Intelligent power plant relay protection fault analysis method based on intelligent algorithm

Abstract

The invention discloses an intelligent power plant relay protection fault analysis method based on an intelligent algorithm. According to the method, double-end current, voltage and switch state signals of an intelligent power plant relay protection system are collected, and original feature data are analyzed; extracting fault feature components by using a Fourier filtering fault feature algorithm to obtain a preliminary fault feature set, and inputting the preliminary fault feature set into a distillation lightweight fault diagnosis model to output a preliminary judgment result of a fault type; constructing a double-end current differential twinborn model, and calculating a double-end current characteristic difference index to obtain fault position associated data; the results are integrated and transmitted to a protection action logic verification platform for verification, and a protection action verification result is generated; and finally, forming a complete fault analysis report by combining data of each link. According to the method, multiple intelligent technologies are fused, the accuracy and timeliness of fault analysis are improved, and a guarantee is provided for safe and stable operation of an intelligent power plant relay protection system.

Description

Technical Field

[0001] This invention relates to the field of relay protection technology for smart power plants, and in particular to a fault analysis method for relay protection in smart power plants based on intelligent algorithms. Background Technology

[0002] With the continuous advancement of smart power plant construction, the complexity of the overall power plant operation system is constantly increasing. As a key component ensuring the safe and stable operation of the power plant, the relay protection system faces more stringent operational requirements. In the current daily operation of power plants, the amount of various operational signal data related to relay protection has increased significantly. Traditional fault analysis methods often rely on manual intervention or single processing methods, which not only struggle to achieve rapid and accurate processing of fault-related information and timely determination of the specific type and location of the fault, but also suffer from low efficiency in verifying the rationality of protection action logic. This situation is severely incompatible with the timeliness and accuracy requirements of smart power plants for fault handling. Therefore, developing more efficient and accurate relay protection fault analysis methods has become an important direction for the current development of smart power plants.

[0003] Existing technologies in the field of fault analysis for relay protection in smart power plants have two significant drawbacks: First, the efficiency of fault feature processing and diagnostic result output is insufficient to meet practical needs. Some technologies, when processing fault-related signals, cannot comprehensively and effectively separate key feature information under fault conditions, resulting in a lack of reliable data support for subsequent diagnostic stages. Furthermore, the overall diagnostic process is slow, failing to provide rapid diagnostic results after a fault occurs, thus affecting the timeliness of fault handling. Second, the coordination between fault location determination and protection action logic verification is insufficient. Traditional technologies lack in-depth analysis and precise comparison of relevant operational data when determining fault locations, making it difficult to accurately pinpoint the fault location. In the protection action logic verification stage, simple rule matching methods are often used, without fully integrating fault diagnosis results and location information for comprehensive judgment, resulting in low reliability of verification results and failing to provide comprehensive and effective reference for the optimization and adjustment of relay protection systems. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a fault analysis method for relay protection in smart power plants based on intelligent algorithms.

[0005] The technical solution adopted in this invention is a fault analysis method for relay protection in smart power plants based on intelligent algorithms, comprising: Step S1, collecting double-ended current signals, voltage signals, and switch status signals during the operation of the relay protection system in the smart power plant, transmitting the collected signals to a data processing module for signal analysis, and extracting original feature data related to relay protection faults; Step S2, calling the Fourier filter fault feature algorithm to extract fault features from the original feature data, decomposing and filtering the frequency components in the original feature data, separating the feature components under the fault state, and obtaining a preliminary fault feature set; Step S3, inputting the preliminary fault feature set into a distillation lightweight fault diagnosis model, which performs feature mapping on the preliminary fault feature set through a pre-trained teacher network, and then compresses and learns the mapped features through a student network, outputting an initial fault type. Step S4: Construct a dual-terminal current differential twin model. Input the current signals collected from both ends of the smart power plant relay protection system into the two symmetrical sub-networks of the model. Simultaneously extract and compare the features of the current signals through the sub-networks, calculate the difference index of the dual-terminal current features, and obtain the fault location association data. Step S5: Integrate the preliminary fault type judgment result and the fault location association data and transmit them to the protection action logic verification platform. The platform performs logical matching and verification on the integrated data according to the preset protection logic rules of the smart power plant relay protection system, and generates the protection action verification result. Step S6: Based on the protection action verification result, combined with the feature components output by the Fourier filter fault feature algorithm, the learning parameters of the distillation lightweight fault diagnosis model, and the difference index of the dual-terminal current differential twin model, form a complete relay protection fault analysis report.

[0006] Furthermore, the Fourier filter fault feature algorithm extracts fault feature components using the following expression: ,in, It is a natural constant. It is the imaginary unit. It is an angular frequency variable. This is the frequency domain expression for the fault characteristic components. The original feature data time-domain signal was collected. The signal angular frequency, This is the frequency response function of the Fourier filter. This is the upper limit of the filter frequency. As a time variable; at the same time, through Calculate the weighted fusion value of the feature components. These are the fused fault characteristic values. For the first angular frequency at each frequency point For the first The weighting coefficients for each frequency point The total number of frequency points. The value of the fault characteristic component in the frequency domain.

[0007] Furthermore, the feature learning process of the distillation lightweight fault diagnosis model satisfies: ,in, The model distillation loss function is... For loss weighting coefficients, Let cross-entropy be the loss function. For fault type labels, For the output of the student network, The input is the preliminary fault feature set. Let the mean squared error loss function be . The output of the teacher network; lightweight compression of the model is achieved through... Perform parameter optimization. The compression ratio of the model parameters. For student network Layer weight matrix, For the number of student network layers, For Teachers' Network Layer weight matrix, For the number of teacher network layers, This represents the Frobenius norm.

[0008] Furthermore, the expression for calculating the difference index of the two-terminal current differential twin model is as follows: ,in, This is an index for the difference in characteristics of two-terminal currents. For the feature mapping function of the Siamese model subnetwork, This refers to the current signal collected at terminal A of the smart power plant. This refers to the current signal collected at the B-end of the smart power plant. It is the L2 norm; at the same time, through Correct the difference index. The corrected fault location correlation index, This is the difference weighting coefficient. This is the weighting coefficient for the current difference. This represents the amplitude difference between the two-terminal current signals.

[0009] Furthermore, the logic matching verification of the protection action logic verification platform adopts the following method: ,in, For logical matching verification values, For the first The trigger threshold for the protection action logic. For the first The actual trigger value of the protection action logic, For the first The reliability coefficient of the protection action logic. To protect the total number of action logics; when When the protection action logic is deemed to meet the requirements, To protect the action logic verification threshold, and pass Calculated.

[0010] Furthermore, in the process of generating a complete relay protection fault analysis report, the following methods are adopted: Integrating multi-dimensional data, among which, This is the comprehensive evaluation value in the fault analysis report. These are the weighting coefficients for fault characteristic values. The weighting coefficients for the fault location-related indicators. For logical matching check value weighting coefficients, The fused fault feature values ​​are output by the Fourier filter fault feature algorithm. The fault location correlation index is the result of correction to the two-terminal current differential twin model. To protect the logic matching verification values ​​of the action logic verification platform; simultaneously, through Calculate the correlation coefficient for fault analysis. The correlation coefficient. For the first The actual trigger value of the protection action logic.

[0011] Further, step S3 includes the following sub-steps: S31, Initialize the teacher network of the distillation lightweight fault diagnosis model, load the weight parameters and bias parameters obtained during pre-training, determine the activation function type and number of neurons in each hidden layer of the teacher network, and establish the feature mapping framework of the teacher network; S32, Divide the preliminary fault feature set obtained in step S2 into multiple feature subsets according to a preset batch size, and input each feature subset into the teacher network in sequence. Through the input layer and hidden layer of the teacher network, the feature subsets are transformed layer by layer to obtain high-dimensional feature vectors; S33, Initialize the student network of the distillation lightweight fault diagnosis model, set the number of layers and the number of neurons in each layer of the student network so that the parameter scale is smaller than that of the teacher network, and use the high-dimensional feature vector output by the teacher network as the reference feature of the student network; S34, The student network extracts features from the input preliminary fault feature set, and calculates the error between the output features of the student network and the reference features of the teacher network. Adjust the weight parameters of the student network according to the error, iterate repeatedly until the error meets the preset conditions, and output the preliminary judgment result of the fault type.

[0012] Further, step S4 includes the following sub-steps: S41, Construct two symmetrical sub-networks of the dual-end current differential twin model. The two sub-networks have the same structure, each containing an input layer, a convolutional layer, a pooling layer, and a fully connected layer. Set the kernel size, stride, and number of output channels for each layer; S42, Preprocess the current signals collected from end A and end B of the smart power plant and input them into the corresponding sub-networks. Extract features from the current signals through the convolutional layers of the sub-networks, reduce the dimensionality of the extracted features through the pooling layers, and convert the dimensionality-reduced features into feature vectors through the fully connected layers; S43, Calculate the Euclidean distance between the output feature vectors of the two sub-networks. Use the Euclidean distance as the initial difference in the dual-end current features, and introduce a time synchronization coefficient to correct the initial difference; S44, Based on the corrected difference, and combined with the line parameters and topology of the smart power plant relay protection system, generate fault location association data to clarify the possible area range of the fault.

[0013] Further, step S5 includes the following sub-steps: S51, unify the data format of the preliminary fault type judgment result output in step S3 and the fault location association data obtained in step S4, and convert it into a feature matrix form that can be recognized by the protection action logic verification platform; S52, retrieve the preset protection logic rules corresponding to the smart power plant relay protection system from the database of the protection action logic verification platform. These rules include protection action thresholds, action delays, and linkage switch information under different fault types; S53, match the unified format feature matrix with the preset protection logic rules row by row, calculate the deviation value between each element in the feature matrix and the corresponding rule threshold, and count the number of elements with deviation values ​​within the allowable range; S54, based on the deviation value statistics, score the compliance of the protection action according to the preset scoring criteria, generate the protection action verification result, and clarify whether the protection action conforms to the preset logic rules.

[0014] A fault analysis method for smart power plant relay protection based on intelligent algorithms is implemented through different units, including: a smart power plant relay protection signal acquisition and parsing unit, used to acquire the dual-terminal current, voltage, and switch status signals during the operation of the smart power plant relay protection system, and to parse and extract fault-related raw feature data; a Fourier filter fault feature extraction unit, used to call the Fourier filter fault feature algorithm to decompose and filter the frequency components of the raw feature data, separating fault feature components to obtain a preliminary fault feature set; and a distillation lightweight fault diagnosis calculation unit, used to receive the preliminary fault feature set and, through the teaching of the distillation lightweight fault diagnosis model... The system includes a teacher network and a student network for computation, outputting preliminary fault type judgment results; a two-terminal current differential twin model construction and computation unit, used to construct a two-terminal current differential twin model, input two-terminal current signals and calculate feature difference index to obtain fault location association data; a protection action logic verification and evaluation unit, used to receive the preliminary fault type judgment results and fault location association data, verify and generate protection action verification results according to preset protection logic rules; and a fault analysis report generation and storage unit, used to combine protection action verification results and multi-algorithm output parameters to generate and store a complete fault analysis report, while also performing data backtracking retrieval.

[0015] Beneficial Effects: This invention proposes a fault analysis method for intelligent power plant relay protection based on intelligent algorithms. This method constructs a full-process analysis system, first collecting relay protection system operating signals and parsing raw feature data, then using specialized algorithms to extract fault features, and combining a lightweight diagnostic model to quickly output fault type judgment results. Simultaneously, a two-terminal current correlation model is used to accurately calculate fault location correlation data. Finally, a protection action logic verification platform generates verification results and forms a complete analysis report, thus improving the overall accuracy and efficiency of fault analysis. This method integrates multiple intelligent algorithms, comprehensively separating fault feature components and accelerating the diagnostic model's computation speed through lightweight optimization to meet real-time analysis requirements. Addressing the poor coordination between fault location determination and protection action logic verification, this method relies on a two-terminal current correlation model to achieve in-depth data analysis and precise comparison to accurately locate faults. Simultaneously, it combines fault diagnosis results with location data for comprehensive verification, improving the reliability of verification results and providing strong support for the safe and stable operation of the relay protection system. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1

[0019] like Figure 1 As shown, the fault analysis method for relay protection in smart power plants based on intelligent algorithms includes: Step S1: Collect the two-terminal current signal, voltage signal and switch status signal during the operation of the intelligent power plant relay protection system, and transmit the collected signals to the data processing module for signal analysis to extract the original feature data related to relay protection faults. Specifically, step S1 uses a device with multi-channel data acquisition capabilities to collect the dual-terminal current signal, voltage signal, and switch status signal during the operation of the smart power plant relay protection system. The sampling frequency for the current signal is set to 2000 Hz to 5000 Hz to ensure that transient signal characteristics during fault occurrence can be captured. The voltage signal acquisition range covers the 6 kV to 500 kV levels commonly used in smart power plants. The switch status signal is acquired using a digital acquisition method with a sampling interval of 1 millisecond to obtain the on / off status of the switch in real time. The acquired signals are transmitted to a dedicated data processing module via industrial Ethernet at a rate of no less than 100 megabits per second. The data processing module analyzes the received signals, first removing pulse interference and DC components, and then extracting original feature data related to faults such as short circuits, grounding, and overloads according to the needs of relay protection fault analysis. This includes information such as signal amplitude changes, phase shifts, frequency fluctuations, and switch state transition times. This provides complete and accurate basic data support for subsequent fault feature extraction. The implementation of this step ensures the authenticity and validity of the data used in subsequent analysis, avoiding deviations in subsequent fault analysis results due to incomplete data acquisition or analysis errors.

[0020] Step S2: Call the Fourier filter fault feature algorithm to extract fault features from the original feature data. By decomposing and filtering the frequency components in the original feature data, the feature components under the fault state are separated to obtain a preliminary fault feature set. Specifically, after obtaining the raw feature data output from step S1, step S2 calls a pre-built Fourier filter fault feature algorithm to extract fault features from the raw feature data. During implementation, the analysis frequency band of the Fourier filter is first determined. Combined with the fault signal characteristics of the smart power plant relay protection system, the analysis frequency band is set to 0 Hz to 1000 Hz to cover the fundamental, harmonic, and transient components that may be generated by the fault. The algorithm decomposes the frequency components of the raw feature data within the set frequency band, using a sliding window method to segment the data. The window length is set to 20 to 50 sampling points to ensure both accurate reflection of frequency component changes and processing efficiency. During the decomposition process, the amplitude and phase information of each frequency component are calculated. Then, based on the difference between fault features and normal operation features, amplitude and phase thresholds are set. The amplitude threshold is determined based on the maximum fluctuation range of the signal during normal operation of the power plant, typically 1.2 to 1.5 times the normal operation amplitude. The phase threshold is set to ±15 degrees. Frequency components exceeding the threshold range are filtered out; these components are the feature components under fault conditions. By integrating the selected feature components and removing redundant information, a preliminary fault feature set is obtained. This step can separate fault-related features from complex raw data, providing targeted data for subsequent fault diagnosis, improving the accuracy of fault diagnosis, and avoiding interference from normal operating data on the diagnostic results.

[0021] Step S3: Input the preliminary fault feature set into the distillation lightweight fault diagnosis model. The model performs feature mapping on the preliminary fault feature set through a pre-trained teacher network, and then compresses and learns the mapped features through a student network to output the preliminary judgment result of the fault type. Specifically, step S3 inputs the preliminary fault feature set obtained in step S2 into the pre-trained distillation lightweight fault diagnosis model. The number of input layer nodes in the model is determined according to the dimensionality of the preliminary fault feature set, typically set to 50 to 200 to adapt to feature data of different scales. During implementation, the teacher network in the model is first activated. The teacher network contains 3 to 5 hidden layers, with 200 to 500 neurons in each hidden layer. The ReLU activation function is used. The teacher network utilizes the parameters learned during pre-training to perform feature mapping on the preliminary fault feature set, converting low-dimensional preliminary features into high-dimensional abstract features. Batch normalization is used during the mapping process to accelerate network convergence and avoid overfitting. The student network is then activated. It has the same number of hidden layers as the teacher network, but each layer contains only 1 / 3 to 1 / 2 the number of neurons, achieving model lightweighting. The student network independently extracts features from the initial fault feature set and calculates the error between its own output features and the high-dimensional abstract features output by the teacher network. Gradient descent is used to adjust the weight parameters of the student network, with 100 to 300 iterations and a learning rate of 0.001 to 0.01, until the error is less than a preset threshold (usually 0.0001). After iteration, the student network outputs a preliminary fault type judgment, including the specific fault type (e.g., single-phase grounding fault, two-phase short-circuit fault) and the confidence level of the fault occurrence. This step utilizes distillation lightweighting technology to significantly reduce the computational load and number of parameters while ensuring diagnostic accuracy. This allows the model to run efficiently on embedded devices in smart power plants, meeting the real-time requirements of fault diagnosis. Simultaneously, the teacher network's guidance improves the diagnostic accuracy of the student network, avoiding performance loss due to lightweighting.

[0022] Step S4: Construct a dual-terminal current differential twin model. Input the current signals collected from both ends of the smart power plant relay protection system into two symmetrical sub-networks of the model. Simultaneously extract and compare the features of the current signals through the sub-networks, calculate the difference index of the dual-terminal current features, and obtain the fault location association data. Specifically, step S4 constructs a dual-terminal current differential twin model. The model contains two structurally identical sub-networks. Each sub-network consists of two convolutional layers, two pooling layers, and two fully connected layers. The convolutional kernel size of the convolutional layers is set to 3×3 with a stride of 1. The pooling layers use max pooling with a kernel size of 2×2. The number of neurons in the fully connected layers is 100 to 300. During implementation, the current signals (after being parsed in step S1) collected from end A and end B of the smart power plant are input into the two sub-networks respectively. The convolutional layers of the sub-networks extract features from the current signals, capturing local features in the signals (such as amplitude abrupt changes, phase jumps, etc.). The pooling layers reduce the dimensionality of the extracted local features, reducing the amount of data while retaining key information. The fully connected layers convert the dimensionality-reduced features into fixed-length feature vectors. Next, the difference index of the output feature vectors of the two sub-networks is calculated. During the calculation, the feature vectors are first normalized to eliminate the influence of dimensions, and then the Euclidean distance formula is used to calculate the distance between the vectors. This distance is compared with a preset reference distance (determined based on the distance between the double-ended current feature vectors during normal operation of the power plant) to obtain the difference index. Based on the magnitude of the difference index, combined with parameters such as the length and impedance of the smart power plant's transmission lines, fault location correlation data is generated. Specifically, the distance range between the fault location and ends A and B is determined through the mapping relationship between the difference index and line parameters, with the error controlled within 50 meters. This step, through synchronous analysis and comparison of the double-ended current signals, can accurately locate the fault location, solving the problem that traditional single-ended signal analysis is difficult to use to determine the fault location. At the same time, the symmetrical structure of the twin model ensures the consistency of double-ended data processing, improving the reliability of location positioning.

[0023] Step S5: After integrating the preliminary judgment result of the fault type with the fault location association data, the data is transmitted to the protection action logic verification platform. The platform performs logical matching and verification on the integrated data according to the preset protection logic rules of the smart power plant relay protection system, and generates protection action verification results. Specifically, step S5 integrates the preliminary fault type judgment result output in step S3 with the fault location association data obtained in step S4. During integration, a data concatenation method is used to combine fault type information (such as fault type code and confidence level) and fault location information (such as distance from point A and distance from point B) into a unified dataset in JSON format to ensure the protection action logic verification platform can parse it correctly. The integrated dataset is then transmitted to the protection action logic verification platform. The platform pre-stores preset protection logic rules for the smart power plant relay protection system. These rules cover protection action types (such as tripping and reclosing), action delays, and action priorities corresponding to different fault types and locations. The number of rules is determined based on the scale of the power plant, typically ranging from 50 to 200. During verification, the platform first matches the corresponding protection action logic rules from the rule base based on the fault type and fault location in the integrated data. Then, it compares the fault characteristic parameters (such as fault current amplitude and duration) in the integrated data with the trigger thresholds in the rules. The comparison process uses an item-by-item verification method to verify the trigger conditions of each rule. After verification, the number of rules that meet the trigger conditions is counted, the rule compliance rate is calculated, and the protection action verification result is generated based on the compliance rate. If the compliance rate is greater than or equal to 90%, the protection action logic is judged to meet the requirements. If it is less than 90%, the non-compliant rule entries and reasons are pointed out. The implementation of this step combines fault diagnosis results and location data for logic verification to ensure the rationality and pertinence of the protection action logic. This avoids misjudgment or omission caused by traditional verification based on a single parameter, and provides a scientific basis for the adjustment of subsequent protection actions.

[0024] Step S6: Based on the protection action verification results, combined with the feature components output by the Fourier filter fault feature algorithm, the learning parameters of the distillation lightweight fault diagnosis model, and the difference index of the dual-terminal current differential twin model, a complete relay protection fault analysis report is generated.

[0025] Specifically, after obtaining the protection action verification result output in step S5, step S6 collects the feature components (including amplitude and phase data of each frequency component) output by the Fourier filter fault feature algorithm in step S2, the learning parameters of the distillation lightweight fault diagnosis model in step S3 (such as the weight parameters and error values ​​of the student network), and the difference index (including the original difference and the corrected difference) of the two-terminal current differential twin model in step S4. During implementation, these multi-dimensional data are first classified and organized according to the logical structure of "fault features - diagnostic parameters - location index - verification result". The fault features section details the specific values ​​of each frequency component and the corresponding fault correlation degree; the diagnostic parameters section records the model iteration number, final error, fault type confidence, etc.; the location index section clearly defines the specific range and error range of the fault location; and the verification result section fully presents the rule compliance rate and details of rule non-compliance. Subsequently, following the standard format for relay protection fault analysis in smart power plants, the processed data is entered into a report template. The report template includes modules for basic fault information (such as fault occurrence time and data acquisition device number), data processing procedures, analysis results, and conclusions / recommendations. In the conclusions / recommendations module, based on the protection action verification results and data from each stage, specific directions for adjusting protection actions are proposed (such as adjusting the protection action delay for a certain type of fault or optimizing the protection logic threshold for a certain line). Finally, the report undergoes format checks and content review to ensure data accuracy and logical coherence, forming a complete relay protection fault analysis report. This step integrates the analysis data and results from each stage, providing power plant operation and maintenance personnel with comprehensive and intuitive fault analysis data. This facilitates quick understanding of the fault situation and the development of targeted handling measures. Furthermore, the standardized report format facilitates subsequent fault data archiving and traceability, providing data support for the long-term optimization of the smart power plant relay protection system.

[0026] Example 2 Preferably, the Fourier filter fault feature algorithm extracts fault feature components using the following expression: ,in, It is a natural constant. It is the imaginary unit. It is an angular frequency variable. This is the frequency domain expression for the fault characteristic components. The original feature data time-domain signal was collected. The signal angular frequency, This is the frequency response function of the Fourier filter. This is the upper limit of the filter frequency. As a time variable; at the same time, through Calculate the weighted fusion value of the feature components. These are the fused fault characteristic values. For the first angular frequency at each frequency point For the first The weighting coefficients for each frequency point The total number of frequency points. The value of the fault characteristic component in the frequency domain.

[0027] Specifically, in step S2, during Fourier filter fault feature extraction, the time-domain signal of the original feature data is first processed to determine the analysis time range, typically covering 0.5 to 2 seconds before and after the fault occurrence to fully capture the transient fault process. When performing the filtering operation, an upper limit for the filtering frequency is set. This upper limit is determined based on the highest frequency of common fault signals in smart power plant relay protection systems, generally between 500 Hz and 1000 Hz, ensuring coverage of harmonics and transient components generated by the fault. After separating fault feature components of different frequencies through filtering, these components need to be weighted and fused. The weight coefficients are set based on the correlation between each frequency component and the fault type. For example, the weight coefficient for frequency components strongly correlated with short-circuit faults is set to 0.6 to 0.8, and for those with weaker correlations, it is set to 0.2 to 0.4. The total number of frequency points is determined based on the frequency resolution after filtering, typically between 100 and 200. During the fusion calculation, the feature component of each frequency point is multiplied by its corresponding weight coefficient, summed, and then divided by the sum of all weight coefficients to obtain the fused fault feature value. By precisely setting the filtering parameters and weighting coefficients, fault features can be extracted more comprehensively and accurately, avoiding the limitations of single frequency component analysis. At the same time, fusion processing reduces the redundancy of feature data, providing more reliable feature basis for subsequent fault diagnosis and improving the accuracy of diagnostic results.

[0028] Preferably, the feature learning process of the distillation lightweight fault diagnosis model satisfies: ,in, The model distillation loss function is... For loss weighting coefficients, Let cross-entropy be the loss function. For fault type labels, For the output of the student network, The input is the preliminary fault feature set. Let the mean squared error loss function be . The output of the teacher network; lightweight compression of the model is achieved through... Perform parameter optimization. The compression ratio of the model parameters. For student network Layer weight matrix, For the number of student network layers, For Teachers' Network Layer weight matrix, For the number of teacher network layers, This represents the Frobenius norm.

[0029] Specifically, in step S3, during the operation of the lightweight fault diagnosis model, the weight coefficient of the model's distillation loss function is first set. This coefficient is determined based on the performance difference between the teacher network and the student network. If the diagnostic accuracy of the teacher network is more than 10% higher than that of the student network, the weight coefficient is set to 0.3 to 0.5; if the difference is small, it is set to 0.5 to 0.7 to balance the influence of cross-entropy loss and mean squared error loss. When calculating the cross-entropy loss function, fault type labels need to be obtained first. The labels are based on historical fault data of the smart power plant, covering common fault types such as single-phase grounding, two-phase short circuit, and three-phase short circuit, totaling 5 to 10 types. When calculating the output difference between the teacher network and the student network using the mean squared error loss function, the output result is the probability value of each fault type, with a value range of 0 to 1. During the model lightweight compression process, the number of layers in the student network remains consistent with that of the teacher network, both being 3 to 5 layers, but the number of neurons in each layer is reduced. For example, the teacher network has 400 to 600 neurons per layer, while the student network has 150 to 250 neurons per layer. The parameter compression ratio is calculated by summing the Frobenius norms of the weight matrices of each layer in both the student and teacher networks, and then dividing the sum of the norms of the student network by the sum of the norms of the teacher network. The compression ratio is typically controlled between 0.3 and 0.5. By reasonably setting the weights of the loss function and the scale of network parameters, the number of model parameters and computational load are significantly reduced while ensuring the model's diagnostic accuracy. This allows the model to adapt to the embedded hardware environment of smart power plants, meet the requirements of real-time fault diagnosis, and avoid computational delays caused by excessive model complexity.

[0030] Preferably, the expression for calculating the difference index of the two-terminal current differential twin model is as follows: ,in, This is an index for the difference in characteristics of two-terminal currents. For the feature mapping function of the Siamese model subnetwork, This refers to the current signal collected at terminal A of the smart power plant. This refers to the current signal collected at the B-end of the smart power plant. It is the L2 norm; at the same time, through Correct the difference index. The corrected fault location correlation index, This is the difference weighting coefficient. This is the weighting coefficient for the current difference. This represents the amplitude difference between the two-terminal current signals.

[0031] Specifically, in step S4, the operation of the two-terminal current differential twin model, firstly, a feature mapping function of the sub-network is constructed. The parameters of this function are determined through training on historical two-terminal current data from a smart power plant. The number of training samples is 1000 to 2000 sets, covering data from normal operation and different fault states. When calculating the two-terminal current feature difference index, the L2 norm of the feature vectors of the current signals at ends A and B after processing by the sub-network is calculated first. The norm result reflects the magnitude of the feature vector. Then, the L2 norm of the difference between the two feature vectors is calculated. Finally, the difference norm is divided by the maximum value of the norms of the two feature vectors to obtain the difference index. The index ranges from 0 to 1, where 0 indicates that the two-terminal features are completely identical, and 1 indicates that they are completely different. When correcting the difference index, the difference weight coefficient is set to 0.6 to 0.7, and the current difference weight coefficient is set to 0.3 to 0.4. The amplitude difference of the two-terminal current signals is obtained by subtracting the maximum amplitude of the current signals at ends A and B, and the units are converted to a unified current unit before calculation. The revised fault location correlation index still ranges from 0 to 1, with higher values ​​indicating a higher probability of fault. Through feature mapping functions and difference calculation rules, it can accurately capture the characteristic differences in the two-terminal current, reflecting the impact of the fault on the current signal. Simultaneously, a correction coefficient is introduced to balance the effects of difference and current difference, improving the accuracy of the fault location correlation index and providing a more reliable quantitative basis for subsequent fault location, thus reducing location errors.

[0032] Preferably, the logic matching verification of the protection action logic verification platform adopts the following method: ,in, For logical matching verification values, For the first The trigger threshold for the protection action logic, For the first The actual trigger value of the protection action logic, For the first The reliability coefficient of the protection action logic. To protect the total number of action logics; when When the protection action logic is deemed to meet the requirements, To protect the action logic verification threshold, and pass Calculated.

[0033] Specifically, in step S5, during the operation of the protection action logic verification platform, the total number of protection action logics is first determined. This number is determined based on the configuration of the smart power plant's relay protection system, typically ranging from 80 to 150, covering protection logics for different equipment and different fault types. The trigger threshold for each protection action logic is set according to the equipment's rated parameters and safe operation requirements. For example, the overcurrent protection trigger threshold for a certain line is set to 1.2 to 1.5 times the rated current, and the actual trigger value is calculated using real-time collected current data. The reliability coefficient is determined based on the historical operational accuracy of the protection logic. Logics with an accuracy rate higher than 95% have a coefficient of 0.9 to 1.0, while those with an accuracy rate lower than 90% have a coefficient of 0.7 to 0.8. When calculating the logic matching verification value, the trigger threshold, actual trigger value, and reliability coefficient of each logic are multiplied and summed to obtain the total verification value. The protection action logic verification threshold is obtained by multiplying the trigger threshold of each logic by the reliability coefficient, summing the results, and then dividing by the total number of logics. The threshold is typically 70% to 80% of the maximum total verification value. When the total verification value is greater than or equal to the threshold, the protection action logic is deemed to meet the requirements; otherwise, it is deemed not to meet the requirements. By setting quantitative verification indicators and thresholds, the verification of protection action logic becomes more objective and scientific, avoiding the subjective errors of traditional qualitative verification. At the same time, the reliability coefficient is combined to reflect the credibility of different logics, improving the reliability of the verification results and ensuring that the protection action can accurately respond to fault conditions.

[0034] Preferably, in the process of generating a complete relay protection fault analysis report, the following is adopted: Integrating multi-dimensional data, among which, This is the comprehensive evaluation value in the fault analysis report. These are the weighting coefficients for fault characteristic values. The weighting coefficients for the fault location-related indicators. For logical matching check value weighting coefficients, The fused fault feature values ​​are output by the Fourier filter fault feature algorithm. The fault location correlation index is the result of correction to the two-terminal current differential twin model. To protect the logic matching verification values ​​of the action logic verification platform; simultaneously, through Calculate the correlation coefficient for fault analysis. The correlation coefficient. For the first The actual trigger value of the protection action logic.

[0035] Specifically, in step S6, when generating a complete fault analysis report, three weighting coefficients for the comprehensive evaluation value are first set. The weighting coefficient for the fault feature value is determined based on the feature extraction accuracy of the Fourier filter algorithm; if the accuracy is higher than 90%, it is set to 0.4 to 0.5. The weighting coefficient for the fault location correlation index is determined based on the positioning error of the dual-end current differential twin model; if the error is less than 50 meters, it is set to 0.3 to 0.4. The weighting coefficient for the logic matching verification value is determined based on the verification accuracy of the verification platform; if the accuracy is higher than 95%, it is set to 0.2 to 0.3. The sum of the three coefficients is 1. When calculating the comprehensive evaluation value, the fault feature value, the fault location correlation index, and the logic matching verification value are multiplied by their respective weighting coefficients and then summed. The resulting evaluation value ranges from 0 to 100; a higher value indicates a more reliable fault analysis result. When calculating the fault analysis correlation coefficient, the sum of the actual trigger values ​​of all protection action logics is first calculated, and then the comprehensive evaluation value is divided by this sum. The correlation coefficient ranges from 0 to 1 and is used to reflect the degree of correlation between the fault analysis result and the actual protection action. By integrating multi-dimensional data with weights, the overall performance of fault analysis results can be comprehensively evaluated, avoiding the one-sidedness of evaluation based on a single indicator. At the same time, the correlation coefficient reflects the degree of matching between the analysis results and actual protection actions, providing a quantitative basis for operation and maintenance personnel to judge the effectiveness of fault analysis, and facilitating the subsequent optimization of the operating parameters of the relay protection system based on the analysis results.

[0036] Preferably, step S3 includes the following sub-steps: S31, Initialize the teacher network of the distillation lightweight fault diagnosis model, load the weight parameters and bias parameters obtained during pre-training, determine the activation function type and number of neurons in each hidden layer of the teacher network, and establish the feature mapping framework of the teacher network; S32, Divide the preliminary fault feature set obtained in step S2 into multiple feature subsets according to a preset batch size, and input each feature subset into the teacher network in sequence. Through the input layer and hidden layer of the teacher network, perform layer-by-layer feature transformation on the feature subsets to obtain high-dimensional feature vectors; S33, Initialize the student network of the distillation lightweight fault diagnosis model, set the number of layers and the number of neurons in each layer of the student network so that the parameter scale is smaller than that of the teacher network, and use the high-dimensional feature vector output by the teacher network as the reference feature of the student network; S34, The student network extracts features from the input preliminary fault feature set, and calculates the error between the output features of the student network and the reference features of the teacher network. Adjust the weight parameters of the student network according to the error, iterate repeatedly until the error meets the preset conditions, and output the preliminary judgment result of the fault type.

[0037] Specifically, step S3 includes four sub-steps: In S31, the teacher network of the distillation lightweight fault diagnosis model is initialized. The pre-trained weight parameters and bias parameters are obtained from the training results of a historical fault dataset from a smart power plant. The historical dataset contains 5000 to 8000 sets of sample data covering different fault types. The ReLU function is used as the activation function for each hidden layer of the teacher network. The number of neurons in the input layer is set to 80 to 150 based on the dimension of the initial fault feature set. The hidden layers are set to four layers, with 500, 400, 300, and 200 neurons in each layer, respectively, thus establishing a feature mapping framework. In S32, the initial fault feature set obtained in step S2 is divided into subsets according to a preset batch size, ranging from 32 to 64. Each subset is input into the teacher network. After being received by the input layer, the hidden layers perform feature transformation layer by layer from the first to the fourth layer. After each transformation, batch normalization is used to stabilize the data distribution, finally outputting a high-dimensional feature vector with a dimension of 200. In step S33, the student network is initialized with the same four layers as the teacher network, but the number of neurons in each layer is reduced, successively set to 200, 150, 100, and 80. The 200-dimensional high-dimensional feature vector output by the teacher network is directly used as the reference feature of the student network. In step S34, the student network receives the preliminary fault feature set and extracts features. At the same time, it calculates the error between its own 80-dimensional output features and the 200-dimensional reference features of the teacher network, using the mean squared error calculation method. Based on the error, the weight parameters of the student network are adjusted using the gradient descent method, with a learning rate of 0.005 and 200 iterations, until the error is less than 0.0001. Finally, the network outputs a preliminary fault type judgment result containing the fault type and corresponding confidence level. By optimizing the network parameters and training process step by step, the model is lightweight while ensuring diagnostic accuracy, meeting the real-time diagnostic needs of smart power plants.

[0038] Preferably, step S4 includes the following sub-steps: S41. Construct two symmetrical sub-networks of the dual-end current differential twin model. The two sub-networks have the same structure, each containing an input layer, a convolutional layer, a pooling layer, and a fully connected layer. Set the kernel size, stride, and number of output channels for each layer. S42. Preprocess the current signals collected from end A and end B of the smart power plant and input them into the corresponding sub-networks. Extract features from the current signals through the convolutional layer of the sub-network, reduce the dimensionality of the extracted features through the pooling layer, and convert the dimensionality-reduced features into feature vectors through the fully connected layer. S43. Calculate the Euclidean distance between the output feature vectors of the two sub-networks. Use the Euclidean distance as the initial difference of the dual-end current features, and introduce a time synchronization coefficient to correct the initial difference. S44. Based on the corrected difference, combined with the line parameters and topology of the smart power plant relay protection system, generate fault location association data to clarify the possible area range of the fault.

[0039] Specifically, step S4 includes four sub-steps: When executing S41, two symmetrical sub-networks of the two-terminal current differential twin model are constructed. Each sub-network contains two convolutional layers, two pooling layers, and two fully connected layers. The first convolutional layer has a kernel size of 3×3, 64 output channels, and a stride of 1. The second convolutional layer has a kernel size of 3×3, 128 output channels, and a stride of 1. Both pooling layers use max pooling with a kernel size of 2×2 and a stride of 2. The first fully connected layer has 256 neurons, and the second fully connected layer has 128 neurons. In step S42, the current signals collected from terminals A and B of the smart power plant are parsed in step S1 and normalized to ensure their values ​​are between 0 and 1. These normalized signals are then input into two sub-networks. Convolutional layers extract transient and steady-state features from the current signals through convolution operations. Pooling layers reduce the dimensionality of the convolutional features to decrease computation. Fully connected layers convert the dimensionality-reduced features into 128-dimensional feature vectors. In step S43, the Euclidean distance between the 128-dimensional feature vectors output by the two sub-networks is calculated as the initial difference. A time synchronization coefficient is also introduced, determined based on the time difference between signal acquisition from terminals A and B. The coefficient is set to 0.95 when the time difference is less than 10 milliseconds, 0.9 when it is between 10 and 20 milliseconds, and 0.85 when it is greater than 20 milliseconds. The corrected difference is obtained by multiplying the initial difference by the time synchronization coefficient. When executing S44, the transmission line parameters of the smart power plant are combined, with the line length set to 50 km to 200 km and the line impedance set to 0.01 ohms / km to 0.05 ohms / km. By mapping the corrected difference degree with the line parameters, the distance between the fault occurrence point and end A and end B is calculated, and the distance error is controlled within 30 meters. Fault location association data containing the fault distance range is generated. By optimizing the sub-network structure and difference degree calculation step by step, the accuracy and reliability of fault location are improved.

[0040] Preferably, step S5 includes the following sub-steps: S51, unifying the data format of the preliminary fault type judgment result output in step S3 and the fault location association data obtained in step S4, converting them into a feature matrix form recognizable by the protection action logic verification platform; S52, retrieving the preset protection logic rules corresponding to the smart power plant relay protection system from the database of the protection action logic verification platform. These rules include protection action thresholds, action delays, and linkage switch information under different fault types; S53, matching the unified format feature matrix with the preset protection logic rules row by row, calculating the deviation value between each element in the feature matrix and the corresponding rule threshold, and counting the number of elements with deviation values ​​within the allowable range; S54, scoring the compliance of the protection action according to the preset scoring criteria based on the deviation value statistics, generating the protection action verification result, and clarifying whether the protection action conforms to the preset logic rules.

[0041] Specifically, step S5 includes four sub-steps: When executing S51, the preliminary fault type judgment result output from step S3, including fault type codes (eight codes corresponding to different faults) and confidence levels (ranging from 0 to 1), along with the fault location association data obtained in step S4, including the distance from fault to end A and the distance to end B, are uniformly converted into a feature matrix. The matrix has 10 rows, each corresponding to a data indicator, and one column, ensuring that the protection action logic verification platform can recognize it. When executing S52, preset protection logic rules are retrieved from the protection action logic verification platform database. The number of rules is set to 120 to 180. Each rule includes information such as the fault type adaptation range, protection action type (tripping, reclosing, etc.), action delay (0.05 seconds to 0.5 seconds), and linkage switch number. The rules are formulated according to the design standards for intelligent power plant relay protection systems. When executing S53, the feature matrix is ​​matched row by row with preset rules. For example, the fault type coding row in the matrix is ​​matched with the fault type adaptation range in the rules. The deviation value of each element in the feature matrix from the corresponding rule threshold is calculated. For example, the action delay deviation value is the difference between the delay data in the matrix and the delay threshold in the rules. The number of elements with deviation values ​​within the allowable range (absolute deviation value less than 0.02 seconds) is counted. When executing S54, according to the preset scoring standard, each element that meets the requirements is scored 1 point, with a total score of 10 points. The protection action verification result is generated based on the total score. A score of 8 or above is considered compliant, and a score below 8 indicates the specific item that does not meet the rules, such as a mismatch in the linkage switch number in a certain rule. By standardizing the data format and verification process step by step, the rigor and accuracy of the protection action logic verification are improved.

[0042] The Fourier filter fault feature algorithm is the core technology in this invention used to extract fault features from the raw signals of a smart power plant relay protection system. It is an algorithm based on Fourier analysis combined with filtering to separate frequency components in a signal and filter out feature components under fault conditions. The implementation process is as follows: First, the analysis frequency band (0 Hz to 1000 Hz, covering the fault fundamental, harmonics, and transient components) is determined. A sliding window (20 to 50 sampling points) is used to segment the raw feature data analyzed in step S1, decomposing each frequency component and calculating its amplitude and phase. Then, an amplitude threshold (1.2 to 1.5 times the normal operating amplitude) and a phase threshold (±15 degrees) are set to filter out frequency components exceeding the thresholds. Finally, weights are assigned to the filtered feature components according to their correlation (0.6 to 0.8 for strong correlations and 0.2 to 0.4 for weak correlations), and a preliminary fault feature set is obtained through weighted fusion. The function of this algorithm is to extract fault features from complex raw data and remove interference from normal operating data. This provides accurate and targeted feature data for subsequent fault diagnosis, avoids the one-sidedness of single signal analysis, lays the foundation for the accuracy of fault analysis, and solves the problem that traditional algorithms cannot fully extract fault features.

[0043] The distillation-based lightweight fault diagnosis model is used to diagnose fault features and output fault types, combining diagnostic accuracy with lightweight characteristics. It is an intelligent diagnostic model based on knowledge distillation technology, where a teacher network guides a student network's learning while compressing model parameters. Its implementation process is as follows: In step S3, the teacher network is first initialized (3 to 5 hidden layers, 200 to 500 neurons per layer, ReLU activation function), and pre-trained parameters are loaded to perform feature mapping on the initial fault feature set, outputting high-dimensional abstract features. Then, the student network is initialized (same number of layers as the teacher network, with 1 / 3 to 1 / 2 the number of neurons), using the teacher network output as a reference, independently extracting features and calculating the error with the teacher network output. The student network parameters are adjusted using gradient descent (learning rate 0.001 to 0.01, iterations 100 to 300) until the error is less than 0.0001, finally outputting the fault type and confidence level. The model's purpose is to reduce the computational load and number of parameters while maintaining diagnostic accuracy. It adapts to the real-time operation requirements of embedded equipment in smart power plants, avoids the problem of computational delay in traditional complex models, improves the accuracy of student networks through online teacher guidance, achieves "high precision + high efficiency" fault diagnosis, and provides support for rapid fault response.

[0044] The two-terminal current differential twin model is used to locate faults in the relay protection system of smart power plants. It achieves accurate location based on the comparison of two-terminal current signals. This model is a twin model containing two structurally identical subnetworks. By analyzing the differences in the characteristics of the two-terminal current signals, it calculates the correlation data of the fault location. The implementation process is as follows: In step S4, a sub-network is first constructed (2 convolutional layers, 2 pooling layers, 2 fully connected layers, 3×3 convolutional kernels with a stride of 1, 2×2 pooling kernels, and 100 to 300 neurons in the fully connected layers); the current signals parsed from ends A and B are input into the sub-network respectively, and local features are extracted through convolution, dimensionality reduction through pooling, and converted into feature vectors through fully connected layers; the L2 norm difference of the vectors is calculated, and divided by the maximum value to obtain the difference index (0 to 1, the larger the value, the greater the difference); then the index is corrected using the difference weight coefficient (0.6 to 0.7) and the current difference weight coefficient (0.3 to 0.4), combined with the line length (50 to 200 kilometers) and impedance (0.01 to 0.05 ohms / kilometer), the fault distance range (error ≤ 30 meters) is generated. The function of this model is to accurately locate the fault location and provide fault location correlation data. To address the issue of large positioning errors in traditional single-ended signal positioning, a symmetrical sub-network is used to ensure consistency in data processing between both ends, thereby improving positioning reliability. This provides maintenance personnel with accurate information for quickly diagnosing faults and shortens troubleshooting time.

[0045] The protection action logic verification platform is used to verify the rationality of relay protection action logic in smart power plants, and performs quantitative verification based on multi-dimensional data. It is a platform that stores preset protection rules, receives fault data, and performs logic matching verification. Its implementation process is as follows: In step S5, the fault type result from step S3 and the location data from step S4 are first received and concatenated into an integrated dataset in JSON format; preset rules (80 to 150 rules, including fault type adaptation range, action type, delay of 0.05 to 0.5 seconds, and linkage switch number) are retrieved from the database; the integrated data is verified item by item according to the rules, and the sum of the products of the trigger threshold (e.g., 1.2 to 1.5 times the rated current for overcurrent protection), the actual trigger value, and the reliability coefficient (0.7 to 1.0, set according to historical accuracy) is calculated to obtain the verification value; then the verification threshold (70% to 80% of the sum of the rule products) is calculated. If the verification value ≥ the threshold, it is determined to meet the requirements. The function of this platform is to quantitatively verify the protection action logic, generate verification results, and indicate items that do not conform to the rules. To avoid the subjective errors of traditional qualitative verification, the reliability of the results is improved by multi-data fusion verification, ensuring that the protection action can accurately respond to the fault, preventing false or missed actions, and ensuring the safe and stable operation of the relay protection system of smart power plants.

[0046] Example 3 like Figure 2As shown, a fault analysis method for intelligent power plant relay protection based on intelligent algorithms is implemented through different units, including: an intelligent power plant relay protection signal acquisition and parsing unit, used to acquire the dual-terminal current, voltage, and switch status signals during the operation of the intelligent power plant relay protection system, and to parse and extract fault-related original feature data; a Fourier filter fault feature extraction unit, used to call the Fourier filter fault feature algorithm to decompose and filter the frequency components of the original feature data, and separate the fault feature components to obtain a preliminary fault feature set; a distillation lightweight fault diagnosis calculation unit, used to receive the preliminary fault feature set, and output the preliminary fault type judgment result through the teacher network and student network calculation of the distillation lightweight fault diagnosis model; a dual-terminal current differential twin model construction and calculation unit, used to construct a dual-terminal current differential twin model, input dual-terminal current signals and calculate the feature difference index to obtain fault location association data; and a protection action logic verification and evaluation unit, used to receive the preliminary fault type judgment result and fault location association data, and, based on the preset protection logic... The system includes a rule verification and generation unit for protection action verification results; a fault analysis report generation and storage unit, which combines protection action verification results and multi-algorithm output parameters to generate and store a complete fault analysis report, and performs data backtracking; the output of the smart power plant relay protection signal acquisition and analysis unit is connected to the input of the Fourier filter fault feature extraction unit, the output of the Fourier filter fault feature extraction unit is connected to the input of the distillation lightweight fault diagnosis calculation unit, the output of the distillation lightweight fault diagnosis calculation unit is connected to the input of the dual-terminal current differential twin model construction and calculation unit and the first input of the protection action logic verification and evaluation unit, the output of the dual-terminal current differential twin model construction and calculation unit is connected to the second input of the protection action logic verification and evaluation unit, the output of the protection action logic verification and evaluation unit is connected to the input of the fault analysis report generation and storage unit, and the fault analysis report generation and storage unit also establishes a data interaction channel with the smart power plant relay protection signal acquisition and analysis unit.

[0047] This paper presents a fault analysis method for relay protection in smart power plants based on intelligent algorithms. The method constructs a complete analysis process covering signal processing, feature extraction, fault diagnosis, location positioning, logic verification, and report generation. It can systematically process various types of data during the operation of the relay protection system, avoiding the limitations of single-stage processing. Furthermore, by integrating multiple intelligent technologies, it achieves comprehensive fault feature extraction, efficient diagnostic models, accurate location positioning, and rigorous logic verification. This method can not only separate key fault features from complex signals and quickly output diagnostic results, but also accurately locate the fault area and verify the rationality of protection actions, thus comprehensively improving the overall performance of relay protection fault analysis and meeting the high requirements of smart power plants for fault handling.

[0048] This method addresses the problem of insufficient efficiency in fault feature processing and diagnosis. Through the synergistic effect of multiple technologies, it utilizes specialized algorithms to comprehensively separate fault feature components, avoiding the one-sidedness of single-algorithm processing and providing a reliable data foundation for diagnosis. Furthermore, it optimizes the diagnostic model to be lightweight, significantly accelerating computation and meeting the timeliness requirements of fault analysis. Addressing the poor coordination between fault location determination and protection action logic verification, it relies on a specialized model to perform in-depth analysis and precise comparison of dual-terminal current data, improving the accuracy of fault location. Simultaneously, it combines fault diagnosis results, location data, and protection action logic verification to achieve multi-dimensional comprehensive verification, enhancing the reliability of verification results and solving the problems of isolated verification and low reliability in traditional methods.

[0049] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fault analysis method for relay protection in smart power plants based on intelligent algorithms, characterized in that, include: Step S1: Collect the two-terminal current signal, voltage signal and switch status signal during the operation of the intelligent power plant relay protection system, and transmit the collected signals to the data processing module for signal analysis to extract the original feature data related to relay protection faults. Step S2: Call the Fourier filter fault feature algorithm to extract fault features from the original feature data. By decomposing and filtering the frequency components in the original feature data, the feature components under the fault state are separated to obtain a preliminary fault feature set. Step S3: Input the preliminary fault feature set into the distillation lightweight fault diagnosis model. The model performs feature mapping on the preliminary fault feature set through a pre-trained teacher network, and then compresses and learns the mapped features through a student network to output the preliminary judgment result of the fault type. Step S4: Construct a dual-terminal current differential twin model. Input the current signals collected from both ends of the smart power plant relay protection system into two symmetrical sub-networks of the model. Simultaneously extract and compare the features of the current signals through the sub-networks, calculate the difference index of the dual-terminal current features, and obtain the fault location association data. Step S5: After integrating the preliminary judgment result of the fault type with the fault location association data, the data is transmitted to the protection action logic verification platform. The platform performs logical matching and verification on the integrated data according to the preset protection logic rules of the smart power plant relay protection system, and generates protection action verification results. Step S6: Based on the protection action verification results, combined with the feature components output by the Fourier filter fault feature algorithm, the learning parameters of the distillation lightweight fault diagnosis model, and the difference index of the dual-terminal current differential twin model, a complete relay protection fault analysis report is generated.

2. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to claim 1, characterized in that, The Fourier filter fault feature algorithm extracts fault feature components using the following expression: ,in, It is a natural constant. It is the imaginary unit. It is an angular frequency variable. This is the frequency domain expression for the fault characteristic components. The original feature data time-domain signal was collected. The signal angular frequency, This is the frequency response function of the Fourier filter. This is the upper limit of the filter frequency. As a time variable; at the same time, through Calculate the weighted fusion value of the feature components. These are the fused fault characteristic values. For the first angular frequency at each frequency point For the first The weighting coefficients for each frequency point The total number of frequency points. The value of the fault characteristic component in the frequency domain.

3. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to claim 1, characterized in that, The feature learning process of the distillation lightweight fault diagnosis model satisfies: ,in, The model distillation loss function is... For loss weighting coefficients, Let cross-entropy be the loss function. For fault type labels, For the output of the student network, The input is the preliminary fault feature set. Let the mean squared error loss function be . The output of the teacher network; lightweight compression of the model is achieved through... Perform parameter optimization. The compression ratio of the model parameters. For student network Layer weight matrix, For the number of student network layers, For Teachers' Network Layer weight matrix, For the number of teacher network layers, This represents the Frobenius norm.

4. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to claim 1, characterized in that, The expression for calculating the difference index of the two-terminal current differential twin model is as follows: ,in, This is an index for the difference in characteristics of two-terminal currents. For the feature mapping function of the Siamese model subnetwork, This refers to the current signal collected at terminal A of the smart power plant. This refers to the current signal collected at the B-end of the smart power plant. It is the L2 norm; at the same time, through Correct the difference index. The corrected fault location correlation index, This is the difference weighting coefficient. This is the weighting coefficient for the current difference. This represents the amplitude difference between the two-terminal current signals.

5. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to claim 1, characterized in that, The logic matching verification of the protection action logic verification platform adopts the following method: ,in, For logical matching verification values, For the first The trigger threshold for the protection action logic, For the first The actual trigger value of the protection action logic, For the first The reliability coefficient of the protection action logic. To protect the total number of action logics; when When the protection action logic is deemed to meet the requirements, To protect the action logic verification threshold, and pass Calculated.

6. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to claim 1, characterized in that, In the process of generating a complete relay protection fault analysis report, the following methods are adopted: Integrating multi-dimensional data, among which, This is the comprehensive evaluation value in the fault analysis report. Here, represents the weighting coefficient for fault characteristic values, and represents the weighting coefficient for fault location correlation indicators. For logical matching check value weighting coefficients, The fused fault feature values ​​are output by the Fourier filter fault feature algorithm. The fault location correlation index is the result of correction to the two-terminal current differential twin model. To protect the logic matching verification values ​​of the action logic verification platform; simultaneously, through Calculate the correlation coefficient for fault analysis. The correlation coefficient. For the first The actual trigger value of the protection action logic.

7. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to claim 1, characterized in that, Step S3 includes the following sub-steps: S31. Initialize the teacher network of the distillation lightweight fault diagnosis model, load the weight parameters and bias parameters obtained during the pre-training process, determine the activation function type and number of neurons in each hidden layer of the teacher network, and establish the feature mapping framework of the teacher network. S32. Divide the preliminary fault feature set obtained in step S2 into multiple feature subsets according to the preset batch size, and input each feature subset into the teacher network in turn. Through the input layer and hidden layer of the teacher network, the feature subsets are transformed layer by layer to obtain high-dimensional feature vectors. S33. Initialize the student network of the distillation lightweight fault diagnosis model, set the number of layers and the number of neurons in each layer of the student network, so that the parameter size is smaller than that of the teacher network, and use the high-dimensional feature vector output by the teacher network as the reference feature of the student network. S34. The student network extracts features from the input preliminary fault feature set, calculates the error between the student network output features and the teacher network reference features, adjusts the weight parameters of the student network according to the error, iterates repeatedly until the error meets the preset conditions, and outputs the preliminary judgment result of the fault type.

8. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to claim 1, characterized in that, Step S4 includes the following sub-steps: S41. Construct two symmetric subnetworks for the dual-terminal current differential twin model. The two subnetworks have the same structure, each containing an input layer, a convolutional layer, a pooling layer, and a fully connected layer. Set the kernel size, stride, and number of output channels for each layer. S42. The current signals collected from the A and B ends of the smart power plant are preprocessed and then input into the corresponding sub-networks. The convolutional layer of the sub-network extracts features from the current signals, the pooling layer reduces the dimensionality of the extracted features, and the fully connected layer converts the dimensionality-reduced features into feature vectors. S43. Calculate the Euclidean distance between the output feature vectors of the two sub-networks, and use the Euclidean distance as the initial difference of the two-terminal current features. At the same time, introduce a time synchronization coefficient to correct the initial difference. S44. Based on the corrected difference degree, combined with the line parameters and topology of the smart power plant relay protection system, generate fault location association data to clarify the possible area range of the fault.

9. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to claim 1, characterized in that, Step S5 It includes the following steps: S51. Unify the data format of the preliminary fault type judgment result output in step S3 and the fault location association data obtained in step S4, and convert them into a feature matrix form that can be recognized by the protection action logic verification platform. S52. Retrieve the preset protection logic rules corresponding to the smart power plant relay protection system from the database of the protection action logic verification platform. The rules include protection action thresholds, action delays and linkage switch information under different fault types. S53. Match the unified format feature matrix with the preset protection logic rules line by line, calculate the deviation value between each element in the feature matrix and the corresponding rule threshold, and count the number of elements with deviation values ​​within the allowable range. S54. Based on the deviation value statistics, score the compliance of the protection actions according to the preset scoring criteria, generate the protection action verification results, and clarify whether the protection actions comply with the preset logical rules.

10. The fault analysis method for intelligent power plant relay protection based on intelligent algorithms according to any one of claims 1-9, characterized in that, This method is implemented through different units, including: a smart power plant relay protection signal acquisition and analysis unit, used to acquire the dual-terminal current, voltage, and switch status signals during the operation of the smart power plant relay protection system, and analyze and extract fault-related original feature data; a Fourier filter fault feature extraction unit, used to call the Fourier filter fault feature algorithm to decompose and filter the frequency components of the original feature data, and separate the fault feature components to obtain a preliminary fault feature set; a distillation lightweight fault diagnosis calculation unit, used to receive the preliminary fault feature set, and output the preliminary fault type judgment result through the teacher network and student network calculation of the distillation lightweight fault diagnosis model; a dual-terminal current differential twin model construction and calculation unit, used to construct a dual-terminal current differential twin model, input dual-terminal current signals and calculate the feature difference index to obtain fault location association data; a protection action logic verification and evaluation unit, used to receive the preliminary fault type judgment result and fault location association data, verify and generate protection action verification results according to preset protection logic rules; and a fault analysis report generation and storage unit, used to combine the protection action verification results and multi-algorithm output parameters to generate a complete fault analysis report and store it, while performing data backtracking retrieval.

Images