Distribution network equipment status evaluation method based on improved rough set optimization fireworks algorithm

Abstract

The present invention discloses a distribution network equipment status evaluation method based on an improved rough set optimized fireworks algorithm. First, historical ledger data, status monitoring records and fault entry data are combined to obtain a balanced data set based on an undersampling method based on K-means clustering in order to solve the problem of data set imbalance in the distribution network equipment status evaluation. Then, the rough set theory is used to discretize and simplify the attribute values ​​of many continuous evaluation indicators in the distribution network equipment status evaluation, and a reduction method based on attribute importance is used to simplify the attributes of the evaluation indicator decision table. Then, the improved fireworks algorithm is introduced to find the global optimal solution in a large number of attribute simplification results, and finally a set of simplified evaluation indicator decision tables are obtained for the status evaluation of distribution network equipment. This method effectively improves the accuracy and reliability of the distribution network equipment status evaluation results, can effectively reduce and avoid the occurrence of distribution network equipment failure accidents, and ensure power supply reliability and customer satisfaction.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network equipment evaluation, and in particular to a distribution network equipment status evaluation method based on an improved rough set optimization fireworks algorithm under cluster undersampling. Background Art

[0002] With the continuous expansion of the asset scale of the power distribution part of my country's power grid and the rapid increase in the total amount of equipment, as well as the gradual improvement of users' requirements for power supply reliability, the original production management model has gradually revealed its limitations. With the expansion of the scale of the distribution network, the contradiction between the traditional maintenance model and the power supply reliability and the shortage of operation and inspection personnel has become increasingly prominent. The existing manpower and material resources are difficult to meet the maintenance needs of the distribution equipment, and it is urgent to transform to condition maintenance. Through the analysis and research of power supply reliability, the relatively weak links of the distribution network are found, and measures are formulated to improve the power supply reliability in connection with the development of the local economy, so as to provide better power quality and create higher social benefits. The main equipment of the distribution network, including distribution transformers, switch equipment, overhead lines, cables, etc., is the final link of power supply to customers and plays an important role in the power supply network. Its safe operation is crucial to the stability of the power grid. Therefore, the status evaluation of distribution network equipment has become a more critical task in power grid operation and maintenance. However, the number of distribution network equipment is large, the types are diverse, the operation status is complex and changeable, and there are too many external disturbance factors such as extreme disaster weather and high load during the peak summer. How to select reasonable key evaluation attribute indicators from the numerous distribution network equipment status evaluation attribute indicators and eliminate the influence of redundant interference attribute indicators is a problem that needs to be solved in the current distribution network equipment status evaluation. At present, power companies have always followed post-maintenance and regular planned maintenance for equipment operation and maintenance, that is, assigning operation and maintenance personnel to conduct on-site evaluation of each important equipment group and feeder section after the equipment failure occurs or at certain fixed time nodes, and evaluate the status of distribution network equipment based on historical equipment failure data and their own experience. The entire evaluation process not only has a large workload for operation and maintenance personnel, high labor costs, and a long time, but also due to subjective factors and the interference of a large number of redundant evaluation indicators, it is impossible to select key and reasonable indicators in the evaluation process, so that the status evaluation results are not scientific and reasonable. Most importantly, most operation and maintenance behaviors occur after the equipment failure occurs, and the economic losses caused by it have become an established fact. All these problems have a great impact on subsequent disaster prevention and equipment operation and maintenance, and may even cause greater economic losses due to failure to correctly assess the status of the equipment in a timely manner.

[0003] Therefore, how to change the traditional evaluation method of distribution network equipment, select scientific and reasonable evaluation attribute indicators, eliminate many irrelevant redundant evaluation indicators, solve the problems of many state parameters, large evaluation workload and low efficiency in the state evaluation of distribution network equipment, and improve the operation and maintenance management capabilities of power grid equipment is an urgent problem to be solved. In order to improve the reliability of equipment operation, strengthen the pertinence of maintenance, reduce production and operation costs, and improve the application efficiency of state evaluation and state maintenance. Summary of the invention

[0004] In order to overcome and solve the problems of many state parameters, large evaluation workload and low efficiency in the state evaluation of distribution network equipment, the purpose of the present invention is to propose a distribution network equipment state evaluation method based on improved rough set optimization fireworks algorithm under cluster undersampling. First, combining historical ledger data, state monitoring records and fault item data, the undersampling method based on K-means clustering is used to obtain a balanced data set for the imbalance problem of data sets in the state evaluation of distribution network equipment; then, the rough set theory is used to realize the discretization and simplification of the attribute values ​​of many continuous evaluation indicators in the state evaluation of distribution network equipment, and the simplification method based on attribute importance is used to simplify the attributes of the evaluation indicator decision table; then, the improved fireworks algorithm is introduced to find the global optimal solution in a large number of attribute simplification results, and finally a set of simplified evaluation indicator decision tables are obtained for the state evaluation of distribution network equipment. The above method greatly reduces the impact of redundant evaluation indicators on the evaluation process and results, effectively improves the accuracy and reliability of the distribution network equipment status evaluation results, makes up for the shortcomings of traditional operation and maintenance manual status evaluation, and provides the evaluation results to the power grid operation and maintenance staff, providing strong assistance for the operation and maintenance and risk prevention of distribution network equipment, which can effectively reduce and avoid the occurrence of distribution network equipment failure accidents, and ensure power supply reliability and customer satisfaction.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A distribution network equipment status evaluation method based on improved rough set optimization fireworks algorithm, the steps are as follows:

[0007] Step 1: Original data acquisition: First, the historical ledger data, condition monitoring records, and fault entry data of the evaluated equipment are combined, and the undersampling method based on K-means clustering is used to obtain a balanced data set to generate original data entries;

[0008] Step 2: Data cleaning: Data cleaning includes missing value processing and outlier processing. Since missing data accounts for a small part, the deletion method is used to directly discard missing data; for abnormal data, the average value of the load data of the same nature is used to interpolate and replace the abnormal data;

[0009] Step 3: Based on the records of the equipment operating status in the data, the acquired raw data items are labeled with abnormal equipment status, i.e., y=1, and good equipment status, i.e., y=0, so as to obtain the raw data set D with all evaluation attribute indicators;

[0010] D={(x1,y1),(x2,y2),...,(x n ,y n )},x i ∈R,y i ∈{0,1}

[0011] Step 4: Split the original data set D with all evaluation attribute indicators, select 80% of it as the training data set D1 and 20% as the test data set D2;

[0012] Step 5: Cluster the dataset T0 with the label of good equipment status in the training dataset D1, and randomly select k cluster centroid points as μ1, μ2, …, μ k ∈R n ;

[0013] Step 6: For (x i ,y i )∈T0, calculate the cluster to which it belongs,

[0014]

[0015] in, is the data (x i ,y i ) belongs to the cluster;

[0016] Step 7: For each cluster j, recalculate the centroid of the cluster;

[0017]

[0018] Step 8: Calculate the maximum moving distance of the cluster center d = max(||μ′ j -μ j ||2); if d>ε, update μ j =μ′ j , jump to step 7;

[0019] Step 9: Randomly undersample each cluster of the clustering results obtained in step 7 according to the ratio α, and finally obtain a balanced training data set D′0;

[0020] Step 10: Use the rough set theory to discretize and simplify the attribute values ​​of the evaluation indicators in the balanced training data set D′0. First, generate a knowledge expression system based on all the contents in the balanced training data set: S = (U, R, V, f); where U is a non-empty set of objects z, called the domain, that is, z is an element of the set U; R = C∪D is an attribute set, C is a conditional attribute set, and D is a set of decision attributes d, that is, d is an element of the set D; V = ∪ a∈R V a is a set of attribute values, V a is the value range of attribute a; f is the mapping of U×R→V, which specifies the attribute value of each object z in U;

[0021] Step 11: Record the value range V of attribute a in the knowledge expression system a breakpoints on R, where each breakpoint is denoted by (a,c), where a∈R,c∈V a ;

[0022] Step 12: In the value range V a =[l a ,r a Any set of breakpoints {(a,c1),(a,c2),…,(a,c k )} define V a A classification on P a , whose expression is:

[0023]

[0024] Step 13: According to the breakpoint set of conditional attributes P = ∪ a∈R P a Define a new decision table, namely the discrete attribute value evaluation index system S P =(U,R,V P ,f P ), where for any object z∈U,i∈{0,1,2,…,k} where z a is the a attribute of object z; that is, after discretization, the original continuous attribute value evaluation index system is replaced by a new discrete attribute value evaluation index system;

[0025] Step 14: Use the rough set attribute reduction method based on attribute importance to evaluate the discrete attribute value index system S P To simplify attributes, first determine the equivalence relation in the discrete attribute value evaluation index system: let A be an equivalence relation determined by the object attribute set R on the domain U, that is, A is a partition of U; for any and Then ∩M is also an equivalence relation on U, and ∩M is an indiscernibility relation on M, denoted as IND(M);

[0026] Step 15: Use the indiscernibility relation IND(M) to form a partition of U, represented by U / IND(M), or simply U / M.

[0027] Step 16: Determine the dependency and independence relationship of the attribute set R: For attribute r∈R, if IND(R)=IND(Rr), then attribute r is called unnecessary in R; otherwise, r is called necessary in R; if every attribute r∈R is necessary in R, then R is called independent; otherwise, R is called dependent;

[0028] Step 17: Calculate the importance of attributes: Calculate the importance Sig of each attribute x in the condition attribute set C to the entire condition attribute set C C-x (x); Sig X (x) is calculated by the following method: The importance of is calculated as:

[0029]

[0030] where |X|=|IND(X)|; assuming U / IND(X)=U / X={X1,X2…X n},So

[0031]

[0032] Step 18: Calculate the attribute core CORE(X): where, If X and Q are independent, and IND(Q) = IND(X), then Q is called a reduction of X, and the reduction of X is denoted by RED(X); the set of all necessary attributes in X is called the core of X, denoted by CORE(X); the relationship between the core and the reduction is as follows: CORE(X) = ∩RED(X); According to the result in step 17, Sig C-x (x) Attributes with values ​​greater than 0 are core attributes;

[0033] Step 19: Let RED(X)=CORE(X);

[0034] Step 20: When IND(RED(X))=IND(X), go to step 23, otherwise go to step 21;

[0035] Step 21: Calculate Sig RED(X) (x), select x1 so that it satisfies:

[0036]

[0037] Step 22: RED(X)=RED(X){x1}, go to step 20;

[0038] Step 23: Output the reduction result RED(X);

[0039] Step 24: Calculate the dependency of the condition attribute set C′ on the decision attribute set D′ in each reduction result:

[0040]

[0041] Where card(·) is the cardinality of the set; POS C′ (D′) is the positive domain of the equivalence class U / D′ with respect to C′;

[0042] Step 25: Define the effectiveness function of the reduction result; the effectiveness function is:

[0043]

[0044] In the formula, m is the total number of conditional attributes; l is the number of breakpoints contained in the individual evaluation object; r C′ (D′) is the dependence of the condition attribute set C′ on the decision attribute set D′; α is the weight adjustment factor;

[0045] Step 26: According to the effectiveness function of the simplified result of step 25, the improved fireworks algorithm is used to seek the global optimal solution with the minimum fitness among all attribute simplified results: first, randomly generate N fireworks in the feasible solution space, and each firework represents a feasible solution in the solution space;

[0046] Step 27: Repeat steps 27.1 to 27.6 until a solution with the global minimum fitness value is found, that is, the objective function value remains at the minimum value for multiple consecutive times:

[0047] Step 27.1: Calculate the fitness value f(q i ), then for each firework q i Generate explosion sparks with explosion radius and explosion spark number parameters, and calculate the number of explosion sparks S for each firework i and explosion spark radius A i :

[0048]

[0049] Where:y min =min(f(q i )), is the minimum fitness value in the current fireworks population, y max =max(f(q i)), is the maximum fitness value in the current fireworks population, where i = 1, 2, …, N; is a constant used to adjust the explosion radius, M is a constant used to adjust the number of explosion sparks generated, and ε is a machine minimum to avoid division by zero.

[0050] Step 27.2: Improve the explosion spark radius in step 27.1 by introducing a minimum explosion spark radius detection mechanism; if the explosion radius of the firework is less than a certain threshold, set it to that value, as shown in the following formula:

[0051]

[0052] How to determine A min,k , a nonlinear decreasing method is introduced:

[0053]

[0054] Where t refers to the number of times the current function is evaluated, evals_max is the maximum number of function evaluations; A init , A final Represents the initial and final explosion radius detection values ​​during the iteration process;

[0055] Step 27.3: Improve the explosion operator of the fireworks algorithm by limiting the number of sparks generated as follows:

[0056]

[0057] Where: a and b are two constants, round(·) is a rounding function based on the rounding principle;

[0058] Step 27.4: Generate Gaussian mutation sparks: First, randomly select a firework q from the firework population. i , and then randomly select a certain number of dimensions for the firework to perform Gaussian mutation operation; for the firework q i The Gaussian mutation operation is performed on a dimension k obtained by a certain choice:

[0059]

[0060] Where: e~N(1,1), N(1,1) represents a Gaussian distribution with a mean of 1 and a variance of 1;

[0061] In the process of explosion operator and mutation operator generating explosion spark and Gaussian mutation spark respectively, the sparks generated may exceed the boundary range of feasible solution space; when the spark q i If the dimension k exceeds the boundary, it will be mapped to a new position using the following mapping rule:

[0062]

[0063] Where: q UB,k ,q LB,k is the upper and lower boundaries of the solution space in dimension k;

[0064] Step 27.5: Improve the Gaussian mutation spark: When some sparks are beyond the boundary range, the following formula is introduced as a mapping rule to handle the sparks beyond the boundary:

[0065]

[0066] Step 27.6: Select N individuals from the fireworks, explosion sparks, and Gaussian mutation spark populations as the next generation of iterative fireworks. The candidate strategies are as follows:

[0067] Assume that the candidate set is K and the fireworks population size is N; the individual with the smallest fitness value in the candidate set will be deterministically selected as the next generation of fireworks, and the remaining N-1 fireworks will be selected from the candidate set using the roulette method; for candidate fireworks q i , the calculation formula of its selection probability is:

[0068]

[0069] Where: R(q i ) is the current individual to the candidate set K divided by q i The sum of the distances between all individuals;

[0070] Step 28: Return the final global optimization result, that is, a set of simplified evaluation index decision tables for status evaluation of distribution network equipment.

[0071] Compared with the prior art, the present invention has the following advantages:

[0072] The present invention proposes a distribution network equipment status evaluation method based on an improved rough set optimized fireworks algorithm. First, combining historical ledger data, status monitoring records and fault entry data, an undersampling method based on K-means clustering is used to obtain a balanced data set, thereby improving the data set imbalance problem in the distribution network equipment status evaluation; then, the rough set theory is used to discretize and simplify the attribute values ​​of many continuous evaluation indicators in the distribution network equipment status evaluation, and a simplification method based on attribute importance is used to simplify the attributes of the evaluation indicator decision table, thereby improving the problems of many state parameters, large evaluation workload and low efficiency in the distribution network equipment status evaluation process; then, the improved fireworks algorithm is introduced to find the global optimal solution in a large number of attribute simplification results, and finally a set of simplified evaluation indicator decision tables are obtained for the status evaluation of distribution network equipment. The above method greatly reduces the impact of redundant evaluation indicators on the evaluation process and results, effectively improves the accuracy and reliability of the distribution network equipment status evaluation results, makes up for the shortcomings of traditional operation and maintenance manual status evaluation, and provides the evaluation results to the power grid operation and maintenance staff, providing strong assistance for the operation and maintenance and risk prevention of distribution network equipment, which can effectively reduce and avoid the occurrence of distribution network equipment failure accidents, and ensure power supply reliability and customer satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1a This is the curve chart before abnormal data processing.

[0074] Figure 1b This is the curve chart after abnormal data processing.

[0075] Figure 2a This is a diagram showing the distribution of an imbalanced data set.

[0076] Figure 2b This is the distribution diagram of the data set after undersampling. DETAILED DESCRIPTION

[0077] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0078] A distribution network outage prediction technology during the peak summer period based on cluster undersampling ensemble learning method, the steps are as follows:

[0079] Step 1: Original data acquisition. First, the historical ledger data, condition monitoring records, and fault entry data of the evaluated equipment are combined, and the undersampling method based on K-means clustering is used to obtain a balanced data set to generate original data entries;

[0080] Step 2: Data cleaning. Data cleaning includes missing value processing and outlier processing. Since missing data accounts for a small part, the deletion method is used to directly discard missing data; for abnormal data, the average value of the load data of the same nature is used to interpolate and replace the abnormal data. See before and after abnormal data processing Figure 1a and Figure 1b As shown, it can be seen that some of the data that stand out in the data value anomaly concentration are replaced by values ​​within the normal range, which helps to improve the accuracy of subsequent status evaluation results.

[0081] Step 3: Based on the records of the equipment operating status in the data, the acquired raw data items are labeled with abnormal equipment status (y=1) and good equipment status (y=0), thereby obtaining the original data set D with all evaluation attribute indicators;

[0082] D={(x1,y1),(x2,y2),...,(x n ,y n )},x i ∈R,y i ∈{0,1}

[0083] Step 4: Split the original data set D with all evaluation attribute indicators, select 80% of it as the training data set D1 and 20% as the test data set D2.

[0084] Step 5: Cluster the dataset T0 with the label of good equipment status in the training dataset D1, and randomly select k cluster centroid points as μ1, μ2, …, μ k ∈R n .

[0085] Step 6: For (x i ,y i )∈T0, calculate the cluster to which it belongs,

[0086]

[0087] in, is the data (x i ,y i ) belongs to the cluster.

[0088] Step 7: For each cluster j, recalculate the centroid of the cluster.

[0089]

[0090] Step 8: Calculate the maximum moving distance of the cluster center d = max(||μ′ j -μ j ||2). If d>ε, update μ j =μ′ j, skip to step 7.

[0091] Step 9: Perform random undersampling of each cluster of the clustering results obtained in step 7 according to a ratio of α, and finally obtain a balanced training data set D′0. The distribution of the unbalanced data set and the distribution of the balanced data set after undersampling are shown in Figure 2a and Figure 2b shown.

[0092] Step 10: Use the rough set theory to discretize and simplify the attribute values ​​of the evaluation indicators in the balanced training data set D′0. First, generate a knowledge expression system based on all the contents in the balanced training data set: S = (U, R, V, f). Among them, U is a non-empty set of objects z, called the domain, that is, z is an element of the set U; R = C∪D is the attribute set, C is the conditional attribute set, D is the set of decision attributes d, that is, d is an element of the set D; V = ∪ a∈R V a is a set of attribute values, V a is the value range of attribute a; f is the mapping of U×R→V, which specifies the attribute value of each object z in U. For example, when evaluating the status of a transformer, the volume fractions of insulation fault characteristic gases H2, CH4, C2H4, C2H6, C2H2, CO2, and CO related to the insulation status of the transformer can be selected as elements of the condition attribute set of the rough set status evaluation, represented by C1 to C7 respectively; types such as normal, low-energy discharge, medium-temperature overheating, arc discharge, and high-temperature overheating can be selected as the set D of decision attribute elements d.

[0093] Step 11: Record the value range V of attribute a in the knowledge expression system a breakpoints on R, each of which can be denoted by (a,c), where a∈R,c∈V a ;

[0094] Step 12: In the value range V a =[l a ,r a Any set of breakpoints {(a,c1),(a,c2),…,(a,c k )} define V a A classification on P a , whose expression is:

[0095]

[0096] Step 13: According to the breakpoint set of conditional attributes P = ∪ a∈R P a Define a new decision table, namely the discrete attribute value evaluation index system S P =(U,R,V P ,f P), where for any object z∈U,i∈{0,1,2,…,k} where z a is the a attribute of object z. That is, after discretization, the original continuous attribute value evaluation index system is replaced by a new discrete attribute value evaluation index system. For example, the volume fraction of the insulating gas related to the insulation condition of the transformer was originally a continuous percentage value. After discretization, it all becomes a discrete value. For example, for a certain transformer, its CH4 content was originally 10.3%, that is, C2 = 10.3%. After discretization, its value becomes 1, that is, C2 = 1.

[0097] Step 14: Use the rough set attribute reduction method based on attribute importance to evaluate the discrete attribute value index system S P To simplify attributes, first determine the equivalence relation in the discrete attribute value evaluation index system: let A be an equivalence relation determined by the object attribute set R on the domain U, that is, A is a partition of U. and Then ∩M is also an equivalence relation on U, and we get ∩M to be an indiscernibility relation on M, denoted as IND(M).

[0098] Step 15: Use the indiscernibility relation IND(M) to form a partition of U, denoted by U / IND(M), or U / M for short.

[0099] Step 16: Determine the dependency and independence relationship of the attribute set R. For an attribute r∈R, if IND(R)=IND(Er), then the attribute r is called unnecessary in R; otherwise, r is called necessary in R. If every attribute r∈R is necessary in R, then R is called independent; otherwise, R is called dependent.

[0100] Step 17: Calculate the importance of attributes. Calculate the importance of each attribute x in the condition attribute set C to the entire condition attribute set C. C-x (x). Sig X (x) is calculated by the following method: The importance of is calculated as:

[0101]

[0102] Where |X|=|IND(X)|. Assume U / IND(X)=U / X={X1,X2…X n},So

[0103]

[0104] Step 18: Calculate the attribute core CORE(X). If X and Q are independent, and IND(Q) = IND(X), then Q is called a reduction of X, and the reduction of X can be denoted as RED(X). The set of all necessary attributes in X is called the core of X, denoted as CORE(X). The relationship between the core and the reduction is as follows: CORE(X) = ∩RED(X). According to the result in step 17, Sig C-x Attributes with values ​​greater than 0 in (x) are core attributes.

[0105] Step 19: Let RED(X)=CORE(X).

[0106] Step 20: When IND(RED(X)) = IND(X), go to step 23, otherwise go to step 21.

[0107] Step 21: Calculate Sig RED(X) (x), select x1 so that it satisfies:

[0108]

[0109] Step 22: RED(X)=RED(X){x1}, go to step 20;

[0110] Step 23: Output the reduction result RED(X).

[0111] Step 24: Calculate the dependency of the condition attribute set C′ on the decision attribute set D′ in each reduction result:

[0112]

[0113] Where card(·) is the cardinality of the set; POS C′ (D') is the positive domain of the equivalence class U / D' with respect to C'.

[0114] Step 25: Define the effectiveness function of the reduction result. The effectiveness function is:

[0115]

[0116] In the formula, m is the total number of conditional attributes; l is the number of breakpoints contained in the individual evaluation object; r C′ (D') is the dependence of the condition attribute set C' on the decision attribute set D'; α is the weight adjustment factor.

[0117] Step 26: According to the effectiveness function of the simplified result of step 25, the improved fireworks algorithm is used to seek the global optimal solution with the minimum fitness among all attribute simplified results: first, randomly generate N fireworks in the feasible solution space, and each firework represents a feasible solution in the solution space;

[0118] Step 27: Repeat steps 27.1 to 27.6 until a solution with the global minimum fitness value is found, that is, the objective function value remains at the minimum value for multiple consecutive times:

[0119] Step 27.1: Calculate the fitness value f(q i ), then for each firework q i Generate explosion sparks with explosion radius and explosion spark number parameters, and calculate the number of explosion sparks S for each firework i and explosion spark radius A i :

[0120]

[0121] Where:y min =min(f(q i )), (i=1,2,…,N) is the minimum fitness value in the current fireworks population, y max =max(f(q i )), (i=1,2,…,N) is the maximum fitness value in the current fireworks population. is a constant used to adjust the explosion radius, M is a constant used to adjust the number of explosion sparks generated, and ε is a machine minimum used to avoid division by zero operations.

[0122] Step 27.2: Improve the explosion spark radius in step 27.1 and introduce a minimum explosion spark radius detection mechanism. If the explosion radius of the firework is less than a certain threshold, it is set to that value, as shown in the following formula:

[0123]

[0124] How to determine A min,k , a nonlinear decreasing method is introduced:

[0125]

[0126] Where t refers to the number of times the current function is evaluated, and evals_max is the maximum number of function evaluations. init , A final Represents the initial and final blast radius detection values ​​during the iteration process.

[0127] Step 27.3: Improve the explosion operator of the fireworks algorithm by limiting the number of sparks generated as follows:

[0128]

[0129] Where a and b are two constants, and round(·) is a rounding function based on the rounding principle.

[0130] Step 27.4: Generate Gaussian mutation sparks. First, randomly select a firework q from the firework population. i , and then randomly select a certain number of dimensions for the firework to perform Gaussian mutation operation. i The Gaussian mutation operation is performed on a dimension k obtained by a certain choice:

[0131]

[0132] In the formula: e~N(1,1), N(1,1) represents a Gaussian distribution with a mean of 1 and a variance of 1.

[0133] In the process of explosion operator and mutation operator generating explosion spark and Gaussian mutation spark respectively, the sparks generated may exceed the boundary range of feasible solution space. i If the dimension k exceeds the boundary, it will be mapped to a new position using the following mapping rule:

[0134]

[0135] Where: q UB,k ,q LB,k are the upper and lower boundaries of the solution space in dimension k.

[0136] Step 27.5: Improve the Gaussian variation spark. When some sparks exceed the boundary range, the following formula is introduced as a mapping rule to handle the sparks beyond the boundary:

[0137]

[0138] Where: x UB,k 、x LB,k are the upper and lower boundaries of the feasible region in dimension k.

[0139] Step 27.6: Select N individuals from the fireworks, explosion sparks, and Gaussian mutation spark populations as the next generation of iterative fireworks. The candidate strategies are as follows:

[0140] Assume that the candidate set is K and the fireworks population size is N. The individual with the smallest fitness value in the candidate set will be deterministically selected as the next generation of fireworks, and the remaining N-1 fireworks will be selected from the candidate set using the roulette method. i , the calculation formula of its selection probability is:

[0141]

[0142] Where: R(q i ) is the current individual to the candidate set K divided by q i The sum of the distances between all individuals.

[0143] Step 28: Return the final global optimization result, that is, a set of simplified evaluation index decision tables for the status evaluation of distribution network equipment. Still taking the transformer status evaluation as an example, the condition attribute set of a transformer before simplification is C1-C7, and the status of the transformer is normal, but at this time, it is necessary to assign values ​​to all 7 condition attributes to obtain the status evaluation result, and the simplified evaluation index decision table obtained by the method of the present invention finally simplifies its evaluation rules, such as 1=1, C3=4 can determine that its status is normal, eliminates redundant indicators and indicators that have little effect on the evaluation results, greatly reduces the workload of the status evaluation process, especially when it involves large-scale equipment status evaluation, this method can significantly improve the efficiency of the status evaluation process.

Claims

1. A distribution network equipment status evaluation method based on improved rough set optimization fireworks algorithm, characterized by: Here are the steps: Step 1: Original data acquisition: First, the historical ledger data, condition monitoring records, and fault entry data of the evaluated equipment are combined, and the undersampling method based on K-means clustering is used to obtain a balanced data set to generate original data entries; Step 2: Data cleaning: Data cleaning includes missing value processing and outlier processing. Since missing data accounts for a small part, the deletion method is used to directly discard missing data; for abnormal data, the average value of the load data of the same nature is used to interpolate and replace the abnormal data; Step 3: Based on the records of the equipment operating status in the data, the acquired raw data items are labeled with abnormal equipment status, i.e., y=1, and good equipment status, i.e., y=0, so as to obtain the raw data set D with all evaluation attribute indicators; D={(x1,y1),(x2,y2),…,(x n ,y n )},x i ∈R,y i ∈{0,1} Step 4: Split the original data set D with all evaluation attribute indicators, select 80% of it as the training data set D1 and 20% as the test data set D2; Step 5: Cluster the dataset T0 with the label of good equipment status in the training dataset D1, and randomly select k cluster centroid points as μ1, μ2, …, μ k ∈R n ; Step 6: For (x i ,y i )∈T0, calculate the cluster to which it belongs, in, is the data (x i ,y i ) belongs to the cluster; Step 7: For each cluster j, recalculate the centroid of the cluster; Step 8: Calculate the maximum moving distance of the cluster center d = max(||μ′ j -μ j ||2); if d>ε, update μ j =μ′ j , jump to step 7; Step 9: Perform random undersampling of each cluster of the clustering results obtained in step 7 according to a ratio α, and finally obtain a balanced training data set D′0; Step 10: Use the rough set theory to discretize and simplify the attribute values ​​of the evaluation indicators in the balanced training data set D′0. First, generate a knowledge expression system based on all the contents in the balanced training data set: S = (U, R, V, f); where U is a non-empty set of objects z, called the domain, that is, z is an element of the set U; R = C∪D is an attribute set, C is a conditional attribute set, and D is a set of decision attributes d, that is, d is an element of the set D; V = ∪ a∈R V a is a set of attribute values, V a is the value range of attribute a; f is the mapping of U×R→V, which specifies the attribute value of each object z in U; Step 11: Record the value range V of attribute a in the knowledge expression system a breakpoints on R, where each breakpoint is denoted by (a,c), where a∈R,c∈V a ; Step 12: In the value range V a =[l a ,r a Any set of breakpoints {(a,c1),(a,c2),…,(a,c k )} define V a A classification on P a , whose expression is: Step 13: According to the breakpoint set of conditional attributes P = ∪ a∈R P a Define a new decision table, namely the discrete attribute value evaluation index system S P =(U,R,V P ,f P ), where for any object z∈U,i∈{0,1,2,…,k} where z a is the a attribute of object z; that is, after discretization, the original continuous attribute value evaluation index system is replaced by a new discrete attribute value evaluation index system; Step 14: Use the rough set attribute reduction method based on attribute importance to evaluate the discrete attribute value index system S P To simplify attributes, first determine the equivalence relation in the discrete attribute value evaluation index system: let A be an equivalence relation determined by the object attribute set R on the domain U, that is, A is a partition of U; for any and Then ∩M is also an equivalence relation on U, and ∩M is an indiscernibility relation on M, denoted as IND(M); Step 15: Use the indiscernibility relation IND(M) to form a partition of U, represented by U / IND(M), or simply U / M. Step 16: Determine the dependency and independence relationship of the attribute set R: For attribute r∈R, if IND(R)=IND(Rr), then attribute r is called unnecessary in R; otherwise, r is called necessary in R; if every attribute r∈R is necessary in R, then R is called independent; otherwise, R is called dependent; Step 17: Calculate the importance of attributes: Calculate the importance Sig of each attribute x in the condition attribute set C to the entire condition attribute set C C-x (x); Sig X (x) is calculated by the following method: The importance of is calculated as: where |X|=|IND(X)|; assuming U / IND(X)=U / X={X1,X2…X n },So Step 18: Calculate the attribute core CORE(X): where X, If X and Q are independent, and IND(Q) = IND(X), then Q is called a reduction of X, and the reduction of X is denoted by RED(X); the set of all necessary attributes in X is called the core of X, denoted by CORE(X); the relationship between the core and the reduction is as follows: CORE(X) = ∩RED(X); According to the result in step 17, Sig C-x (x) Attributes with values ​​greater than 0 are core attributes; Step 19: Let RED(X)=CORE(X); Step 20: When IND(RED(X))=IND(X), go to step 23, otherwise go to step 21; Step 21: Calculate Sig RED(X) (x), select x1 so that it satisfies: Step 22: RED(X)=RED(X){x1}, go to step 20; Step 23: Output the reduction result RED(X); Step 24: Calculate the dependency of the condition attribute set C' on the decision attribute set D' in each reduction result: Where card(·) is the cardinality of the set; POS C′ (D′) is the positive domain of the equivalence class U / D′ with respect to C′; Step 25: Define the effectiveness function of the reduction result; the effectiveness function is: In the formula, m is the total number of conditional attributes; l is the number of breakpoints contained in the individual evaluation object; r C′ (D') is the dependence of the condition attribute set C' on the decision attribute set D'; α is the weight adjustment factor; Step 26: According to the effectiveness function of the simplified result of step 25, the improved fireworks algorithm is used to seek the global optimal solution with the minimum fitness among all attribute simplified results: first, randomly generate N fireworks in the feasible solution space, and each firework represents a feasible solution in the solution space; Step 27: Repeat steps 27.1 to 27.6 until a solution with the global minimum fitness value is found, that is, the objective function value remains at the minimum value for multiple consecutive times: Step 27.1: Calculate the fitness value f(q i ), then for each firework q i Generate explosion sparks with explosion radius and explosion spark number parameters, and calculate the number of explosion sparks S for each firework i and explosion spark radius A i : Where:y min =min(f(q i )), is the minimum fitness value in the current fireworks population, y max =max(f(q i )), is the maximum fitness value in the current fireworks population, where i=1,2,…,N is a constant used to adjust the explosion radius, M is a constant used to adjust the number of explosion sparks generated, and ε is a machine minimum to avoid division by zero. Step 27.2: Improve the explosion spark radius in step 27.1 by introducing a minimum explosion spark radius detection mechanism; if the explosion radius of the firework is less than a certain threshold, set it to that value, as shown in the following formula: How to determine A min,k , a nonlinear decreasing method is introduced: Where t refers to the number of times the current function is evaluated, evals_max is the maximum number of function evaluations; A init , A final Represents the initial and final explosion radius detection values ​​during the iteration process; Step 27.3: Improve the explosion operator of the fireworks algorithm by limiting the number of sparks generated as follows: Where: a and b are two constants, round(·) is a rounding function based on the rounding principle; Step 27.4: Generate Gaussian mutation sparks: First, randomly select a firework q from the firework population. i , and then randomly select a certain number of dimensions for the firework to perform Gaussian mutation operation; for the firework q i The Gaussian mutation operation is performed on a dimension k obtained by a certain choice: Where: e~N(1,1), N(1,1) represents a Gaussian distribution with a mean of 1 and a variance of 1; In the process of explosion operator and mutation operator generating explosion spark and Gaussian mutation spark respectively, the sparks generated may exceed the boundary range of feasible solution space; when the spark q i If the dimension k exceeds the boundary, it will be mapped to a new position using the following mapping rule: Where: q UB，k ,q LK，k is the upper and lower boundaries of the solution space in dimension k; Step 27.5: Improve the Gaussian mutation spark: When some sparks are beyond the boundary range, the following formula is introduced as a mapping rule to handle the sparks beyond the boundary: Step 27.6: Select N individuals from the fireworks, explosion sparks, and Gaussian mutation spark populations as the next generation of iterative fireworks. The candidate strategies are as follows: Assume that the candidate set is K and the fireworks population size is N; the individual with the smallest fitness value in the candidate set will be deterministically selected as the next generation of fireworks, and the remaining N-1 fireworks will be selected from the candidate set using the roulette method; for candidate fireworks q i , the calculation formula of its selection probability is: Where: R(q i ) is the current individual to the candidate set K divided by q i The sum of the distances between all individuals; Step 28: Return the final global optimization result, that is, a set of simplified evaluation index decision tables for status evaluation of distribution network equipment.

Images