DC Distribution Network Fault Location Method Based on Multivariate Information Fusion

Abstract

A DC distribution network fault location method based on multi-source information fusion belongs to the technical field of DC power distribution networks. The purpose of the present invention is to normalize the sparse characteristics of fault current and the fault probability to obtain the fault degree, and finally use an improved DS evidence theory fusion algorithm to fuse these two fault degrees to obtain a more accurate fault location result for the DC distribution network fault location method based on multi-source information fusion. First, the compressed sensing algorithm is used to reconstruct the fault transient electrical quantity data to obtain the sparse characteristics of the fault high-frequency current, and the fault degree of each component is obtained through normalization. Secondly, the relay protection device generates protection action information for the fault component, and a Bayesian network is established under the protection action information to obtain the fault probability of the fault component. At the same time, the sparse characteristics of the fault current and the fault probability are normalized to obtain the fault degree, and finally an improved DS evidence theory fusion algorithm is used to fuse these two fault degrees to obtain a more accurate fault location result. The simulation shows that this fusion method can improve the accuracy of fault location.

Description

Technical Field

[0001] The present invention belongs to the technical field of DC power distribution networks. Background Art

[0002] DC power distribution networks are conducive to meeting the flexible access and efficient operation of various types of distributed power sources, energy storage, and flexible loads. Their structure will also change from the traditional radial type to a multi-terminal closed interconnected network and further develop towards a complex network with multiple layers, levels, and loops, thus presenting a new form of multi-element integration and multi-state mixing. However, as the structure of DC power distribution networks develops towards a diversified and multi-level network structure, the frequency of faults occurring becomes higher and higher, bringing new challenges to future research on DC power distribution networks.

[0003] After a fault occurs in a DC power distribution network, electrical quantity information such as the current and voltage of components will change accordingly. At the same time, the relay protection device will generate action information and corresponding circuit breaker action information to cut off the faulty component and ensure the safe and reliable operation of the DC power distribution network.

[0004] The fault characteristics of DC distribution networks are complex, and the accuracy of fault location methods based on the analysis of electrical quantities or switch quantities is restricted, making it difficult to effectively address the above challenges. On the one hand, it is very difficult to accurately collect the complex frequency components in the fault electrical quantity data, and the characteristics of single-pole grounding faults are not obvious, resulting in the difficulty of meeting the applicability requirements for various fault types in terms of the location accuracy based on electrical quantity data. On the other hand, the method of switch quantity data analysis depends on accurate historical information on component reliability. In the absence of long-term actual engineering tests, it also cannot meet the requirements for precise fault location in DC distribution networks under the protection action and circuit breaker action logic relationships after the expansion of the switch quantity category. The adoption of multi-source information fusion technology can make full use of the measurement device information of each feeder in the DC distribution network, enabling each measurement information to complement and verify each other, establishing a corresponding fault-tolerant mechanism, and avoiding the problems of incorrect or inaccurate location caused by a single type of measurement. Document 1: Convert switch quantities into the fuzzy fault degree of components through a fuzzy Petri net, analyze electrical quantities using wavelet transform to obtain the wavelet fault characterization of components, and then make an information fusion decision on the faulty component based on an improved DS evidence theory. Document 2: First, use the compressive sensing algorithm to reconstruct the electrical quantity signal twice to obtain the fault range and the electrical quantity fault degree, then use a Bayesian network to obtain the fault range and the switch quantity fault degree of each component, and finally use a multi-sensor fusion algorithm to fuse these two fault degrees to achieve fault location. Document 3: Use a one-dimensional convolutional neural network to extract the fault data characteristics of each branch, use a feature fusion algorithm to fuse the characteristics of prior information with the line fault characteristics, and input this feature vector into the neural network to achieve fault line location. Document 4: First, establish a corresponding network tree diagram through searching for fault indicator information, distribution transformer alarm information, and telephone complaint information generated during DC distribution network faults for preliminary fault location, and then use an improved DS evidence theory to fuse the location results of each type of fault information to obtain the final location result. Summary of the Invention

[0005] The object of the present invention is a DC distribution network fault location method based on multi-source information fusion, which normalizes the sparse characteristics of fault current and the fault probability to obtain the fault degree, and finally uses an improved DS evidence theory fusion algorithm to fuse these two fault degrees to obtain a more accurate fault location result.

[0006] The steps of the present invention are as follows:

[0007] S1. Processing of electrical quantities by compressive sensing

[0008] The high-frequency impedance matrix of nodes can be obtained by inverting the high-frequency admittance matrix of nodes:

[0009] (9)

[0010] Where: N is the number of nodes in the DC distribution network;

[0011] The nodal voltage equation obtained from the nodal impedance matrix is as follows:

[0012] (10)

[0013] Where: is the column vector of nodal high-frequency voltages; is the column vector of nodal high-frequency currents;

[0014] The number and location of measurement points are reasonably configured, and the number of measurement points is as follows:

[0015] (11)

[0016] Where: R is the number of measurement points; is the sparsity of the signal to be reconstructed; M is the length of the signal to be reconstructed;

[0017] Finally, the OMP algorithm is used to reconstruct the high-frequency nodal current, and the implementation process steps are as follows:

[0018] (1) First, according to the high-frequency information formed by the faults at the nodes of the DC distribution network, collect the high-frequency admittance matrices at R measurement nodes, and then perform an inverse operation to obtain the sensing matrix ;

[0019] (2) Read the voltage data of the measurement points 2 ms before and after the fault, extract its high-frequency voltage components, and calculate the observation signal with a continuous window length ;

[0020] (3) Send the obtained sensing matrix and the observation signal as parameters into the OMP reconstruction algorithm for calculation to obtain the sparse vector composed of the high-frequency current amplitudes of each node ;

[0021] (4) Judge according to the number of iterations until the algorithm ends, and output the sparse characteristics of the fault current;

[0022] S2. DC distribution network fault location method based on multi-source information fusion

[0023] S2.1 Multi-data source information fusion technology

[0024] Suppose is the exhaustion of all result sets for a specific event, and each event is mutually exclusive, that is can be expressed as: Where there are hypothetical events in the set, and each subset contains kinds of events, and there is a mapping relationship from to [0,1] : → [0, 1] and satisfies , then is called the basic probability assignment function of; where there exists when it is called the focal element;

[0025] Under the same identification framework there exist pieces of evidence, and each piece of evidence is independent of each other. The basic probability assignment function is , satisfying:

[0026] (12)

[0027] In the formula, is the conflict factor;

[0028] When, the combination holds, When, the conflict is too high, otherwise the combination rule cannot be used;

[0029] S2.2 Fault Location Based on Multi-Sensor Data Fusion Algorithm

[0030] The current amplitude fault degrees of each suspected faulty component are obtained through normalization. The normalization formula is as follows:

[0031] (13)

[0032] In: is the current characteristic of each faulty component, , representing the credibility of the fault location result based on electrical quantity data;

[0033] The probability values of each faulty component are used to obtain the switch quantity fault degrees of each suspected faulty component through normalization. The normalization formula is as follows:

[0034] (14)

[0035] In the formula: is the probability value of each faulty component, , representing the credibility of the fault location result based on switch quantity;

[0036] During the information fusion process, the obtained electrical quantity fault degree and switch quantity fault degree are used as two pieces of evidence, and the DS evidence theory is used for fusion. At the same time, each component is an independent identification framework , and this framework includes three different states, = {faulty, normal, uncertain}, assuming is the basic probability assignment function for two independent evidence bodies of electrical quantities and switching quantities; the electrical quantities and switching quantities are fused using Equation (12);

[0037] Make a decision on the fusion result of the component according to the obtained fusion result Make a decision through Equation (15):

[0038] (15)

[0039] where and are threshold values;

[0040] When making a decision on the state of the fused component, if the component state satisfies the above formula, then the component is determined to be a faulty component; the fusion process is as follows:

[0041] Solve the problem of how to redistribute and manage conflicts as follows:

[0042] (16)

[0043] where , represents the frame of discernment contains the sum of elements, represents the frame of discernment the set of all elements, when is a singleton subset, it satisfies the operation rules, when is an uncertain subset, its basic probability is 0;

[0044] Under the frame of discernment and the evidence body using the Bayes estimation method, adopt the weighted method. First, preprocess the conflicting evidence, and then fuse the evidence using the evidence combination rule:

[0045] (17)

[0046] where: belongs to the subset of the frame of discernment ;

[0047] After obtaining a new evidence body using the weighted method, adopt the secondary fusion method:

[0048] (18)

[0049] In the formula , represents the evidence generated under the Bayes estimation method; represents the evidence generated under the weighted method.​​​

[0050] When a fault occurs in the DC distribution network, the present invention can improve the accuracy of fault location. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a partial protection configuration diagram of the DC distribution network;

[0052] Figure 2 It is the DC bus B The Bayesian network diagram of 14;

[0053] Figure 3 It is the calculation flow chart of the OMP reconstruction algorithm;

[0054] Figure 4 It is the fusion flow chart of the DS evidence theory;

[0055] Figure 5 It is the fusion flow chart of the improved DS evidence theory;

[0056] Figure 6 It is the signal reconstruction result diagram of Fault Scenario 1;

[0057] Figure 7 It is the fault range area diagram of Fault Scenario 1;

[0058] Figure 8 It is for Fault Scenario 1 L 14-15 The Bayesian network model diagram of;

[0059] Figure 9 It is the positioning result diagram of three positioning algorithms in Fault Scenario 1;

[0060] Figure 10 It is for Fault Scenario 1 line L 14-15 The fault location result diagram of three methods of;

[0061] Figure 11 It is the signal reconstruction result diagram of Fault Scenario 2;

[0062] Figure 12 It is the fault range area diagram of Fault Scenario 2;

[0063] Figure 13 It is for Fault Scenario 2 L 14-15 The Bayesian network model diagram of;

[0064] Figure 14 It is the positioning result diagram of three positioning algorithms in Fault Scenario 2;

[0065] Figure 15 It is for Fault Scenario 2 line L 26-27 The fault location result diagram of three methods of;

[0066] Figure 16For four algorithms for line L 26-27 Comparison chart of positioning results Specific implementation manner

[0067] Fault information digital quantity processing method

[0068] 1. Construction and inference of Bayesian network

[0069] In terms of fault location based on digital quantities, establishing a Bayesian network is a commonly used location method. A Bayesian network (BN) is a probabilistic graphical model that combines Bayesian theory and graph theory methods and is an important technology for processing probability problems in the field of artificial intelligence. A Bayesian network can visually reflect the causal and independent relationships between variables, make full use of prior knowledge for reasoning, and make uncertain reasoning clearer and easier to understand logically.

[0070] The Bayesian network modeling needs to follow the following steps: (1) First, construct a Bayesian network structure in the form of a "directed acyclic graph" according to the causal relationships between variables; (2) Then, train the CPT of each variable node based on prior knowledge; (3) Finally, make predictions or locate according to the specific mechanism in combination with the known evidence.

[0071] Bayesian network inference is to obtain the probability distribution of an unknown node according to the states of known variable nodes given a Bayesian network structure model.

[0072] In the mathematical principle of the Bayesian network, assume is a directed acyclic graph, where E is the set of directed edges, A is the set of nodes, represents the node the random variable represented, represents the set of parent nodes of node Therefore, the joint probability of can be expressed as:

[0073] (1)

[0074] For a certain node variable in the Bayesian network, it contains basic events . If the set of observation results of all variables related to it except node is known, then the conditional probability of the occurrence of the th event of node variable is: ​

[0075] (2).

[0076] In fault diagnosis decision-making, according to the collected fault information, determine the state of the corresponding nodes in the Bayesian network, and use the state values of these nodes as the evidence values for diagnosis. Assign them to the corresponding nodes in the Bayesian network, and calculate the probability distribution of the unknown nodes in different states in the network through Bayesian probability reasoning. This method is called the semi-tensor product Bayesian network method, and its mathematical expression is:

[0077] (3).

[0078] 2. Switch quantity processing method based on Bayesian network

[0079] The fault location of the DC distribution network based on switch quantities is to determine the fault components by constructing a Bayesian network according to the relay protection action principle. It is necessary to use the switch quantity information collected by the dispatching center as evidence and substitute it into the Bayesian network of each suspected fault component. The fault probability of the suspected fault component is calculated through Bayesian reverse reasoning. However, since the relay protection information received by the dispatching center is not completely accurate, it brings great difficulties to fault diagnosis. Apply the actual topological structure information of the power grid and the action information of relay protection devices to construct a Bayesian network. Expert system rules are formulated to realize the identification of protection refusal and misoperation, and the action sequence of each layer of protection devices and circuit breakers during a fault is deduced, realizing the deduction of complex faults in the power grid. The calculation and reasoning process of this Bayesian network reasoning method is relatively complex, and the diagnosis efficiency is not high when facing a large-scale distribution network. A Bayesian network reasoning method based on the matrix semi-tensor product theory is proposed. This analysis method helps to study the mathematical properties of the Bayesian network, and it can perform probability network modeling and reasoning more efficiently. Therefore, this method is adopted in the present invention to process switch quantity information.

[0080] When using the Bayesian network to process switch quantity information, the nodes of the Bayesian network are each component in the power grid and the corresponding levels of protection. The present invention uses L, B, CB to represent the bus, line, and circuit breaker respectively ， The main protection, near-backup protection, and far-backup protection of the component correspond to m , p and s respectively. The left-end protection of the line is marked with L, and the right-end protection is marked with R. A local system is established and a Bayesian network is constructed using the relay protection principle. The DC distribution network protection configuration model is as Figure 1 shown. According to the structure of the DC distribution network and its relay protection principle, a corresponding Bayesian network can be established for each component in the system. Assume that the DC bus B 14 has a fault. According to the associated relay protection action principle, the DC bus B can be constructed14 The fault location Bayesian network topology structure, the topology diagram is as Figure 2 shown.

[0081] In the Bayesian network, the arcs connected to the bus B 14 represent the logical relationship between protection and action after a fault occurs. For the bus B 14 after a fault occurs, first the main protection B 14m acts, causing the circuit breakers at both ends CB 14-13 , CB 14-15 to trip. If the circuit breaker CB 14-13 refuses to operate, then the near-backup protection L 13-14 of the remote line L 13-14 L s acts, causing the circuit breaker CB 13-14 to trip.

[0082] Accurately assigning values to the Bayesian network is the key to fault location. Since the Bayesian network is used for fault location in DC distribution networks, there is a lack of data on DC-related prior probabilities and conditional probability tables. DC circuit breakers are relatively new in technology and have been in engineering applications for a short time, without actual engineering data for reference. There is rich historical operation data and expert knowledge on AC circuit breakers. The present invention refers to the relevant prior probability data of AC for the DC prior probability data. It is reasonably assumed that the probability that the protection and circuit breaker in the DC distribution network do not refuse to operate and do not malfunction is 90% of that in the AC distribution network where the protection and circuit breaker do not refuse to operate and do not malfunction. The present invention sets the prior probability of component node faults and the fault probabilities of relay protection devices and circuit breakers as shown in Table 1 and Table 2 respectively. Set a threshold. If the component fault probability, then judge that the component is faulty.

[0083] Table 1 Prior Probability Table of Component Node Faults

[0084] .

[0085] Table 2 Fault Probability Table of Relay Protection Devices and Circuit Breakers

[0086] .

[0087] According to the prior probabilities of component faults in Table 1 and Table 2 and the probabilities of protection (circuit breaker) refusal to operate and malfunction, the fault probabilities of each component are obtained based on Bayes' theorem.

[0088] Fault Information Electrical Quantity Processing Method

[0089] 1. Analysis of Electrical Quantity Characteristics of DC Distribution Network Faults

[0090] In a DC distribution network, a rich high-frequency mutation signal will be generated instantaneously during a bipolar short-circuit fault. This invention analyzes the high-frequency signal; at the same time, the sparse measurement method is used to reconstruct the current characteristics of the mutation at the fault point. The sparse measurement processing method is based on the theory of compressive sensing, which will be introduced in detail in the next section.

[0091] 2. Principle of Compressive Sensing

[0092] The proposed theory of Compressed Sensing (CS) breaks through the limitation of the Nyquist sampling theorem and has become a new research hotspot in the field of signal processing. Its idea is that a signal that is sparse in a certain transform space is multiplied by an observation matrix that satisfies the RIP condition to complete sampling and compression, and then the original signal is reconstructed with high probability by solving a highly nonlinear optimization problem.

[0093] The mathematical model of this theory is expressed as follows: Assume that the signal s is sparse in the domain, and is sparsely represented by the sparse basis :

[0094] (4).

[0095] Among them, is the coefficient in the sparse domain. The sparse signal is measured through the constructed measurement matrix to obtain an observation value y of length M:

[0096] (5).

[0097] Let , and the observation matrix y is transformed into:

[0098] (6).

[0099] Since equation (6) is an underdetermined equation, the original signal s cannot be directly reconstructed using the observation value y. For the sparse signal , when the matrix A satisfies the (Restricted Isometry Property, RIP) restricted isometry condition, the original signal s can be reconstructed by solving the optimization problem (7)

[0100] (7).

[0101] Since The norm minimization problem is an NP (non-deterministic polynomial) hard problem and is often transformed into the norm minimization problem for solution:

[0102] (8)

[0103] Finally, the reconstructed signal is obtained through the reconstruction algorithm.

[0104] 3. Electrical quantity processing method based on compressive sensing

[0105] When obtaining the discrete quantity information, there are often situations of "failure to operate" and "maloperation" of the protection and circuit breaker. The relay protection information received by the dispatching center is not completely accurate, which brings great difficulties to fault diagnosis. It is difficult to accurately locate the DC distribution network fault only relying on a single discrete quantity information. Therefore, combining the electrical quantity information to achieve accurate positioning. Since the acquisition cost of electrical quantity information is relatively high, combined with the compressive sensing algorithm, a small amount of electrical quantity can be collected to accurately reconstruct the original data, and the fault can be quickly located after the power grid fails to obtain an accurate positioning result.

[0106] When a node fault occurs in the DC distribution network, only the fault node will generate high-frequency information, and other nodes are non-fault nodes. Therefore, the high-frequency currents of each node in the network are sparse. The node high-frequency impedance matrix can be obtained by inverting the node high-frequency admittance matrix, that is

[0107] (9)

[0108] In the formula: N is the number of nodes in the DC distribution network.

[0109] The node voltage equation obtained according to the node impedance matrix is:

[0110] (10)

[0111] In the formula: is the column vector of node high-frequency voltages; is the column vector of node high-frequency currents.

[0112] According to the basic theory of compressive sensing, to perform fault location, the number and location of measurement points need to be reasonably configured. The number of measurement points is determined according to the minimum number of measurement points that can reflect the overall characteristics of the compressive sensing algorithm. If it is less than this defined value, it is difficult to perform reconstruction. The calculation of the number of measurement points is shown in formula (11):

[0113] (11)

[0114] In the formula: R is the number of measurement points; is the sparsity of the signal to be reconstructed; M is the length of the signal to be reconstructed.

[0115] Finally, the OMP algorithm is used to reconstruct the high-frequency node current, and the implementation process is as Figure 3 shown:

[0116] The specific steps are as follows:

[0117] (1) First, according to the high-frequency information formed by the fault at the nodes of the DC distribution network, collect the high-frequency admittance matrices at R measurement nodes, and then perform an inverse operation to obtain the sensing matrix .

[0118] (2) Read the measured point voltage data for 2 ms before and after the fault, extract its high-frequency voltage component, and calculate the observation signal with a continuous window length .

[0119] (3) Feed the obtained sensing matrix and the observation signal as parameters into the OMP reconstruction algorithm for calculation to obtain the sparse vector composed of the high-frequency current amplitudes of each node .

[0120] (4) Judge according to the number of iterations until the algorithm ends, and output the sparse characteristics of the fault current.

[0121] DC Distribution Network Fault Location Method Based on Multi-Source Information Fusion

[0122] 1. Multi-Data Source Information Fusion Technology

[0123] D-S evidence theory can effectively solve some uncertain information in unknown environments. This special advantage makes it widely used in information fusion. However, the traditional DS evidence theory has the problem of evidence conflict. Therefore, in order to reduce the evidence conflict problem, an improved DS evidence theory fusion method is proposed.

[0124] Suppose is an exhaustive list of all the results of a specific event, and each event is mutually exclusive, that is can be expressed as: where there are hypothesis events in the set, and each subset contains kinds of events, and there is a mapping relationship from to [0,1] and satisfies , then is called 's basic probability assignment function. Among them, when there is it is called as the focal element.

[0125] The most core of the DS evidence theory is the combination rule, which reflects the fusion process among various evidences. In the same identification framework there are evidences, and each evidence is independent of each other. The basic probability assignment function is , satisfying:

[0126] (12)

[0127] In the formula, is the conflict factor.

[0128] When the combination is valid, when

[0129] the conflict is too high, otherwise the combination rule cannot be used.

[0130] 2 Fault location method based on multi-sensor data fusion algorithm

[0131] First, the compressed sensing algorithm is used to reconstruct the transient electrical quantity data at the measurement component to obtain the sparse characteristics of the fault high-frequency current, and the fault degree of each component is obtained through normalization. Secondly, the relay protection device generates protection action information for the fault component, and a Bayesian network is established under the protection action information to obtain the fault probability of the fault component. At the same time, the sparse characteristics of the fault current and the fault probability are normalized to obtain the fault degree.

[0132] (13)

[0133] In the formula: is the current characteristic of each fault component, , indicating the credibility of the fault location result based on the electrical quantity data, .

[0134] For the switch quantity fault information, the Bayesian network is used to calculate the probability of each device, and the probability value of each fault component is obtained. The switch quantity fault degree of each suspected fault component is obtained through normalization. The normalization formula is as follows:

[0135] (14)

[0136] In the formula: is the probability value of each fault component, , indicating the credibility of the fault location result based on the switch quantity, .

[0137] Secondly, during the information fusion process, the obtained electrical quantity fault degree and switch quantity fault degree are used as two pieces of evidence, and DS evidence theory is utilized for fusion. Meanwhile, each component is an independent recognition framework , and this framework includes three different states = {fault, normal, uncertain}. Assume is the basic probability assignment function. For the two independent evidence bodies of electrical quantity and switch quantity, Equation (12) is adopted to fuse the electrical quantity and switch quantity.

[0138] Finally, a decision is made on the fusion result of the component. According to the obtained fusion result A decision is made through Equation (15)

[0139] (15)

[0140] Where is the threshold value, taking 0.95 and 0.05 respectively. When making a decision on the state of the fused component, if the component state satisfies the above formula, then it is determined that the component is a faulty component.

[0141] The flow of its fusion is as follows:

[0142] Although the traditional DS evidence theory can accurately locate faults, the problem of evidence conflict in the traditional evidence theory still exists. Therefore, in order to reduce the problem of evidence conflict, an improved DS evidence theory fusion method is proposed, and its flow chart is as Figure 6 shown.

[0143] Firstly, for the Bayes estimation method, it mainly solves the problem of how to redistribute and manage conflicts, as follows:

[0144] (16)

[0145] Where , represents the recognition framework contains the sum of elements, represents the set of all elements of the recognition framework . When is a singleton subset, it satisfies the operation rules. When is an uncertain subset, its basic probability is 0.

[0146] Under the recognition framework and the evidence body through the Bayes estimation method, a weighted method is adopted. Firstly, the conflicting evidence is preprocessed, and then the evidence combination rule is used to fuse the evidence, as Figure 5 shown:

[0147] (17)

[0148] Among them: belongs to the subset of the recognition framework .

[0149] After obtaining a new evidence body by using the weighting method and adopting the secondary fusion method, which mainly combines these two methods. Taking the two methods as new evidence sources for fusion is not only suitable for conflict situations but also applicable to situations where the evidence is relatively consistent, and at the same time effectively improves the fault location accuracy.

[0150] Its mathematical expression is:

[0151] (18)

[0152] In the formula , represents the evidence generated under the Bayes estimation method; represents the evidence generated under the weighting method.

[0153] Simulation experiment

[0154] 1. Fault location of multi-source information fusion under typical fault scenarios

[0155] The present invention verifies the effectiveness of the proposed method through the IEEE33-node system. In order to systematically analyze whether the method proposed by the present invention is effective, first, a DC distribution network model is built in the PSCAD power system simulation software, and then fault settings are made and verified.

[0156] 1.1 Example 1

[0157] (1) Fault setting

[0158] The first fault scenario is set as follows: A bipolar short-circuit fault occurs on line L 14-15 , the circuit breaker CB 14-15 refuses to operate, and the remote backup protection of its line L 13-14 starts, prompting the circuit breaker at the left end of line L 13-14 to operate. The protection and circuit breaker operation information is shown in the following table.

[0159] Table 3 Protection and circuit breaker operation information

[0160]

[0161] (2) Fault range determination

[0162] The sampling frequency of this time is 1KHZ. At the same time, in order to meet the minimum number of measurement points condition, the measurement points , the obtained sensing matrix and the observed signal are used to reconstruct the sparse characteristics of the fault current through the compressive sensing algorithm as Figure 6 shown, where the abscissa represents the node number and the ordinate represents the reconstructed fault current amplitude.

[0163] According to the reconstruction result, the fault interval can be obtained, as Figure 7 shown.

[0164] As can be seen from the figure, the set fault line L 14-15 is included in it, thus verifying the effectiveness of the fault area determination.

[0165] (3) Establishment of electrical quantity amplitude and fault degree

[0166] According to the signal reconstruction result, the reconstructed amplitudes of each bus can be obtained. For the line amplitude between two buses, the weighted peak method is used for definition. After obtaining the amplitude, the electrical quantity fault degree is solved by formula (13) as shown in the following table.

[0167] Table 4 Electrical quantity fault degree of components

[0168]

[0169] (4) Establishment of Bayesian network

[0170] The information located by the electrical quantity is calibrated. Through the establishment of the Bayesian network, the topological structure of line L 14-15 is as Figure 8 shown.

[0171] (5) Establishment of switch quantity probability and switch quantity fault degree

[0172] According to the prior probabilities of component faults and the probabilities of protection (circuit breaker) refusal and misoperation in Tables 1 and 2 in the second section, and then combined with Bayes' theorem, the fault probabilities of each component are obtained. After obtaining the component probabilities, the switch quantity fault degree is determined by formula (14). As shown in the following table.

[0173] Table 5 Switch quantity fault degree of each component

[0174]

[0175] (6) DS evidence fusion and decision-making

[0176] DS evidence theory is used for fusion, and Table 6 shows the fusion result.

[0177] Table 6 Fusion result of fault degree of each component

[0178]

[0179] The evidence fusion of Table 4 and Table 5 is carried out using Equation (18), and the fusion result is shown in Table 6. Among them, the decision threshold is based on Equation (15), and the thresholds and are taken as 0.95 and 0.05 respectively. According to the set threshold, the faulty line is L 14-15 , which is consistent with the set faulty line, thus verifying the effectiveness of the DS evidence theory. Figure 9 are the results of three location algorithms for locating all components.

[0180] From Figure 9 , it can be seen that if the fault location is only based on electrical quantities, there will be problems such as the similar fault degrees of busbars B 13 , B 15 , line L 13-14 , line L 14-15 , and it is difficult to determine the fault node; if the switch quantity is used for fault location, under the condition of the refusal of circuit breaker CB 14-15 to operate, although the fault location range is reduced, the location accuracy is not high. By fusing the fault degrees of electrical quantities and switch quantities using the DS evidence theory, the fusion result is more accurate than the single quantity location, verifying the accuracy of the DS evidence theory. Figure 10 is the comparison of the fault degree location results of line L 14-15 using three location methods. According to Figure 10 , it can be seen that the fault location result of line L 14-15 using the DS evidence theory fusion is more accurate.

[0181] 1.2 Example 2

[0182] (1) Fault setting

[0183] The second fault scenario is set as follows: A bipolar short-circuit fault occurs on line L 26-27 , the circuit breaker CB 27-26 refuses to operate, and its remote backup protection of line L 26-27 starts, prompting the circuit breaker CB 27-28 at the right end of line L 28-27 to operate. The protection and circuit breaker action information is shown in Table 7.

[0184] Table 7 Protection and circuit breaker action information

[0185]

[0186] (2) Fault range determination

[0187] Similarly, the sampling frequency of this time is 1KHZ. At the same time, in order to meet the minimum number of measurement points condition, the measurement points, and the obtained sensing matrix and observation signal are reconstructed through the compressive sensing algorithm to obtain the sparse characteristics of the fault current as Figure 11As shown, where the abscissa represents the node number and the ordinate represents the reconstructed fault current amplitude.

[0188] According to the reconstruction result, the fault interval can be obtained, such as Figure 12 shown.

[0189] As can be seen from the figure, the faulty line L 14-15 is included in it, thus verifying the effectiveness of the fault area determination.

[0190] (3) Establishment of electrical quantity amplitude and fault degree

[0191] According to the signal reconstruction result, the reconstructed amplitudes of each bus can be obtained. For the line amplitude between two buses, the weighted peak method is used for definition. After obtaining the amplitude, the electrical quantity fault degree is solved by formula (13) as shown in Table 8.

[0192] Table 8 Electrical quantity fault degree of components

[0193]

[0194] (4) Establishment of Bayesian network

[0195] The information located by the electrical quantity is calibrated. Through the establishment of the Bayesian network, the topological structure of line L 26-27 is as Figure 13 shown.

[0196] (5) Establishment of switch quantity probability and switch quantity fault degree

[0197] According to the prior probabilities of component faults and the probabilities of protection (circuit breaker) refusal and misoperation in Tables 1 and 2 in the second section, and then combined with Bayes' theorem, the fault probabilities of each component are obtained. After obtaining the component probabilities, the switch quantity fault degree is determined by formula (14). As shown in Table 9.

[0198] Table 9 Switch quantity fault degree of each component

[0199]

[0200] Table 10 Fusion results of fault degrees of each component

[0201]

[0202] Formula (18) is used to fuse the evidence in Tables 8 and 9, and the fusion result is shown in Table 10. Among them, the decision threshold is based on formula (15). According to the set threshold, the faulty line is L 26-27 , which is consistent with the set faulty line, thus verifying the effectiveness of the DS evidence theory. Figure 14 is the result of the three positioning algorithms for all component positioning and Figure 15 is line L26-27 Fault location results of three methods

[0203] From Figure 14 it can be seen that if fault location is only based on electrical quantities, there will be a problem that the fault degrees of bus B 27 and line L 26-27 are almost the same, and it is difficult to determine the fault node; if switch quantities are used for fault location, under the condition that circuit breaker CB 27-26 refuses to operate, although the fault location range is reduced, the location accuracy is not high. Fusing the fault degrees of electrical quantities and switch quantities using the DS evidence theory, the fusion result is more accurate than the location using a single quantity, verifying the accuracy of the DS evidence theory, Figure 15 is the comparison of the fault degree location results of line L 26-27 using three location methods. According to Figure 15 it can be seen that for line L 26-27 the fault location result fused using the DS evidence theory is more accurate.

[0204] 2. Algorithm comparison

[0205] Under the conditions of Example 2, the algorithm of the present invention is analyzed and compared with the traditional DS evidence fusion, as well as the Bayes estimation method and the weighting method. The obtained data is shown in Table 11.

[0206] Table 11 Comparison of the fusion results of the fault degrees of each component

[0207]

[0208] Using equations (12), (16), (17), and (18) to fuse the obtained fault degrees of electrical quantities and switch quantities, the fusion results are shown in Table 11. It can be seen from the table that when a bipolar short circuit occurs in line L 26-27 and circuit breaker CB 27-26 refuses to operate, the difference in the fusion results of the traditional DS fusion algorithm, the Bayes estimation method, and the improved weighting method is not significant, and the improved DS fusion algorithm has the highest accuracy. Figure 16 is the comparison of the location results of line L 26-27 using four algorithms.

[0209] It can be seen from the figure that although the Bayes estimation method reduces the conflict between evidences compared with the traditional DS fusion algorithm, it is only limited to the size of the uncertain subset, so the improvement of accuracy is not obvious. The weighting method can better correct the data sources and treat each piece of evidence equally, but the location accuracy is not high. The improved DS combines the Bayes estimation method and the weighting method, reduces the conflict between evidences, and improves the location accuracy.

[0210] Conclusion and outlook

[0211] When a fault occurs in a DC distribution network, the fault information obtained by the dispatching center is mainly the information of electrical quantities and switch quantities. Therefore, this paper introduces the analysis of electrical quantities and switch quantities. First, the compressive sensing algorithm is used to reconstruct the electrical quantities to obtain the amplitude of the fault current. Then, the relay protection device sets the electrical quantity information to generate the information instruction for protection action, and establishes a Bayesian network for the protection action information generated by the fault component to obtain the switch quantity fault probability of the component. At the same time, preprocessing of the fault degree is carried out on the fault current amplitude and the switch quantity fault probability. Finally, the improved DS evidence theory fusion algorithm is used to fuse these two fault degrees to obtain a more accurate result for fault location.

Claims

1. A DC distribution network fault location method based on multi-source information fusion, characterized in that: The steps are as follows: S1. Electrical quantity processing of compressive sensing The node high-frequency impedance matrix can be obtained by inverting the node high-frequency admittance matrix: （9） In the formula: is the number of nodes in the DC distribution network; The node voltage equation obtained from the node impedance matrix is: （10） Where: is the column vector of the high-frequency voltage of the node; is the column vector of the high-frequency current of the node; Reasonably configure the number and location of measurement points. The number of measurement points is as follows: （11） In the formula: is the number of measurement points; is the sparsity of the signal to be reconstructed; is the length of the signal to be reconstructed; Finally, use the OMP algorithm to reconstruct the high-frequency node current. The implementation process steps are as follows: (1) First, according to the high-frequency information formed by a fault occurring at a node in the DC distribution network, collect the high-frequency admittance matrices at R measurement nodes, and then perform an inversion operation to obtain the sensing matrix ; (2)Read the measured point voltage data for 2 ms before and after the fault, extract its high-frequency voltage component, and calculate the observation signal with a continuous window length ; (3) Feed the obtained sensing matrix and the observed signal as parameters into the OMP reconstruction algorithm for calculation to obtain the sparse vector composed of the high-frequency current amplitudes of each node ; (4) Make judgments according to the number of iterations until the algorithm ends, and output the sparse characteristics of the fault current; S2. DC distribution network fault location method based on multi-source information fusion S2.1 Multi-data source information fusion technology Hypothesis is an exhaustion of all result sets for a specific event, and each event is mutually exclusive, that is can be expressed as: where there are hypothetical events in the set, and each subset contains kinds of events, and there is a mapping relationship from to [0, 1] : → [0, 1] and satisfies , then is called 's basic probability assignment function; where there is when it is called the focal element; within the same recognition framework there are pieces of evidence, each of which is independent of the others, and the basic probability assignment function is , satisfying: (12) In the formula , is the conflict factor; The combination is valid when the conflict is too high at other times, otherwise the combination rule cannot be used; S2.2 Fault location based on multi-sensor data fusion algorithm The current amplitude fault degree of each suspected fault component is obtained through normalization. The normalization formula is as follows: (13) Chinese: is the current characteristic of each faulty component, , indicating the confidence level of the fault location result based on electrical quantity data; The probability value of each fault component is used to obtain the switch quantity fault degree of each suspected fault component through normalization. The normalization formula is as follows: (14) Wherein: is the probability value of each faulty component, , representing the confidence level based on the switching quantity fault location result; In the process of information fusion, the obtained electrical quantity fault degree and switch quantity fault degree are used as two pieces of evidence, and DS evidence theory is used for fusion. At the same time, each component is an independent identification framework , and there are three different states included in this framework {fault, normal, uncertain}. Suppose is the basic probability assignment function. For the two independent evidence bodies of electrical quantity and switch quantity, Equation (12) is used to fuse the electrical quantity and switch quantity Make a decision on the fusion result of the component, based on the obtained fusion result , , Make a decision through Equation (15): (15) wherein and are threshold values; When making a decision on the fused component state, if the component state satisfies the above formula, then it is determined that the component is a faulty component; the fusion process is as follows: Solve the problem of how to reallocate and manage conflicts, as follows: (16) Among them , represents the recognition framework including the sum of elements, represents the recognition framework the set of all elements, which satisfies the operation rules when is a singleton subset, and when is an indeterminate subset, its basic probability is 0; By using the Bayes estimation method in the identification framework and the body of evidence under the condition, a weighted method is adopted. First, the conflicting evidence is preprocessed, and then the evidence combination rule is used to fuse the evidence: (17) Wherein: , , are subsets of the recognition framework ; A new evidence body is obtained by using the weighting method After that, a secondary fusion method is adopted: (18) In the formula , represents the evidence generated under the Bayes estimation method; represents the evidence generated under the weighting method.

Images