A Method for Transferring Power Grid Fault Diagnosis Models Based on CycleGAN

Abstract

The present invention proposes a method for migrating a power grid fault diagnosis model based on CycleGAN, which obtains measured and simulated data samples of power grid fault alarm information; converts the measured and simulated alarm information texts into temporal density maps based on the temporal density of the alarm information as the model input; performs feature learning and migration on the temporal density maps of the measured and simulated alarm information based on CycleGAN; constructs a power grid fault diagnosis model based on VGG; and inputs the temporal density map of the measured alarm information converted by CycleGAN into the fault diagnosis model pre-trained based on simulation data, and still can accurately identify faults. This method performs feature learning and migration on the power grid alarm information through deep learning methods, and improves the model migration performance of the power grid fault diagnosis model in practical applications.

Description

Technical Field

[0001] The present invention belongs to the field of power grid fault diagnosis, and specifically provides a method for migrating a power grid fault diagnosis model based on CycleGAN, which is applicable to a fault diagnosis system with the fault information source being power grid fault alarm information. Background Art

[0002] Power grid fault diagnosis is the basis for ensuring the safe and stable operation of the power grid. The data sources for power grid fault diagnosis are diverse. In the early stage, the data acquisition and monitoring system (SCADA, Supervisory Control And Data Acquisition) was used as the fault information source for power grid fault diagnosis. The SCADA system is a computer-based DCS and power automation monitoring system, which can realize various functions such as data acquisition, equipment control, measurement, parameter adjustment, and various signal alarms, and is an indispensable tool for power dispatching. In recent years, deep learning has made great breakthroughs in the fields of image, speech, and natural language processing. Deep learning designs corresponding deep network structures for different tasks and discovers complex structures in large data sets through the backpropagation algorithm. Therefore, the application of deep learning to power grid fault diagnosis has increasingly become a research hotspot.

[0003] During the actual operation of the power grid, the operation information of a large number of various devices is sent to the SCADA system in real time. When a power grid fault occurs, the information such as the oscillograph information, protection action information, and circuit breaker tripping related to the faulty device is also sent. However, during the actual normal operation of the power grid, the normal operation time is long, and faults only occur in a few cases. The limited fault information is often submerged in the normal operation and abnormal operation information. Therefore, it is necessary to extract fault samples, and the number of extracted fault samples is limited.

[0004] In the case of a lack of measured samples, it is very difficult to train a fault diagnosis model with excellent performance relying on measured data. Therefore, a diagnosis model with excellent performance is trained by generating simulation data. However, when the measured data is input into the diagnosis model trained based on the simulation data for testing, the diagnosis effect is not optimistic, and accurate fault discrimination cannot be achieved. The reason lies in the subtle differences in features between the measured data and the simulation data. Therefore, it is considered to realize feature learning between the measured and simulation data and optimize the migration performance of the fault diagnosis model. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a method for migrating a power grid fault diagnosis model based on CycleGAN, and the method includes:

[0006] S1. Obtain the measured and simulation data samples of the power grid fault alarm information;

[0007] S2. Convert the measured and simulated data samples of the alarm information into a time series density map based on the time series density of the alarm information;

[0008] S3. Perform feature learning and migration on the measured and simulated alarm information time series density maps based on the Cycle Generative Adversarial Network (CycleGAN);

[0009] S4. Construct a power grid fault diagnosis model based on VGG;

[0010] S5. Input the measured alarm information time series density map converted by CycleGAN into the fault diagnosis model pre-trained with simulation data to obtain the fault diagnosis result.

[0011] Preferably, the step S1 includes:

[0012] After a fault occurs in the power grid, record the information of the protection and switch actions of electrical equipment and upload it to the SCADA system in real time. Intercept the corresponding measured samples of fault alarm information from the SCADA system, and conduct fault simulation experiments on the power system simulation platform to obtain the simulation samples of fault alarm information.

[0013] Preferably, the step S2 includes:

[0014] For a single electrical equipment, classify the alarm information into four categories: protection action information, circuit breaker action information, other information, and global information; form a discrete digital sequence for each alarm information within a certain time window, and superimpose a certain type of alarm information within this time window to form a summary sequence of the sum of discrete points at each moment; perform graphic encoding on the four categories of alarm information after superimposing and calculating according to the time series characteristics to form a time series density map.

[0015] Preferably, in the step S3:

[0016] The CycleGAN consists of two generators and two discriminators. Take the measured alarm information time series density map sample set as domain A and the simulated alarm information time series density map sample set as domain B. Through the game training of the generators and discriminators in CycleGAN, domain A and domain B learn each other's distribution characteristics, so as to realize the mutual feature migration and transformation between domain A and domain B.

[0017] Preferably, the generator consists of an encoder, a converter, and a decoder; the encoder realizes the extraction of features through each convolutional layer in the CNN, the converter combines and extracts the feature vectors given by the encoder, and uses these features to realize the conversion learning of the feature vectors of the alarm information time series density map from the measured data to the target domain simulation data. The decoder uses the transposed convolutional layer to gradually restore the input features of the converter to the low-level features of the time series density map, and finally generates the time series density map after feature migration.

[0018] Preferably, in the step S4:

[0019] The VGG-based power grid fault diagnosis model includes 13 convolutional layers, 5 pooling layers and 3 fully connected layers; the convolutional layers extract the features of the time series density map of the alarm information through convolutional kernels, and the pooling layers reduce the dimensions of the features extracted by the convolutional layers, remove redundant information, compress the features, and form a two-dimensional feature map. The fully connected layer converts the previously extracted two-dimensional feature map into a one-dimensional feature vector, and outputs the fault diagnosis result through the Softmax activation function, that is, discriminates five fault types: instantaneous fault, permanent fault, circuit breaker refusal to operate, circuit breaker malfunction, and protection malfunction.

[0020] Preferably, in the step S5:

[0021] Input the measured alarm information time series density map into the trained CycleGAN model to learn the distribution characteristics of the simulated alarm information time series density map, and input the measured alarm information time series density map after feature migration into the power grid fault diagnosis model trained based on VGG and the simulated alarm information time series density map to realize the fault discrimination of the measured data in the simulation data training model.

[0022] The present invention also proposes a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method according to the present invention.

[0023] The present invention also proposes a computer-readable storage medium, on which a computer program is stored, characterized in that when the program is executed by a processor, the steps of the method according to the present invention are realized.

[0024] The present invention uses the variant CycleGAN of the generative adversarial network to enable the measured and simulated alarm information to perform feature learning and migration with each other, solves the problem that it is impossible to train a diagnostic model based on a large number of measured samples due to insufficient measured fault samples, and the diagnostic model trained based on a large number of simulation data cannot accurately diagnose the measured data. In this process, the alarm information text is also converted into an alarm information time series density map, making the data source of fault diagnosis more rich and diverse, which is conducive to the wide application of deep learning methods in fault diagnosis. Finally, based on the VGG network, the power grid fault diagnosis is transformed into a classification problem. This method optimizes the migration performance of the power grid fault diagnosis model, enabling the measured data to still be accurately tested in the simulation data model, which is conducive to the development of intelligent power grid regulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flowchart of the method for optimizing the migration performance of the CycleGAN-based power grid fault diagnosis model according to a preferred embodiment of the present invention.

[0026] Figure 2 It is the principle structure diagram and cyclic consistency representation of the CycleGAN model described in a preferred embodiment of the present invention.

[0027] Figure 3 It is the structure diagrams of the CycleGAN model generator and discriminator described in a preferred embodiment of the present invention.

[0028] Figure 4 It is the structure diagram of the VGG-based power grid fault diagnosis model described in a preferred embodiment of the present invention.

[0029] Figure 5 It is the confusion matrix diagram of the power grid fault diagnosis results after feature migration by the CycleGAN model described in a preferred embodiment of the present invention.

[0030] Figure 6 It is the confusion matrix diagram of the power grid fault diagnosis results without feature migration by the CycleGAN model described in a preferred embodiment of the present invention. Detailed implementation manners

[0031] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. The embodiments described in this application are only a part of the embodiments of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.

[0032] The present invention provides a method for migrating a power grid fault diagnosis model based on CycleGAN (Cycle-Consistent Generative Adversarial Networks). This method first obtains the measured data and simulation data of power grid fault alarm information through the SCADA system and the power system simulation software TS2000 respectively, and converts the alarm information text into an alarm information time series density map. The alarm information time series density map is input into the CycleGAN model for learning and migrating the distribution characteristics between the measured and simulation data, which provides a new idea for optimizing the migration performance of the fault diagnosis model. A VGG-based power grid fault diagnosis model is constructed, and the model is trained with the simulation alarm information time series density map as the data basis. The measured alarm information time series density map processed by CycleGAN is input into the aforementioned trained diagnosis model to obtain accurate fault diagnosis results, that is, to distinguish five fault types: instantaneous fault, permanent fault, circuit breaker refusal to operate, circuit breaker misoperation, and protection misoperation.

[0033] SeeFigure 1 , the method specifically includes the following steps:

[0034] S1. Obtain the measured and simulated data samples of power grid fault warning information from the SCADA system and the power system simulation software TS2000;

[0035] Specifically, after a power grid fault occurs, a large amount of information recording the protection and switch actions of electrical equipment is uploaded to the SCADA system in real time, and the corresponding fault warning information samples are intercepted from the SCADA system. A large number of fault simulation experiments are carried out on the power system simulation software TS2000 platform to obtain a sufficient number of fault warning information simulation samples.

[0036] S2. Convert the measured and simulated warning information texts into a time series density map based on the time series density of the warning information as the model input;

[0037] Specifically, analyze the time series characteristics of the warning information, consider the information type and time series distribution characteristics, propose a graphical data representation method of time series superposition for warning information classification oriented to equipment, form the time series density map of the warning information of each equipment, and provide a unified data representation form for model feature learning and migration and the deep learning model of fault diagnosis.

[0038] The warning information for a single electrical equipment is divided into four categories: protection action information, circuit breaker action information, other information, and global information. In milliseconds (ms), the start and end times of the warning information are defined as t start and t end .

[0039] For each warning information, within a 15s time window between t start and t end , a discrete digital sequence X i with a time interval Δt of 10ms and a length of 1500 is formed as follows:

[0040] X i =[x i,0 x i,△t x i,2△t … x i,t (1)

[0041] In the formula, the subscript i represents the i-th warning information, and x i,t represents the quantity of the i-th warning information at time t, and its expression is as follows:

[0042]

[0043] In the formula, t i,S represents the start time of the warning information, and t i,EIndicates the end time of the alarm information action. If t ∈ [t i,S , t i,E , the number of the i-th alarm information at time t is 1, otherwise it is 0.

[0044] At time t start ~t end , a certain type of alarm information is superimposed to form a summary sequence N of the discrete point superposition summation at each moment as follows:

[0045]

[0046] In the formula, n is the total number of a certain type of alarm information of a certain electrical equipment in the fault event.

[0047] The above four types of alarm information are superimposed and calculated according to the time series characteristics, and the numerical and color coding correspondence is calculated based on the prism module of the miscellaneous - colormap function in the matplotlib toolbox. Through such graphic coding, a time series density map of alarm information is formed. Using this numerical and color coding correspondence method is clear at a glance, and some alarm information with short duration can be clearly distinguished, which meets the expected input of CycleGAN and is beneficial for the model to extract the time series characteristics of such alarm information.

[0048] S3. Feature learning and migration are performed on the time series density maps of measured and simulated alarm information based on CycleGAN;

[0049] Specifically, feature learning and migration are performed on the time series density maps of measured and simulated alarm information based on CycleGAN. Refer to Figure 2 , different from ordinary GAN, CycleGAN is composed of two generators and two discriminators in total, including the generator G from domain X to domain Y and the generator F from domain Y to domain X; the two discriminators D X and D Y , which respectively judge the quality of the pictures generated by the two generators. Assume that the generator F represents the mapping X→Y between the source domain X and the target domain Y. The source domain sample x ∈ X is generated by F(X) to obtain a sample y ∈ Y close to the target domain. Then the loss function L of the generator F and the discriminator D Y is:

[0050]

[0051] In the formula, P data (x) represents the distribution of real data samples, ~ represents the relationship of obedience, and E represents the mathematical expectation. Because learning from the source domain to the target domain cannot be achieved only through one objective function, CycleGAN also introduces the reverse mapping G Y→X from Y→X, D XDetermine whether the sample y ∈ Y in the target domain generated by G(y) to be a sample x ∈ X close to the source domain is a real sample, G Y→X and D X The mapping loss function is defined as follows:

[0052]

[0053] See Figure 2 , cycle consistency is the core idea of CycleGAN, which prevents all samples in the source domain from being converted into a single sample in the target domain. According to F(G(y)) ≈ y and G(F(x)) ≈ x, the cycle consistency loss L CCL can be expressed as follows:

[0054]

[0055] From formulas (1)-(3), the objective function of CycleGAN is:

[0056]

[0057] In the formula, λ CCL controls the weight coefficient of L CCL in the overall loss.

[0058] Refer to Figure 3 , the generator consists of an encoder, a transformer, and a decoder. The encoder extracts features through each convolutional layer in the CNN. The transformer combines and extracts the feature vectors given by the encoder, and uses these features to realize the transformation learning of the feature vectors of the time series density map of alarm information from the source domain (measured data) to the target domain (simulation data). Refer to Figure 3 , the function of the decoder is the opposite of that of the encoder. It uses transposed convolutional layers to gradually restore the input features of the transformer to the low-level features of the time series density map until the time series density map is finally generated. The discriminator uses a PatchGAN classifier. PatchGAN maps the input to an N×N patch (matrix) X, and the value of X ij represents the probability that each patch is a real sample. Take the mean of X ij to get the final output of the discriminator.

[0059] Take the time series density map sample set of measured alarm information as domain X, and the time series density map sample set of simulated alarm information as domain Y. Through the game training of the generator and discriminator in CycleGAN, domain X and domain Y learn each other's distribution characteristics, so as to realize the mutual feature transfer and transformation between domain X and domain Y.

[0060] Among them, the inputs of the two generators are respectively the real measured and simulated alarm information time series density maps, and the outputs are respectively the fake measured and simulated alarm information time series density maps generated after learning the features of each other. The discriminator D X takes as inputs the real measured alarm information time series density map and the fake measured alarm information time series density map generated after learning the features of the simulated alarm information time series density map. The discriminator D Y takes as inputs the real simulated alarm information time series density map and the fake simulated alarm information time series density map generated after learning the features of the measured alarm information time series density map. The discriminator D X outputs the probability that the generated measured alarm information time series density map is close to the real measured alarm information time series density map. The discriminator D Y outputs the probability that the generated simulated alarm information time series density map is close to the real simulated alarm information time series density map.

[0061] S4. Construct a power grid fault diagnosis model based on VGG (Visual Geometry Group);

[0062] Specifically, refer to Figure 4 , the VGG-based power grid fault diagnosis model includes 13 convolutional layers, 5 pooling layers and 3 fully connected layers. Among them, the 13 convolutional layers are divided into 5 segments, and the number of convolutional kernels in each segment of convolutional layers is the same. All convolutional layers use 3×3 small convolutional kernels and ReLU activation functions. The network constructed by cascading small convolutional kernels has a deeper depth, stronger nonlinearity and fewer parameters, and cascading multiple small-sized convolutional kernels can obtain the same receptive field as a large-sized convolutional kernel. The pooling layers are distributed between each segment of convolutional layers, and all pooling layers use max pooling with a window size of 2×2 and a stride of 2. The alarm information time series density map is input into the input layer, and after feature extraction through convolutional layers, pooling layers and fully connected layers, the fault diagnosis result is output through the Softmax activation function, that is, five fault types are discriminated: instantaneous fault, permanent fault, circuit breaker refusal to operate, circuit breaker malfunction and protection malfunction.

[0063] Among them, the convolutional layer extracts the features of the alarm information time series density map through a convolutional kernel, and its expression is:

[0064]

[0065] In the formula, y is an M×N matrix, and y mn is the element in its m-th row and n-th column; m = 0, 1,..., M - 1; n = 0, 1,..., N - 1; w is a J×I convolutional kernel, and w ij is the element in its i-th row and j-th column; x m+i,n+jis the element at the (m + i)-th row and (n + j)-th column of the input matrix x; b is the bias variable; f is the activation function.

[0066] The pooling layer reduces the dimension of the features extracted by the convolutional layer, removes redundant information, and compresses the features. Its sampling equation is:

[0067]

[0068] where S1 and S2 are the dimensions of the pooling region in rows and columns respectively; C is the output matrix of order (M / S1)×(N / S2), and C ab is the element at its a-th row and b-th column; a = 0, 1,..., M / S 1-1 , b = 0, 1,..., N / S 2-1 ; y aS1+i,bS2+j is the element at the aS-th row and bS-th column of the output matrix y. 1+i row and bS-th 2+j column.

[0069] The fully connected layer converts the previously extracted two-dimensional feature map into a one-dimensional feature vector and then outputs the classification through the Softmax activation function.

[0070] The expression of the fully connected layer is as follows:

[0071]

[0072] where e = [e1, e2,..., ei,..., en] is the n-dimensional input variable; k = [k1, k2,..., ki,..., kn] is the connection weight; g is the bias; o is the output.

[0073] The expression of the Softmax activation function is as follows:

[0074]

[0075] where z is the output value of the neuron, and z j is the output value of the j-th neuron, K is the total number of categories, and e is the natural base.

[0076] Refer to Figure 4 , the size of the input alarm information time series density map is 256×256×3, and the sizes of the output feature maps of each convolutional layer and pooling layer are as Figure 4As shown on the right, it then passes through two fully connected layers of 1×1×4096, and after activation by the ReLU function, the output size is 1×1×4096. Finally, it passes through a fully connected layer of 1×1×5 (5 is determined by the final classification quantity. Since five types of faults need to be discriminated, a fully connected layer of 1×1×5 is required), and then the probabilities of the five prediction results are output through Softmax (the sum of the probabilities is 1), that is, the five types of faults: instantaneous fault, permanent fault, breaker refusal to operate, breaker malfunction, and protection malfunction. The fault type with the highest probability value among the five types of fault prediction results is the finally determined fault type, thus realizing fault diagnosis.

[0077] S5. Input the time series density map of the measured alarm information transformed by CycleGAN into the fault diagnosis model pre-trained based on simulation data to obtain the fault diagnosis result.

[0078] Specifically, input the time series density map of the measured alarm information into the trained CycleGAN model to learn the distribution characteristics of the time series density map of the simulated alarm information, and then input the time series density map of the measured alarm information with feature migration into the power grid fault diagnosis model trained based on VGG and the time series density map of the simulated alarm information. Refer to Figure 5 , the accuracy rate of the power grid fault diagnosis after feature migration by the CycleGAN model can reach 90.5%; refer to Figure 6 , the accuracy rate of the power grid fault diagnosis without feature migration by the CycleGAN model is only 78.1%. The purpose of accurately discriminating faults for the measured data in the model trained with simulation data is realized, thus enhancing the migration performance of the diagnosis model.

[0079] The power grid fault diagnosis model migration method based on CycleGAN provided by the present invention optimizes the migration performance of the power grid fault diagnosis model and improves the accuracy of testing the measured data in the fault diagnosis model trained with simulation data.

[0080] This disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to implement various aspects of this disclosure.

[0081] A computer-readable storage medium can be a tangible device that can hold and store instructions for use by an instruction execution device. A computer-readable storage medium may be, for example—but not limited to—an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device, such as a punched card or raised structures in grooves storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage medium used herein is not construed as an instantaneous signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated through a waveguide or other transmission medium (e.g., an optical pulse through an optical fiber cable), or an electrical signal transmitted through a wire.

[0082] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to respective computing / processing devices, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include a copper transmission cable, an optical fiber transmission, a wireless transmission, a router, a firewall, a switch, a gateway computer, and / or an edge server. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0083] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine - related instructions, microcode, firmware instructions, state - setting data, or source code or object code written in any combination of one or more programming languages, including object - oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as the "C" language or similar programming languages. The computer - readable program instructions may be executed entirely on the user's computer, partially on the user's computer, executed as a stand - alone software package, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider). In some embodiments, by using the state information of the computer - readable program instructions to customize an electronic circuit, such as a programmable logic circuit, a field - programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute the computer - readable program instructions to implement various aspects of the present disclosure.

[0084] Aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer - readable program instructions.

[0085] These computer - readable program instructions can be provided to a processor of a general - purpose computer, a special - purpose computer, or other programmable data - processing apparatus to produce a machine such that the instructions, when executed by the processor of the computer or other programmable data - processing apparatus, create a means for implementing the functions / acts specified in one or more blocks of the flowchart and / or block diagram. These computer - readable program instructions can also be stored in a computer - readable storage medium, which causes a computer, a programmable data - processing apparatus, and / or other devices to operate in a particular manner, so that the computer - readable medium storing the instructions includes a manufacture that includes instructions for implementing various aspects of the functions / acts specified in one or more blocks of the flowchart and / or block diagram.

[0086] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device, causing a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process such that the instructions executed on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in one or more boxes of the flowchart and / or block diagram.

[0087] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a portion of an instruction, which contains one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two consecutive blocks may in fact be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block of the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or acts, or by a combination of dedicated hardware and computer instructions.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent substitutions can still be made to the specific embodiments of the present invention, and any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A method for migrating a power grid fault diagnosis model based on CycleGAN, characterized in that, The method includes: S1. Obtain the measured and simulated data samples of grid fault alarm information; S2. Based on the time series density of the alarm information, convert the text of the measured and simulated data samples of the alarm information into a time series density map; for a single electrical device, divide the alarm information into four categories: protection action information, circuit breaker action information, other information, and global information; for each alarm information, form a discrete digital sequence within a certain time window, and superimpose a certain type of alarm information within this time window to form a summary sequence of the sum of discrete points at each moment; after superimposing and calculating the four types of alarm information according to the time series characteristics, perform graphic coding to form a time series density map; S3. Based on the Cycle Generative Adversarial Network (CycleGAN), perform feature learning and migration on the measured and simulated alarm information time series density maps; the CycleGAN consists of two generators and two discriminators. Take the measured alarm information time series density map sample set as domain A and the simulated alarm information time series density map sample set as domain B. Through the game training of the generators and discriminators in the CycleGAN, domain A and domain B learn each other's distribution characteristics, so as to realize the mutual feature transfer and transformation between domain A and domain B; the generator consists of an encoder, a converter, and a decoder; the encoder extracts features through each convolutional layer in the CNN, the converter combines and extracts the feature vectors given by the encoder, and uses these features to realize the conversion learning of the feature vectors of the alarm information time series density map from the measured data to the target domain simulated data. The decoder uses the deconvolution layer to gradually restore the input features of the converter to the low-level features of the time series density map, and finally generates the time series density map after feature migration; S4. Build a grid fault diagnosis model based on VGG; S5. Input the measured alarm information time series density map transformed by CycleGAN into the fault diagnosis model pre-trained based on simulation data to obtain the fault diagnosis result; including: input the measured alarm information time series density map into the trained CycleGAN model to learn the distribution characteristics of the simulated alarm information time series density map, and input the measured alarm information time series density map after feature migration into the grid fault diagnosis model trained based on VGG and the simulated alarm information time series density map to realize the fault discrimination of the measured data in the simulation data training model.

2. The method according to claim 1, characterized in that The step S1 includes: After a grid fault occurs, record the information of the protection and switch actions of electrical equipment and upload it to the SCADA system in real time. Intercept the corresponding measured samples of the fault alarm information from the SCADA system, and conduct a fault simulation experiment on the power system simulation platform to obtain the simulated samples of the fault alarm information.

3. The method according to claim 2, characterized in that, In the step S4: The VGG-based power grid fault diagnosis model includes 13 convolutional layers, 5 pooling layers, and 3 fully connected layers; the convolutional layers extract the features of the temporal density map of alarm information through convolutional kernels, the pooling layers reduce the dimensions of the features extracted by the convolutional layers, remove redundant information, compress the features, and form a two-dimensional feature map, and the fully connected layers convert the aforementioned extracted two-dimensional feature map into a one-dimensional feature vector, and the fault diagnosis result is output through the Softmax activation function, that is, five fault types are discriminated: instantaneous fault, permanent fault, circuit breaker refusal to operate, circuit breaker misoperation, and protection misoperation.

4. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used for storing instructions; The processor is used for operating according to the instructions to execute the steps of the method according to any one of claims 1-3.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, the steps of the method according to any one of claims 1-3 are implemented.

Images