Small Current Single-Phase Grounding Fault Line Selection Method Based on CIM and Bi-LSTM

Abstract

According to the characteristics of dispatching data in the dispatching center, the present invention provides a single-phase grounding fault line selection method for small current based on CIM and Bi-LSTM. It includes three parts: power system CIM model parsing, single-line grounding fault diagnosis based on Bi-LSTM, and result fusion. First, after the CIM model is parsed, the connection relationships of various power equipment in the power system are obtained, and then a traversal algorithm is designed according to the information of each power equipment to obtain the distribution network topology structure; secondly, a single-line small current single-phase grounding fault diagnosis system based on Bi-LSTM is designed to obtain the fault probability of each line; finally, the load lines connected to the same bus are compared and analyzed through the distribution network topology structure, and the line with single-phase grounding fault is finally obtained. The present invention does not require injecting external signals, nor does it require strict oscillogram data based on substations. It is equipped in the power dispatching center and can monitor the fault of load lines of hundreds of substations based on dispatching data at the same time.

Description

Technical Field

[0001] The present invention relates to the technical fields of power dispatching automation and machine learning, and particularly to a single-phase grounding fault line selection method for small current based on CIM and Bi-LSTM. Background Art

[0002] At present, single-phase grounding faults with small current can be divided into two categories according to the signals they utilize: indirect recognition methods using external signals and direct recognition methods using fault signals. The external signal recognition method is the injection method, and the most representative one is the S injection method. However, the resonant circuit grounded through an arc suppression coil will compensate or even cancel out the injected signal, resulting in poor detection effect. Moreover, the injection method requires adding new signal generation and acquisition equipment to the power grid. The direct recognition methods using fault signals, such as the current amplitude ratio method and current phase method of zero-sequence current, active component method, harmonic method, zero-sequence admittance method, first half-wave method, zero-sequence energy method, waveform similarity method, traveling wave method, and time-frequency analysis method, as well as the multi-criterion line selection method and deep learning method that combine these methods based on fuzzy theory. Although the above methods can achieve the accuracy of fault line selection, they are all based on the recorded wave data of the distribution substation, with very strict requirements for data, requiring millisecond level, and small selection devices need to be installed in each distribution substation.

[0003] The above methods can be referred to the following literatures:

[0004] Jiang Jian, Bao Guanghai. Review of single-phase grounding fault line selection methods in small current grounding systems [J]. Electrical Engineering Technology, 2015, (12): 1-5.

[0005] Gao Ming, Xie Qing. Strategy design of a small current grounding fault location device based on S injection method [J]. Scientific and Technological Innovation and Application, 2015, (15): 13-14.

[0006] Chen Qiufeng. Exploration of single-phase grounding faults and line selection in small current grounding systems [J]. Electronic Test, 2021, (16): 115-116+7

[0007] Zou Chaoxun. Research and application of single-phase grounding fault line selection technology in small current grounding systems [J]. Industrial Science and Technology Innovation, 2019, 1(29): 29-30.

[0008] Wang Jianyuan, Zhu Yongtao, Qin Siyuan. Fault line selection method for small current grounding system based on direction traveling wave energy [J].

[0009] Transactions of China Electrotechnical Society, 2021, 36(19): 12.

[0010] Xu Siyang. Research on fault line selection device for small current grounding system based on multi-criterion fusion [D]. Harbin University of Science and Technology, 2021.

[0011] Hongshan Chen, Yong Shi, Zebing Shi, Kaida Dong, Ming Lu, Jianlong Zhao. Fault Line Selection Method for Single-Phase Grounding Fault with Small Current

[0012] [J]. Power Grid and Clean Energy, 2020, 36(05): 42-48+57 Summary of the Invention

[0013] The purpose of the present invention is to provide a fault line selection method for single-phase grounding fault with small current based on CIM and Bi-LSTM, which is equipped in the power dispatching center. Based on the dispatching data, it can monitor the fault of the load lines of hundreds of substations at the same time, which is of great significance to the field of power dispatching automation.

[0014] The technical solutions to achieve the purpose of the present invention are as follows:

[0015] A fault line selection method for single-phase grounding fault with small current based on CIM and Bi-LSTM, including:

[0016] Step 1, parse the CIM model file to obtain the power grid topology without standby busbars, that is

[0017] {Substation 1: Busbar 1 [Load 1, Load 2, Load 3...]; Busbar 2 [Load 1, Load 2, Load 3...];...; Substation 2: Busbar 1 [Load 1, Load 2, Load 3...]; Busbar 2 [Load 1, Load 2, Load 3...];...;...}; Step 2, construct and train a double-layer Bi-LSTM neural network model; the double-layer Bi-LSTM neural network model includes an input layer, a first hidden layer, a second hidden layer and an output layer connected in sequence; among them, the first hidden layer and the second hidden layer are both Bi-LSTM;

[0018] Step 3, detect 5 features of Load 1 on Busbar 1 of Substation 1 at m time points to obtain feature data with m rows and 5 columns; where m is greater than or equal to 3, and the 5 features are active power, reactive power, effective value of phase current of phase A, effective value of phase current of phase B, and effective value of phase current of phase C; input the feature data with m rows and 5 columns into the double-layer Bi-LSTM neural network model to obtain the line fault probability P1 of Load 1 on Busbar 1 of Substation 1; in the same way, obtain the line fault probabilities P2...Pn of other loads on Busbar 1 of Substation 1 at the same time point; if the maximum value P_max of the fault probabilities P1, P2...Pn of all lines on Busbar 1 of Substation 1 is greater than or equal to the fault probability threshold, it is determined that the line corresponding to P_max has a single-phase grounding fault with small current;

[0019] Step 4: In the same way as in Step 3, determine whether there is a single-phase grounding fault with small current in the load lines of other buses of Substation 1 in the power grid topology, and whether there is a single-phase grounding fault with small current in the load lines of all buses of other substations in the power grid topology.

[0020] A further technical solution is that parsing the CIM model file to obtain a power grid topology without standby buses specifically includes: The CIM model file includes substations, and its attributes include name and identifier; the CIM model file also includes devices; the devices include transformers, transformer windings, circuit breakers, earthing trolleys, buses, loads, transmission lines, and compensators; the devices are connected to nodes through terminals; the attributes of the devices include name, identifier, and terminal; among them, the attributes of the transformer also include the substation to which it belongs; the attributes of the terminal include name, identifier, and node; the attributes of the node include name and identifier.

[0021] 1.1 Read the CIM, store the name of each substation in the list station_all; let i = 1;

[0022] 1.2 Read the identifiers of the transformer windings included in the transformer of the i-th substation in station_all, and store them in the list TW_all; let j = 1;

[0023] 1.3 Read the identifier x1 of the j-th transformer winding in TW_all and the identifier y1 of the node connected through the terminal, input x1 and y1 into the algorithm for obtaining the next-level device to get next_eqID1 and next_cnID1; 1.4 Case 1: If next_eqID1 is empty or next_cnID1 is empty, the traversal of all devices under the current transformer winding ends, let j = j + 1, and return to 1.3;

[0024] Case 2: If there are more than one device identifier in next_eqID1, it means that the subsequent devices of the current device have branched, that is, more than one device is connected; let m = 1, indicating the start of traversing the first branch at the bifurcation, take the m-th device identifier and node identifier in next_eqID1 and next_cnID1, if the m-th node identifier in next_cnID1 is not empty, input it into the algorithm for obtaining the next-level device to get next_eqID2 and next_cnID2; return to 1.4, if Case 1 occurs, m = m + 1, indicating the end of traversing this branch, if Case 2 occurs, it means that there is another bifurcation under this branch, and continue to traverse the subordinate branches at the new bifurcation according to Case 2; when m is greater than the total number of device identifiers in next_eqID2, let j = j + 1, and return to 1.3;

[0025] Case 3: If next_eqID1 contains only one device identifier and next_cnID1 is not empty, it means that only one device is connected after the current device. Input the device identifier and node identifier in next_eqID1 and next_cnID1 into the algorithm for obtaining the next-level devices to continue the traversal, and store the results in next_eqID1 and next_cnID1; return 1.4;

[0026] The algorithm for obtaining the next-level devices is specifically as follows:

[0027] 2.1 Let the identifier x1 of the current device be preveqID, and let the identifier y1 of the node connected by the current device through the terminal be prevcnID;

[0028] 2.2 Screen out all devices connected to the device with the identifier prevcnID, and the identifier of the connected device is not equal to prevcnID; store the screening result in the set next_eqID;

[0029] 2.3 Traverse all device identifiers in next_eqID. If the number of nodes contained in device k is 1, or the number of nodes is 2 and the device name is a transmission line, store an empty list in the set next_cnID, indicating that no other device is connected after this device; if the number of nodes contained in device k is 2 and the device name is not a transmission line, store the identifier in device k that is not equal to precnID into next_cnID;

[0030] 1.5 Store the devices of the bus and load during the traversal process into the set of the i-th substation. After traversing all elements in TW_all, let i = i + 1; return 1.2

[0031] 1.6 After traversing all substations, store the results in station_T: {substation 1[], substation 2[], substation 3[]...};

[0032] 1.7 If there is a spare bus in the substation data, delete the spare bus.

[0033] For a further technical solution, the training of the double-layer Bi-LSTM neural network model is specifically as follows:

[0034] 3.1 Build a medium- and low-voltage distribution network simulation model through Simulink to simulate the actual grid operation status;

[0035] 3.2 Generate a data set by changing the neutral grounding method and the fault line, and divide it into a training set and a test set;

[0036] 3.3 Use the training set to train the double-layer Bi-LSTM neural network model, use the test set to verify the model accuracy rate, and obtain the trained double-layer Bi-LSTM neural network model.

[0037] A further technical solution further includes 3.4: Put the trained double-layer Bi-LSTM neural network model into trial operation on the network, that is, generate a data set by the transfer learning method for the prediction results of the simulation data and the actual power grid fault results, and divide it into a training set and a test set; then use the training set to train the double-layer Bi-LSTM neural network model, use the test set to verify the model accuracy rate, and obtain the optimized double-layer Bi-LSTM neural network model.

[0038] Aiming at the problem that there is a lack of a small current single-phase grounding fault line selection device based on dispatching data in the medium and low voltage distribution network in the field of power dispatching automation, according to the characteristics of the dispatching data of the dispatching center, the present invention provides a small current single-phase grounding fault line selection method based on CIM and Bi-LSTM. Specifically speaking, the present invention includes three parts: the CIM model parsing of the power system, the single-line grounding fault diagnosis based on Bi-LSTM, and the result fusion. First, after the CIM model is parsed, the connection relationship of each power equipment in the power system is obtained, and then the traversal algorithm is designed according to the information of each power equipment to obtain the distribution network topology structure; secondly, a single-line small current single-phase grounding fault diagnosis system based on Bi-LSTM is designed to obtain the fault probability of each line; finally, the load lines connected to the same bus are compared and analyzed through the distribution network topology structure, and finally the line with a single-phase grounding fault is obtained. The present invention does not require injecting external signals, nor does it require strict oscillogram data based on the substation. After being equipped in the power dispatching center, it can monitor the fault of the load lines of hundreds of substations based on the dispatching data. Brief Description of the Drawings

[0039] Figure 1 It is the flow chart of the present invention.

[0040] Figure 2 It is the schematic diagram of the connection relationship between the main categories in the CIM model.

[0041] Figure 3 It is the flow chart of the algorithm for obtaining the next-level equipment for obtaining the power grid topology based on the CIM model.

[0042] Figure 4 It is the flow chart of the traversal algorithm of the substation topology structure for obtaining the power grid topology based on the CIM model.

[0043] Figure 5 It is the schematic diagram of the double-layer Bi-LSTM model structure.

[0044] Figure 6Schematic diagram of the training result of the Bi-LSTM model based on actual dispatching data of the present invention. Detailed implementation manners

[0045] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0046] Aiming at the problem that there is a lack of a small current single-phase grounding fault line selection device based on dispatching data in the low-voltage distribution network in the field of power dispatching automation, the present invention designs a small current single-phase grounding fault line selection method based on CIM (Common Information Model) and Bi-LSTM (Bi-directional Long Short-Term Memory). This method includes three parts: parsing the CIM model of the power system, diagnosing the grounding fault of a single line based on Bi-LSTM, and result fusion. First, after parsing the CIM model, the connection relationships of various power equipment in the power system are obtained, and then a traversal algorithm is designed according to the information of each power equipment to obtain the topology structure of the distribution network; secondly, a small current single-phase grounding fault diagnosis system based on Bi-LSTM is designed to obtain the fault probabilities of each line; finally, the load lines connected to the same bus are compared and analyzed through the topology structure of the distribution network, and the line where the single-phase grounding fault occurs is finally obtained. The technical problems that the present invention needs to actually solve are:

[0047] 1. Parse the CIM model, obtain the connection relationships between power equipment, and design an algorithm to traverse the entire medium- and low-voltage distribution network system to finally obtain the topological relationships of the substation, bus, and load lines.

[0048] 2. Obtain the Bi-LSTM small current single-phase grounding fault diagnosis model based on simulation data, and mount it to the power grid for actual fault dispatching data expansion.

[0049] 3. Optimize the Bi-LSTM small current single-phase grounding fault diagnosis model based on actual dispatching data.

[0050] 4. A small current single-phase grounding fault line selection system based on the fusion of the CIM parsing result and the Bi-LSTM model.

[0051] Among them, the specific steps for generating the power grid topology structure based on CIM are as follows:

[0052] (a) Traverse all the CIM classes and their respective attribute information required to be used in the CIM XML / RDF model file by enumeration as shown in Table 1 (√ represents having this attribute, × represents not having this attribute).

[0053] Table 1:

[0054]

[0055] In Table 1, it includes categories such as Substation (substation name), PowerTransformer (transformer), TransformerWinding (transformer winding), Disconnector (switchgear), Breaker (circuit breaker), Groundisconnector (ground switchgear), BusbarSection (busbar), EnergyConsumer (load), ACLineSegment (transmission line), Compensator (compensator), ConnectivityNode (node), Terminal (terminal). Among them, the categories except Substation, ConnectivityNode, and Teriminal are collectively referred to as electrical equipment; the attribute information includes mRID (identifier), Terminal (electrical equipment terminal), Name (name), ConnectivityNode (terminal node). Each category has its corresponding mRID number as the unique identifier. Terminal in the attributes is a unique attribute of electrical equipment. An electrical equipment has at most two Terminals. ConnectivityNode in the attributes is a unique attribute of Terminal, and one Terminal corresponds to one ConnectivityNode.

[0056] (b) Design an algorithm to obtain lower-level electrical equipment. There are usually more than one device connected to the lower level of the current device. Through this algorithm, the mRID and ConnectivityNode of the lower-level device can be obtained.

[0057] (c) Design an algorithm to traverse the topological structure of the distribution substation. Based on the algorithm for obtaining lower-level electrical equipment, all devices in the distribution network are traversed. When there is more than one lower-level device, the current state is defined as a fork, and the mRID numbers of all electrical equipment in the current state are saved. Every time a fork state appears later, it will be compared with the previous one. If they are the same, the traversal will not be repeated, thus solving the situation of repeated traversal caused by the loop network in the medium- and low-voltage distribution network.

[0058] (d) The Substations, BusbarSections, and EnergyConsumers obtained during the cache traversal are finally used to obtain the topological structure of the medium- and low-voltage distribution network as Substation1{BusbarSection1[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]; BusbarSection2[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]…}; Substation2{BusbarSection1[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]; BusbarSection2[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]…}….

[0059] (e) Solve the problem that the standby BusbarSection is also obtained. There will be standby BusbarSections in the medium- and low-voltage distribution network, resulting in the appearance of Substation1{BusbarSection1, BusbarSection2[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]} in the final traversal result, where BusbarSection1 is the standby busbar and no load is connected to it, which has no meaning for the final topological structure. Therefore, the standby BusbarSections in the topological structure are deleted through the designed algorithm to obtain the complete and effective topological structure of the medium- and low-voltage distribution network.

[0060] The specific steps of the small current single-phase grounding fault diagnosis model based on Bi-LSTM are as follows:

[0061] (a) Enrich the data set. Build a simulation model of the medium- and low-voltage distribution network through Simulink to simulate the actual grid operation state. Generate a large number of data sets for training the model by changing the neutral grounding method and the fault line. Each data has a length of 20 and a width of 5. The length represents 20 time points, and the width represents 5 features, namely active power, reactive power, and three-phase phase current.

[0062] (b) Bi-LSTM deep learning model. The model includes an input layer, 2 hidden layers, and an output layer. The hidden layers contain 36 and 72 neurons respectively. Adam is selected as the model optimizer, and the output result is the fault probability of a single load line.

[0063] (c) Train the pre-trained model based on simulation data. 2767 pieces of data with the transition resistance at the neutral point less than 1000 ohms obtained from the simulation are divided into a training set and a test set in a ratio of 8:2 for training the Bi-LSTM model. The final model achieved a fault diagnosis accuracy rate of 98.69% in the low-resistance state.

[0064] (d) Verify the fault diagnosis accuracy rate of the trained model in the case of high-resistance grounding at the neutral point. Input 294 pieces of high-resistance grounding data with a transition resistance of 1500 ohms at the grounding point and 288 pieces of high-resistance grounding data with a transition resistance of 2000 ohms at the grounding point into the model respectively. The fault diagnosis accuracy rates of the model reached 95.11% and 91.02% respectively, which is of great significance for the small-current single-phase grounding fault based on dispatching data.

[0065] Combine the CIM parsing result and the Bi-LSTM model. The specific steps are as follows:

[0066] Generate the connection relationship between the bus and the load line in the distribution network topology according to the CIM-based power grid topology structure, and obtain the fault probability of each load line through the small-current single-phase grounding fault diagnosis model based on Bi-LSTM. Screen out the load line with the highest fault probability connected to the same bus by voting, thereby eliminating the situation where two lines are detected as faulty at the same time and preventing the misjudgment of normal lines when there are too many load lines, effectively improving the accuracy rate of fault line selection of the model. Specific embodiments

[0068] As Figure 1 shown, the small-current single-phase grounding fault line selection method based on the CIM and Bi-LSTM models mainly includes the generation of the power grid topology structure based on the CIM XML / RDF model file, the small-current single-phase grounding fault diagnosis of the load line based on the Bi-LSTM model, and the small-current single-phase grounding fault line selection that fuses the results of the previous two parts. The specific technical solutions are as follows:

[0069] 1. Generate the power grid topology structure based on the CIM XML / RDF model file

[0070] By analyzing the CIM XML / RDF model file and studying the relationships between classes and the attributes contained in each class in the CIM model, the connection relationships between the main classes in the CIM model can be obtained as Figure 2As shown, in order to finally obtain the topological structure in the format of Substation1{BusbarSection1[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]; BusbarSection2[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]…}; Substation2{BusbarSection1[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]; BusbarSection2[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…]…}…, it is necessary to traverse middleware such as VoltageLevel, PowerTransformer, TransformerWinding, Disconnector, Breaker, Groundisconnector, etc., and also prevent traversing to the next substation across the ACLineSegment. Among them, the connection relationship between each power equipment starting from the PowerTransformer is: Power Equipment 1 → Terminal → ConnectivityNode → Power Equipment 2 → Terminal → ConnectivityNode → Power Equipment 3 → Terminal → ConnectivityNode →… Usually, a ConnectivityNode is connected to more than one power equipment. Therefore, an algorithm for obtaining the next equipment is designed, and the algorithm flowchart is as Figure 3 shown, and the specific steps are as follows:

[0071] (1) Input the equipment number (mRID number of the equipment) of the current equipment as preveqID and the number of its backend node (mRID number of the ConnectivityNode) as prevcnID;

[0072] (2) Filter out all equipment with the node number prevcnID and the equipment number not equal to preveqID, and store the equipment numbers of these equipment in the set next_eqID;

[0073] (3) Traverse all devices in next_eqID. If the number of ConnectivityNodes contained in device i is 1, then store an empty list in the set next_cnID, indicating that the traversal of this branch has ended; if the number of ConnectivityNodes contained in device i is 2 and the device name is not ACLineSegment, then store the device number in device i that is not equal to precnID into next_cnID.

[0074] (4) Finally, output next_eqID and next_cnID, representing the set of device numbers of the next-level devices and the set of node numbers at the back end of the next-level devices respectively.

[0075] Based on the above algorithm for obtaining the next level, design a traversal algorithm for the topological structure of the distribution substation. The algorithm flow chart is as Figure 4 shown, and the specific steps are as follows:

[0076] (1) Read the parsing result of the CIM / XML file, traverse all Substations, obtain all site names name and store them in the list station_all, and set i = 1;

[0077] (2) Read the TransformerWindings contained in the PowerTransformer of the i-th site in station_all, and store them in the list TW_all, and set j = 1;

[0078] (3) Read the j-th TransformerWinding in TW_all, with the device number x1 and the back-end node number y1. Input x1 and y1 into the algorithm for obtaining the next-level devices to get next_eqID1 and next_cnID1.

[0079] (4) The obtained results of the next-level devices are divided into three cases;

[0080] Case 1: If next_eqID1 is empty or next_cnID1 is empty, then the traversal of all devices under the current transformer winding ends. Let j = j + 1 and return to (3);

[0081] Case 2: If there are more than one device identifier in next_eqID1, it means that the current device has bifurcated subsequently, that is, more than one device is connected; let m = 1, indicating the start of traversing the first branch at the bifurcation. Take the m-th device identifier and node identifier in next_eqID1 and next_cnID1. If the m-th node identifier in next_cnID1 is not empty, input it into the algorithm for obtaining the next-level device to get next_eqID2 and next_cnID2; return 1.4. If Case 1 occurs, m = m + 1, indicating the end of traversing this branch. If Case 2 occurs, it means that there is another bifurcation under this branch, and continue to traverse the subordinate branches at the new bifurcation according to Case 2; when m is greater than the total number of device identifiers in next_eqID2, let j = j + 1 and return (3);

[0082] Case 3: If next_eqID1 contains only one device identifier and next_cnID1 is not empty, it means that only one device is connected after the current device. Input the device identifier and node identifier in next_eqID1 and next_cnID1 into the algorithm for obtaining the next-level device to continue traversing, and store the results in next_eqID1 and next_cnID1; return (4);

[0083] (5) Store all devices of categories BusbarSection and EnergyConsumer encountered during the traversal into the set of the i-th Substation. After traversing all elements in T W _all, increment i by 1 and return (2); (6) After traversing all Substations, store the results in station_T and output the result as station_T{Substation1[], Substation2[], Substation3[]…}, where Substation1[] expands to Substation1[BusbarSection1, EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…, BusbarSection2, EnergyConsumer1, EnergyConsumer2, EnergyConsumer3…].

[0084] After the traversal, BusbarSection1, BusbarSection2, EnergyConsumer1, EnergyConsumer2, EnergyConsumer3... still exist in the result, that is, the situation where two BusbarSections are connected together. The reason for this situation is that BusbarSection1 is a standby busbar used for subsequent expansion of the power grid scale and needs to be removed from the model for subsequent algorithm implementation. Traverse all the indexes of BusbarSection in station_T, delete the previous BusbarSection in adjacent indexes, and store the subsequent EnergyConsumers in the set of the latter BusbarSection. Finally, the model parsing result is: station_T{Substation1:BusbarSection1[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3...]; BusbarSection2[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3...];...; Substation2:BusbarSection1[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3...]; BusbarSection2[EnergyConsumer1, EnergyConsumer2, EnergyConsumer3...];...;...}.

[0085] 2. Single-phase Grounding Fault Diagnosis of Load Lines Based on Bi-LSTM

[0086] Build a two-layer Bi-LSTM model, and the model structure is as Figure 5 shown. The two-layer Bi-LSTM neural network structure model is composed of 4 LSTM models. The input sequence is input into 2 LSTM neural networks for feature extraction in forward and reverse order respectively. The forward and reverse LSTMs do not interfere with each other. The feature vector formed by splicing the 2 output vectors (i.e., the extracted feature vectors) is used as the final feature expression of the signal. Therefore, the feature data of the Bi-LSTM model has information between the past and the future at the same time. Since there is no fault detection device in the actual power grid dispatching center, the model cannot be trained with actual data at the beginning. Therefore, it is necessary to pre-train the model with simulation data, then mount the model to the dispatching center to capture data suspected of faults during a period of time, then organize the actual data set, and optimize the model again based on transfer learning. The specific steps are as follows:

[0087] (1) Enrich the dataset. Build a simulation model of medium- and low-voltage distribution networks through Simulink to simulate the actual operation status of the power grid. Generate a large number of datasets for training the model by changing the neutral grounding method and the faulty line. Each piece of data has a length of 20 (the minimum is set to 3) and a width of 5. The length represents 20 time points, and the width represents 5 features, namely active power, reactive power, and the effective values of three-phase phase currents.

[0088] (2) Build a two-layer Bi-LSTM deep learning model. The model includes an input layer, two hidden layers, and an output layer. The hidden layers contain 36 and 72 neurons respectively (the number of neurons before and after is a multiple of 4, and the effect is better when the ratio of the number of neurons before and after is 1:2). Select Adam (other optimizers can also be applied) as the model optimizer, and the output result is the fault probability of a single load line.

[0089] (3) Train the pre-trained model based on simulation data. Divide the 2,767 pieces of data with the transition resistance at the neutral point lower than 1,000 ohms (low-resistance case) obtained from the simulation into a training set and a test set in a ratio of 8:2 for training the Bi-LSTM model. The final model achieved a fault diagnosis accuracy rate of 98.69% in the low-resistance state.

[0090] (4) Verify the fault diagnosis accuracy rate of the trained model under the condition of high-resistance grounding at the neutral point. Input the 294 pieces of high-resistance grounding data with a transition resistance of 1,500 ohms at the grounding point and the 288 pieces of high-resistance grounding data with a transition resistance of 2,000 ohms at the grounding point into the model respectively, and the fault diagnosis accuracy rates of the model reached 95.11% and 91.02% respectively.

[0091] (5) Conduct on-grid trial operation and capture suspected fault data in actual operation data. Embed the Bi-LSTM model trained based on simulation data into the power grid dispatching center for on-grid trial operation. Two months later, a total of 218 fault records were finally screened out based on the operation logs of the dispatching center and the prediction results of the Bi-LSTM model, with a prediction accuracy rate of 73.4%. Manually screen out 360 pieces of normal data, and a total of 578 pieces of data are used to train the model.

[0092] (6) Optimize the Bi-LSTM single-phase grounding fault diagnosis model for load lines. Based on the model weights obtained in steps (3) and (5) and 578 pieces of actual dispatching data, through transfer learning, take out 463 pieces of data as the training set and 115 pieces of data as the test set from the 578 pieces of data in a ratio of 8:2. The final fault recognition accuracy rate is 96.52%, as shown in Figure 6 shown.

[0093] 3. Small current single-phase grounding fault line selection system based on the fusion of CIM parsing results and Bi-LSTM model

[0094] The fault diagnosis completed in two stages only targets a single line. When a single-phase grounding fault actually occurs, the phase currents of the load lines connected under the same busbar will all change significantly. At this time, multiple lines under the same busbar will be diagnosed as having faults. Therefore, it is necessary to analyze the fault conditions of each load line under the current busbar based on the parsing results of the CIM model, that is, the power grid topology, and finally obtain the line where the actual fault occurs. The specific steps are as follows:

[0095] (1) Detect each site in turn according to the parsing results of the CIM model. The detection order is EnergyConsumer1 of BusbarSection1 under Substation1. Take the active power, reactive power, and RMS values of the three-phase phase currents at the current time point and the previous 19 time points of EnergyConsumer1, and input a total of 20 rows and 5 columns of data into the Bi-LSTM small-current single-phase grounding fault diagnosis model to obtain the fault probability P1. Then detect EnergyConsumer2 of BusbarSection1 under Substation1 in the same way to obtain the fault probability P2.

[0096] (2) After detecting all EnergyConsumers under BusbarSection1, obtain the fault probabilities P = {P1, P2, P3...Pn} of all load lines under the same busbar. Find the maximum value P_max in P. If P_max is greater than or equal to 0.5 (the fault probability threshold), then the line corresponding to P_max has a small-current single-phase grounding fault; otherwise, no fault occurs and the line operates normally.

[0097] (3) Then start to detect the fault probabilities of all EnergyConsumers under BusbarSection2 of Substation1 in turn. When all BusbarSections under Substation1 are detected, start to detect Substation2 until all sites included in the CIM model are detected. After detecting all sites, repeat steps (1)(2)(3) and start to detect the operating conditions of the lines at the next moment.

[0098] (4) Verify the reliability of the method. Put the optimized small-current single-phase grounding fault line selection system into operation on the network. After 1 month, based on the operation logs of the power dispatching center and the prediction results of the Bi-LSTM model, 44 fault records are finally selected. The small-current single-phase grounding fault line selection method based on CIM and Bi-LSTM selects the correct lines for 43 fault records, and the accuracy rate reaches 97.7%, which is of great significance for small-current single-phase grounding faults based on dispatching data.

[0099] The specific implementation methods of the line selection method include the following points:

[0100] (1) Data acquisition. This small current single-phase grounding fault line selection method uses the data of three zones in the power system as the original data, pushes the data through Kafka and stores it in the SQL database. It totally includes 6 categories, namely BusbarSection, EnergyConsumer, AclineSegment, Disconnector, Breaker, and Compensator. And the data in each category is classified according to different stations. The present invention applies the information in the EnergyConsumer database to select the line for small current single-phase grounding fault.

[0101] (2) Data processing. The actual data is real-time data. Based on the current time point, the present invention selects the data of the previous 20 time points as the model input. The data of each time point includes 5 feature vectors, namely the active power, reactive power, and the effective values of three-phase phase currents of the load line. After dividing each column of data by the maximum value of that column (data normalization), it is input into the Bi-LSTM small current single-phase grounding fault diagnosis model to obtain the line fault probability.

[0102] (3) Selection of the line for small current single-phase grounding fault. Detect each station in turn according to the parsing result of the CIM model. The detection order is EnergyConsumer1 of BusbarSection1 under Substation1 to obtain the fault probability P1. After detecting all EnergyConsumers under BusbarSection1, the fault probabilities of all lines under the same bus are obtained as P = {P1, P2, P3... Pn}, and the maximum value P_max in P is calculated.

[0103] (4) Storing the line selection result into the database. If P_max is greater than or equal to 0.5 (fault probability threshold), then the line corresponding to P_max has a small current single-phase grounding fault. Then, the data of all running EnergyConsumers connected under the same BusbarSection at the current moment are stored in the fault log, which is convenient for subsequent expansion of the model data set and fault query. If P_max is less than 0.5, no fault occurs and the line operates normally, and the system continues to detect the subsequent stations.

Claims

1. A small - current single - phase grounding fault line - selection method based on CIM and Bi - LSTM, characterized by comprising: Step 1, parse the CIM model file to obtain the power grid topology without the spare bus, that is {Substation 1: Bus 1[Load 1, Load 2, Load 3...]; Bus 2[Load 1, Load 2, Load 3...];...; Substation 2: Bus 1[Load 1, Load 2, Load 3...]; Bus 2[Load 1, Load 2, Load 3...];...;...}; Step 2, construct and train a double - layer Bi - LSTM neural network model; the double - layer Bi - LSTM neural network model includes an input layer, a first hidden layer, a second hidden layer, and an output layer connected in sequence; among them, both the first hidden layer and the second hidden layer are Bi - LSTM; Step 3, detect 5 features of Load 1 on Bus 1 of Substation 1 at m time points to obtain feature data with m rows and 5 columns; where m is greater than or equal to 3, and the 5 features are active power, reactive power, effective value of phase - A current, effective value of phase - B current, and effective value of phase - C current; input the feature data with m rows and 5 columns into the double - layer Bi - LSTM neural network model to obtain the line fault probability P1 of Load 1 on Bus 1 of Substation 1; in the same way, obtain the line fault probabilities P2...Pn of other loads on Bus 1 of Substation 1 at the same time points; if the maximum value P_max of the fault probabilities P1, P2...Pn of all lines on Bus 1 of Substation 1 is greater than or equal to the fault probability threshold, it is determined that the line corresponding to P_max has a small - current single - phase grounding fault; Step 4, in the same way as Step 3, determine whether the load lines of other buses in Substation 1 in the power grid topology have small - current single - phase grounding faults, and whether the load lines of all buses in other substations in the power grid topology have small - current single - phase grounding faults; The parsing of the CIM model file to obtain the power grid topology without the spare bus is specifically: The CIM model file includes substations, and its attributes include name and identifier; The CIM model file also includes devices; devices include transformers, transformer windings, circuit breakers, earthing trolleys, buses, loads, transmission lines, and compensators; devices are connected to nodes through terminals; the attributes of devices include name, identifier, and terminals; among them, the attributes of transformers also include the substations to which they belong; the attributes of terminals include name, identifier, and nodes; the attributes of nodes include name and identifier; 1.1 Read CIM, store the name of each substation in the list station_all; let i = 1; 1.2 Read the identifiers of the transformer windings included in the transformer of the i-th substation in station_all, and store them in the list TW_all; let j = 1; 1.3 Read the identifier x1 of the j-th transformer winding in TW_all and the identifier y1 of the node connected through the terminal, and input x1 and y1 into the algorithm for obtaining the next-level device to get next_eqID1 and next_cnID1; 1.4 Case 1: If next_eqID1 is empty or next_cnID1 is empty, the traversal of all devices under the current transformer winding ends. Let j = j + 1, and return to 1.3; Case 2: If next_eqID1 contains more than one device identifier, it means that there is a bifurcation after the current device, that is, more than one device is connected. Let m = 1, indicating the start of traversing the first branch at the bifurcation. Take the m-th device identifier and node identifier from next_eqID1 and next_cnID1. If the m-th node identifier in next_cnID1 is not empty, input it into the algorithm for obtaining the next-level device to get next_eqID2 and next_cnID2; return to 1.

4. If Case 1 occurs, m = m + 1, indicating the end of traversing this branch. If Case 2 occurs, it means that there is another bifurcation under this branch, and continue to traverse the subordinate branches at the new bifurcation according to Case 2; when m is greater than the total number of device identifiers in next_eqID2, let j = j + 1, and return to 1.3; Case 3: If next_eqID1 contains only one device identifier and next_cnID1 is not empty, it means that only one device is connected after the current device. Input the device identifier and node identifier in next_eqID1 and next_cnID1 into the algorithm for obtaining the next-level device to continue the traversal, and store the results in next_eqID1 and next_cnID1; return to 1.4; The algorithm for obtaining the next-level device is specifically as follows: 2.1 Let the identifier x1 of the current device be preveqID, and let the identifier y1 of the node connected through the terminal of the current device be prevcnID; 2.2 Screen out all devices connected to the device with the identifier prevcnID, and the identifier of the connected device is not equal to prevcnID; store the screening result in the set next_eqID; 2.3 Traverse all device identifiers in next_eqID. If the number of nodes contained in device k is 1, or the number of nodes is 2 and the device name is transmission line, then store an empty list in the set next_cnID, indicating that there are no other devices connected to this device subsequently; if the number of nodes contained in device k is 2 and the device name is not transmission line, then store the identifier in device k that is not equal to precnID into next_cnID; 1.5 Store the devices of busbars and loads during the traversal process into the set of the i-th substation. After traversing all elements in TW_all, let i = i + 1; return to 1.2 1.6 After traversing all substations, store the results into station_T: {substation 1[], substation 2[], substation 3[]...}; 1.7 If there are spare busbars in the substation data, delete the spare busbars.

2. The small current single-phase grounding fault line selection method based on CIM and Bi-LSTM as claimed in claim 1, wherein, The described training of the double-layer Bi-LSTM neural network model is specifically as follows: 3.1 Build a medium- and low-voltage distribution network simulation model through Simulink to simulate the actual grid operation status; 3.2 Generate a data set by changing the neutral grounding method and the fault line, and divide it into a training set and a test set; 3.3 Use the training set to train the double-layer Bi-LSTM neural network model, and use the test set to verify the model accuracy to obtain the trained double-layer Bi-LSTM neural network model.

3. The small current single-phase grounding fault line selection method based on CIM and Bi-LSTM as claimed in claim 2, wherein, It also includes 3.4: Put the trained double-layer Bi-LSTM neural network model into trial operation on the grid, that is, generate a data set through the transfer learning method from the prediction results of the simulation data and the actual grid fault results, and divide it into a training set and a test set; Then use the training set to train the double-layer Bi-LSTM neural network model, and use the test set to verify the model accuracy to obtain the optimized double-layer Bi-LSTM neural network model.

Images