A GIS device-based detection method for the state of a knife switch and the synchronization of three-phase knife switch actions

Abstract

The application discloses a GIS device-based detection method for the synchronization of the state of a knife switch and the action of a three-phase knife switch, comprising the following steps: 1, preprocessing current data and dividing data sets; 2, designing a CNN network based on a double-layer LSTM neural network with an attention mechanism and a dynamic time warping algorithm; 3, constructing the CNN network based on the double-layer LSTM neural network with the attention mechanism and the dynamic time warping algorithm; and 4, obtaining the output result of a test set sample based on the CNN network based on the double-layer LSTM neural network with the attention mechanism and the dynamic time warping algorithm. The application can provide multi-scale and multi-resolution analysis of time series, dynamically capture the characteristics of data in the time domain and the frequency domain, improve the detection accuracy of the synchronization of the state of a GIS device knife switch and the action of a three-phase knife switch, and thus meet the actual requirements of accuracy and rapidity.

Description

Technical Field

[0001] This invention pertains to the power industry, specifically a method for detecting the synchronization of disconnector status and three-phase disconnector operation in GIS equipment. The invention relates to LSTM technology, CNN technology, attention mechanism technology, and DTW algorithm, specifically a neural network based on a parallel structure of a two-layer LSTM neural network with attention mechanism and a CNN network with dynamic time warping algorithm. This method is suitable for detecting the opening and closing status of GIS equipment and the synchronization of disconnector operation. Background Technology

[0002] SF6 gas-insulated switchgear (GIS) is a reliable power transmission and transformation equipment that encloses primary equipment (excluding transformers), such as circuit breakers, disconnectors, voltage transformers, current transformers, busbars, surge arresters, cable terminal boxes, and grounding switches, in several compartments filled with SF6 gas. It is organically combined through optimized design. Compared to conventional open-type switchgear, GIS features a smaller footprint, higher reliability, stronger safety, less maintenance, and shorter construction period.

[0003] Circuit breakers, as one of the core components of GIS (Gas Integrator System), are the most critical monitoring components for condition monitoring of GIS. Domestically, according to statistics from the Electric Power Research Institute of the Ministry of Energy, from 1989 to 1997, a total of 4,632 high-voltage circuit breaker failures occurred. Mechanical failures accounted for 39.3% of the total failures; insulation failures accounted for 839 failures (18.1%); opening and closing failures accounted for 212 failures (4.6%); and other types of failures accounted for 38%. This shows that mechanical failures are the main form of high-voltage circuit breaker failures, and operating mechanism failures are the main cause of mechanical failures. Therefore, condition monitoring of the operating mechanisms of GIS circuit breakers is of great significance for ensuring the stable operation of the power grid, reducing national economic losses, and maintaining social stability.

[0004] The main faults of high-voltage circuit breakers originate from electrical faults in the operating mechanism. Therefore, the operating mechanism is a key target for condition monitoring of high-voltage circuit breakers. Condition monitoring of the operating mechanism mainly includes monitoring the current of the opening (closing) coil and measuring the synchronicity of opening (closing) time. Existing technologies typically analyze the waveform characteristics of the opening and closing grounding currents to determine the status of the disconnector and detect the synchronicity of its operation. These methods mainly include the following:

[0005] The comparison-based technique determines the state and synchronization of a disconnector by comparing indicators such as phase difference and frequency difference between two voltage or current waveforms. This method is simple and reliable, but it is not suitable for situations where harmonics exist in the current waveform.

[0006] Time-series analysis-based techniques analyze the timing information of current waveforms, such as extracting specific patterns, crossover points, and zero-crossing points, to accurately determine the state and synchronization of the disconnector. This method offers high accuracy but requires the acquisition and processing of a large amount of waveform data, resulting in high costs.

[0007] Frequency domain analysis-based techniques: By performing a Fourier transform on the current waveform to convert it to the frequency domain, and then analyzing the frequency distribution information, such as the content, frequency, and phase difference of each harmonic, the status and synchronization of the disconnector can be determined. This method is effective for judging current waveforms containing a large number of harmonics.

[0008] Model-based prediction techniques predict the state and operation of disconnectors by establishing a mathematical model of the power grid and comparing the switching and grounding current signals with the model. This method offers high accuracy but requires precise measurement and value acquisition of the power grid's physical parameters and model parameters, resulting in high modeling costs.

[0009] Existing technologies have some problems, shortcomings, and deficiencies. For example, time-series analysis-based and model-based prediction technologies require a large amount of data acquisition, processing, and model building, resulting in high costs; comparison-based technologies are greatly affected by harmonic interference in waveforms; and frequency domain analysis-based technologies require accurate measurement and analysis of various harmonic characteristics, making the process complex. Summary of the Invention

[0010] The present invention addresses the shortcomings of the existing technology by proposing a method for detecting the synchronization of disconnector status and three-phase disconnector operation of GIS equipment. This method aims to simultaneously detect the synchronization of disconnector status and three-phase disconnector operation of GIS equipment through dual channels, thereby obtaining complete feature information on the synchronicity of disconnector status and operation of GIS equipment, and achieving high-precision detection of the synchronization of disconnector status and three-phase disconnector operation of GIS equipment.

[0011] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0012] The present invention provides a method for detecting the synchronization of disconnector status and three-phase disconnector operation based on GIS equipment, characterized by the following steps:

[0013] Step 1: Construct the training set T tr ;

[0014] Step 1.1: Collect the current data of the three-phase M circuits of the GIS equipment and construct a three-phase current dataset X = {x1, ..., x}. k ,…,x K}, x k Let x represent the k-th sampled three-phase current data, and x k={x k,1 ,...,x k,m ,...,x k,M}; where x k,m This represents the m-th three-phase current data sampled from the k-th sample, and x k,m ={x k,1,1 ,...,x k,m,t ,…,x k,M,T}, x k,m,t This represents the m-th three-phase current data sampled at time t in the k-th data; 1≤m≤M, 1≤k≤K, 1≤t≤T, K represents the total number of sampled data; M represents the total number of data channels collected; T represents the total sampling time.

[0015] Step 1.2: Construct a label information set Y = {y1, ..., y2} for the three-phase current dataset X. k ,…,y K}, where y k x represents the k-th sampled three-phase current data. k The label value, and y k ∈[1,N], where N is the number of state types;

[0016] Step 1.3: Use the labeled three-phase current dataset P = (X, Y) as the training set T tr ;

[0017] Step 2: Construct a state synchronization detection network consisting of a state detection branch and a synchronization detection branch. The state detection branch includes a two-layer LSTM neural network layer, a one-layer Attention mechanism layer, and a one-layer Softmax layer. The synchronization detection branch includes a one-layer CNN layer and a one-layer dynamic time warping algorithm layer.

[0018] Step 2.1: Construct a two-layer LSTM neural network, including a shallow LSTM neural network and a deep LSTM neural network. The shallow LSTM neural network includes a shallow forget gate, a shallow input gate, a shallow update unit, and a shallow output gate. The deep LSTM neural network includes a deep forget gate, a deep input gate, a deep update unit, and a deep output gate.

[0019] Step 2.2: Construct the CNN layers, including: an input layer, two convolutional layers, two pooling layers, and an output layer;

[0020] Step 3: Transfer the three-phase current data x of the m-th channel. k,m,t The input synchronization detection network's state detection branch processes the data.

[0021] Step 3.1: The shallow LSTM neural network in the two-layer LSTM neural network processes the m-th three-phase current data x. k,m,tAfter processing, we obtain x. k,m,t Shallow state information hidden state at time step t

[0022] Step 3.2: In the two-layer LSTM neural network, the deep LSTM neural network hides the state information of the shallow layer. Processing is performed to obtain Hidden state of deep state information at time step t

[0023] Step 3.3: The Attention mechanism layer uses equations (11)-(16) to hide the state of the deep state information. After processing, the multi-resolution time-frequency features S are obtained. k,m ;

[0024]

[0025]

[0026]

[0027] e k,m =Q k,m Ke k,m T (14)

[0028]

[0029]

[0030] In equations (11)-(16), Q k,m V k,m Ke k,m They are respectively The query value, truth value, and key value, w Qh It is the linear transformation matrix of the query value, w Vl It is the truth-valued linear transformation matrix, w Kl It is the key value linear transformation matrix, e k,m for Attention score, Ke k,m T for Ke k,m The transpose of α k,m for Attention weights;

[0031] Step 3.4, the multi-resolution time-frequency feature S k,m The data is input into the Softmax layer, and the three-phase current data x is obtained using equation (17). k,m,t The probability values ​​P(x) corresponding to all statesk,m,t |N);

[0032] P(x k,m,t |N)=softmax(W p,k,m ·S k,m +b p,k,m (17)

[0033] In equation (17), W P,k,m and b P,k,m S represents the multi-resolution time-frequency features respectively. k,m The probability weight matrix and probability bias vector; softmax represents the activation function;

[0034] Step 3.5: Construct the cross-entropy loss function L(y) for the state detection branch. k,m ,P k,m,n ), where P k,m,n x represents the grounding current data of the k-th line and the m-th path. k,m The label value y k The probability of predicting the Nth state;

[0035] Step 4: Processing of synchronization detection branches in the state synchronization detection network:

[0036] Step 4.1: The CNN layer processes the m-th three-phase current data x. k,m,t The m-th feature vector D is obtained through processing. k,m,t ;

[0037] Step 4.2: The dynamic time warping algorithm layer processes the m-th feature vector D. k,m,t The time difference of the three-phase disconnector operation synchronization is obtained through processing.

[0038] Step 4.3: Construct the cross-entropy loss function L(x) for the synchronization detection branch. k,m ,Δ k,m ), where Δ k,m x represents the grounding current data of the k-th line and the m-th path. k,m The corresponding time difference of the synchronization of the disconnector action;

[0039] Step 5: Training and processing of the state synchronization detection network:

[0040] Step 5.1: Construct the loss function Loss = ɑL(y) for the state synchronization detection network. k,m ,P k,m,n )+βL(x k,m ,Δ k,m ); where ɑ represents L(y k,m ,P k,m,n The weights of L(x), where β represents L(x). k,m ,Δ k,mThe weights of ) are given, and α + β = 1;

[0041] Step 5.2: Based on the training set T tr The network is trained using backpropagation and gradient descent to detect the state synchronization of GIS disconnectors. The loss function Loss is calculated to update the network parameters. Training is stopped when the training epoch reaches the maximum or the loss function Loss reaches the minimum. This results in a trained GIS disconnector operation state detection network, which is used to map the input three-phase M-path grounding data set to the corresponding state category label. Finally, the disconnector state of the GIS equipment and the time difference of the three-phase disconnector action synchronization are output.

[0042] The method for detecting the synchronization of disconnector status and three-phase disconnector operation based on GIS equipment, as described in this invention, is also characterized in that step 3.1 includes:

[0043] Step 3.1.1: The shallow forget gate of the shallow LSTM neural network applies the m-th three-phase current data x. k,m,t Selective discarding is performed to obtain x. k,m,t Shallow fault selection information

[0044] Step 3.1.2: The shallow input gate will select shallow fault information. The memory information at time step t-1 output by the shallow update unit After multiplying, we get x k,m,t Shallow fault retention information at time step t When t=1, let

[0045] Step 3.1.3: The shallow input gate obtains x using equations (1) and (2) respectively. k,m,t Shallow input fault information at time step t and shallow fault modulation information

[0046]

[0047]

[0048] In equations (1) and (2), W pl,k,m and U pl,k,m They represent x respectively k,m,t and The shallow input weight matrix, b pl,k,m x represents k,m,t The shallow input bias vector; x represents k,m,tThe hidden state of the shallow state information at time step t-1; when t=1, let W gl,k,m and U g1,k,m They represent x respectively k,m,t and The shallow modulation weight matrix, b gl,k,m x represents k,m,t The shallow modulation bias vector; tanh is the activation function;

[0049] Step 3.1.4, the shallow update unit will and After multiplying, we get Shallow state information pending update Therefore, x can be obtained using equation (3). k,m,t Memory information at the t-th time step

[0050]

[0051] Step 3.1.5: The shallow output gate obtains x using equation (4). k,m,t Shallow state information hidden state at time step t Thus, the three-phase current data x are obtained. k,m,t The hidden state of the shallow state information at all time steps in a shallow LSTM neural network.

[0052]

[0053] In equation (4), ⊙ represents element-wise multiplication. Represents three-phase current data x k,m The shallow composite signal at time step t in the shallow memory unit And obtained from equation (5);

[0054]

[0055] In equation (5), W ol,k,m and U ol,k,m They represent x respectively k,m,t and The shallow output weight matrix, b ol,k,,m This represents the shallow layer output bias vector.

[0056] Step 3.2 includes:

[0057] Step 3.2.1: The deep forget gate of the deep LSTM neural network hides the shallow state information at the t-th time step. After selective discarding, the following is obtained Deep fault selection information at time step t

[0058] Step 3.2.2: The deep input gate will select deep fault information. With the output of the deep update unit Memory information at time step t-1 After multiplying, we get Deep fault retention information at time step t When t=1, let

[0059] Step 3.2.3: The deep channel input gate is obtained using equations (6)-(7). Deep input fault information at time step t and deep fault modulation information

[0060]

[0061]

[0062] In equations (6)-(7), W pg,k,m and U pg,k,m They represent and The deep input weight matrix, b pg,k,m express The deep input bias vector; express The hidden state of the deep state information at the (t-1)th time step; when t=1, let W gg,k,m and U gg,k,m They represent and The deep modulation weight matrix, b gg,k,m express The deep input bias vector;

[0063] Step 3.2.4: The deep update unit will input deep fault information. and deep fault modulation information After multiplying, we get Deep state information awaiting update Thus, by using equation (8), we can obtain Memory information at the t-th time step

[0064]

[0065] Step 3.2.5: The deep output gate is obtained using equation (9). Hidden state of deep state information at time step t Thus obtain Hidden states of deep state information at all time steps in a deep LSTM neural network

[0066]

[0067] In equation (9), Indicates hidden state information Deep composite signal at time step t in deep memory unit And obtained from equation (10);

[0068]

[0069] In equation (10), W og,k,m and U og,k,m They represent and The deep output weight matrix, b og,k,m express The deep output bias vector.

[0070] Step 4.1 includes:

[0071] Step 4.1.1, the m-th three-phase current data x k,m,t The input layer of the CNN is passed to the first convolutional layer for convolution operation, thus obtaining the feature vector output by the first convolutional layer.

[0072] Step 4.1.2: The first pooling layer processes the feature vector. After performing max pooling, the feature vector output by the first pooling layer is obtained.

[0073] Step 4.1.3: The second convolutional layer processes the feature vector. Perform a second convolution operation to obtain the feature vector output by the second convolutional layer.

[0074] Step 4.1.4: The second pooling layer processes the feature vector. After max pooling, the feature vector D output by the second pooling layer is obtained. k,m,t ={D k,m,a,t D k,m,b,t D k,m,c,t}, where D k,m,a,t D k,m,b,t D k,m,c,t These represent the time series of phases a, b, and c at time steps of the kth path and mth path, respectively.

[0075] Step 4.2 includes:

[0076] Step 4.1: The dynamic time warping algorithm layer uses equations (18)-(20) to obtain the a-phase time series D. k,m,a,t and b phase time series D k,m,b,t Distance between the t-th time steps Phase a time series D k,m,a,t and c-phase time series D k,m,c,t Distance between the t-th time steps b-phase time series D k,m,b,t and c-phase time series D k,m,c,t Distance between the t-th time steps Thus, we obtain D k,m,a,t and D k,m,b,t Distance between all time steps D k,m,a,t and D k,m,c,t Distance between all time steps D k,m,b,t and D k,m,c,t Distance between all time steps

[0077]

[0078]

[0079]

[0080] In equations (18)-(20), D k,m,a,t-1 D k,m,b,t-1 D k,m,c,t-1 Let A, B, and C represent the time series of phases a, b, and c at time steps t-1 of the k-th path m-th path, respectively. Let D(,) represent the normalized path distance between the two sequences, and min{} represent the minimum value.

[0081] Step 2.4.7: Using equations (21)-(23), obtain the time difference Δ of the synchronization of the a-phase disconnector action of the k-th and m-th sampled data respectively. k,m,a The time difference Δ of the synchronization of phase b disconnect switch operation k,m,b The time difference Δ of the c-phase disconnector's action synchronization k,m,c ;

[0082]

[0083]

[0084]

[0085] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing any of the synchronization detection methods, and the processor is configured to execute the program stored in the memory.

[0086] The present invention provides a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs any step of the synchronization detection method.

[0087] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0088] 1. This invention proposes a method for detecting the status of disconnectors and the synchronization of three-phase disconnector actions in GIS equipment, based on a parallel two-layer LSTM neural network with an attention mechanism and a CNN network with a dynamic time warping algorithm. Grounding current data is first pre-processed through a single-layer LSTM neural network. The processed output is then input into both the CNN and LSTM networks for further processing. In the CNN network, the dynamic time warping algorithm is used to analyze the time-frequency characteristics of the three-phase current data, effectively extracting the differences in the synchronization of three-phase disconnector actions and the disconnector status characteristics. Simultaneously, to improve the model's robustness, a two-layer LSTM neural network with an attention mechanism is employed to enhance the model's memory capacity. In the LSTM network, a recurrent neural network structure is used for time-series data modeling, segmenting the grounding current data to improve the model's accuracy and stability. By using the CNN and LSTM networks in parallel and performing time-frequency analysis at different scales, this method effectively improves the status detection performance of GIS equipment while reducing false alarm and false negative rates, demonstrating good practicality and feasibility.

[0089] 2. The proposed method for detecting the status of disconnectors and the synchronization of three-phase disconnector actions in GIS equipment, based on a parallel two-layer LSTM neural network with attention mechanism and a CNN network with dynamic time warping algorithm, effectively analyzes the time-frequency characteristics of grounding current data and extracts relevant features of disconnector operation, thereby enabling the detection of GIS equipment status. In the method designed in this invention, an LSTM network is used to preprocess the grounding current data and obtain corresponding output results. These output results are then input into the CNN network and LSTM network for subsequent processing. The CNN network uses a dynamic time warping algorithm to perform time-frequency analysis between two-to-three phases, further extracting the differences in the synchronization of three-phase disconnector actions and disconnector status features. This allows for the determination of whether the equipment operation is abnormal, enabling high-precision detection of the status of disconnectors and the synchronization of three-phase disconnector actions in GIS equipment, demonstrating excellent practicality and feasibility. Attached Figure Description

[0090] Figure 1 This is a neural network structure diagram of the synchronization detection method for the disconnector status and three-phase disconnector operation of GIS equipment proposed in this invention. Detailed Implementation

[0091] In this embodiment, a synchronization detection method for the disconnector status and three-phase disconnector operation of GIS equipment utilizes a deep learning network framework and comprehensively considers the characteristics of the grounding current of the disconnector in the GIS equipment. Taking the three-phase disconnector grounding current data as input, a two-layer LSTM neural network is used to extract the disconnector grounding current features to obtain hidden state values. Then, an Attention layer is used to obtain weighted average time-frequency features, and finally, a Softmax layer is used to determine the state category. Simultaneously, the disconnector grounding current features extracted by the first-layer LSTM neural network obtain multi-scale, multi-resolution hidden state values, which are then subjected to two convolution and pooling operations through a CNN neural network. Finally, a dynamic time warping algorithm is used to obtain the synchronization time difference of the three-phase disconnector operation. Figure 1 As shown. Specifically, it is done according to the following steps:

[0092] Step 1: Construct the training set T tr ;

[0093] Step 1.1: Collect the three-phase M-circuit current data of the GIS equipment. This involves sampling the three-phase disconnector grounding current data, normalizing it, and constructing a three-phase current dataset X = {x1, ..., x}. k ,…,x K}, x k Let x represent the k-th sampled three-phase current data, and x k ={x k,1 ,...,x k,m ,...,x k,M}; where x k,m This represents the m-th three-phase current data sampled from the k-th sample, and x k,m ={x k,1,1 ,...,x k,m,t ,…,x k,M,T}, x k,m,t This represents the m-th three-phase current data sampled at time t in the k-th data; 1≤m≤M, 1≤k≤K, 1≤t≤T, K represents the total number of sampled data; M represents the total number of data channels collected; T represents the total sampling time.

[0094] Step 1.2: Construct a label information set Y = {y1, ..., y2} for the three-phase current dataset X. k ,…,y K}, where y k x represents the k-th sampled three-phase current data.k The label value, and y k ∈[1,N], where N is the number of state types;

[0095] Step 1.3: Use the labeled three-phase current dataset P = (X, Y) as the training set T tr ;

[0096] Step 2: Construct a state synchronization detection network consisting of state detection branches and synchronization detection branches, such as... Figure 1 As shown, the state detection branch includes: a two-layer LSTM neural network layer, a one-layer Attention mechanism layer, and a one-layer Softmax layer; the synchronization detection branch includes: a one-layer CNN layer and a one-layer dynamic time warping algorithm layer.

[0097] Step 2.1: Construct a two-layer LSTM neural network, such as... Figure 1 As shown, it includes: a shallow LSTM neural network and a deep LSTM neural network. The shallow LSTM neural network includes: a shallow forget gate, a shallow input gate, a shallow update unit, and a shallow output gate. The deep LSTM neural network includes: a deep forget gate, a deep input gate, a deep update unit, and a deep output gate.

[0098] Step 2.2: Construct the CNN layer, as follows Figure 1 As shown, it includes: an input layer, two convolutional layers, two pooling layers, and an output layer;

[0099] Step 3: Transfer the three-phase current data x of the m-th channel. k,m,t The input synchronization detection network's state detection branch processes the data.

[0100] Step 3.1: In the two-layer LSTM neural network, the shallow LSTM layer processes the m-th three-phase current data x. k,m,t After processing, we obtain x. k,m,t Shallow state information hidden state at time step t

[0101] Step 3.1.1: The function of the shallow forget gate in the shallow LSTM neural network is to determine which information should be discarded or retained. For the m-th channel three-phase current data x... k,m,t Selective discarding is performed to obtain x. k,m,t Shallow fault selection information

[0102] Step 3.1.2: The shallow input gate selects shallow fault information. The memory information at time step t-1 output by the shallow update unit After multiplying, we get x k,m,t Shallow fault retention information at time step t When t=1, let

[0103] Step 3.1.3: The shallow input gate obtains x using equations (1) and (2) respectively. k,m,t Shallow input fault information at time step t and shallow fault modulation information

[0104]

[0105]

[0106] In equations (1) and (2), W pl,k,m and U pl,k,m They represent x respectively k,m,t and The shallow input weight matrix, b pl,k,m x represents k,m,t The shallow input bias vector; x represents k,m,t The hidden state of the shallow state information at time step t-1; when t=1, let W gl,k,m and U g1,k,m They represent x respectively k,m,t and The shallow modulation weight matrix, b gl,k,m x represents k,m,t The shallow modulation bias vector; tanh is the activation function;

[0107] Step 3.1.4: The function of the shallow update unit is to output shallow update memory information based on the shallow input gate, shallow input gate, and shallow input gate forget gate, and then... and After multiplying, we get Shallow state information pending update Therefore, x can be obtained using equation (3). k,m,t Memory information at the t-th time step

[0108]

[0109] Step 3.1.5: The function of the shallow output gate is to determine the hidden state of the shallow state information at the current time step, and to obtain x using equation (4). k,m,t Shallow state information hidden state at time step t Thus, the three-phase current data x are obtained. k,m,t The hidden state of the shallow state information at all time steps in a shallow LSTM neural network.

[0110]

[0111] In equation (4), ⊙ represents element-wise multiplication. Represents three-phase current data x k,m The shallow composite signal at time step t in the shallow memory unit And obtained from equation (5);

[0112]

[0113] In equation (5), W ol,k,m and U ol,k,m They represent x respectively k,m,t and The shallow output weight matrix, b ol,k,,m This represents the shallow layer output bias vector.

[0114] Step 3.2: In a two-layer LSTM neural network, the deep LSTM neural network hides the state information of the shallow layer. Processing is performed to obtain Hidden state of deep state information at time step t

[0115] Step 3.2.1: The function of the deep forget gate in the deep LSTM neural network is to determine which information should be discarded or retained, thus hiding the shallow state information at time step t. After selective discarding, the following is obtained Deep fault selection information at time step t

[0116] Step 3.2.2: The function of the deep input gate is to input the hidden state information of the shallow state at time step t into the shallow output gate. Select deep fault information With the output of the deep update unit Memory information at time step t-1 After multiplying, we get Deep fault retention information at time step t When t=1, let

[0117] Step 3.2.3, the deep channel input gate is obtained using equations (6)-(7) respectively. Deep input fault information at time step t and deep fault modulation information

[0118]

[0119]

[0120] In equations (6)-(7), W pg,k,m and U pg,k,m They represent and The deep input weight matrix, b pg,k,m express The deep input bias vector; express The hidden state of the deep state information at the (t-1)th time step; when t=1, let W gg,k,m and U gg,k,m They represent and The deep modulation weight matrix, b gg,k,m express The deep input bias vector;

[0121] Step 3.2.4: The function of the deep update unit is to output deep update memory information based on the deep input gate and the deep input gate forget gate, and to update the deep input fault information. and deep fault modulation information After multiplying, we get Deep state information awaiting update Thus, by using equation (8), we can obtain Memory information at the t-th time step

[0122]

[0123] Step 3.2.5: The function of the deep output gate is to determine the hidden state of the deep state information at the current time step, and obtain it using equation (9). Hidden state of deep state information at time step t Thus obtain Hidden states of deep state information at all time steps in a deep LSTM neural network

[0124]

[0125] In equation (9), Indicates hidden state information Deep composite signal at time step t in deep memory unit And obtained from equation (10);

[0126]

[0127] In equation (10), W og,k,m and Uog,k,m They represent and The deep output weight matrix, b og,k,m express The deep output bias vector.

[0128] Step 3.3: The Attention mechanism layer functions by weighting and averaging the state information contained in the hidden state at each time step to obtain a multi-resolution time-frequency feature vector. Equations (11)-(16) are then used to hide the deep state information. After processing, the multi-resolution time-frequency features S are obtained. k,m ;

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135] In equations (11)-(16), Q k,m V k,m Ke k,m They are respectively The query value, truth value, and key value, w Qh It is the linear transformation matrix of the query value, w Vl It is the truth-valued linear transformation matrix, w Kl It is the key value linear transformation matrix, e k,m for Attention score, Ke k,m T for Ke k,m The transpose of α k,m for Attention weights;

[0136] Step 3.4, Multi-resolution time-frequency features S k,m In the input Softmax layer, the function of the Softmax layer is to calculate the probability of different state label values ​​corresponding to the classification feature vector, and to obtain the three-phase current data x using equation (17). k,m,t The probability values ​​P(x) corresponding to all states k,m,t |N);

[0137] P(x k,m,t |N)=softmax(Wp,k,m ·S k,m +b p,k,m (17)

[0138] In equation (17), W P,k,m and b P,k,m S represents the multi-resolution time-frequency features respectively. k,m The probability weight matrix and probability bias vector; softmax represents the activation function;

[0139] Step 3.5: Construct the cross-entropy loss function L(y) for the state detection branch. k,m ,P k,m,n ), where P k,m,n x represents the grounding current data of the k-th line and the m-th path. k,m The label value y k The probability of predicting the Nth state;

[0140] Step 4: Processing of synchronization detection branches in the state synchronization detection network:

[0141] Step 4.1: The CNN layer processes the m-th three-phase current data x. k,m,t The m-th feature vector D is obtained through processing. k,m,t ;

[0142] Step 4.1.1, m-th three-phase current data x k,m,t The input layer of the CNN is fed to the first convolutional layer for convolution operations. The function of the convolutional layer is to extract features from the data, thereby obtaining the feature vector output by the first convolutional layer.

[0143] Step 4.1.2: The first pooling layer processes the feature vector. Max pooling is performed, and the function of the convolutional layer is to reduce the dimension of the feature map, thereby reducing the number of model parameters, resulting in the feature vector output by the first pooling layer.

[0144] Step 4.1.3: The second convolutional layer processes the feature vector. Perform a second convolution operation to obtain the feature vector output by the second convolutional layer.

[0145] Step 4.1.4: The second pooling layer processes the feature vector. After max pooling, the feature vector D output by the second pooling layer is obtained. k,m,t ={D k,m,a,t D k,m,b,t D k,m,c,t}, where D k,m,a,t D k,m,b,t D k,m,c,tThese represent the time series of phases a, b, and c at time steps of the kth path and mth path, respectively.

[0146] Step 4.2: The dynamic time warping algorithm layer processes the m-th feature vector D. k,m,t The time difference of the three-phase disconnector operation synchronization is obtained through processing.

[0147] Step 4.2.1: The dynamic time warping algorithm layer uses equations (18)-(20) to obtain the a-phase time series D. k,m,a,t and b phase time series D k,m,b,t Distance between the t-th time steps Phase a time series D k,m,a,t and c-phase time series D k,m,c,t Distance between the t-th time steps b-phase time series D k,m,b,t and c-phase time series D k,m,c,t Distance between the t-th time steps Thus, we obtain D k,m,a,t and D k,m,b,t Distance between all time steps D k,m,a,t and D k,m,c,t Distance between all time steps D k,m,b,t and D k,m,c,t Distance between all time steps

[0148]

[0149]

[0150]

[0151] In equations (18)-(20), D k,m,a,t-1 D k,m,b,t-1 D k,m,c,t-1 Let A, B, and C represent the time series of phases a, b, and c at time steps t-1 of the k-th path m-th path, respectively. Let D(,) represent the normalized path distance between the two sequences, and min{} represent the minimum value.

[0152] Step 4.2.2: Using equations (21)-(23), obtain the time difference Δ of the synchronization of the a-phase disconnector action of the k-th and m-th sampled data respectively. k,m,a The time difference Δ of the synchronization of phase b disconnect switch operation k,m,b The time difference Δ of the c-phase disconnector's action synchronization k,m,c ;

[0153]

[0154]

[0155]

[0156] Step 4.3: Construct the cross-entropy loss function L(x) for the synchronization detection branch. k,m ,Δ k,m ), where Δ k,m x represents the grounding current data of the k-th line and the m-th path. k,m The corresponding time difference of the synchronization of the disconnector action;

[0157] Step 5: Training and processing of the state synchronization detection network:

[0158] Step 5.1: Construct the loss function Loss = ɑL(y) for the state synchronization detection network. k,m ,P k,m,n )+βL(x k,m ,Δ k,m ); where ɑ represents L(y k,m ,P k,m,n The weights of L(x), where β represents L(x). k,m ,Δ k,m The weights of ) are given, and α + β = 1;

[0159] Step 5.2, based on training set T tr The network is trained using backpropagation and gradient descent to detect the state synchronization of GIS disconnectors. The loss function is calculated to update the network parameters. Training is stopped when the training epoch reaches the maximum or the loss function reaches the minimum. This results in a trained GIS disconnector operation state detection network, which is used to map the input three-phase M-path grounding data set to the corresponding state category label. Finally, the disconnector status of the GIS equipment and the time difference of the three-phase disconnector action synchronization are output.

[0160] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0161] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A method for detecting the synchronization of disconnector status and three-phase disconnector operation based on GIS equipment, characterized in that, The procedure is as follows: Step 1: Construct the training set T tr ; Step 1.1: Collect the current data of the three-phase M circuits of the GIS equipment and construct a three-phase current dataset X = {x1, ..., x}. k ,…,x K }, x k Let x represent the k-th sampled three-phase current data, and x k ={x k,1 ,...,x k,m ,...,x k,M }; where x k,m This represents the m-th three-phase current data sampled from the k-th sample, and x k,m ={x k,1,1 ,...,x k,m,t ,…,x k,M,T }, x k,m,t This represents the m-th three-phase current data sampled at time t in the k-th data; 1≤m≤M, 1≤k≤K, 1≤t≤T, K represents the total number of sampled data; M represents the total number of data channels collected; T represents the total sampling time. Step 1.2: Construct a label information set Y = {y1, ..., y2} for the three-phase current dataset X. k ,…,y K }, where y k x represents the k-th sampled three-phase current data. k The label value, and y k ∈[1,N], where N is the number of state types; Step 1.3: Use the labeled three-phase current dataset P = (X, Y) as the training set T tr ; Step 2: Construct a state synchronization detection network consisting of a state detection branch and a synchronization detection branch. The state detection branch includes a two-layer LSTM neural network layer, a one-layer Attention mechanism layer, and a one-layer Softmax layer. The synchronization detection branch includes a one-layer CNN layer and a one-layer dynamic time warping algorithm layer. Step 2.1: Construct a two-layer LSTM neural network, including a shallow LSTM neural network and a deep LSTM neural network. The shallow LSTM neural network includes a shallow forget gate, a shallow input gate, a shallow update unit, and a shallow output gate. The deep LSTM neural network includes a deep forget gate, a deep input gate, a deep update unit, and a deep output gate. Step 2.2: Construct the CNN layers, including: an input layer, two convolutional layers, two pooling layers, and an output layer; Step 3: Transfer the three-phase current data x of the m-th channel. k,m,t The input synchronization detection network's state detection branch processes the data. Step 3.1: The shallow LSTM neural network in the two-layer LSTM neural network processes the m-th three-phase current data x. k,m,t After processing, we obtain x. k,m,t Shallow state information hidden state at time step t Step 3.2: In the two-layer LSTM neural network, the deep LSTM neural network hides the state information of the shallow layer. Processing is performed to obtain Hidden state of deep state information at time step t Step 3.3: The Attention mechanism layer uses equations (11)-(16) to hide the state of the deep state information. After processing, the multi-resolution time-frequency features S are obtained. k,m ; e k,m =Q k,m Ke k,m T (14) In equations (11)-(16), Q k,m V k,m Ke k,m They are respectively The query value, truth value, and key value, w Qh It is the linear transformation matrix of the query value, w Vl It is the truth-valued linear transformation matrix, w Kl It is the key value linear transformation matrix, e k,m for Attention score, Ke k,m T for Ke k,m The transpose of α k,m for Attention weights; Step 3.4, the multi-resolution time-frequency feature S k,m The data is input into the Softmax layer, and the three-phase current data x is obtained using equation (17). k,m,t The probability values ​​P(x) corresponding to all states k,m,t |N); P(x k,m,t |N)=softmax(W p,k,m ·S k,m +b p,k,m ) (17) In equation (17), W P,k,m and b P,k,m S represents the multi-resolution time-frequency features respectively. k,m The probability weight matrix and probability bias vector; softmax represents the activation function; Step 3.5: Construct the cross-entropy loss function L(y) for the state detection branch. k,m ,P k,m,n ), where P k,m,n x represents the grounding current data of the k-th line and the m-th path. k,m The label value y k The probability of predicting the Nth state; Step 4: Processing of synchronization detection branches in the state synchronization detection network: Step 4.1: The CNN layer processes the m-th three-phase current data x. k,m,t The m-th feature vector D is obtained through processing. k,m,t ; Step 4.2: The dynamic time warping algorithm layer processes the m-th feature vector D. k,m,t The time difference of the three-phase disconnector operation synchronization is obtained through processing. Step 4.3: Construct the cross-entropy loss function L(x) for the synchronization detection branch. k,m ,Δ k,m ), where Δ k,m x represents the grounding current data of the k-th line and the m-th path. k,m The corresponding time difference of the synchronization of the disconnector action; Step 5: Training and processing of the state synchronization detection network: Step 5.1: Construct the loss function Loss = ɑL(y) for the state synchronization detection network. k,m ,P k,m,n )+βL(x k,m ,Δ k,m ); where ɑ represents L(y k,m ,P k,m,n The weights of L(x), where β represents L(x). k,m ,Δ k,m The weights of ) are given, and α + β = 1; Step 5.2: Based on the training set T tr The network is trained using backpropagation and gradient descent to detect the state synchronization of GIS disconnectors. The loss function Loss is calculated to update the network parameters. Training is stopped when the training epoch reaches the maximum or the loss function Loss reaches the minimum. This results in a trained GIS disconnector operation state detection network, which is used to map the input three-phase M-path grounding data set to the corresponding state category label. Finally, the disconnector state of the GIS equipment and the time difference of the three-phase disconnector action synchronization are output.

2. The synchronization detection method for disconnector status and three-phase disconnector operation based on GIS equipment according to claim 1, characterized in that, Step 3.1 includes: Step 3.1.1: The shallow forget gate of the shallow LSTM neural network applies the m-th three-phase current data x. k,m,t Selective discarding is performed to obtain x. k,m,t Shallow fault selection information Step 3.1.2: The shallow input gate will select shallow fault information. The memory information at time step t-1 output by the shallow update unit After multiplying, we get x k,m,t Shallow fault retention information at time step t When t=1, let Step 3.1.3: The shallow input gate obtains x using equations (1) and (2) respectively. k,m,t Shallow input fault information at time step t and shallow fault modulation information In equations (1) and (2), W pl,k,m and U pl,k,m They represent x respectively k,m,t and The shallow input weight matrix, b pl,k,m x represents k,m,t The shallow input bias vector; x represents k,m,t The hidden state of the shallow state information at time step t-1; when t=1, let W gl,k,m and U g1,k,m They represent x respectively k,m,t and The shallow modulation weight matrix, b gl,k,m x represents k,m,t The shallow modulation bias vector; tanh is the activation function; Step 3.1.4, the shallow update unit will and After multiplying, we get Shallow state information pending update Therefore, x can be obtained using equation (3). k,m,t Memory information at the t-th time step Step 3.1.5: The shallow output gate obtains x using equation (4). k,m,t Shallow state information hidden state at time step t Thus, the three-phase current data x are obtained. k,m,t The hidden state of the shallow state information at all time steps in a shallow LSTM neural network. In equation (4), ⊙ represents element-wise multiplication. Represents three-phase current data x k,m The shallow composite signal at time step t in the shallow memory unit And obtained from equation (5); In equation (5), W ol,k,m and U ol,k,m They represent x respectively k,m,t and The shallow output weight matrix, b ol,k,m This represents the shallow layer output bias vector.

3. The synchronization detection method for disconnector status and three-phase disconnector operation based on GIS equipment according to claim 2, characterized in that, Step 3.2 includes: Step 3.2.1: The deep forget gate of the deep LSTM neural network hides the shallow state information at the t-th time step. After selective discarding, the following is obtained Deep fault selection information at time step t Step 3.2.2: The deep input gate will select deep fault information. With the output of the deep update unit Memory information at time step t-1 After multiplying, we get Deep fault retention information at time step t When t=1, let Step 3.2.3: The deep channel input gate is obtained using equations (6)-(7). Deep input fault information at time step t and deep fault modulation information In equations (6)-(7), W pg,k,m and U pg,k,m They represent and The deep input weight matrix, b pg,k,m express The deep input bias vector; express The hidden state of the deep state information at the (t-1)th time step; when t=1, let W gg,k,m and U gg,k,m They represent and The deep modulation weight matrix, b gg,k,m express The deep input bias vector; Step 3.2.4: The deep update unit will input deep fault information. and deep fault modulation information After multiplying, we get Deep state information awaiting update Thus, by using equation (8), we can obtain Memory information at the t-th time step Step 3.2.5: The deep output gate is obtained using equation (9). Hidden state of deep state information at time step t Thus obtain Hidden states of deep state information at all time steps in a deep LSTM neural network In equation (9), Indicates hidden state information Deep composite signal at time step t in deep memory unit And obtained from equation (10); In equation (10), W og,k,m and U og,k,m They represent and The deep output weight matrix, b og,k,m express The deep output bias vector.

4. The synchronization detection method for disconnector status and three-phase disconnector operation based on GIS equipment according to claim 3, characterized in that, Step 4.1 includes: Step 4.1.1, the m-th three-phase current data x k,m,t The input layer of the CNN is passed to the first convolutional layer for convolution operation, thus obtaining the feature vector output by the first convolutional layer. Step 4.1.2: The first pooling layer processes the feature vector. After performing max pooling, the feature vector output by the first pooling layer is obtained. Step 4.1.3: The second convolutional layer processes the feature vector. Perform a second convolution operation to obtain the feature vector output by the second convolutional layer. Step 4.1.4: The second pooling layer processes the feature vector. After max pooling, the feature vector D output by the second pooling layer is obtained. k,m,t ={D k,m,a,t D k,m,b,t D k,m,c,t }, where D k,m,a,t D k,m,b,t D k,m,c,t These represent the time series of phases a, b, and c at time steps of the k-th path and m-th path, respectively.

5. The synchronization detection method for disconnector status and three-phase disconnector operation based on GIS equipment according to claim 4, characterized in that, Step 4.2 includes: Step 4.1: The dynamic time warping algorithm layer uses equations (18)-(20) to obtain the a-phase time series D. k,m,a,t and b phase time series D k,m,b,t Distance between the t-th time steps Phase a time series D k,m,a,t and c-phase time series D k,m,c,t Distance between the t-th time steps b-phase time series D k,m,b,t and c-phase time series D k,m,c,t Distance between the t-th time steps Thus, we obtain D k,m,a,t and D k,m,b,t Distance between all time steps D k,m,a,t and D k,m,c,t Distance between all time steps D k,m,b,t and D k,m,c,t Distance between all time steps In equations (18)-(20), D k,m,a,t-1 D k,m,b,t-1 D k,m,c,t-1 Let A, B, and C represent the time series of phases a, b, and c at time steps t-1 of the k-th path m-th path, respectively. Let D(,) represent the normalized path distance between the two sequences, and min{} represent the minimum value. Step 2.4.7: Using equations (21)-(23), obtain the time difference Δ of the synchronization of the a-phase disconnector action of the k-th and m-th sampled data respectively. k,m,a The time difference Δ of the synchronization of phase b disconnect switch operation k,m,b The time difference Δ of the c-phase disconnector's action synchronization k,m,c ; 。 6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing any of the synchronization detection methods of claims 1-5, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the synchronization detection method according to any one of claims 1-5.

Images