A smart substation secondary equipment fault positioning method and system

Abstract

A kind of intelligent substation secondary equipment fault positioning method and system, method includes the following steps: collecting original message information when secondary equipment fails and parsing, constructs secondary equipment fault dataset;Training sample in secondary equipment fault dataset is inputed into the generator and discriminator pre-constructed in turn, and the abnormal score of training sample is output;Normal sample is collected and inputed into discriminator, and the abnormal score of normal sample is obtained, reference abnormal threshold is determined and is revised;When abnormal score is greater than the threshold after revision, it is judged as fault;The loss of training sample is calculated and generator and discriminator are trained;Real-time input sample is collected, and inputed into trained model, when secondary equipment is judged as fault, the fault position of secondary equipment is positioned.The present application realizes the accurate detection and accurate positioning of intelligent substation secondary equipment fault by fusing multi-source operation data and constructing adaptive anomaly detection and feature contribution analysis mechanism.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology for intelligent substations in power systems, and specifically relates to a method and system for locating faults in secondary equipment of intelligent substations. Background Technology

[0002] With the intelligent upgrading of my country's power industry, the deep integration of digital technology, communication technology, and automation control technology is driving the power system towards high integration and intelligence. As a core component of the new power system, smart substations, while achieving network-based transformation of secondary equipment, also face the challenge of a surge in massive operational data. Currently, the causes of various secondary equipment failures in smart substations are mainly determined by technicians based on their fault-handling experience and device message information. However, due to the complexity of the secondary system network and the lack of effective means to analyze the correlation between fault characteristics, it is difficult to quickly locate and analyze faults. Therefore, accurate location of secondary equipment faults is crucial. Traditional methods are insufficient for intelligent diagnosis and fault tracing of equipment status in complex network environments. To improve the accuracy and timeliness of fault location in smart substation secondary equipment, deep learning algorithms, with their significant advantages in multivariate data correlation analysis, can achieve accurate identification and intelligent judgment of abnormal states of secondary equipment. This enables the construction of a deep learning-based secondary equipment fault diagnosis and precise handling system.

[0003] Currently, existing methods for locating secondary equipment faults include: Existing technology 1 (Cao Haiou, Wu Di, Xue Fei, et al. Fault location method for secondary equipment in smart substations based on improved BA-PNN [J]. Smart Power, 2024, 52(04):32-39) addresses the problems of high training complexity and strong sensitivity to smoothing factors in the fault location of probabilistic neural networks (PNN) in smart substations. In the summation layer of PNN, the Laplace distribution is used instead of the Gaussian distribution, and the BA algorithm is used to obtain the optimal smoothing factor. Thus, an improved bat algorithm is proposed to optimize the probabilistic neural network (BA-PNN). Secondly, based on the feature analysis of secondary equipment in smart substations, fault feature quantities are selected and mapped to them, which significantly improves the accuracy and efficiency of fault location. Patent application CN119691619A discloses a method and system for secondary system fault location based on a dual-channel convolutional neural network. The method obtains historical fault data by dividing the secondary system into intervals, trains a dual-channel convolutional neural network model, monitors and preprocesses secondary system alarm signals in real time, and inputs the preprocessed alarm signals into the model to obtain fault location results. This achieves secondary system fault location without relying on the physical port information of secondary equipment, and the model has high location accuracy and robustness. Patent application CN119557740A discloses a method for substation secondary equipment fault location based on a spatiotemporal graph convolutional network model. It extracts and constructs a fault feature set based on historical fault data of substation secondary equipment, trains an STGCN model after feature selection and multimodal feature fusion, and combines dynamic thresholds based on adaptive weather conditions and fault inference rules. The feature data is input into the trained model to complete fault location. This method can simultaneously learn the spatiotemporal characteristics of the data, improving the accuracy and efficiency of substation secondary equipment fault diagnosis and solving the problem of excessive reliance on maintenance personnel experience in fault diagnosis.

[0004] In the field of secondary equipment fault location in smart substations, existing methods typically process self-test information, SV / GOOSE messages, and electrical quantities independently, which cannot effectively identify composite faults caused by the superposition of multiple data anomalies, resulting in missed detection of composite faults. In addition, existing methods focus more on a single device or signal (smart terminal self-test), ignoring the topological correlation of secondary circuits, and cannot analyze the chain reaction between virtual terminal logic errors, fiber optic interruptions, and equipment anomalies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for fault location of secondary equipment in intelligent substations. The method includes the following steps: collecting and parsing the original message information when a secondary equipment fault occurs to construct a secondary equipment fault dataset; sequentially inputting training samples from the secondary equipment fault dataset into a pre-built generator and discriminator, outputting anomaly scores for the training samples; collecting normal samples and inputting them into the discriminator to obtain anomaly scores for the normal samples, determining and correcting a baseline anomaly threshold; determining a fault when the anomaly score exceeds the corrected threshold; calculating the loss of the training samples and training the generator and discriminator; collecting real-time input samples and inputting them into the trained model; and locating the fault location of the secondary equipment when it is determined to be faulty. This invention achieves accurate detection and precise location of secondary equipment faults in intelligent substations by fusing multi-source operational data and constructing an adaptive anomaly detection and feature contribution analysis mechanism.

[0006] The present invention adopts the following technical solution: The first aspect of the present invention provides a method for fault location of secondary equipment in an intelligent substation, comprising the following steps: Collect and parse the original message information when secondary equipment fails, extract current and voltage sampling sequences and switch status; process the extracted data, filter out fault feature information and determine the corresponding fault equipment type, and construct a secondary equipment fault dataset; The training samples from the secondary equipment fault dataset are input into a pre-built generator to generate a fault distribution matrix. The fault distribution matrix is ​​then input into a pre-built discriminator to output anomaly scores for the training samples. Normal samples are collected and input into the discriminator to obtain anomaly scores for the normal samples, and a baseline anomaly threshold is determined based on these scores. The average load rate and time-series volatility within the corresponding time window of the training samples are calculated to determine the load level factor and operational volatility intensity factor, respectively, and the baseline anomaly threshold is corrected to obtain an adaptive anomaly threshold. When the anomaly score of a training sample is greater than the adaptive anomaly threshold, the discriminator determines it as a fault. Based on the discriminator's determination result and the fault distribution matrix, the loss for each training sample is calculated, and the generator and discriminator are trained. The system collects and parses real-time message information from the secondary equipment in the substation, generates real-time input samples, and inputs them into a trained generator and discriminator to determine whether the secondary equipment is faulty. When a secondary equipment fault occurs, the system calculates the feature anomaly contribution of different features to the anomaly score. Based on the physical relationship between the features and the equipment and links, and combined with the feature anomaly contribution, the system calculates the fault contribution of the equipment and locates the fault position of the secondary equipment.

[0007] Preferably, the process of constructing the secondary equipment fault dataset is as follows: Each secondary equipment failure is collected as a training sample, and the failure type in each secondary equipment failure is used as the label of the corresponding training sample. For each training sample, the secondary equipment operating status data, SV / GOOSE receiving status data and sampled values ​​at each moment within a predetermined time window before and after the secondary equipment failure are selected as the failure feature data of the training sample.

[0008] Preferably, the process of determining the baseline anomaly threshold is as follows: Normal samples from different operating scenarios of smart substations are collected and input into a pre-built discriminator to obtain anomaly scores for normal samples. For normal samples in each scenario, the inverse function value of the cumulative probability function of kernel density estimation at the predefined normal probability is calculated as the benchmark anomaly threshold.

[0009] Preferably, the process of determining the load level factor is as follows: The mean value of the main transformer load rate of the substation within the corresponding time window of the training samples is obtained as the corresponding average load rate. Based on the predetermined linear coefficient and quadratic coefficient of the low load area, the average load rate is calculated by a quadratic binomial to obtain the low load growth term. The average load is subtracted from the predefined heavy load threshold and multiplied by the predefined exponential weight to obtain the exponent of the exponential function with the natural constant as the base, thus obtaining the high load growth term. The low load growth term and the low load growth term are weighted and normalized to obtain the load level factor.

[0010] Preferably, the process of determining the operating volatility intensity factor is as follows: The coefficients of variation of different sampled values ​​in the training samples within the corresponding time window are calculated and averaged to obtain the current coefficient of variation and the voltage coefficient of variation. The coefficient of variation of the GOOSE message frequency within the corresponding time window of the training samples is calculated as the message frequency coefficient of variation. Based on the predefined fluctuation weight coefficient, the current coefficient of variation, the voltage coefficient of variation, and the message frequency coefficient of variation are weighted to obtain the time series volatility. The time series volatility is set with a segmented volatility intensity factor by using a predefined segmented function.

[0011] Preferably, the process of correcting the baseline anomaly threshold is as follows: Based on the real-time operation scenario of the substation, predefine the corresponding scenario correction factor; The baseline anomaly threshold is multiplied by the difference between the predefined load level threshold and the load level factor, and then multiplied by the fluctuation intensity factor and the scenario correction factor in turn to obtain the adaptive anomaly threshold, which is then limited to the predefined upper and lower limits.

[0012] Preferably, the process of calculating the feature anomaly contribution of different features to the anomaly score is as follows: The real-time input samples are sequentially input into the trained generator and discriminator to obtain the corresponding anomaly scores. When the anomaly score is greater than the corresponding adaptive anomaly threshold, the secondary device malfunctions. The partial derivatives of the anomaly score output by the discriminator with respect to each feature of the real-time input sample at different times are calculated respectively, which are used as the feature anomaly sensitivity of the corresponding feature to the anomaly score at the corresponding time, and a feature anomaly sensitivity matrix is ​​generated. The feature anomaly sensitivity matrix is ​​normalized to obtain the standard feature anomaly sensitivity matrix. For each feature, the standard feature anomaly sensitivity of the corresponding feature at different times is accumulated as the numerator, and the standard feature anomaly sensitivity of all features at all times is accumulated as the denominator to calculate the feature anomaly contribution of the corresponding feature.

[0013] Preferably, the process of locating the fault location of the secondary equipment is as follows: Based on the physical association between features and devices and links, features corresponding to different secondary devices and different links are divided, and feature sets for each secondary device and each link are generated respectively. For each secondary device and link, the anomaly contribution of all features in the corresponding feature set is accumulated and averaged to obtain the corresponding fault contribution. The features are sorted in descending order of their contribution to anomaly, and the first predetermined number of features are selected as key features. If the key features include links, the fault contribution of all links is compared, and the link with the highest fault contribution is identified as the link fault location. If the key features include devices, the fault contribution of all devices is compared, and the device with the highest fault contribution is identified as the secondary device fault location.

[0014] A second aspect of the present invention provides a fault location system for secondary equipment in an intelligent substation, and a fault location method for secondary equipment in an intelligent substation, comprising: The fault dataset construction module collects and parses the original message information when secondary equipment fails, extracts current and voltage sampling sequences and switch status; processes the extracted data, filters out fault feature information and determines the corresponding fault equipment type, and constructs a secondary equipment fault dataset. The fault location model construction module inputs training samples from the secondary equipment fault dataset into a pre-built generator to generate a fault distribution matrix; inputs the fault distribution matrix into a pre-built discriminator to output anomaly scores for the training samples; collects normal samples and inputs them into the discriminator to obtain anomaly scores for the normal samples, and determines a baseline anomaly threshold based on these scores; calculates the average load rate and time-series volatility within the corresponding time window of the training samples, determines the load level factor and operational volatility intensity factor respectively, and corrects the baseline anomaly threshold to obtain an adaptive anomaly threshold; when the anomaly score of a training sample is greater than the adaptive anomaly threshold, the discriminator determines it as a fault; based on the discriminator's determination result and the fault distribution matrix, calculates the loss for each training sample and trains the generator and discriminator. The equipment fault location module collects and parses real-time message information from the secondary equipment of the substation, generates real-time input samples, and inputs them into the trained generator and discriminator to determine whether the secondary equipment is faulty. When a secondary equipment fault occurs, it calculates the feature anomaly contribution degree of different features to the anomaly score. Based on the physical association between the features and the equipment and links, combined with the feature anomaly contribution degree, it calculates the fault contribution degree of the equipment and locates the fault location of the secondary equipment.

[0015] A third aspect of the present invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of a method for locating faults in secondary equipment of a smart substation.

[0016] A fourth aspect of the invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of a method for locating faults in secondary equipment of an intelligent substation.

[0017] Compared with the prior art, the beneficial effects of the present invention include at least the following: 1. Traditional time series models have limited ability to capture long-term dependencies and struggle to analyze the cross-time-step propagation characteristics of secondary equipment faults. This invention utilizes a Bi-LSTM bidirectional gating mechanism and, through multi-layer nonlinear transformation, can automatically extract high-order interaction features from the data, forward model the evolution of historical data, reverse predict potential abnormal chain reactions, deeply integrate time series context information, and combine Bi-LSTM's WGAN to capture long-term dependencies, enhancing sensitivity to latent faults (such as synchronization error accumulation and communication interruption hysteresis effects), and providing more accurate fault tracing capabilities for complex time series dependency scenarios.

[0018] 2. This invention collects and parses the original message information when secondary equipment malfunctions, extracting multi-source operational data such as current and voltage sampling sequences and switch status, and filters these data to form fault feature information, thereby constructing a secondary equipment fault dataset. Based on this, a model consisting of a generator and a discriminator is used to learn and discriminate the fault features, achieving automatic identification of abnormal states in secondary equipment. Compared to traditional methods relying on human experience or single monitoring indicators, this invention can comprehensively analyze multi-dimensional operational data, improving the accuracy of fault identification and reducing the false positive rate.

[0019] 3. This invention introduces an adaptive anomaly threshold mechanism based on operating conditions during the anomaly detection process. By calculating the average load rate and time-series volatility within the training sample time window, a load level factor and an operating volatility intensity factor are obtained, respectively. The baseline anomaly threshold is then dynamically corrected, enabling the anomaly judgment threshold to adaptively adjust according to the substation load level and operating volatility. Compared with fixed threshold detection methods, this method effectively reduces false alarms and false negatives under different operating scenarios, enhances the model's adaptability to complex operating environments, and thus improves the stability and reliability of secondary equipment fault detection.

[0020] 4. After detecting an anomaly, this invention calculates the feature anomaly sensitivity by taking the partial derivative of the anomaly score with respect to each feature, and further calculates the feature anomaly contribution. Then, combining this with the physical correlation between the feature and the device and communication link, it calculates the fault contribution of the device or link, achieving precise location of the fault in the secondary device or communication link. Compared with existing technologies that can only determine whether an anomaly has occurred, this invention can not only detect faults but also locate the fault source, improving the interpretability and accuracy of fault analysis. Attached Figure Description

[0021] Picture 1 The flowchart of a method for locating faults in secondary equipment of an intelligent substation provided by the present invention is shown. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0023] Example 1 Embodiment 1 of the present invention provides a method for fault location of secondary equipment in intelligent substations, see reference. Picture 1 This includes the following steps: S1. Collect and parse the original message information when secondary equipment fails, extract current and voltage sampling sequences and switch status; process the extracted data, filter out fault feature information and determine the corresponding fault equipment type, and construct a secondary equipment fault dataset.

[0024] When secondary equipment (such as merging units, relay protection devices, and measurement and control devices) malfunctions, its fault information is transmitted to the relay protection information system via the station control layer network and bay layer network of the smart substation in the form of SV (sampled value) messages and GOOSE (General Object-Oriented Substation Event) messages. The information system processes and analyzes the fault information through the following process: The system captures the original messages and records the timestamp, message type, and key parameters. Based on the IEC 61850 standard, it parses the message structure, extracts the current / voltage sampling sequence of SV and the switching status of GOOSE. Combining historical fault databases and technical specifications, it filters messages through pre-configured rules, removes redundant data, and retains fault characteristic information associated with faulty components (such as abnormal SV sampling, GOOSE communication interruption, etc.). Based on the typical fault characteristic information listed in Table 1 (such as SV waveform distortion, GOOSE timing limit exceeding), it locates the faulty equipment type and potential causes (such as merging unit synchronization failure, switch port abnormality).

[0025] Table 1 Typical Fault Types and Fault Characteristics

[0026] In this embodiment, the selected secondary equipment fault information is encoded or standardized. Numerical data is directly standardized, and switch quantities (e.g., secondary equipment operating status data, SV / GOOSE link status data) are encoded. Alarm information is collected, and the data type is switch quantity. The value is 1 when an alarm message is received and 0 when no alarm message is received.

[0027] As a preferred implementation method, the process of constructing a secondary equipment fault dataset is as follows: Each secondary equipment failure is collected as a training sample, and the failure type in each failure is used as the label for the corresponding training sample. For each training sample, the selected failure feature information, including the secondary equipment operating status data, SV / GOOSE reception status data, and sampled values ​​at each moment within a predetermined time window before and after the failure, is used as the failure feature data of the training sample. Specifically, it is represented as follows: ; In the formula, Let X be the fault characteristic matrix of the s-th fault accident; ZTiFor the secondary equipment operating status data at time i, this embodiment generates switch data (0, 1) of the equipment operating status through hardware / software self-test features, providing a data source for fault diagnosis. See Table 2 for specific data; X SGi X represents the sampled value at time i and the received status data of the General Object-Oriented Substation Event (SV / GOOSE); CYi The sampled data is at time i; N is the total number of times within the predetermined time window before and after the equipment failure. Table 2. Operating Status Data of Secondary Equipment

[0028] The secondary equipment operation status data mainly consists of the self-test information status of the smart terminal, merging unit, and protection device, specifically represented as follows: ; In the formula, k, m, and u represent the total number of intelligent terminals, merging units, and protection devices in the secondary system, respectively. This refers to the hardware and software self-test information of the m-th merging unit during a fault, including self-test anomalies. Synchronization anomaly Version error Fiber optic port malfunction Power fluctuation alarm Sampling frequency out of tolerance Synchronization source switching anomaly ; This includes the hardware and software self-test information of the u-th protection device during a fault, including self-test anomalies. Device interlock Memory error AD sampling board malfunction Abnormal input / output circuit Error in the fixed value area verification GOOSE timeout ; This refers to the hardware and software self-test information of the k-th smart terminal during a fault, including self-test anomalies. Memory error Verification error 1. Switch input jitter Power redundancy failure Error prevention logic check timeout SV subscription interrupted ; SV / GOOSE link status information includes the SV / GOOSE link message reception status of secondary devices. and As shown in the following formula: ; In the formula, For the first Status information of each SV link; This refers to the status information of the w-th GOOSE link; The sampled data consists of the A, B, and C phase voltage and current sampled values ​​of the protection device via dual channels, as shown in the following formula: ; In the formula, and These are the sampled values ​​of the three-phase voltage and current in channels 1 and 2, respectively. , For current and voltage; subscripts 1 and 2 represent channel 1 and channel 2, respectively.

[0029] S2. Input the training samples from the secondary equipment fault dataset into the pre-built generator to generate a fault distribution matrix; input the fault distribution matrix into the pre-built discriminator to output the anomaly score of the training samples; collect normal samples and input them into the discriminator to obtain the anomaly score of the normal samples, and determine the baseline anomaly threshold based on this; calculate the average load rate and time-series volatility within the corresponding time window of the training samples, determine the load level factor and volatility intensity factor respectively, and correct the baseline anomaly threshold to obtain the adaptive anomaly threshold; when the anomaly score of the training sample is greater than the adaptive anomaly threshold, the discriminator determines it as a fault; based on the discriminator's determination result and the fault distribution matrix, calculate the loss of each training sample and train the generator and discriminator.

[0030] Generative Adversarial Networks (GANs) follow the ideas of Boolean theory, defined by a zero-sum between two competing neural networks. The generator network receives an input noise vector and generates fake data to deceive the discriminator; the discriminator network receives batches of real data and fake data generated by the generator, calculates the loss between the fake and real data samples, and then backpropagates the loss back to the generator network. GANs have shown great potential in generating realistic samples, but they also expose problems with training difficulty. First, the generator often cannot generate sufficiently diverse samples, limiting its ability to simulate complex data distributions; second, as the discriminator's performance improves, the gradient vanishing problem faced by the generator becomes more severe. For the temporal characteristics of substation secondary equipment operation data, this embodiment proposes a generator G and discriminator D network structure based on a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network. Bi-LSTM is used to solve the gradient vanishing problem and enhance the ability to capture long-term dependencies.

[0031] Bi-LSTM's multi-level feature abstraction mechanism can parse multi-dimensional correlation patterns in substation secondary equipment data: capture timing jitter in SV / GOOSE messages (such as synchronization errors in merging units); and model the nonlinear coupling between protection actions and communication states (such as protection maloperation caused by GOOSE link failure).

[0032] As a preferred implementation, WGAN (Wasserstein Generative Adversarial Network) is used as the fault localization model, which includes a generator and a discriminator. Both the generator and the discriminator use Bi-LSTM networks. In the hidden layer of the recurrent neural network in Bi-LSTM, each neuron node consists of a forget gate, a candidate memory unit, an input gate, and an output gate. The forget gate determines the degree to which the output of the previous hidden layer is forgotten; specifically: ; In the formula, The output of the forget gate at time t; Use the Sigmoid activation function; The input vector is the fault feature vector of the training sample's fault feature matrix at time t. The forget gate weight matrix is ​​the input vector. The weight matrix for the forget gate in the hidden state; Forget gate bias term; This represents the hidden state at time t-1; The current candidate memory unit value is: ; In the formula, The state of the candidate memory cell at time t; The candidate memory unit weight matrix; The weight matrix for candidate memory units in the hidden layer state; Candidate memory unit bias terms; The input gate controls the extent to which the input and the current computational state are updated to the memory unit; the specific formula is: ; In the formula, The output of the input gate at time t; The input vector is the input gate weight matrix; The hidden state input gate weight matrix; For input gate bias terms; The output gate controls the input and the current output, which depends on the size of the current memory cell.

[0033] ; In the formula, The output of the output gate at time t; Output the gate weight matrix to output the vector; The hidden state output gate weight matrix; This is the output gate bias term; The current forget gate and input gate outputs, candidate memory cell states, and the previous memory cell state are input into the memory cell to update the current memory cell state; specifically, this is expressed as: ; In the formula, The updated memory cell state at time t; The state of the memory cell at time t-1; Based on the output of the output gate at the current time and the state of the memory unit at time t, the hidden state at the current time is calculated; specifically, it is expressed as: ; Where tanh is the hyperbolic tangent function; In Bi-LSTM, the forward and backward hidden states are concatenated and then input into the output layer to obtain the fault distribution matrix for each sample.

[0034] In this embodiment, the joint architecture of Bi-LSTM and WGAN provides dual enhancement in fault detection. Within the WGAN framework, the generator G generates pseudo-normal data with temporal coherence through Bi-LSTM, while the discriminator D accurately identifies abnormal distribution shifts through bidirectional feature comparison. In the Bi-LSTM-WGAN network design, the Bi-LSTM neural network is embedded in the WGAN framework as both a generator and a discriminator to analyze the temporal data in the substation's secondary equipment. The Bi-LSTM consists of two inverse recurrent networks: a forward LSTM neural network and a backward LSTM neural network. The output layer merges the outputs of the forward and backward LSTM neural networks to overcome the limitations of unidirectional recurrent networks. The final Bi-LSTM can reasonably evaluate past and future information within a certain time range.

[0035] As a preferred implementation method, the process of determining the baseline anomaly threshold is as follows: This embodiment abandons the traditional fixed threshold anomaly determination method. Based on the output distribution of the discriminator's normal operation samples, a dynamic quantile threshold model is constructed. Combined with substation load level, operating fluctuation intensity and scenario classification information, multiple corrections are performed to generate an adaptive anomaly determination threshold. This allows the model to achieve optimal anomaly detection sensitivity and accuracy under different operating conditions such as light load, heavy load, steady state and transient state.

[0036] Normal samples from different operating scenarios of smart substations are collected and input into a pre-built discriminator to obtain anomaly scores for the normal samples. These scores serve as the base distribution sample library for dynamic thresholds and are stored according to scenario classification (e.g., daily steady-state, peak load, equipment switching, grid disturbance, etc.). For normal samples in each scenario, the inverse function value of the cumulative probability function of kernel density estimation at the predefined normal probability is calculated and used as the baseline anomaly threshold. The specific formula is as follows: ; In the formula, The baseline anomaly threshold; The normal probability is predefined; in this embodiment, it is taken as 0.997. This indicates that 99.7% of the normal samples have abnormal scores below the baseline abnormality threshold; The cumulative probability function for kernel density estimation at the baseline anomaly threshold; It represents the inverse function of the cumulative probability function; This represents the normal sample size. A set of abnormal scores for normal samples. The abnormal score for the i-th normal sample; Bandwidth is used to control the smoothness of the density curve; Kernel functions include, but are not limited to, Gaussian kernel function, rectangular kernel function, and trigonometric kernel function; The anomaly score for the sample to be estimated.

[0037] As a preferred implementation method, the process of correcting the baseline anomaly threshold is as follows: Three correction factors are introduced: load level factor, fluctuation intensity factor, and scenario correction factor, to dynamically correct the basic threshold. All correction factors are calculated using real-time operation monitoring data of the substation. When the factor value is less than 1, the threshold is lowered (to improve detection sensitivity), and when the factor value is greater than 1, the threshold is raised (to reduce the false detection rate).

[0038] The calculation is based on the substation main transformer load rate. A higher load rate indicates a greater probability of equipment failure, necessitating increased detection sensitivity and a lower threshold. The mean load rate of the substation main transformer within the corresponding time window of the training samples is obtained as the corresponding average load rate. Based on predetermined linear and quadratic coefficients for the low-load zone, a quadratic binomial calculation is performed on the average load rate to obtain the low-load growth term. The average load is then subtracted from a predefined heavy-load threshold and multiplied by a predefined exponential weight to obtain the exponent of an exponential function with a base of the natural constant, yielding the high-load growth term. The low-load growth term and the high-load growth term are weighted and normalized to obtain the load level factor. The specific formula is as follows: ; In the formula, The average loading rate of the training samples Load level factor; , These are the linear and quadratic coefficients for the low-load region, respectively, both of which are greater than 0, thus amplifying the risk trend of increasing load under low load conditions. It is an exponential function with the natural constant as its base; The predefined overload threshold is set to 80% in this embodiment; The predefined index weights control the steepness of the risk increase during high-load phases. A predefined high-load exponential growth weight is used to ensure that the exponential term dominates the risk and achieves a sharp increase under heavy load.

[0039] As an optional implementation, the load level factor can also be directly set in stages based on the average load rate of the training samples, specifically expressed as follows: .

[0040] Calculate the time-series volatility of real-time operational data (such as the coefficient of variation of SV sample values ​​and GOOSE message frequency). Higher volatility indicates greater grid instability, necessitating suppression of false detections and raising the threshold. Calculate the coefficient of variation of different sample values ​​within the corresponding time window of the training samples, and take the average to obtain the current and voltage coefficients of variation. Calculate the coefficient of variation of the GOOSE message frequency within the corresponding time window of the training samples, using this as the message frequency coefficient of variation. Based on a predefined volatility weighting coefficient, weight the current, voltage, and message frequency coefficients of variation to obtain the time-series volatility. The specific formula is as follows: ; In the formula, For time series volatility; The coefficient of variation of current is the mean of the coefficients of variation of current in different phases of different channels in the training sample within the time window. The voltage coefficient of variation is the mean of the voltage coefficients of variation of different phases in different channels of the training sample within the time window. The coefficient of variation of message frequency; , and These are the weighting coefficients corresponding to different coefficients of variation, all of which take values ​​within the interval (0,1). ; Using a predefined piecewise function, a piecewise volatility intensity factor is set for time series volatility, specifically expressed as follows: ; In the formula, This is the volatility intensity factor.

[0041] As a preferred implementation method, the fault risk level of the substation in real-time operation is assigned. The higher the fault risk level, the higher the detection sensitivity and the lower the threshold need to be. According to the substation in real-time operation, the corresponding scenario correction factor is predefined, as shown in Table 3.

[0042] Table 3 Mapping Table of Substation Operation Scenarios and Scenarios Correction Factors

[0043] The adaptive anomaly threshold is calculated by multiplying the baseline anomaly threshold by the difference between the predefined load level threshold and the load level factor, followed by multiplying by the fluctuation intensity factor and the scenario correction factor, and then limiting it to a predefined upper and lower limit range. The calculation formula is as follows: ; In the formula, An adaptive anomaly threshold; This is a pruning function used to limit the range of values ​​for the adaptive anomaly threshold; For predefined load level thresholds; This is a scene correction factor; and These are the lower and upper limits of the adaptive anomaly threshold, respectively, in this embodiment. ; The abnormality score of the training sample is compared with the adaptive abnormality threshold. If the abnormality score of the training sample is greater than the adaptive abnormality threshold, it is judged as a fault; otherwise, it is judged as normal.

[0044] In this embodiment, the detection sensitivity is dynamically adjusted by updating the adaptive anomaly threshold in real time: when the substation is in a high fault risk scenario (heavy load, grid disturbance, small fluctuation), the threshold is lowered and the detection sensitivity is improved, which can accurately capture minor hidden faults (such as synchronization error accumulation, minor fiber optic port anomalies); when the substation is in a low fault risk, high fluctuation scenario (light load, grid disturbance, high fluctuation), the threshold is increased and the detection specificity is improved, which can effectively suppress false detections caused by operational fluctuations.

[0045] As a preferred implementation, the loss of each training sample is calculated based on the discriminator's judgment result and the fault distribution matrix, and the generator and discriminator are trained. To address the vanishing and exploding gradient problems during generative adversarial network (GAN) training, as well as the poor diversity and accuracy of generated data samples, Wassertein distance is introduced to construct the discriminator loss function. To improve training stability, reduce oscillations and convergence difficulties, and accelerate network convergence, a gradient penalty term is introduced into the discriminator loss function, and Lipschitz constraints are applied to the generated samples. The discriminator loss function is defined as follows: ; in, The discriminator loss function; This is the original data sample; It is a sample of random noise data; It is the distribution of real data; It is the distribution of the generated data; This indicates that the discriminator has a good understanding of the samples. The output; Indicates the penalty weight; This represents a sample randomly sampled between real samples and generated samples; This represents the probability distribution of random noise data; Represents the probability distribution of the original data sample; This represents the probability distribution of randomly sampled data. This represents calculating the gradient of the discriminator network with respect to randomly sampled samples; It is an L2 norm; The generator loss function for guiding generator training is as follows: ; In the formula, The generator loss function; The loss of each training sample is calculated using the generator loss function and the discriminator loss function, thereby training the fault location model.

[0046] The training process is a minima / maxima optimization process, and ultimately the data generated by G cannot be distinguished by D. In this embodiment, the discriminator network is trained using normal operating data, and the trained discriminator is used to determine whether the real-time data is close to the distribution of normal data, thereby identifying faults.

[0047] S3. Collect and parse real-time message information from the secondary equipment of the substation, generate real-time input samples, and input them into the trained generator and discriminator to determine whether the secondary equipment is faulty. When the secondary equipment is faulty, calculate the feature anomaly contribution of different features to the anomaly score. Based on the physical association between the features and the equipment and links, and combined with the feature anomaly contribution, calculate the fault contribution of the equipment and locate the fault location of the secondary equipment.

[0048] Real-time message information from substation secondary equipment is collected and parsed, current and voltage sampling sequences and switch status are extracted and processed to generate real-time input samples. The real-time input samples include: secondary equipment operating status characteristics, SV / GOOSE link status characteristics, and sampled value characteristics; a total of r-dimensional time-series features, denoted as... ,in This refers to the operating status characteristics of secondary equipment (corresponding to the hardware and software self-test status of merging units, intelligent terminals, and protection devices). This refers to the SV / GOOSE link status characteristics (corresponding to the message reception status of each SV link and GOOSE link). The characteristics of the sampled values ​​(corresponding to the sampled values ​​of three-phase voltage and current); Real-time input samples are sequentially fed into the trained generator and discriminator to obtain corresponding anomaly scores. When the anomaly score exceeds the corresponding adaptive anomaly threshold, the secondary device malfunctions. The partial derivatives of the anomaly score output by the discriminator with respect to each feature dimension of the real-time input sample at different times are calculated as the feature anomaly sensitivity of the corresponding feature to the anomaly score at that time, generating a feature anomaly sensitivity matrix. Specifically, it is expressed as follows: ; In the formula, This is the feature anomaly sensitivity matrix; Let be the feature anomaly sensitivity of the i-th feature to the anomaly score at time t; Let i be the i-th feature at time t; The total number of moments within the corresponding time window for the real-time input sample; For generator; For discriminators; Indicates the discriminator pair The partial derivative is the gradient; the absolute value of the partial derivative represents the sensitivity of the feature to anomaly detection, and the positive / negative partial derivative represents the positive / negative driving effect of the abnormal change of the feature on the anomaly score. The feature anomaly sensitivity matrix is ​​normalized to obtain the standard feature anomaly sensitivity matrix. For each feature, the standard feature anomaly sensitivity at different times is summed as the numerator, and the standard feature anomaly sensitivity of all features at all times is summed as the denominator to calculate the feature anomaly contribution of the corresponding feature. The specific formula is as follows: ; In the formula, The contribution of the feature anomaly to the i-th feature; In this embodiment, column normalization normalizes the feature anomaly sensitivity of each dimension by performing maximum and minimum value normalization over time steps to eliminate differences in the dimensions of different features; row normalization normalizes the feature anomaly sensitivity of all features at each time step to quantify the contribution ratio of each feature within a single time step.

[0049] Based on the physical association between features and devices / links, features are categorized for different secondary devices and links, generating feature sets for each secondary device and each link. For each secondary device and link, the average of the anomaly contributions of all features within the corresponding feature set is calculated to obtain the corresponding fault contribution. The specific formula is as follows: ; In the formula, For equipment de The fault contribution rate is the equipment-level fault contribution rate, and its value is in the range [0,1]. The equipment includes different intelligent terminals, merging units and protection devices. For equipment de feature set The total number of features in; For link l The fault contribution is the link-level fault contribution, and its value is in the range [0,1]. The larger the value, the higher the probability that the link abnormality is the cause of the fault. The link includes different SV links and GOOSE links. For link l feature set The total number of features in; The features are sorted in descending order of their contribution to anomaly, and the first predetermined number of features are selected as key features. If the key features include links, the fault contribution of all links is compared, and the link with the highest fault contribution is identified as the link fault location. If the key features include devices, the fault contribution of all devices is compared, and the device with the highest fault contribution is identified as the secondary device fault location.

[0050] Example 2 Embodiment 2 of the present invention provides a fault location system for secondary equipment in an intelligent substation, and uses a fault location method for secondary equipment in an intelligent substation, including: The fault dataset construction module collects and parses the original message information when secondary equipment fails, extracts current and voltage sampling sequences and switch status; processes the extracted data, filters out fault feature information and determines the corresponding fault equipment type, and constructs a secondary equipment fault dataset. The fault location model construction module inputs training samples from the secondary equipment fault dataset into a pre-built generator to generate a fault distribution matrix; inputs the fault distribution matrix into a pre-built discriminator to output anomaly scores for the training samples; collects normal samples and inputs them into the discriminator to obtain anomaly scores for the normal samples, and determines a baseline anomaly threshold based on these scores; calculates the average load rate and time-series volatility within the corresponding time window of the training samples, determines the load level factor and operational volatility intensity factor respectively, and corrects the baseline anomaly threshold to obtain an adaptive anomaly threshold; when the anomaly score of a training sample is greater than the adaptive anomaly threshold, the discriminator determines it as a fault; based on the discriminator's determination result and the fault distribution matrix, calculates the loss for each training sample and trains the generator and discriminator. The equipment fault location module collects and parses real-time message information from the secondary equipment of the substation, generates real-time input samples, and inputs them into the trained generator and discriminator to determine whether the secondary equipment is faulty. When a secondary equipment fault occurs, it calculates the feature anomaly contribution degree of different features to the anomaly score. Based on the physical association between the features and the equipment and links, combined with the feature anomaly contribution degree, it calculates the fault contribution degree of the equipment and locates the fault location of the secondary equipment.

[0051] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0052] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0053] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0054] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status 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 execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via 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., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0055] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for fault location of secondary equipment in an intelligent substation, characterized in that, Includes the following steps: Collect and parse the original message information when secondary equipment fails, and extract the current and voltage sampling sequences and switch status; The extracted data is processed to filter out fault feature information and determine the corresponding faulty equipment type, and a secondary equipment fault dataset is constructed. The training samples from the secondary equipment fault dataset are input into a pre-built generator to generate a fault distribution matrix; the fault distribution matrix is ​​then input into a pre-built discriminator to output anomaly scores for the training samples. Normal samples are collected and input into the discriminator to obtain the abnormality score of the normal samples, and the baseline abnormality threshold is determined based on this score. Calculate the average load rate and time-series volatility within the corresponding time window of the training samples, determine the load level factor and the operational volatility intensity factor respectively, and correct the benchmark anomaly threshold to obtain the adaptive anomaly threshold; When the anomaly score of a training sample is greater than the adaptive anomaly threshold, the discriminator determines it to be faulty; Based on the discriminator's judgment results and the fault distribution matrix, the loss of each training sample is calculated, and the generator and discriminator are trained. Collect and parse real-time message information from the secondary equipment of the substation, generate real-time input samples, and input them into the trained generator and discriminator to determine whether the secondary equipment is faulty; when the secondary equipment is faulty, calculate the feature anomaly contribution of different features to the anomaly score. Based on the physical association between features and devices / links, and combined with the anomaly contribution of features, the fault contribution of devices is calculated, and the fault location of secondary devices is located.

2. The method for fault location of secondary equipment in an intelligent substation according to claim 1, characterized in that: The process of constructing a secondary equipment fault dataset is as follows: Each secondary equipment failure incident collected is used as a training sample, and the failure type in each secondary equipment failure incident is used as the label of the corresponding training sample. For each training sample, the selected fault feature information, including the secondary equipment operating status data, SV / GOOSE receiving status data, and sampled values ​​at each moment within a predetermined time window before and after the secondary equipment failure, is used as the fault feature data of the training sample.

3. The method for fault location of secondary equipment in an intelligent substation according to claim 1, characterized in that: The process of determining the baseline anomaly threshold is as follows: Normal samples from different operating scenarios of smart substations are collected and input into a pre-built discriminator to obtain anomaly scores for normal samples. For normal samples in each scenario, the inverse function value of the cumulative probability function of kernel density estimation at the predefined normal probability is calculated as the benchmark anomaly threshold.

4. The method for fault location of secondary equipment in an intelligent substation according to claim 1, characterized in that: The process of determining the load level factor is as follows: The mean value of the main transformer load rate of the substation within the corresponding time window of the training samples is obtained as the corresponding average load rate. Based on the predetermined linear coefficient and quadratic coefficient of the low load area, the average load rate is calculated by a quadratic binomial to obtain the low load growth term. The average load is subtracted from the predefined heavy load threshold and multiplied by the predefined exponential weight to obtain the exponent of the exponential function with the natural constant as the base, thus obtaining the high load growth term. The low load growth term and the low load growth term are weighted and normalized to obtain the load level factor.

5. The method for fault location of secondary equipment in an intelligent substation according to claim 1, characterized in that: The process of determining the running volatility intensity factor is as follows: The coefficients of variation of different sampled values ​​in the training samples within the corresponding time window are calculated and averaged to obtain the current coefficient of variation and the voltage coefficient of variation. The coefficient of variation of the GOOSE message frequency within the corresponding time window of the training samples is calculated as the message frequency coefficient of variation. Based on the predefined fluctuation weight coefficient, the current coefficient of variation, the voltage coefficient of variation, and the message frequency coefficient of variation are weighted to obtain the time series volatility. The time series volatility is set with a segmented volatility intensity factor by using a predefined segmented function.

6. The method for fault location of secondary equipment in an intelligent substation according to claim 1, characterized in that: The process of correcting the baseline anomaly threshold is as follows: Based on the real-time operation scenario of the substation, predefine the corresponding scenario correction factor; The baseline anomaly threshold is multiplied by the difference between the predefined load level threshold and the load level factor, and then multiplied by the fluctuation intensity factor and the scenario correction factor in turn to obtain the adaptive anomaly threshold, which is then limited to the predefined upper and lower limits.

7. The method for fault location of secondary equipment in an intelligent substation according to claim 1, characterized in that: The process of calculating the feature anomaly contribution of different features to anomaly scoring is as follows: The real-time input samples are sequentially input into the trained generator and discriminator to obtain the corresponding anomaly scores. When the anomaly score is greater than the corresponding adaptive anomaly threshold, the secondary device malfunctions. The partial derivatives of the anomaly score output by the discriminator with respect to each feature of the real-time input sample at different times are calculated respectively, which are used as the feature anomaly sensitivity of the corresponding feature to the anomaly score at the corresponding time, and a feature anomaly sensitivity matrix is ​​generated. The feature anomaly sensitivity matrix is ​​normalized to obtain the standard feature anomaly sensitivity matrix. For each feature, the standard feature anomaly sensitivity of the corresponding feature at different times is accumulated as the numerator, and the standard feature anomaly sensitivity of all features at all times is accumulated as the denominator to calculate the feature anomaly contribution of the corresponding feature.

8. The method for fault location of secondary equipment in an intelligent substation according to claim 1, characterized in that: The process of locating the fault in secondary equipment is as follows: Based on the physical association between features and devices and links, features corresponding to different secondary devices and different links are divided, and feature sets for each secondary device and each link are generated respectively. For each secondary device and link, the anomaly contribution of all features in the corresponding feature set is accumulated and averaged to obtain the corresponding fault contribution. The features are sorted in descending order of their contribution to anomaly, and the first predetermined number of features are selected as key features. If the key features include links, the fault contribution of all links is compared, and the link with the highest fault contribution is identified as the link fault location. If the key features include devices, the fault contribution of all devices is compared, and the device with the highest fault contribution is identified as the secondary device fault location.

9. A fault location system for secondary equipment in an intelligent substation, using the method described in any one of claims 1-8, characterized in that, include: The fault dataset construction module collects and parses the original message information when secondary equipment fails, and extracts current and voltage sampling sequences as well as switch status. The extracted data is processed to filter out fault feature information and determine the corresponding faulty equipment type, and a secondary equipment fault dataset is constructed. The fault location model building module inputs training samples from the secondary equipment fault dataset into a pre-built generator to generate a fault distribution matrix; it then inputs the fault distribution matrix into a pre-built discriminator to output anomaly scores for the training samples. Normal samples are collected and input into the discriminator to obtain the abnormality score of the normal samples, and the baseline abnormality threshold is determined based on this score. Calculate the average load rate and time-series volatility within the corresponding time window of the training samples, determine the load level factor and the operational volatility intensity factor respectively, and correct the benchmark anomaly threshold to obtain the adaptive anomaly threshold; When the anomaly score of a training sample is greater than the adaptive anomaly threshold, the discriminator determines it to be faulty; Based on the discriminator's judgment results and the fault distribution matrix, the loss of each training sample is calculated, and the generator and discriminator are trained. The equipment fault location module collects and parses real-time message information from the secondary equipment of the substation, generates real-time input samples, and inputs them into the trained generator and discriminator to determine whether the secondary equipment is faulty. When the secondary equipment is faulty, it calculates the feature anomaly contribution of different features to the anomaly score. Based on the physical association between features and devices / links, and combined with the anomaly contribution of features, the fault contribution of devices is calculated, and the fault location of secondary devices is located.

10. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-8.

Images