Secondary system fault location method for smart substation based on Transformer-GRU

Abstract

The present invention discloses a fault location method for the secondary system of an intelligent substation based on Transformer-GRU, comprising the following steps: for each fault scenario of each interval secondary device, collect alarm signals from related equipment to form an alarm signal set and a corresponding fault position; for each fault scenario of each interval, construct a feature matrix; for a small sample in each interval, copy it; considering network delay, for each fault scenario of each interval, disrupt the order of the row vectors at each moment under the fault scenario, and obtain multiple feature matrices; construct the Transformer-GRU fault location model of each interval respectively; train the Transformer-GRU fault location model of each interval respectively; send the test set of each interval into the Transformer-GRU fault location model for testing, complete the fault location of the secondary system, and count the test accuracy of each model. The present invention can accurately locate the fault of the secondary system of the substation, and provide efficient and accurate auxiliary decision-making for operation and maintenance personnel.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault detection, and in particular to a secondary system fault locating method for an intelligent substation based on Transformer-GRU. Background Art

[0002] Smart substations can collect in-station equipment information through network communication platforms and monitor the operating status of primary and secondary equipment. Their safe and reliable operation is crucial to the stable operation of the power system. Traditional smart substation fault location methods are mainly based on expert systems, combining communication devices and the status information of each secondary device to determine the fault location. When an abnormality occurs in the secondary system of a smart substation, a large number of alarm signals will be generated. It is difficult for operation and maintenance personnel to extract useful information from a large number of alarm signals and accurately determine the fault location.

[0003] Reference 1: Ye Yuanbo, Sun Yueqin, Huang Taigui, et al.; Online monitoring and fault diagnosis technology for relay protection secondary circuit of smart substation [J], Power System Protection and Control, 2016, 44(20): 148-153. A method for online monitoring and fault diagnosis of relay protection secondary circuit based on multi-parameter information is proposed. Through a network message analyzer or protection status detection device, the alarm signals generated when a fault occurs in the secondary system of the smart substation are collected for fault diagnosis. However, the operation and maintenance personnel need to extract relevant alarm signals based on experience.

[0004] Reference 2: LIU YQ, GAO HL, GAO WC, et al. Development of a substation-area backup protective relay for smart substation[J]. IEEE Transactions on SmartGrid, 2017, 8(06): 2544-2553. A substation area backup protection is proposed, which uses the current differential principle to locate the faulty component.

[0005] Reference 3: Wang Yueyue, Chen Minyu, Jiang Zhenchao, et al., State assessment of secondary equipment in smart substation based on cloud theory [J]. Power System Protection and Control, 2018, 46(01): 71-77; proposed a state assessment model for secondary equipment based on variable weight theory and trapezoidal cloud model, using the intermediate node information of relay protection device.

[0006] Reference 4: Dai Zhihui, Lu Hao, Liu Yuan, et al., Intelligent substation protection secondary circuit fault diagnosis method based on improved DS evidence theory [J]. Power System Protection and Control, 2020, 48(09): 59-67. By parsing the (Substation configuration description, SCD) file and analyzing the mapping relationship between the virtual circuit and the physical circuit of the secondary circuit, the fault location of the secondary circuit of the smart substation is carried out based on the improved DS evidence theory. However, there are differences in the SCD files of different smart substations, which makes it difficult to apply to other smart substations.

[0007] In recent years, artificial intelligence and big data technologies have developed rapidly and become a new type of intelligent substation fault location method. Reference 5: Wang Lei, Chen Qing, Gao Hongyu, et al., Intelligent substation fault tracking architecture based on big data mining technology [J]. Automation of Electric Power Systems, 2018, 42(03): 84-91. Spark is used to mine fault information and the decision tree theory is used to find the cause of the fault, but a large amount of monitoring data is required to accurately determine the fault location.

[0008] Reference 6: REN B, LI JS, ZHENG YK, et al. Research on Fault Location of Process-Level Communication Networks in Smart Substation Based on Deep Neural Networks[J]. IEEE Access, 2020, 8: 109707-109718. A process-level communication network fault location method based on deep neural network is proposed. Through redundant monitoring of the fault state, the fault characteristic information of monitoring nodes at different locations is summarized, but the complex communication network structure is not taken into account.

[0009] Reference 7: Sun Yuyan, Cai Zexiang, Guo Caishan, et al.; A method for fault diagnosis and location of smart substation communication network based on deep learning [J]. Power System Technology, 2019, 43(12): 4306-4314; Using the message reception status of the sink and the flow status information of the switch, a smart substation communication network fault location model based on deep belief network is proposed, but the fault feature information used is relatively small.

[0010] Reference 8: Ren Bo, Zheng Yongkang, Wang Yongfu, et al., Research on fault location of secondary equipment in smart substation based on deep learning [J]. Power System Technology, 2021, 45(02): 713-721; proposed a fault location model for secondary equipment in smart substation based on recurrent neural network, but did not take into account the communication network delay and the time sequence of alarm signals arriving at the station control layer. The alarm signals with disordered sequence could not be accurately identified, and the analog value of current was used for each fault scenario, resulting in the need to collect too much information.

[0011] The existing deep learning network fault location model for smart substations has certain problems: (1) The number of alarm signal samples corresponding to different fault locations in the same interval is different, which may lead to serious imbalance of alarm signal samples at each fault location, causing the neural network to misjudge due to insufficient training of small samples; (2) The communication delay of secondary equipment in each interval will cause the time sequence of the alarm signal arriving at the station control layer equipment to be disordered, which will also affect the accuracy of fault location.

[0012] In view of the problems existing in the fault location of the above-mentioned smart substation, the present invention collects the alarm signals from the relevant equipment when the secondary equipment of each interval of the smart substation fails, combines the Transformer network with the gated recurrent unit (GRU) network, constructs the deep learning network model of the fault location of each interval, and trains and tests the model. By comparing with other traditional deep learning fault location models, the effectiveness of the fault location model based on Transformer-GRU is verified. Summary of the invention

[0013] In view of at least one defect of the prior art, the object of the present invention is to provide a smart substation secondary system fault locating method based on Transformer-GRU, which can be used to accurately locate substation secondary system faults and provide auxiliary decision-making for operation and maintenance personnel.

[0014] In order to achieve the above object, the present invention adopts the following technical solution: a fault location method for the secondary system of an intelligent substation based on Transformer-GRU, comprising the following steps:

[0015] Step A: For each fault scenario of each secondary device in the secondary system of the line bay, transformer bay, bus bay, and central switch bay, collect alarm signals from related devices to form an alarm signal set and corresponding fault location; for the CT disconnection fault scenario of the line merging unit under the line bay, collect current sampling values ​​to obtain a judgment value, which is used as an auxiliary variable and added to the alarm signal set;

[0016] Step B: For each fault scenario in each interval, construct a corresponding feature matrix;

[0017] Step C: for a small sample fault scenario of a secondary device in each interval, replicate it; considering network delay, for each fault scenario in each interval, disrupt the order of the row vectors at each moment under the fault scenario to obtain multiple feature matrices; in a specific embodiment, the small sample fault scenario may refer to a sample set with a sample number of less than 10;

[0018] Step D: normalize and preprocess the feature matrix corresponding to each fault scenario in each interval;

[0019] Step E: Construct the Transformer-GRU fault localization models for each interval, design their network structures, and set training parameters and evaluation indicators;

[0020] Step F: Train the Transformer-GRU fault localization model of each interval separately and optimize the parameters of each model;

[0021] Step G: Send the test sets of each interval into their respective Transformer-GRU fault location models for testing, complete the fault location of the secondary system of the smart substation, and count the test accuracy of each model.

[0022] The normalized feature matrix is ​​used as the training fault sample of the Transformer-GRU fault localization model; the training fault samples are divided into a training set and a test set.

[0023] In step A, the fault scenarios include: the fault scenarios include: the self-test information of the secondary device; the SV / GOOSE receiving status of the secondary device; the sending / receiving light intensity status of the network port of the secondary device and the switch; the judgment amount of the protection device from the sampling value;

[0024] The fault location includes secondary equipment such as protection equipment, measurement and control equipment, merging units, and their sending and receiving network ports, optical fiber ports, and switch ports;

[0025] According to the above fault scenarios, construct the alarm signal set S under each fault scenario:

[0026] S={S ZJ ,S SG ,S GQ ,S FZ} (3);

[0027] Where: S ZJ S is the self-test information of the secondary equipment; SG It is the receiving status of the secondary equipment SV / GOOSE; S GQ S is the sending / receiving light intensity status of the secondary device and switch port; FZIt is an auxiliary judgment quantity for protecting the sampling value of the equipment.

[0028] In the step A, for the CT disconnection fault scenario of the line merging unit under the line interval, the line protection phase A, phase B, and phase C current sampling values ​​are obtained, and it is checked whether the current amplitudes of phase A, phase B, and phase C are zero. If so, the corresponding judgment value is set to 1, otherwise, the corresponding judgment value is set to 0, and these judgment values ​​are used as analog signal auxiliary variables and added to the alarm signal set of the corresponding fault scenario.

[0029] In step B, for each fault scenario in each interval, constructing a corresponding feature matrix includes:

[0030] For all alarm signal sets of each interval, first count the total number of alarm signals and sampled value judgment quantities of protection equipment in the interval as M, and count the total number of fault locations in the interval as N; then number each alarm signal and analog signal auxiliary variable from 1 to M, and number each fault location from 1 to N, and each number corresponds to a fault location vector with a length of N;

[0031] In a fault scenario where a secondary device fails in each interval, the station control layer device successively receives alarm signals sent by multiple related devices at different times, and regards the alarm signal sent by the same device as the alarm signal at that time; the maximum number of fault times of each fault scenario in the interval is recorded as T, and the maximum number of alarm signals at each fault time of each fault scenario in the interval is recorded as A;

[0032] For each fault moment in each fault scenario in each interval, the numbers of the alarm signals and analog signal auxiliary variables are combined into a 1×A row vector; if the number of alarm signals at this moment is less than A, the missing elements in the row vector are filled with 0;

[0033] The row vectors of each fault moment in each fault scenario are combined to form a T×A feature matrix for the fault scenario; if the number of fault moments in the fault scenario is less than T, the elements in the insufficient row vectors in the feature matrix are filled with 0.

[0034] In the step D, normalizing the feature matrix for preprocessing includes:

[0035] The elements of the characteristic matrix of each interval are the alarm signal numbers. The alarm signal numbers are normalized and mapped to x using the Min-Max normalization method. new , in the interval [0,1]:

[0036]

[0037] Where: x is the alarm signal number in the alarm signal set S; x min 、x max are the minimum and maximum values ​​of the alarm signal number respectively; x new Maps the alarm signal number to a value in the interval [0, 1].

[0038] In the step E, the network structure of the Transformer-GRU fault localization model includes an input layer, an encoding network, a GRU network, and an output layer connected in sequence;

[0039] The input layer includes the normalized feature matrix plus the position encoding matrix; the encoding network is the Encoders part of the Transformer network, which includes N L The GRU network consists of the first GRU layer, the second GRU layer, the first fully connected Dense layer, the Dropout layer, and the second fully connected Dense layer connected in sequence.

[0040] The output layer is a Softmax classification layer, and the data passes through the output layer to obtain a fault location vector.

[0041] The position encoding matrix is ​​a part of the Transformer network and is generated according to the formula based on the dimension of the feature matrix. The result of adding the position encoding matrix to the feature matrix is ​​the input of the entire fault location model.

[0042] Significant effect: The present invention provides a fault location method for the secondary system of a smart substation based on Transformer-GRU, which collects alarm signals from related equipment when faults occur in the secondary equipment of each interval of the smart substation, combines the Transformer network with the gated recurrent unit (GRU) network, and constructs a deep learning network model for fault location of each interval, which can be used to accurately locate faults in the secondary system of the substation and provide efficient and accurate auxiliary decision-making for operation and maintenance personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is the structure diagram of each sub-layer in the encoder Encoders;

[0044] Figure 2 This is the basic structure diagram of GRU;

[0045] Figure 3 This is the structure diagram of the Transformer-GRU fault location model;

[0046] Figure 4 This is the line interval Transformer-GRU fault location model training accuracy curve;

[0047] Figure 5 This is a curve of the training loss value of the line interval Transformer-GRU fault location model;

[0048] Figure 6 This is the structure diagram of the original Transformer model;

[0049] Figure 7 is a flow chart of the method of the present invention;

[0050] Figure 8 Schematic diagram of the center switch interval fault location process based on expert rules;

[0051] Fig. 9 The training accuracy curve of the fault location model for different line intervals;

[0052] Fig.10 The graph of the training loss value of the fault location model for different line intervals. DETAILED DESCRIPTION

[0053] like Figure 1-Figure 10 As shown, the present invention aims at the problems that the order of the alarm messages of the secondary equipment in each interval of the smart substation arrives at the station control layer is disordered, and the number of fault samples of different secondary equipment in the same interval is unbalanced, which may cause the misjudgment of the fault location of the deep learning network. For the lines, busbars, main transformer intervals, and central switch intervals of the smart substation, the alarm signals from the relevant equipment when the secondary equipment fails are collected, and the Transformer network and the gated recurrent unit (GRU) network are combined to construct the fault location model of each interval, and the neural network location model is trained and tested. The experimental results show that compared with the traditional deep learning fault location model, the fault location model based on Transformer-GRU can more accurately locate the fault of the secondary system of the smart substation.

[0054] The following is combined with Figure 1-Figure 10 The present invention is further described in detail with reference to the accompanying drawings and specific embodiments.

[0055] 1. Characteristic quantities of the secondary system of smart substation:

[0056] When a fault occurs in the secondary equipment of each compartment of a smart substation, a large number of alarm signals will be generated, and each fault scenario has corresponding fault characteristics. This paper refers to the fault cases that have occurred in actual smart substations, and collects and forms the fault characteristic information required for locating the fault of the secondary system of the smart substation, mainly including: (1) self-test information of secondary equipment; (2) SV / GOOSE receiving status of secondary equipment; (3) sending / receiving light intensity status of secondary equipment and switch ports; (4) judgment quantity of protection equipment from sampling value.

[0057] For each fault scenario of the secondary equipment in the line bay, bus bay, main transformer bay, and central switch bay of the smart substation, the alarm signals from the relevant equipment are collected to form an alarm signal set and the corresponding fault location. The fault location includes secondary equipment such as protection equipment, measurement and control equipment, merging units, and their sending and receiving network ports, optical fiber ports, switch ports, etc.

[0058] According to the above fault characteristic information, construct the alarm signal set S under each fault scenario:

[0059] S={S ZJ ,S SG ,S GQ ,S FZ} (5);

[0060] Where: S ZJ S is the self-test information of the secondary equipment; SG It is the receiving status of the secondary equipment SV / GOOSE; S GQ Send / receive light intensity status for secondary devices and switch ports; S FZ It is an auxiliary judgment quantity for protecting the sampling value of the equipment.

[0061] For a small number of fault scenarios at different intervals, obtain the current sampling values ​​of phase A, phase B, and phase C of the protection device, check whether the current amplitudes of phase A, phase B, and phase C are zero, if so, set the corresponding judgment value to 1, otherwise, set the corresponding judgment value to 0, and set these judgment values ​​S FZ As simulation signal auxiliary variables, they are added to the alarm signal set S of the corresponding fault scenario.

[0062] For all alarm signal sets in each interval, the total number of alarm signals and sampled value judgment quantities of protection equipment in the interval is first counted as M, and the total number of fault locations in the interval is counted as N. Then, each alarm signal and analog signal auxiliary variable is numbered from 1 to M, and each fault location is numbered from 1 to N, and each number corresponds to a fault location vector with a length of N. Taking number j as an example, in the fault location vector corresponding to the number, the jth last element is 1, and the rest of the elements are all 0.

[0063] In a fault scenario where a secondary device fails in each interval, the station control layer device successively receives alarm signals from multiple related devices at different times, and regards the alarm signal sent by the same device as the alarm signal at that time. The maximum number of fault moments of each fault scenario in the interval is recorded as T, and the maximum number of alarm signals at each fault moment of each fault scenario in the interval is recorded as A.

[0064] For each fault moment in each fault scenario in each interval, the numbers of the alarm signals and analog signal auxiliary variables are combined into a 1×A row vector. If the number of alarm signals at this moment is less than A, the missing elements in the row vector are filled with 0.

[0065] The row vectors of each fault moment in each fault scenario are combined to form a T×A feature matrix for the fault scenario. If the number of fault moments in the fault scenario is less than T, the missing row vectors in the feature matrix are filled with 0.

[0066] 2 Transformer and GRU network:

[0067] 2.1 Transformer Network

[0068] The Transformer network was first proposed by Ashish Vaswani et al. from the Google Brain team in 2017 and was initially used in the field of machine translation for natural language processing. The original model structure is shown in the attached Figure 6 After improvement, the neural network model can be applied to text classification tasks.

[0069] The Transformer network mainly consists of two parts: encoders and decoders. Encoders are composed of multiple identical encoder sublayers (ES), and decoders are composed of multiple identical decoder sublayers. Encoders and decoders contain the same number of sublayers.

[0070] The core component of Encoders is the Self Attention layer (SA), which pays attention to the similarity between the current element and elements at other positions in the input sequence, and learns the long-distance dependency features of long sentences. Each encoder sublayer contains two important parts, one is the Multi Head Self Attention layer (MHSA), and the other is the Feed Forward Neural Network (FFNN). The encoder sublayer also uses a residual network, which can effectively solve the gradient disappearance problem and has strong adaptability to deep models. By adding the data before a single submodule and the data after the submodule, the obtained data is layer normalized (LN). Unlike traditional batch normalization (Batch Normalization, BN), LN can effectively improve the training speed of neural networks. The structure of each sublayer in encoders is as follows Figure 1 shown.

[0071] Assume that the number of encoder sublayers contained in Encoders is N L , the output of the previous encoder sublayer is the input of the next encoder sublayer, which are connected in sequence to form Encoders.

[0072] Depend on Figure 1 It can be seen that Transformer does not contain any loops and convolutions, and uses parallel processing for the input matrix. The input matrix X can be regarded as n row vectors. After a single self-attention layer operation, an output matrix Y consisting of n row vectors is generated. This parallel processing method speeds up the data operation speed, but ignores the position relationship between the elements in the input matrix, so position encoding is required to obtain the position encoding matrix P.

[0073] The constituent elements of P are calculated according to the following formula:

[0074]

[0075]

[0076] Where: p is the position of the element in the input matrix; i is the dimension number; d model is the number of input features. When i is an even number, the encoding of the position is calculated by formula (2); when i is an odd number, the encoding of the position is calculated by formula (3). The dimension of P is n×d model .

[0077] For a single self-attention layer in Transformer, its input consists of dimensions d k The query vector q of dimension dk The key vector k of dimension d v These three vectors are randomly initialized by the Transformer network and can be trained and optimized.

[0078] For each row vector in the input matrix, there is a corresponding q, k, v. The q of this subsequence will be dot-multiplied with all other k. To keep the gradient stable, the dot product result needs to be divided by The weight vector w corresponding to each v is obtained through the softmax function, and w is multiplied by v to obtain the output vector corresponding to the input row vector.

[0079] In actual operations, the above processes are all operated in matrix form. For the input matrix X, the output matrix Y after the self-attention layer is:

[0080]

[0081] Where: Q is the query matrix; K is the key matrix; V is the value matrix.

[0082] The dimension of the input matrix X is n×d model , the dimension of the output matrix Y is n×d v .

[0083] The core of MHSA is to add a multi-head mechanism to the self-attention mechanism, which enables the input vector to be mapped to different subspaces to obtain more input features and make the model easier to train. In Transformer, each head is a self-attention layer. Assume that MHSA contains N H The matrix Z obtained by MHSA is:

[0084]

[0085] Where: W O is(N H ×d v )×d model The matrix is ​​randomly initialized by the Transformer network. The dimension of Z is n×d model The result of adding matrix X to matrix Z needs to be layer-normalized to get n×d model The matrix L 1 , L 1 Each row vector of will pass through a forward neural network.

[0086] The forward neural network consists of two fully connected Dense layers. The number of neurons in the first fully connected Dense layer is d FF, the activation function uses the ReLU function; the number of neurons in the second fully connected Dense layer is d model , no activation function is required.

[0087] The output of the feedforward neural network plus L 1 Layer standardization is required to obtain n×d model The matrix L 2 .L 2 is the output result of a single encoder sublayer. The output of this encoder sublayer will be used as the input of the next encoder sublayer. They are connected and operated in sequence to finally obtain an n×d model The matrix E, E is the output result of the entire Encoders.

[0088] 2.2 GRU Network

[0089] Recurrent neural networks are very suitable for text classification tasks, but they have problems with gradient vanishing and gradient exploding, and can only process information of limited length. The long short-term memory (LSTM) network introduces a gating mechanism based on the recurrent neural network (RNN), including a forget gate, input gate, and output gate, which can solve the problems of recurrent neural networks. However, LSTM training has more parameters and takes longer time.

[0090] In order to reduce training parameters, scholars simplified the LSTM structure based on the most important forget gate mechanism of LSTM and proposed the GRU network. This network not only has the same accuracy as LSTM, but also effectively speeds up the training speed.

[0091] GRU combines the internal state vector and output vector in LSTM into a unified state vector h t , the number of gates is also reduced to 2, namely the reset gate and the update gate. The basic structure of GRU is as follows Figure 2 shown.

[0092] Depend on Figure 2 It can be seen that the forward propagation formula of the GRU network is as follows:

[0093] r t =σ(W r [h t-1 ,x t ]) (10);

[0094] u t =σ(W u [h t-1 ,x t ]) (11);

[0095]

[0096]

[0097] Where: r t 、u t are the gate values ​​of the reset gate and the update gate respectively; W r , W r , W r are weight matrices respectively; σ is the sigmoid activation function; h t-1 、h t are the hidden unit states at the previous moment and the current moment respectively; x t is the input at the current moment; is the intermediate state of the hidden unit at the current moment; * is the multiplication operator; tanh is the hyperbolic tangent activation function.

[0098] 3. Secondary system fault location model based on Transformer and GRU:

[0099] 3.1 Feature Matrix Preprocessing and Training Strategy

[0100] The elements of the feature matrix of each interval are the alarm signal numbers. Normalizing the alarm signal numbers helps to increase the accuracy of neural network classification. Considering that the elements of the feature matrix in this paper do not need to follow the normal distribution law, the common Min-Max normalization method is used to map the alarm signal numbers to x new , in the interval [0,1]:

[0101]

[0102] Where: x is the number of the alarm signal in the alarm signal set S; min 、x max They are the minimum and maximum values ​​of the alarm signal number respectively. new Maps the alarm signal number to a value in the interval [0,1].

[0103] There are two training strategies when the feature matrix is ​​sent to the model for training. The first training strategy is to eliminate the serious imbalance in the number of alarm signal samples at different fault locations in the same interval. The samples with a small number of alarm signals in each interval are copied and their number is expanded, so that the neural network can enhance the training of small samples, reduce the misjudgment caused by insufficient training of small samples, and improve the accuracy of neural network classification.

[0104] The second training strategy is to consider that there is a network delay when each secondary device uploads the alarm signal to the station control layer device, and some alarm signals may lag, that is, the alarm signal that should arrive first arrives later, resulting in a disordered order. By disrupting the order of the row vectors at each moment in each fault scenario of each interval, multiple T×A feature matrices can be obtained, so that they all point to the corresponding fault location, fully considering the problem of disordered order of alarm signals and fully training them.

[0105] 3.2 Secondary System Fault Location Model Based on Transformer-GRU

[0106] In order to obtain more spatial location information and time information of the feature matrix and maintain the stability of the gradient during model training, corresponding Transformer-GRU fault location models are established for line intervals, bus intervals, main transformer intervals, and central switch intervals. The specific structure is as follows: Figure 3 shown.

[0107] When processing text sentence classification tasks, Transformer divides a complete sentence into individual words to obtain the semantic information of the sentence. Each word is represented by a multi-dimensional vector through the embedding layer. The vectors of multiple words form a matrix to represent a complete sentence. The position encoding is added, and the sum is used as the input of the Transformer network. This paper considers that there is no semantic information in the alarm signal number, removes the embedding layer, and directly adds the feature matrix to the position encoding matrix. The summed matrix is ​​used as the input of the Transformer network.

[0108] Considering that Decoders are generative models used for natural language generation, and the fault location of the secondary system of smart substations belongs to a text classification task, the Transformer network proposed in this paper only includes the Encoders part, without Decoders. Figure 3 It can be seen that the entire Transformer network consists of N L The GRU network consists of two sequentially connected GRU layers, a fully connected Dense layer, and a Dropout layer. This network structure can obtain more input features for classification training and suppress overfitting. The output layer is a Softmax classification layer, and the data passing through the output layer will obtain a fault location vector.

[0109] 3.3 Secondary system fault location process based on Transformer and GRU:

[0110] The secondary system fault location flowchart of each interval Transformer-GRU is shown in Figure 7The specific process is as follows: Step A: For each fault scenario of each secondary device in the secondary system of the line bay, transformer bay, bus bay, and central switch bay, collect the alarm signals from the relevant equipment to form an alarm signal set and the corresponding fault location. For the CT disconnection fault scenario of the line merging unit under the line bay, collect the current sampling value, obtain the judgment value, and add it as an auxiliary variable to the alarm signal set;

[0111] Step B: For each fault scenario in each interval, construct a corresponding feature matrix;

[0112] Step C: For a small sample fault scenario of the secondary equipment in each interval, replicate it; considering the network delay, for each fault scenario in each interval, scramble the order of the row vectors at each moment under the fault scenario to obtain multiple feature matrices;

[0113] Step D: normalize and preprocess the feature matrix corresponding to each fault scenario in each interval;

[0114] Step E: construct the Transformer-GRU fault localization model of each interval respectively, design their network structures, set training parameters and evaluation indicators; Step F: train the Transformer-GRU fault localization model of each interval respectively, and optimize the parameters of each model;

[0115] Step G: Send the test sets of each interval into their respective Transformer-GRU fault location models for testing, complete the fault location of the secondary system of the smart substation, and count the test accuracy of each model.

[0116] 4 Example analysis:

[0117] Tensorflow 2.5.0 was used to build the Transformer-GRU fault location model for line bays, bus bays, main transformer bays, and central switch bays. The CPU used for model training was AMD Ryzen 7 5800H, and the GPU was NVIDIA GeForce RTX 3060 Laptop GPU.

[0118] 4.1 Calculation example

[0119] Taking the Transformer-GRU fault location model of line interval as an example, each alarm signal and auxiliary variable is numbered from 1 to 81 as the elements of the feature matrix. The fault location includes secondary equipment, its sending network port and receiving network port, optical fiber, switch port, etc. The fault location is numbered from 1 to 11, and each number corresponds to a fault location vector with a length of 11.

[0120] The fault locations and their corresponding numbers are shown in Table 1.

[0121] Table 1 Fault location and number of line bays;

[0122]

[0123] The total fault samples are divided into the original training set and the test set, and the small samples in the original training set are appropriately copied so that the number of samples corresponding to each fault location in the training set is balanced. The copied training set is used as the real training set of the Transformer-GRU fault localization model for this interval.

[0124] The training set of line intervals contains 256 samples and the test set contains 74 samples.

[0125] There are 47 alarm signals and auxiliary variables for the bus bay, 6 fault locations, 186 samples in the training set, and 45 samples in the test set. There are 27 alarm signals and auxiliary variables for the main transformer bay, 8 fault locations, 142 samples in the training set, and 20 samples in the test set.

[0126] There are 13 alarm signals and auxiliary variables for the central switch interval, 9 fault locations, 158 samples in the training set, and 25 samples in the test set.

[0127] 4.2 Model hyperparameter optimization and positioning results

[0128] 4.2.1 Parameter settings in the Transformer-GRU model

[0129] In the parameter optimization process of the Transformer-GRU fault localization model at different intervals, it is necessary to set the basic parameters of the Transformer network and the GRU network, determine the number of neural network layers and the number of neurons in each layer, the activation function, the batch size, adjust the learning rate, consider the evaluation index, set the loss function category, and select the gradient descent algorithm. Taking the Transformer-GRU fault localization model of line interval as an example, the specific parameters of the model are shown in Table 2.

[0130] Table 2 Transformer-GRU fault localization model parameters for line intervals

[0131]

[0132]

[0133] The basic parameters of the other three types of intervals are shown in Tables 3 to 5.

[0134] Table 3 Transformer-GRU fault localization model parameters for bus bays;

[0135]

[0136] Table 4 Transformer-GRU fault localization model parameters for the main transformer bay;

[0137]

[0138]

[0139] Table 5 Transformer-GRU fault localization model parameters for central switch interval;

[0140]

[0141]

[0142] 4.2.2 Model training and testing results:

[0143] For the Transformer-GRU fault localization model of each interval, the training set is sent to the corresponding model for training. After the parameter optimization is completed, the test set is sent to the model under the corresponding interval for testing to obtain the training results and test results of each interval.

[0144] Taking the line interval as an example, the total number of training times for the Transformer-GRU fault location model is 700. The curves of training accuracy and loss value changing with the number of iterations are shown in the figure below: Figure 4 , Figure 5 shown.

[0145] Figure 4 This is the line interval Transformer-GRU fault location model training accuracy curve;

[0146] Figure 5 This is a curve of the training loss value of the line interval Transformer-GRU fault location model;

[0147] Depend on Figure 4 , Figure 5 As shown in the figure, the training loss value of the Transformer-GRU fault localization model of the line interval decreases rapidly with the number of iterations and gradually converges, and the training accuracy increases rapidly with the number of iterations and gradually converges. After 650 iterations, the training loss value drops to 0.1591, close to 0 with a small decrease, and the training accuracy reaches 95.31% and basically no longer increases, indicating that the model has been fully trained.

[0148] The test data of each interval is sent to the respective Transformer-GRU fault localization model for testing. The test accuracy of the model at each interval is shown in Table 6.

[0149] Table 6 Test accuracy of Transformer-GRU fault localization model at different intervals

[0150]

[0151]

[0152] As can be seen from Table 6, the test accuracy of the Transformer-GRU fault localization model in each interval is relatively high, which shows that each model has a strong generalization ability and can solve the problems of the disordered order of the alarm message arriving at the station control layer equipment and the misjudgment caused by the imbalanced number of fault samples of different secondary equipment in the same interval.

[0153] When a central switch fails, the process of finding the fault location based on traditional expert rules is rather cumbersome. The specific process can be seen in Figure 8 The Transformer-GRU fault location model with central switch interval proposed in this paper can efficiently and accurately find the fault location.

[0154] 4.3 Comparison of results of different models

[0155] In order to verify the advantages of the Transformer-GRU fault localization model of each interval proposed in this paper, the Transformer-GRU fault localization model of each interval is compared with the fault localization models of four deep learning networks: RNN, LSTM, GRU, and Transformer.

[0156] Taking the line interval as an example, for different models, the main parameters such as learning rate, number of hidden layers, gradient descent algorithm, batch size and loss function category are kept consistent, and the model parameters are optimized. The training results are obtained after 700 iterations. The curves of the training accuracy and test loss value of the five fault location neural network models with the number of iterations are shown in Fig. 9 , Fig.10 .

[0157] Depend on Fig. 9 , Fig.10 It can be seen that after 650 iterations, all models have converged. The training accuracy and test accuracy of the fault location models with different line intervals are shown in Table 7.

[0158] Table 7 Training accuracy and test accuracy of fault location models with different line intervals

[0159]

[0160]

[0161] From Table 7 and Fig. 9 , Fig.10 It can be seen that compared with RNN and LSTM, the GRU-based fault location model converges faster and has higher training accuracy, which verifies that the GRU network can effectively reduce training time and maintain high accuracy. The training accuracy of the Transformer-based fault location model and the GRU-based fault location model is not much different, and both can quickly converge to a higher accuracy. The Transformer-GRU-based fault location model proposed in this paper has the highest training accuracy, which shows that the combination of Transformer and GRU can obtain more feature information of the feature matrix, making the feature matrix easier to be classified and trained. The Transformer-GRU fault location model in this paper has the highest test accuracy, which shows that the model has stronger generalization ability and can better solve the problems of the disordered order of the alarm message arriving at the station control layer equipment and the misjudgment caused by the imbalanced number of fault samples of different secondary equipment in the same interval.

[0162] For the line bay, bus bay, main transformer bay, and central switch bay of the smart substation, this paper collects the alarm signals of secondary equipment failures in each bay of the smart substation, and establishes a fault location model combining Transformer and GRU in each bay. The following conclusions can be drawn:

[0163] 1) For different intervals, the present invention establishes a Transformer-GRU fault location model, which has a wide range of fault location and can complete the fault location of major equipment in the secondary system of the smart substation.

[0164] 2) The Transformer-GRU fault localization model of each interval has a high positioning accuracy, which can solve the problems of disordered order of alarm messages arriving at the station control layer equipment and misjudgment caused by the imbalanced number of fault samples of different secondary equipment in the same interval.

[0165] 3) Compared with the fault localization models based on Transformer, GRU, LSTM, and RNN for each interval, the fault localization model based on Transformer-GRU has the highest positioning accuracy and better model generalization ability.

[0166] 4) For different intervals, this paper only requires the current sampling values ​​of phase A, phase B, and phase C of the protection equipment for a few fault scenarios. The remaining fault scenarios can be accurately located without sampling values, reducing the amount of information collected.

[0167] Figure 6 This is the structure diagram of the original Transformer model;

[0168] Figure 7 This is a flow chart of the method of the present invention, that is, a fault locating process of the secondary system of a smart substation.

[0169] Figure 8 The figure is a schematic diagram of the fault location process of the central switch interval based on expert rules. In this interval, assuming that the "220kV fault recording device A" on the central switch cannot receive the SV message of the "220kV line 2 merging unit A", there are 12 possible fault locations:

[0170] ①Inside the central switch;

[0171] ②Central switch optical port 3M;

[0172] ③Fiber optic loop between the central switch and the line 2 interval switch;

[0173] ④Inside the switch of line 2 interval;

[0174] ⑤ Line 2 interval switch optical port 1M;

[0175] ⑥ Line 2 interval switch optical port 6M (assuming 6M access to the merging unit SV networking port);

[0176] ⑦220kV line 2 merging unit output plug-in;

[0177] ⑧220kV line 2 merging unit SV output plug-in SV networking optical port;

[0178] ⑨ The optical fiber loop between the line 2 interval switch optical port 6M and the 220kV line 2 merging unit SV output plug-in SV networking optical port;

[0179] ⑩Fault recording SV acquisition plug-in;

[0180] Central switch optical port 9M;

[0181] Optical fiber loop between the fault recorder and the central switch optical port 9M;

[0182] Because the probability of multiple faults occurring at the same time is very low, under the assumption that multiple faults are not considered, the fault location rules are as follows:

[0183] 1) If both the line protection and the busbar protection cannot receive the direct sampling SV of the branch, it is judged as ⑦ fault; otherwise, ⑦ is not faulty.

[0184] 2) If the line protection and busbar protection can both receive the SV directly collected from the branch normally. At this time, the 220kV line 2 measurement and control device can also receive the SV networking message of "220kV line 2 merging unit A" normally, then ④⑥⑦⑧⑨ are normal; on the contrary, if the light intensity status monitoring of the optical port at ⑧ is abnormal, then ⑧ is faulty; if the light intensity status monitoring of the optical port at ⑧ is normal, and the light intensity sent at ⑥ is normal and the light intensity received is abnormal, then ⑨ is faulty; if the light intensity sent and the light intensity received at ⑥ are both abnormal, then ⑥ is faulty.

[0185] 3) If ④, ⑥, ⑦, ⑧ and ⑨ are all normal, the switch port is usually full-duplex communication. When the sending and receiving light intensities at ⑤ are normal, and the sending light intensity at ② is normal and the receiving light intensity is abnormal (usually, when the optical module is damaged, both the sending and receiving fail at the same time. If either the sending or receiving is normal, it can be judged that the optical module is normal), then ③ is faulty; if the sending light intensity at ⑤ is abnormal, then ⑤ is faulty; if the sending and receiving light intensities at ② are both abnormal, then ② is faulty.

[0186] 4) If the "network message recording device 1" on the central switch receives the SV message of "220kV line 2 merging unit A" normally, the faults ①②③④⑤⑥⑦⑧⑨ can be eliminated at one time. If the light intensity is abnormal, Fault; if If the light intensity is normal, then ⑩ or Fault.

[0187] 5) When the switch "device failure" hard contact signal is issued, the corresponding ① or ④ fault.

[0188] Fig. 9 and Fig.10 is the training result of the fault location model with different line intervals; Fig. 9 The training accuracy curve of the fault location model for different line intervals; Fig.10 This is a curve diagram of the training loss value of the fault location model for different line intervals. Fig. 9 After 700 iterations, the top curve is the training accuracy curve of the fault location model of the present invention; Fig.10 After 700 iterations, the bottom curve is the fault location model training loss value curve of the present invention.

[0189] Finally, it should be noted that the above examples are only specific implementation examples of the present invention. Of course, those skilled in the art can make changes and modifications to the present invention. If these modifications and variations fall within the scope of the claims of the present invention and their equivalent technologies, they should be considered to be within the scope of protection of the present invention.

Claims

1. A fault location method for the secondary system of a smart substation based on Transformer-GRU, characterized in that: The steps include: Step A: For each fault scenario of each secondary device in the secondary system of the line bay, transformer bay, bus bay, and central switch bay, collect alarm signals from related devices to form an alarm signal set and corresponding fault location; for the CT disconnection fault scenario of the line merging unit under the line bay, collect current sampling values ​​to obtain a judgment value, which is used as an auxiliary variable and added to the alarm signal set; Step B: For each fault scenario in each interval, construct a corresponding feature matrix; Step C: Replicate a small sample failure scenario of the secondary equipment in each interval; Considering the network delay, for each fault scenario in each interval, the order of the row vectors at each moment in the fault scenario is disrupted to obtain multiple feature matrices; Step D: normalize and preprocess the feature matrix corresponding to each fault scenario in each interval; Step E: Construct the Transformer-GRU fault localization models for each interval, design their network structures, and set training parameters and evaluation indicators; Step F: Train the Transformer-GRU fault localization model of each interval separately and optimize the parameters of each model; Step G: Send the test sets of each interval into their respective Transformer-GRU fault location models for testing, complete the fault location of the secondary system of the smart substation, and calculate the test accuracy of each model; In the step A, for the CT disconnection fault scenario of the line merging unit under the line interval, the line protection phase A, phase B, and phase C current sampling values ​​are obtained, and it is checked whether the current amplitudes of phase A, phase B, and phase C are zero. If so, the corresponding judgment value is set to 1, otherwise, the corresponding judgment value is set to 0, and these judgment values ​​are used as analog signal auxiliary variables, and they are added to the alarm signal set of the corresponding fault scenario; In step B, for each fault scenario of each interval, a corresponding feature matrix is ​​constructed, including: For all alarm signal sets of each interval, first count the total number of alarm signals and sampled value judgment quantities of protection equipment in the interval as M, and count the total number of fault locations in the interval as N; then number each alarm signal and analog signal auxiliary variable from 1 to M, and number each fault location from 1 to N, and each number corresponds to a fault location vector with a length of N; In a fault scenario where a secondary device fails in each interval, the station control layer device successively receives alarm signals sent by multiple related devices at different times, and regards the alarm signal sent by the same device as the alarm signal at that time; the maximum number of fault times of each fault scenario in the interval is recorded as T, and the maximum number of alarm signals at each fault time of each fault scenario in the interval is recorded as A; For each fault moment in each fault scenario in each interval, the numbers of the alarm signals and analog signal auxiliary variables are combined into a 1×A row vector; if the number of alarm signals at this moment is less than A, the missing elements in the row vector are filled with 0; The row vectors of each fault moment in each fault scenario are combined to form a T×A feature matrix for the fault scenario; if the number of fault moments in the fault scenario is less than T, the elements in the insufficient row vectors in the feature matrix are filled with 0.

2. According to claim 1, the Transformer-GRU-based smart substation secondary system fault location method is characterized by: In step A, the fault scenario includes: self-test information of the secondary device; SV / GOOSE receiving status of the secondary device; the sending / receiving light intensity status of the network port of the secondary device and the switch; the judgment value of the protection device from the sampling value; The fault location includes the protection equipment, measurement and control equipment, merging unit and their sending and receiving network ports, optical fiber ports, and switch ports in the secondary system; According to the above fault scenarios, construct the alarm signal set S under each fault scenario: S={S ZJ ,S SG ,S GQ ,S FZ } (1); Where: S ZJ S is the self-test information of the secondary equipment; SG It is the receiving status of the secondary equipment SV / GOOSE; S GQ S is the sending / receiving light intensity status of the secondary device and switch port; FZ It is an auxiliary judgment quantity for protecting the sampling value of the equipment.

3. The fault location method for the secondary system of a smart substation based on Transformer-GRU according to claim 1 is characterized in that: In the step D, normalizing the feature matrix for preprocessing includes: The elements of the characteristic matrix of each interval are the alarm signal numbers. The alarm signal numbers are normalized and mapped to x using the Min-Max normalization method. new , in the interval [0,1]: Where: x is the alarm signal number in the alarm signal set S; x min 、x max are the minimum and maximum values ​​of the alarm signal number respectively; x new Maps the alarm signal number to a value in the interval [0, 1].

4. The fault location method for the secondary system of a smart substation based on Transformer-GRU according to claim 1 is characterized in that: In the step E, the network structure of the Transformer-GRU fault localization model includes an input layer, an encoding network, a GRU network, and an output layer connected in sequence; The input layer includes the normalized feature matrix plus the position encoding matrix; the encoding network is the Encoders part of the Transformer network, which includes N L The GRU network consists of the first GRU layer, the second GRU layer, the first fully connected Dense layer, the Dropout layer, and the second fully connected Dense layer connected in sequence. The output layer is a Softmax classification layer, and the data passes through the output layer to obtain a fault location vector.

5. The fault location method for the secondary system of a smart substation based on Transformer-GRU according to claim 1 is characterized in that: In the step F, during the parameter optimization process of the Transformer-GRU fault localization model at each interval, the basic parameters of the Transformer-GRU fault localization model are set, the number of neural network layers and the number of neurons contained in each layer, the activation function, the batch size are determined, the learning rate is adjusted, the evaluation index is considered, the loss function category is set, and the gradient descent algorithm is selected.

6. The fault location method for the secondary system of a smart substation based on Transformer-GRU according to claim 1 is characterized in that: In the step F, the total fault samples are divided into an original training set and a test set, and the small samples in the original training set are appropriately copied so that the number of samples corresponding to each fault position in the training set is balanced, and the copied training set is used as the real training set of the Transformer-GRU fault localization model of the interval; the Transformer-GRU fault localization model of each interval is trained, and the parameters of each model are optimized.

Images