes GOOSE and SV information at the process level.

[0161] Table 4 Standardization Standards for Protection Device Data

[0162]

[0163]

[0164] Taking a specific device within the station on June 1, 2020 as an example, the operational status assessment calculation for the 110kV circuit I protection of the Qingshan transformer is as follows: Figure 4 As shown. Figure 4 a shows the correspondence between each indicator of the protection device. Figure 4 b shows the relationship after considering the weights. The dashed line represents the upper limit of the ideal upper limit of the evaluation index, and the solid line represents the actual evaluation result.

[0165] Taking a certain device as an example on June 1, 2020, the operational status assessment of the 110kV I-circuit protection of the Qingshan substation was calculated as follows: Figure 4 As shown (the dashed line represents the upper limit of the ideal upper limit of the evaluation index. The solid line represents the actual evaluation result).

[0166] The training set provided for the machine learning algorithm consists of evaluation results of 80 protection devices, as shown in Table 5.

[0167] S represents the size of the machine learning training set, indicating the proportion of the entire sample. The training set contains samples with the sequence [1, 8000s]. Table 5 shows the test set consisting of samples with the sequence [8000s + 1, 8000].

[0168] Table 5 Evaluation Results of Protection Devices

[0169] Assessment time Indicator 1 … Indicator 24 Score 2020.06.01 0 … 4 94.99 2020.06.13 0 … 10 90.93 2020.07.02 0 … 5 88.36 … … … … … 2020.07.06 0 … 20 94.29

[0170] 2. Machine learning algorithm for the operating status of relay protection system

[0171] The AHP fuzzy comprehensive evaluation method was used to evaluate four devices, with d being the number of each device. i ={80,10,24,33}. The evaluation period is c. i pi d represents the proportion of the total sample. i c i p i d represents the number of training sets per device. i c i (1-p i () refers to the number of test sets for each type of device. To overcome the subjectivity inherent in fuzzy synthesis algorithms, this invention employs a novel machine learning method to make the calculation results more objective and realistic.

[0172] 2.1 Supervised Multiple Regression Analysis Algorithm

[0173] For different types of equipment, the multiple regression equations for protection devices, MC devices, merging devices, and smart terminals are as follows:

[0174] Y i =f(X) i B i ,ε i (i = 1, 2, 3, 4) (24)

[0175] Where Y i These are the evaluation results from different devices; X i It is an evaluation indicator; B i ε is the regression coefficient; i It is random error.

[0176] To minimize the sum of squares over the historical evaluation period, the following optimization objective function is proposed:

[0177]

[0178] Among them, Y 1j The observed values ​​of the protection device; It is an estimated value for the protection device.

[0179] Considering the principle of extrema, the constraint function is:

[0180]

[0181] Where j is the training set number.

[0182] Using statistical data from 100 evaluation cycles of 80 protection devices as a training set (p1 = 80%), a supervised multiple regression analysis algorithm based on the training set can be used to obtain a residual level map, such as... Figure 5 As shown.

[0183] When the significance level α = 0.05, there are 308 outlier points in the residual plot, such as... Figure 5 As shown, the pass rate is 95.19%.

[0184] After removing the red outliers in the graph Figure 6 The image shows the corrected residual plot after data cleaning. The number of outliers in the corrected regression model is 274, and the pass / fail rate has improved to 95.50%.

[0185] 2.2 Unsupervised k-means algorithm based on ED

[0186] Taking a protection device as an example, an unsupervised k-means algorithm based on Euclidean distance is used to process unlabeled data, and the evaluation index is X1={x1,x2,…,x 24}

[0187] The difference between samples a and b in the training set is:

[0188]

[0189] Where n = 24, the objective function is:

[0190]

[0191] Where k is the cluster number; n k It is the sample number; m k It is the average value of the categorized samples.

[0192] The unsupervised k-means algorithm based on ED can be described as follows:

[0193] (1) The initial number of center clusters is k = 4. To select the initial center point of the protection device training set, the two samples x that are furthest apart from each other are selected from the training set. i1 and x i2 As the first and second initial cluster points that satisfy the objective function:

[0194]

[0195] Where x ij and x il These are any two samples in the training set T. Then, a third initial cluster point is selected that satisfies the following conditions:

[0196]

[0197] Similarly, select the fourth initial cluster point x that satisfies the following conditions. i4 :

[0198]

[0199] Therefore, the initial values ​​of the training set are expressed as:

[0200] (2) Initial classification is obtained through ED calculation.

[0201] (3) Recalculate the new cluster centers based on the initial classification results obtained in step (2).

[0202] (4) Obtain new clustering results that satisfy the condition function:

[0203] (5) Repeat the above steps t times until two consecutive calculations are identical, satisfying the following condition: If Then t = t + 1 and return to step (3); if End the iteration process and record the final grouping results.

[0204] 2.3 Semi-supervised MD algorithm

[0205] Compared to ED, the MD-based algorithm can utilize the covariance σ ij =cov(x i ,x j This method describes the correlation between variables without being affected by dimensionality. It has advantages in learning distance metrics between multivariate, high-dimensional samples.

[0206] To fully utilize the useful information contained in unlabeled data, this invention proposes Online Sequential ELM-MD (OSELM-MD). By using a semi-supervised learning algorithm with reversible reasoning capabilities, correctly labeled samples are applied to the expansion of new half-labeled samples, incorrectly labeled training samples are identified as invalid samples, and then the parameters are corrected.

[0207] The training set T contains a labeled set T1 and an unlabeled set T2. The cost function of the semi-supervised OSELM-MD is expressed as:

[0208]

[0209] In the formula, N0 is the number of labeled samples; It represents the number of implicit layer nodes; is the labeled sample vector; n is the dimension of the sample vector; yes The class-labeled sample vector; C is the number of output nodes of the network; w i =[w i1 ,…,w in ] T It is the input weight vector between the input node and the implicit layer node; b i β is the offset of node i; g is the activation function of the implicit layer node; β i =[β i1,β i2 ,…,β iC ] T N represents the output weight vector between implicit layer nodes and output nodes. e It is the number of expanded samples; It is an extended sample vector; yes The class-labeled sample vector.

[0210] Through calculation, the result satisfies the equation (33).

[0211]

[0212] In the formula, H0 is the hidden layer output matrix of the labeled sample set; β is the weight output matrix; T0 is the class label matrix of the labeled samples; H e The hidden layer output matrix for expanding the sample; T e It is the class label matrix of the expanded samples.

[0213] The population sample G-index is m = 24, and the mean vector of the high-dimensional data sample is μ = (μ1, μ2, ..., μ...). 24 The population sample G and the individual sample X = (x1, x2, ..., x) 24 The MD between ')' is:

[0214] d 2 (X,G)=(X-μ)'∑(X-μ) -1 (34)

[0215] The covariance between the evaluation indicators is:

[0216]

[0217] The solution formula is expressed as:

[0218]

[0219] 2.4 Comparative Analysis

[0220] Compared to ED, the MD-based algorithm can utilize the covariance σ ij =cov(x i ,x j This method describes the correlation between variables without being affected by dimensionality. It has advantages in learning distance metrics between multivariate, high-dimensional samples.

[0221] To verify the effectiveness and accuracy of these algorithms, the supervised, unsupervised, and semi-supervised machine learning methods described above were compared and analyzed. To ensure the fairness of the comparative analysis, simulation tests were performed on the same computer using the same sample.

[0222] Taking a protection device as an example, when the size of the training set is s = 80%, the ROC (Receiver-Rising Characteristic) evaluation curves of the three algorithms are as follows: Figure 7 As shown in the figure (the black dashed line is the random reference curve, which is the reference curve for 50% probability classification), the closer the ROC curve is to the upper left corner, the closer the area under the curve (AUC) is to 1, meaning that the corresponding machine learning method has higher predictive performance. Through calculation and analysis, the supervised multivariate regression analysis algorithm (blue curve) has an AUC value of 0.6084 ∈ (0.5, 1), indicating that its prediction results are better than random predictions and have some reference value. The unsupervised k-means algorithm based on ED (yellow curve) has an AUC value of 0.4952–0.5, meaning that its prediction results are similar to random predictions and have no reference value. The semi-supervised MD learning algorithm has an AUC value of 0.9464–1, meaning that the algorithm's prediction results are very good and relatively accurate. Table 6 shows the comparative analysis results of the three algorithms in terms of accuracy, processing time, and reliability.

[0223] Table 6 Evaluation Results of Protection Devices

[0224]

[0225]

[0226] As shown in Table 6, the supervised multivariate regression analysis algorithm takes 1.8864–2.0159 s to complete, the unsupervised k-means algorithm based on ED takes 2.0135–2.9846 s, while the semi-supervised MD learning algorithm requires 460.1249–501.3124 seconds to obtain results. Compared with the semi-supervised MD learning algorithm, the supervised multivariate regression analysis algorithm and the unsupervised k-means algorithm based on ED have shorter simulation computation times, but their accuracy is much lower than that of the semi-supervised MD learning algorithm (AUC value around 0.95, significantly better than the other two methods).

[0227] Furthermore, considering that the size of the training set is one of the main factors affecting the accuracy of machine learning algorithms, nine training subsets with sizes between 10% and 90% were selected from the same sample set. The accuracy analysis results are as follows: Figure 8 As shown. Observing the changing trends of regression analysis algorithms and semi-supervised MD learning algorithms, it can be seen that the size of the training set affects the computational accuracy. When the sample size is insufficient, the accuracy is sensitive to the scale of the training set. Δr=1-r i / r max(i = 10, 20, ..., 90%) represents the sensitivity threshold of the algorithm. When the training set size exceeds 70%, the evaluation accuracy of Algorithm 1 (multivariate regression analysis algorithm) becomes insensitive to the training set size (Δr1 < 5%). When the training set size exceeds 60%, the evaluation accuracy of Algorithm 3 (semi-supervised MD learning algorithm) becomes insensitive to the training set size (Δr3 < 5%). Looking at the accuracy trend of Algorithm 2 (unsupervised k-means algorithm based on ED), its accuracy is not limited by the training set size. As can be seen from the figure, Algorithm 3 has the lowest requirement for training set size. When the scale reaches 60%, the requirement for excellent evaluation accuracy is met.

[0228] 3. Conclusion

[0229] In summary, this invention first provides a training set for machine learning methods, establishing state assessment models based on the AHP fuzzy synthesis algorithm for MC devices, merging units, protection devices, and intelligent terminals. Secondly, to make the assessment results more objective and accurate, supervised multivariate regression analysis, unsupervised k-means, and semi-supervised MD algorithms are applied to the state assessment of relay protection systems, and comparisons are made. Simulation results show that the machine learning algorithm proposed in this invention has high accuracy and low data size requirements.

[0230] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for risk assessment of a protection system based on semi-supervised MD algorithm, characterized in that, It comprises the following steps: S1, selecting the static and real-time parameters of the M-C device, merging unit, protection device and intelligent terminal as evaluation indexes, calculating the weight of the evaluation indexes, and establishing a state evaluation model based on AHP fuzzy comprehensive algorithm; the method for calculating the weight of the evaluation indexes is as follows: using a nine-level judgment table to obtain an importance comparison matrix through pairwise comparison, and processing the importance comparison matrix to obtain a judgment matrix; checking the consistency of the judgment matrix by using a safety random index, and obtaining the judgment matrix of the M-C device, merging unit, protection device and intelligent terminal through a transformation matrix, and calculating the evaluation index weight of the M-C device, merging unit, protection device and intelligent terminal; S2, a risk assessment model of the relay protection system based on semi-supervised MD algorithm is established to overcome the subjective defects of the fuzzy comprehensive algorithm, and the specific method is as follows: The cost function of the semi-supervised MD algorithm is: The value of the following equation is calculated to satisfy 0.0001. The G index of the overall sample is m = 24, the mean vector of the high-dimensional data sample is μ = (μ1, μ2, …, μ 24 ), and the MD between the overall sample G and the single sample X = (x1, x2, …, x 24 ) is: d 2 (X, G) = (X - μ)'∑(X - μ) -1 The covariance between the evaluation indexes is: The solution formula is: where N0is the number of labeled samples; is the number of hidden layer nodes; is the labeled sample vector; n is the dimension of the sample vector; is is the class-labeled sample vector; C is the number of output nodes of the network; w i = [w i1 ,…,w in ] T is the input weight vector between the input nodes and the hidden layer nodes; b i is the bias of node i; g is the activation function of the hidden layer nodes; β i = [β i1 , β i2 ,..., β iC ] T is an output weight vector between the hidden layer node and the output node; N e is the number of extended samples; is an extended sample vector; is a class label sample vector of ; H0 is a hidden layer output matrix of the labeled sample set; β is an output matrix of the weight; T0 is a class label matrix of the labeled sample; H e is a hidden layer output matrix of the extended sample; T e is a class label matrix of the extended sample.

2. The method of risk assessment of a protective relaying system based on semi-supervised MD algorithm according to claim 1, wherein, The formula of the judgment matrix is as follows: The maximum eigenvalue of the judgment matrix is: The eigenvector of the judgment matrix is: W = [ω1', ω'2,... ω'N] n ] wherein the comparison coefficient is defined as a ij = x i / x j , i = 1, 2, 3,..., m; j = 1, 2, 3,..., m.

3. The method of risk assessment of a protective relaying system based on semi-supervised MD algorithm according to claim 2, wherein, The method for checking the consistency of the judgment matrix by using a safety random index is as follows: consistency definition is: CR=CI / RI, RI is a random index, CI is a consistency check standard, CI=(λ-N) / (N-1), λ is the maximum eigenvalue of the judgment matrix, and N is the non-zero eigenvalue of an n-order consistent matrix.

4. The method of risk assessment of a protective relaying system based on semi-supervised MD algorithm according to claim 3, wherein, The method for establishing the state evaluation model based on AHP fuzzy comprehensive algorithm is as follows: S11, considering the fuzzy relationship between the evaluation index variables, the fuzzy relationship of the continuous interval [0, 1] is used to represent the fuzzy feature domain of the relay protection system when evaluating the operation state of the relay protection system, and the operation state of the relay protection system includes: normal, warning, error and hidden danger; the fuzzy feature domain of the relay protection system includes: index set, weight set and evaluation set; S12, the data of the fuzzy feature domain of the relay protection system is standardized, the actual value is converted to the interval [0, 1], and the state evaluation indexes are divided into: benefit type, cost type and interval type according to the exchange mode; S13, the membership degree matrix of each state evaluation index is calculated by using the ridge distribution; S14, according to the statistical data, setting the index set, weight set and evaluation set, and taking them as a training set, the protection device evaluation result of the training set is calculated.

5. A risk assessment system for a protective relaying system based on a semi-supervised MD algorithm, characterized in that, It comprises: an operation state evaluation module and a risk assessment module; The running state evaluation module is used for selecting static and real-time parameters of the M-C device, the merging unit, the protection device and the intelligent terminal as evaluation indexes, performing weight calculation on the evaluation indexes, and establishing a state evaluation model based on an AHP fuzzy comprehensive algorithm; the weight calculation method on the evaluation indexes is as follows: an important degree comparison matrix is obtained through pairwise comparison by using a nine-level judgment table, and the judgment matrix is obtained by processing the important degree comparison matrix; the judgment matrix is checked for consistency by using a safety random index, and the judgment matrix of the M-C device, the merging unit, the protection device and the intelligent terminal is obtained by using a transformation matrix, and the evaluation index weight of the M-C device, the merging unit, the protection device and the intelligent terminal is calculated; The risk evaluation module is used for establishing a relay protection system risk evaluation model based on a semi-supervised MD algorithm to overcome the subjective defects of the fuzzy comprehensive algorithm; The method for establishing the relay protection system risk evaluation model based on the semi-supervised MD algorithm to overcome the subjective defects of the fuzzy comprehensive algorithm is as follows: The cost function of the semi-supervised MD algorithm is as follows: The value of the following equation is calculated to satisfy 0.0001. The G index of the overall sample is m = 24, the mean vector of the high-dimensional data sample is μ = (μ1, μ2, …, μ 24 ), and the MD between the overall sample G and the single sample X = (x1, x2, …, x 24 ) is: d 2 (X,G) = (X - μ)'∑(X - μ) -1 The covariance between the evaluation indexes is as follows: The solution formula is as follows: where N0is the number of labeled samples; is the number of hidden layer nodes; is the labeled sample vector; n is the dimension of the sample vector; is the class-labeled sample vector; C is the number of output nodes of the network; w i = [w i1 ,…,w in ] T is the input weight vector between the input nodes and the hidden layer nodes; b i is the bias of node i; g is the activation function of the hidden layer nodes; β i = [β i1 , β i2 ,..., β iC ] T is an output weight vector between the hidden layer node and the output node; N e is the number of extended samples; is an extended sample vector; is a class label sample vector of ; H0is a hidden layer output matrix of the labeled sample set; β is an output matrix of the weight; T0is a class label matrix of the labeled sample; H e is a hidden layer output matrix of the extended sample; T e is a class label matrix of the extended sample.

6. The risk assessment system for a protective relaying system based on semi-supervised MD algorithm of claim 5, wherein, The formula of the judgment matrix is as follows: The maximum eigenvalue of the judgment matrix is as follows: The eigenvector of the judgment matrix is as follows: W = [ω1', ω'2,... ω'N] n ] wherein the comparison coefficient is defined as a ij = x i / x j , i = 1, 2, 3,..., m; j = 1, 2, 3,..., m.

7. The risk assessment system for a protective relaying system based on semi-supervised MD algorithm of claim 6, wherein, The method for checking the consistency of the judgment matrix by using the safety random index is as follows: the consistency is defined as CR=CI / RI, RI is a random index, CI is a consistency check standard, CI=(λ-N) / (N-1), λ is the maximum eigenvalue of the judgment matrix, and N is a non-zero eigenvalue of an n-order consistent matrix.

8. The risk assessment system for a protective relaying system based on semi-supervised MD algorithm of claim 7, wherein, The method for establishing the state evaluation model based on the AHP fuzzy comprehensive algorithm is as follows: (1) considering the fuzzy relationship between the evaluation index variables, the fuzzy relationship of the continuous interval [0, 1] is used to represent the fuzzy feature domain of the relay protection system when evaluating the running state of the relay protection system, the running state of the relay protection system includes normal, warning, error and hidden danger, and the fuzzy feature domain of the relay protection system includes an index set, a weight set and an evaluation set; (2) the data of the fuzzy feature domain of the relay protection system is standardized to convert the actual value to the interval [0, 1], and the state evaluation indexes are divided into benefit type, cost type and interval type according to the exchange mode; (3) the membership degree matrix of each state evaluation index is calculated by using the ridge distribution; (4) according to the statistical data, the index set, the weight set and the evaluation set are set as a training set, and the protection device evaluation result of the training set is calculated.

Images