Intelligent anti-islanding method based on transient characteristics

Abstract

The invention provides an intelligent anti-islanding method based on transient characteristics, and the method comprises the steps: recognizing a link reconfiguration state of a core switch when an instantaneous power failure occurs, obtaining a link reconfiguration timestamp, calculating the time sequence correlation between reconfiguration and power failure in combination with an instantaneous power failure occurrence moment, and obtaining the triggering characteristics of link reconfiguration; performing mode matching according to the abnormal transient feature vector and a load adjustment model established based on historical operation data statistical analysis to obtain a matching degree index, and performing clustering analysis on the matching degree index to obtain an island state classification result; and carrying out weighted fusion on the classification result of the island state to obtain a final judgment result, and carrying out dynamic correction on the power supply frequency offset according to the final judgment result to obtain a corrected frequency stable signal.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an intelligent anti-islanding method based on transient characteristics. Background Art

[0002] In complex scenarios such as disaster recovery drills, the mass offline of server clusters can expose the power system to transient shocks. Dynamic identification of multi-dimensional characteristic signals is crucial for optimizing anti-islanding protection and has become a key issue in operations and maintenance management. Traditional solutions primarily rely on static threshold monitoring and single-signal analysis, making them incapable of handling dynamic scenarios characterized by sudden load changes and concurrent service migration. This is particularly true during off-grid UPS self-powered switching. Misjudging abnormal transients associated with load adaptive adjustments and islanding states can easily lead to malfunction or failure of anti-islanding protection devices, effectively failing to address the reliability issues of power supply anti-islanding protection. Existing methods, when processing complex signals such as multi-level current spikes, high-frequency pulse noise, frequency offsets, voltage fluctuations, and phase jumps, face challenges in accurately identifying power system transient characteristics and optimizing anti-islanding protection strategies in the complex dynamic environment of sudden load changes and service migration. Specifically, temperature control strategies for high-performance computing nodes can trigger instantaneous power outages, link reconfiguration in core switches can introduce load fluctuations, and adaptive adjustments in cooling systems can further exacerbate current spikes and noise. These factors, coupled with each other, can cause abnormal transient conditions such as frequency offsets, voltage fluctuations, and phase jumps in power systems when islanding. Traditional anti-islanding protection devices struggle to dynamically distinguish between normal service adjustments and true islanding abnormalities. Optimizing the response strategy of power supply anti-islanding protection devices based on dynamic identification of multi-level spikes, high-frequency noise, and anti-islanding signature signals in scenarios where O&M batch scheduling and sudden load changes coexist is a key challenge that needs to be addressed. This ensures accurate identification and response to islanding conditions under various operating conditions. Summary of the Invention

[0003] The present invention provides an intelligent anti-islanding method based on transient characteristics, which mainly includes:

[0004] Acquire real-time current data from UPS off-grid switching, analyze the high-frequency pulse noise spectrum in the current data, obtain spectrum characteristics, and extract the amplitude and time interval of the current spike. If the amplitude exceeds the preset amplitude threshold and the time interval is less than the preset interval threshold, it is determined that the node has experienced a momentary power outage;

[0005] Identify the core switch link reconfiguration status when a transient power outage occurs, obtain the link reconfiguration timestamp, calculate the timing correlation between reconfiguration and power outage based on the moment the transient power outage occurs, and obtain the triggering characteristics of the link reconfiguration;

[0006] If the rate of change of each operating parameter during the cooling system adjustment process exceeds a corresponding preset threshold, it is determined that the cooling system responds to a momentary power outage, and the operating parameters of the cooling system in response to the momentary power outage are analyzed to obtain the dynamic characteristics of the cooling system adjustment;

[0007] When an islanding signal is detected that is temporally correlated with a link reconfiguration trigger feature or a cooling system adjustment dynamic feature, the abnormal electrical parameter values ​​of the anti-islanding protection device are obtained, and the time-frequency distribution of the power supply voltage and current characteristic signals are extracted to obtain an abnormal transient characteristic vector.

[0008] Pattern matching is performed between the abnormal transient characteristic vector and the load adjustment model established based on statistical analysis of historical operating data to obtain a matching index. Cluster analysis is performed on the matching index to obtain the islanding state classification result.

[0009] The classification results of the islanding state are weighted and fused to obtain the final discrimination result. The power supply frequency offset is dynamically corrected according to the final discrimination result to obtain a corrected frequency stability signal.

[0010] By analyzing the corrected frequency stability signal and voltage fluctuations, the coupling relationship between frequency and voltage is obtained, and the voltage fluctuation standard deviation of the voltage sampling within the preset time window is calculated. If the standard deviation is less than the preset threshold, the protection setting of the anti-islanding protection device is corrected based on the historical operating condition data and the islanding state classification results.

[0011] Furthermore, real-time current data of the UPS during off-grid switching is obtained, the high-frequency pulse noise spectrum in the current data is analyzed to obtain spectral characteristics, and the amplitude and time interval of the current spike are extracted. If the amplitude exceeds a preset amplitude threshold and the time interval is less than a preset interval threshold, a node is determined to have experienced a momentary power outage. This includes: sampling and obtaining real-time current data of the uninterruptible power supply during off-grid switching operation using a high-precision current sampler, continuously collecting current data according to a preset sampling frequency and sampling interval, and quantizing the collected current data according to a preset resolution to obtain a first current data sequence. The first current data sequence is digitally preprocessed by filtering, using a median filter to eliminate outliers, and band-limiting the preprocessed current data using a bandpass filter to obtain a second current data sequence. The second current data sequence is decomposed using a multi-scale wavelet transform, with the wavelet basis function set to a Debyeqin wavelet. The current data is subjected to a five-layer wavelet decomposition to obtain high-frequency pulse noise spectral components, segmenting the spectral components according to a preset frequency component threshold, and calculating the amplitude distribution density function for each frequency component interval to obtain a first spectral characteristic value. For the first spectrum eigenvalue, an adaptive threshold peak detection algorithm is used to extract the current spike signal sequence, and the current spike signal sequence is segmented according to the preset time window. The peak amplitude and the time interval between adjacent peaks are extracted in each time window to obtain the second spectrum eigenvalue. A feature vector is constructed for the extracted second spectrum eigenvalue, which includes four feature components: the mean value of the peak amplitude, the mean value of the peak time interval, the standard deviation of the peak amplitude, and the standard deviation of the peak time interval. A pre-trained neural network classifier is used to identify the power-off state of the feature vector. The neural network classifier adopts a multi-layer perceptron structure, the number of input layer nodes is four feature components, the hidden layer adopts a two-layer structure, and the output layer is a power-off state identifier. If the mean value of the peak amplitude in the feature vector exceeds the preset amplitude threshold and the mean value of the peak time interval is less than the preset time interval threshold, and the neural network classifier outputs the power-off state identifier as true, it is determined that an instantaneous power outage occurs at the node in the time window.

[0012] Furthermore, the link reconfiguration status of the core switch when a transient power outage occurs is identified, the link reconfiguration timestamp is obtained, and the timing correlation between the reconfiguration and the power outage is calculated in combination with the moment the transient power outage occurs, thereby obtaining the triggering characteristics of the link reconfiguration. This includes: polling and collecting the status of all link ports of the core switch through a real-time communication agent, with the collection period set to 100 milliseconds, normalizing the collected link traffic, port bandwidth, and link load data, and calculating link status feature values ​​for the normalized data using a sliding window of 10 sampling points to obtain first link status data. The first link status data is validated, and data points with zero link traffic or abnormal port bandwidth are eliminated. The validated data is subjected to a discrete wavelet transform to extract link status mutation features including port status change rate and link topology change rate to obtain second link status data. The second link status data is classified using a gradient clustering algorithm, with clustering features including port status change rate and link topology change rate, identifying link reconfiguration events, and recording the start and end times of the reconfiguration to obtain a first timestamp sequence. The duration of the reconfiguration is calculated based on the first timestamp sequence, and the reconfiguration timing feature vector is established in combination with the port status and link topology data before and after the reconfiguration. The reconfiguration timing feature vector is subjected to data normalization processing to obtain the first reconfiguration feature data. The first reconfiguration feature data is subjected to outlier detection, and the data points whose reconfiguration duration exceeds the normal range are eliminated to obtain the second reconfiguration feature data. The timestamp data of the moment when the instantaneous power outage occurs is obtained from the external data interface, and the random forest classifier is used to calculate the timing correlation between the second reconfiguration feature data and the power outage timestamp data to extract the time correlation characteristics between the reconfiguration and the power outage. If the time interval between the reconfiguration occurrence moment and the power outage moment is less than the preset time threshold and the reconfiguration duration is greater than the preset duration threshold, the reconfiguration event is determined to be a link reconfiguration event triggered by the instantaneous power outage.

[0013] Furthermore, if the rate of change of each operating parameter during the cooling system adjustment process exceeds a corresponding preset threshold, the cooling system is determined to have responded to a momentary power outage. The operating parameters of the cooling system in response to the momentary power outage are analyzed to obtain dynamic characteristics of the cooling system adjustment, including: real-time acquisition of four sets of cooling system operating parameters (temperature, pressure, flow rate, and cooling power) using a multi-channel collector, with the acquisition frequency set to 10 times per second; digital filtering of the acquired data to remove high-frequency noise; and normalization of the filtered data to obtain a first cooling parameter sequence. Data validity is verified for the first cooling parameter sequence, abnormal data points exceeding the physical limit range are eliminated, and a differential algorithm is used to calculate four dynamic characteristic quantities (temperature change rate, pressure change rate, flow rate change rate, and cooling power change rate) to obtain first dynamic characteristic data. Preset change rate thresholds are set based on the physical characteristics of each parameter, including a temperature change rate threshold, a pressure change rate threshold, a flow rate change rate threshold, and a cooling power change rate threshold. Threshold determination is performed on the first dynamic characteristic data. If any dynamic characteristic quantity exceeds the corresponding preset change rate threshold, it is determined that the cooling system has responded to a momentary power outage, obtaining first response characteristic data. The wavelet transform is used to perform time-frequency decomposition on the four sets of parameter change curves in the time series corresponding to the first response characteristic data. The Debyeqin wavelet is selected as the basis function, and the spectral energy characteristics of the decomposed signal are extracted to obtain the second response characteristic data. The peak detection method is used to locate the spectral energy characteristics in the second response characteristic data, and the timestamps of the mutation points of each parameter curve are recorded. The response start time and the stabilization time are calculated to obtain the third response characteristic data. The cooling response time is calculated based on the third response characteristic data, and the adjustment amplitude is calculated based on the maximum fluctuation amplitude of each parameter in the response time period. A feature vector containing the response time and the adjustment amplitude is constructed. A five-layer convolutional neural network is used to extract features from the feature vector. The input layer corresponds to the four sets of parameter features, the three hidden layers contain 16, 32, and 16 convolution kernels respectively, and the output layer corresponds to the dynamic feature value of the cooling adjustment.

[0014] Furthermore, when an islanding signal temporally correlated with a link reconfiguration trigger feature or a cooling system adjustment dynamic feature is detected, anomalies in the electrical parameters of the anti-islanding protection device are acquired, the time-frequency distribution of the power supply voltage and current characteristic signals is extracted, and an abnormal transient characteristic vector is obtained. This includes: sampling three sets of electrical parameters of the anti-islanding protection device (power supply frequency, voltage amplitude, and phase angle) using a high-speed data acquisition component, eliminating high-frequency noise from the electrical parameters using a digital filter, and normalizing the filtered data to obtain a first electrical parameter sequence. An offset is calculated for the first electrical parameter sequence based on a preset standard range, and three sets of anomalies (power supply frequency offset, voltage fluctuation amplitude, and phase jump amplitude) are extracted. An adaptive threshold detection algorithm is used to classify the anomalies to obtain first abnormal characteristic data. Hilbert transform is used to extract the envelope of the power supply voltage and current signals, the signal amplitude change rate is calculated based on the envelope curve, wavelet decomposition is used to extract the transient component in the voltage and current signals, and spectral analysis of the transient component is performed to obtain second abnormal characteristic data. Based on the time coordinates of the first abnormal feature data and the second abnormal feature data, a time correlation analysis is performed on the link reconfiguration trigger feature or the cooling adjustment dynamic feature. A recursive neural network is used to extract the voltage and current characteristics of the abnormal transient process. The time-frequency spectrum distribution of each component in the feature vector is calculated to obtain the abnormal transient feature vector.

[0015] Furthermore, pattern matching is performed between the abnormal transient characteristic vector and the load adjustment model established based on statistical analysis of historical operating data to obtain a matching index, and cluster analysis is performed on the matching index to obtain the island state classification result, including: normalizing the historical operating data through a data preprocessing component, using a sliding window with a length of 60 sampling points to intercept the voltage and current characteristic sequence during the load adjustment process, and calculating the voltage and current mean, standard deviation, and volatility within the window to obtain the first load characteristic data. The first load characteristic data is validated, and data segments with voltage and current means exceeding the nominal range or abnormal standard deviations are eliminated. The verified data is reduced in dimension using principal component analysis, and the characteristic components with a cumulative contribution rate of 95% are selected to construct the first characteristic matrix. The load adjustment pattern is calculated based on the first characteristic matrix, and the matching index between the abnormal transient characteristic vector and the load adjustment pattern is calculated using three methods: Euclidean distance, cosine similarity, and Pearson correlation coefficient, to obtain the first matching data. The first matching data is normalized, and the normalized data is nonlinearly mapped using a support vector machine with a radial basis kernel function. The three-dimensional matching index is mapped to a high-dimensional feature space to obtain the second matching data. The clustering validity of the second matching data is verified, and the distance matrix and local density distribution between samples are calculated. Based on the verification results, the optimal number of clusters is determined to obtain the third matching data. The third matching data is clustered using a density peak clustering algorithm. The local density of the cluster center and the distance to the nearest high-density point are calculated, and the cluster boundaries are divided according to the local density threshold and the distance threshold. The samples are classified according to the cluster boundaries, and the average density within the class and the minimum distance between classes are calculated. If both indicators meet the preset threshold conditions, the classification result is determined as an island state classification result.

[0016] Furthermore, cluster analysis is performed on the abnormal transient characteristic vectors to obtain initial clusters. The average matching index, volatility, and duration of the initial clusters are calculated and classified. If the average matching index, volatility, and duration of the initial clusters are all higher than the corresponding preset risk threshold range, it is determined to be an island state and the island operation control instruction is activated. If it is between the preset risk range and is classified as a high-risk critical island, a warning signal is triggered and preventive control instructions are initiated. This includes: using an adaptive fuzzy clustering method to perform initial clustering on the abnormal transient characteristic vectors, calculating the optimal number of cluster centers based on the Euclidean distance between samples and the local density distribution, calculating the silhouette coefficient using the ratio of the distance between samples to the density within the class, and selecting the cluster result corresponding to the largest silhouette coefficient as the first cluster result data. A sliding window with a length of 60 sampling points is used to extract features from the first cluster result data, calculating the three indicators of fluctuation amplitude, fluctuation frequency, and fluctuation period of the characteristic vector within the window, and normalizing these three indicators to obtain the first fluctuation characteristic data. The time series characteristics of the clusters are calculated based on the first fluctuation characteristic data, including three statistics: fluctuation duration, fluctuation amplitude mean, and fluctuation frequency standard deviation. The second fluctuation characteristic data is obtained by calculating the degree of dispersion of the samples within the cluster using Euclidean distance. The average matching degree of each cluster is calculated for the second fluctuation characteristic data. A risk assessment index is constructed based on the distance between samples within the cluster and the minimum distance between clusters. The risk score is calculated using a weighted summation method to obtain the first risk assessment data. The first risk assessment data is graded according to the preset risk threshold. The upper and lower boundary values ​​of the risk threshold are set, and the risk score is classified and determined to obtain the second risk assessment data. If the risk score in the second risk assessment data is higher than the preset upper threshold, an island isolation control code including a circuit breaker tripping instruction and a load transfer instruction is generated. If the risk score in the second risk assessment data is between the upper and lower thresholds, a preventive control code including a frequency adjustment instruction and a voltage adjustment instruction is generated, and an early warning signal is triggered.

[0017] Furthermore, the islanding state classification results are weighted and fused to obtain a final discrimination result. Based on the final discrimination result, the power supply frequency offset is dynamically corrected to obtain a corrected frequency stability signal. This process includes: extracting data from the islanding state classification results using a sliding window of 60 sampling points; calculating a confidence score based on the three indicators of the classification results, namely, fluctuation amplitude, duration, and state stability; assigning a fusion weight coefficient based on the confidence score; and normalizing the weighted fusion result to obtain first discrimination data. The first discrimination data is then validated, and abnormal data points with confidence scores below a preset threshold are removed. A five-layer adaptive neural network is used to extract features from the validated data, and the network parameters are trained based on labeled historical data to obtain second discrimination data. The frequency offset is calculated based on the second discrimination data, and the frequency offset sequence is dynamically updated using the recursive least squares method with a preset sampling interval. The frequency compensation increment is calculated to obtain first frequency compensation data. The first frequency compensation data is then smoothed, and a low-pass filter with a cutoff frequency twice the power frequency is used to eliminate high-frequency noise. The filtered data is then normalized to obtain second frequency compensation data. The power supply frequency is dynamically corrected based on the second frequency compensation data. A correction coefficient is calculated based on the frequency change rate and frequency deviation, and the original frequency signal is compensated to obtain third frequency compensation data. The third frequency compensation data is segmented using a preset frequency stability interval. The frequency stability state is determined based on the mean, variance, and fluctuation rate of the frequency signal, and a frequency stability signal is generated, which includes a frequency value, a timestamp, and a status indicator.

[0018] Furthermore, by analyzing the corrected frequency stability signal and voltage fluctuations, the frequency-voltage coupling relationship is determined. The standard deviation of voltage fluctuations within a preset time window is calculated. If the standard deviation is less than a preset threshold, the protection setting of the anti-islanding protection device is corrected based on historical operating condition data and islanding status classification results. This includes synchronously sampling the frequency stability signal and voltage fluctuation data using a sliding time window with a length of 100 sampling points, calculating the Pearson correlation coefficient between the frequency deviation sequence and the voltage deviation sequence within the window, and normalizing the correlation coefficient sequence to obtain first coupling characteristic data. A frequency-voltage coupling matrix is ​​constructed based on the first coupling characteristic data. The coupling matrix is ​​de-noised using an adaptive Kalman filter with a noise covariance of 0.01 and a measurement covariance of 0.1. The main characteristic components with eigenvalues ​​greater than a preset threshold are extracted to obtain second coupling characteristic data. The second coupling characteristic data is then time-series segmented, the voltage sampling data is segmented using a preset time window, the voltage fluctuation standard deviation is calculated for each data segment, and the mean and variance of the fluctuation standard deviation are calculated based on historical data of normal operating conditions to obtain first standard deviation data. The voltage stability interval is calculated based on the first standard deviation data. Data segments with fluctuation standard deviations less than a preset threshold are marked as stable intervals. The frequency-voltage coupling characteristics within the stable interval are extracted to obtain the second standard deviation data. A steady-state feature sequence is established based on the second standard deviation data. This feature sequence is extracted using a long short-term memory network. Combined with the islanding state classification results, dynamic protection setting correction rules are constructed to obtain the first correction data. Based on the first correction data, a protection setting correction is generated, consisting of three components: the frequency protection setting value, the voltage protection setting value, and the protection delay setting value. The protection parameter configuration signal is encapsulated in a preset data format.

[0019] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0020] The present invention discloses an intelligent anti-islanding method based on transient characteristics. By analyzing the high-frequency pulse noise spectrum and current spike characteristics in the current data, instantaneous power outage events are identified, and the core switch link reconfiguration and cooling system response characteristics are combined to achieve accurate identification of the islanding state. The present invention collects the abnormal amount of electrical parameters of the anti-islanding protection device, extracts the time-frequency distribution of the power supply voltage and current characteristic signals, obtains the abnormal transient characteristic vector, and implements the classification of the islanding state by performing pattern matching and cluster analysis with the load adjustment model. The present invention performs weighted fusion on the classification results, dynamically corrects the power supply frequency offset, analyzes the coupling relationship between frequency and voltage, and corrects the protection setting of the anti-islanding protection device according to the voltage fluctuation, thereby improving the accuracy and reliability of islanding detection, effectively preventing misjudgment and missed judgment, and ensuring the safe and stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1This is a flow chart of an intelligent anti-islanding method based on transient characteristics of the present invention.

[0022] Figure 2 Schematic diagram of an intelligent anti-islanding method based on transient characteristics of the present invention. DETAILED DESCRIPTION

[0023] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments derived by those skilled in the art based on the embodiments in this specification without creative effort shall fall within the scope of protection of this specification.

[0024] like Figure 1-2 In this embodiment, an intelligent anti-islanding method based on transient characteristics may specifically include:

[0025] S101 collects real-time current data, analyzes the spectrum characteristics of high-frequency pulse noise, extracts the amplitude and time interval of the current spike, and determines that a momentary power outage has occurred if the peak amplitude exceeds a preset threshold and the time interval is less than the preset threshold.

[0026] In the embodiment of the present invention, the uninterruptible power supply will generate significant current fluctuations when switching off the grid. To accurately capture these transient characteristics, it is necessary to perform high-precision current sampling on the power supply system and analyze its spectral characteristics to ensure the accuracy and reliability of feature extraction.

[0027] S1011 collects real-time current data of the uninterruptible power supply in the off-grid switching state through a high-precision current sampler, and generates a first current data sequence according to a preset sampling frequency and interval time.

[0028] In this embodiment of the present invention, the sampling frequency is set to 20 kHz, the sampling interval is 0.05 milliseconds, and the quantization accuracy is 16 bits to capture subtle changes in the current signal. The sampler can be deployed at the output of the uninterruptible power supply to record current data in real time. The collected data undergoes analog-to-digital conversion to form a first current data sequence.

[0029] S1012 pre-processes the first current data sequence, removes outliers using a median filter, and performs band-limiting processing using a band-pass filter to generate a second current data sequence.

[0030] In this embodiment of the present invention, to eliminate interfering signals such as power frequency interference and switching noise, a median filter with a window length of 5 sampling points is used to remove sudden pulse interference. Subsequently, a bandpass filter with a passband range of 1 kHz to 10 kHz is used to band-limit the data, retaining the frequency band associated with power outage characteristics and generating a second current data sequence. This effectively improves the signal-to-noise ratio, laying the foundation for feature extraction.

[0031] S1013 performs multi-scale decomposition based on the second current data sequence using wavelet transform, extracts the spectrum characteristics of the high-frequency pulse noise, and constructs a feature vector.

[0032] In an embodiment of the present invention, the Debyechen wavelet is selected as the basis function, and a five-layer wavelet decomposition is performed on the second current data sequence, and the decomposition level covers the frequency band from 10 kHz to 0.625 kHz. For high-frequency pulse noise, the amplitude distribution density function of the spectral components of each layer is calculated to generate the first spectral eigenvalue. Subsequently, an adaptive threshold algorithm is used to detect the current spike sequence, and the mean value of the peak amplitude in each time window and the mean time interval between adjacent spikes are extracted to construct a feature vector containing the amplitude mean, interval mean, amplitude standard deviation and interval standard deviation. This step utilizes the time-frequency positioning characteristics of the wavelet transform to ensure the accuracy of feature extraction.

[0033] S1014 uses a pre-trained classification model to identify the power outage status of the feature vector. If the mean value of the peak amplitude exceeds the preset amplitude threshold and the mean value of the time interval is less than the preset time interval threshold, it is determined that the node has a momentary power outage.

[0034] In an embodiment of the present invention, the classification model utilizes a multi-layer perceptron architecture. The input layer contains four feature components, the hidden layer comprises eight and six nodes, and the output layer generates a power outage status indicator. The model was trained on 1,000 sets of normal operation data and 200 sets of power outage data, achieving a classification accuracy exceeding 95%. The preset amplitude threshold is 6 amps, and the time interval threshold is 10 milliseconds. When the feature vector indicates a mean peak amplitude of 7.5 amps, a mean time interval of 6 milliseconds, and the classification model outputs a power outage probability of 0.98, a momentary power outage is determined to have occurred. Intelligent classification effectively improves the reliability of power outage identification, providing a key basis for islanding state analysis.

[0035] In this embodiment of the present invention, high-precision sampling and multi-scale feature extraction are used to accurately capture the current spikes and high-frequency noise characteristics during off-grid switching of an uninterruptible power supply. Combined with an intelligent classification model, this allows for rapid identification of transient power outages, reducing the risk of misjudgment and providing reliable data support for anti-islanding protection. This step enables efficient identification of transient characteristics, laying the foundation for stable power system operation.

[0036] S102 monitors the link status of the core switch in real time, analyzes the timing characteristics of link reconfiguration, calculates its timing correlation based on the power-off moment, and extracts the triggering characteristics of link reconfiguration.

[0037] In this embodiment of the present invention, momentary power outages can trigger link reconfiguration in core switches. To accurately identify the correlation between the two, real-time monitoring of the switch's link port status is required, and reconfiguration triggering features are extracted through timing analysis. Multi-dimensional feature extraction and intelligent classification enable precise correlation between power outages and reconfiguration events, providing a key basis for island state analysis and ensuring comprehensive feature extraction and reliable timing analysis.

[0038] S1021 periodically collects the link port status of the core switch through a real-time communication agent, normalizes the collected data based on a sliding window, and generates a first link status sequence.

[0039] In this embodiment of the present invention, the communication agent polls the status of all link ports on the core switch every 100 milliseconds, collecting data including port traffic, bandwidth utilization, and link load. To standardize the dimensions of different indicators, a linear normalization method is used to map the data to a range of 0 to 1. Next, a sliding window is set with a length of 10 sampling points, covering a time span of 1 second. The mean and standard deviation of the data within the window are calculated to generate the first link status sequence. This step effectively smooths short-term fluctuations, preserves the trend characteristics of the link status, and lays the foundation for subsequent analysis.

[0040] S1022 verifies the validity of the first link state sequence, removes abnormal data points, and uses wavelet transform to extract link state mutation features to generate a second link state sequence.

[0041] In this embodiment of the present invention, to ensure data quality, the validity of the first link state sequence is verified, and data points with zero traffic or abnormal bandwidth utilization, such as invalid records caused by equipment failure, are eliminated. Subsequently, a discrete wavelet transform (DWT) is used, using the Symlet wavelet basis function, to perform a three-layer decomposition of the valid data, extracting mutation features such as the port state change rate and link topology change rate. These features reflect the dynamic changes in link state during the reconfiguration process, generating the second link state sequence. This step enhances the accuracy of feature extraction by leveraging the time-frequency analysis capabilities of the wavelet transform.

[0042] S1023: Based on the second link state sequence, a clustering algorithm is used to identify link reconfiguration events, record event timestamps, and generate a reconfiguration timing feature vector.

[0043] In an embodiment of the present invention, a K-means clustering algorithm is used to classify the second link state sequence. The clustering features include the port state change rate and the link topology change rate. The number of clusters is set to 3 to distinguish between normal operation, reconfiguration, and abnormal states. The start and end timestamps of the identified reconfiguration events are recorded to form a first timestamp sequence. Based on the timestamp sequence, the duration of the reconfiguration is calculated, and combined with the port state data before and after the reconfiguration, a time series feature vector containing the change rate, duration, and topology change amplitude is constructed. The reconfiguration events are effectively separated through cluster analysis, providing reliable input for timing correlation analysis.

[0044] S1024 normalizes and detects outliers on the timing feature vector, generates a reconfiguration feature data set, and calculates timing correlation in combination with the power-off timestamp.

[0045] In this embodiment of the present invention, to eliminate dimensional differences, the time series feature vectors are Z-score normalized. A density-based outlier detection algorithm is then used to remove data points whose durations exceed the normal range, such as those with durations exceeding 2 seconds or less than 50 milliseconds, to generate a reconfiguration feature dataset. The timestamps of instantaneous power outages are obtained through an external interface, and a support vector machine classifier is used to calculate the temporal correlation between the reconfiguration feature dataset and the power outage timestamps. The time interval and reconfiguration duration are extracted as correlation features, ensuring the accuracy of the temporal correlation.

[0046] S1025: If the time interval between the reconfiguration occurrence time and the power-off time is less than the preset threshold and the reconfiguration duration is greater than the preset threshold, it is determined that the reconfiguration event is triggered by a momentary power-off.

[0047] In an embodiment of the present invention, the time interval threshold is set to 100 milliseconds, and the duration threshold is set to 300 milliseconds. When the timestamp of the reconfiguration event shows that the interval between the time of occurrence and the time of power outage is less than 100 milliseconds, and the duration of the reconfiguration process exceeds 300 milliseconds, it is determined that the event is triggered by a momentary power outage. For example, in a certain data center scenario, it is monitored that the power outage occurs at 10:30:20:125 milliseconds, the reconfiguration starts at 10:30:20:160 milliseconds, lasts for 485 milliseconds, and the time interval is 35 milliseconds, which meets the threshold condition. The system confirms that the reconfiguration is a power outage trigger event. The accurate extraction of trigger features is achieved through threshold judgment.

[0048] In this embodiment of the present invention, through real-time monitoring and multi-scale feature extraction, the link reconfiguration behavior of core switches after a transient power outage can be accurately captured. Combined with timing correlation analysis and intelligent classification, the system can effectively determine the triggering source of the reconfiguration event, significantly improving the reliability of feature extraction. Compared with traditional static threshold methods, this method fully utilizes the timing characteristics of dynamic features, reduces the risk of misjudgment, and provides key triggering features for subsequent anti-islanding protection.

[0049] If a momentary power outage event is detected in S103, the operating parameters of the cooling system are collected in real time, their change rates are analyzed to determine whether a power outage is responded to, and the dynamic characteristics of the cooling system adjustment are extracted.

[0050] In this embodiment of the present invention, a momentary power outage can trigger adaptive adjustments in the cooling system, resulting in significant changes in parameters such as temperature, pressure, flow rate, and cooling power. To accurately identify the cooling system's response to a power outage, high-frequency acquisition and multidimensional feature analysis of operating parameters are required. Through time-frequency decomposition and neural network feature extraction, dynamic features of the cooling system's adjustments are generated, providing a key basis for islanding state analysis and ensuring the accuracy and reliability of feature extraction.

[0051] S1031 performs high-frequency sampling of the temperature, pressure, flow, and cooling power of the cooling system through a multi-channel collector, and generates a first operating parameter sequence using digital filtering and normalization processing.

[0052] In this embodiment of the present invention, the collector samples the cooling system's temperature, pressure, flow rate, and cooling power in real time at a frequency of 10 times per second to ensure that rapid changes in parameters are captured. The collected data is filtered through a Butterworth low-pass filter to remove high-frequency noise. The filter cutoff frequency is set to 5 Hz to retain signals related to power outage response. Subsequently, the minimum and maximum normalization method is used to map each parameter to a range of 0 to 1 to generate the first operating parameter sequence. This effectively reduces sensor noise interference and provides a high-quality data foundation for subsequent feature extraction.

[0053] S1032 verifies the validity of the first operating parameter sequence, removes abnormal data points, and uses a differential algorithm to calculate the parameter change rate to generate a first dynamic feature sequence.

[0054] In this embodiment of the present invention, to ensure data reliability, it is necessary to verify whether the first operating parameter sequence is within the physical limit range, such as temperature between -10°C and 40°C and pressure between 0 and 2.5 MPa. Data points that fall outside this range, such as outliers caused by sensor failure, are eliminated. Next, a first-order difference algorithm is used to calculate the rate of change of temperature, pressure, flow rate, and cooling power at adjacent sampling points to generate a first dynamic feature sequence. This difference calculation highlights the dynamic change trend of the parameters, providing a direct basis for response determination.

[0055] S1033 performs a threshold determination on the first dynamic characteristic sequence. If the rate of change of any parameter exceeds a preset threshold, it is determined that the cooling system responds to an instantaneous power outage, and a first response characteristic sequence is generated.

[0056] In this embodiment of the present invention, based on historical operating data statistics, the temperature change rate threshold is set at 2°C per second, the pressure change rate threshold is set at 0.5 MPa per second, the flow rate change rate threshold is set at 0.3 cubic meters per second, and the cooling power change rate threshold is set at 5 kilowatts per second. If the change rate of any parameter in the first dynamic feature sequence exceeds the corresponding threshold, the cooling system is determined to have responded to the momentary power outage. The parameter data within the response time period is extracted to generate the first response feature sequence. This threshold determination enables rapid response identification and ensures real-time feature extraction.

[0057] S1034 performs time-frequency analysis based on the first response feature sequence using wavelet transform, extracts spectrum energy features, and records parameter mutation timestamps to generate a second response feature sequence.

[0058] In an embodiment of the present invention, the Debyechen wavelet is selected as the basis function, and a four-layer wavelet decomposition is performed on the first response feature sequence, with the decomposition frequency band covering 0.625 Hz to 5 Hz. For the decomposed signal, the spectral energy distribution of each frequency band is calculated, and the main frequency characteristics reflecting the power outage response are extracted. For example, the main frequency of temperature fluctuation is between 0.1 Hz and 1 Hz, and the main frequency of pressure fluctuation is between 2 Hz and 3 Hz. A peak detection algorithm is used to locate the mutation point of each parameter, and the start and end timestamps of the mutation are recorded to generate a second response feature sequence containing spectral energy and timestamps. This step enhances the diversity of features through time-frequency analysis.

[0059] S1035 calculates the response time and the adjustment amplitude of the second response feature sequence, constructs a feature vector, and extracts the dynamic feature value of the cooling adjustment through a convolutional neural network.

[0060] In an embodiment of the present invention, the earliest parameter mutation moment is taken as the response start time, and the last parameter restoration stability moment is taken as the end time to calculate the response duration. Based on the maximum fluctuation amplitude of each parameter in the response time period, the adjustment amplitude is calculated, for example, the temperature fluctuation is 5°C, and the pressure fluctuation is 0.8 MPa. A feature vector containing the response time and adjustment amplitude is constructed and input into a five-layer convolutional neural network for feature extraction. The network structure includes an input layer, three hidden layers and an output layer. The hidden layers contain 16, 32 and 16 convolution kernels respectively. The time series features are extracted through convolution and pooling operations, and the dynamic feature values ​​of the cooling adjustment are output. For example, in a certain industrial scenario, the response time is 15 seconds and the temperature adjustment amplitude is 5°C. The network output reflects the dynamic response characteristics of the cooling system. This step improves the accuracy of feature extraction through deep learning.

[0061] Through high-frequency sampling and multidimensional feature analysis, we can accurately capture the cooling system's response to momentary power outages. Compared to traditional static thresholding methods, this step, combined with time-frequency analysis and neural networks, significantly improves the robustness and adaptability of feature extraction. For example, in a data center application, the cooling system achieved pressure stabilization within 8 seconds and temperature regulation within 12 seconds after a power outage. The system accurately extracted the 15-second response time and the adjustment range of each parameter, providing reliable dynamic feature support for islanding protection.

[0062] When S104 detects an islanding signal that is temporally associated with a link reconfiguration trigger feature or a cooling system adjustment dynamic feature, it collects electrical parameters of the anti-islanding protection device, analyzes its abnormal quantities, and extracts time-frequency features of the voltage and current signals to generate an abnormal transient feature vector.

[0063] In this embodiment of the present invention, the appearance of islanding signals is often accompanied by abnormal fluctuations in the power system's electrical parameters, such as frequency offsets, voltage variations, and phase jumps. To accurately capture these transient characteristics and correlate them with link reconfiguration or cooling system adjustments, high-precision sampling and multi-dimensional analysis of the electrical parameters of the anti-islanding protection device are required. This step extracts abnormal transient feature vectors through time-frequency analysis and neural network processing, providing a key basis for islanding status determination, ensuring comprehensive feature extraction and accurate timing correlation.

[0064] S1041 uses a high-speed data acquisition component to perform high-frequency sampling on the power frequency, voltage amplitude and phase angle of the anti-islanding protection device, and generates a first electrical parameter sequence through digital filtering and normalization processing.

[0065] In this embodiment of the present invention, a high-speed data acquisition component samples the power supply frequency, voltage amplitude, and phase angle in real time at a sampling rate of 10,000 times per second, ensuring that instantaneous changes in electrical parameters are captured. To remove high-frequency interference introduced during the sampling process, such as switching noise or electromagnetic interference, a Chebyshev bandpass filter is used with a passband range of 20 Hz to 100 Hz. Subsequently, the filtered data is linearly normalized to map the frequency, voltage, and phase angle to a range of 0 to 1, generating a first electrical parameter sequence. This step provides high-quality data support for anomaly extraction through high-precision sampling and preprocessing.

[0066] S1042 calculates the abnormality amount based on the first electrical parameter sequence, uses an adaptive threshold detection algorithm for classification, and generates a first abnormal feature sequence.

[0067] In an embodiment of the present invention, abnormal quantities include power frequency offset, voltage fluctuation amplitude, and phase jump amplitude. The offset of each parameter is calculated based on the standard operating range, such as a frequency of 50 Hz ± 0.5 Hz, a voltage rated value ± 10%, and a phase angle deviation of ± 5 degrees. For example, in a certain photovoltaic power station scenario, a frequency offset of 0.7 Hz, a voltage fluctuation of 12% of the rated value, and a phase jump of 6 degrees were detected. An adaptive threshold detection algorithm is used to dynamically adjust the threshold to distinguish between normal fluctuations and abnormal states, generating a first abnormal feature sequence. This step improves the robustness of anomaly detection through adaptive classification and adapts to parameter changes under different operating conditions.

[0068] S1043 performs Hilbert transform on the power supply voltage and current signals to extract envelope features, combines wavelet decomposition to analyze the spectral characteristics of the transient component, and generates a second abnormal feature sequence.

[0069] In an embodiment of the present invention, the Hilbert transform is used to extract the instantaneous envelope of the voltage and current signals and calculate the amplitude change rate of the envelope curve. For example, during normal operation, the voltage envelope fluctuation rate is less than 3%, and the current envelope fluctuation rate is less than 6%. However, in the island state, the voltage fluctuation rate can reach 9%, and the current fluctuation rate rises to 14%. Subsequently, the envelope curve is subjected to a five-layer wavelet decomposition using the Coiflet wavelet basis function. The decomposition frequency band covers 10 Hz to 80 Hz, and the spectral energy distribution of the transient component is extracted. Analysis shows that in the island state, the voltage signal has significant harmonic components at 30 Hz and 60 Hz, while the current signal has concentrated energy at 40 Hz. A second abnormal feature sequence containing spectral features is generated. This step enhances the characterization capability of transient features through time-frequency analysis.

[0070] S1044 analyzes the timing correlation with link reconfiguration or cooling system adjustment based on the time coordinates of the first abnormal feature sequence and the second abnormal feature sequence, and uses a recursive neural network to extract a final abnormal transient feature vector.

[0071] In an embodiment of the present invention, the first abnormal feature sequence and the second abnormal feature sequence are correlated with the link reconfiguration or cooling system adjustment features through time coordinates, and the time difference is calculated. For example, a data center records that the link reconfiguration occurs at 10:30:20, and the island signal appears at 10:30:20 180 milliseconds, with a time difference of 180 milliseconds. A long short-term memory recursive neural tart network is used. The input layer receives standardized abnormal feature data, the hidden layer contains 64 cyclic units, and the timing dependency is captured through cyclic connections. The output layer generates an abnormal transient feature vector containing 16 time domain features and 24 frequency domain features. The vector reflects the timing law that the voltage anomaly precedes the current anomaly by about 40 milliseconds, as well as the significant increase in the harmonic components at 35 Hz and 60 Hz. This step achieves accurate feature extraction and timing association through deep learning.

[0072] Through high-frequency sampling and multidimensional time-frequency analysis, it is possible to accurately capture the electrical anomaly characteristics of the power system during islanding. Compared to traditional static threshold methods, this step, combined with Hilbert transform, wavelet decomposition, and recursive neural networks, significantly improves the accuracy and robustness of feature extraction. For example, in an industrial park application, the system detected a voltage fluctuation amplitude reaching 13% of the rated value, a phase jump of 7 degrees, and an increase in the 40 Hz harmonic component. The system successfully correlated the anomaly characteristics with the link reconfiguration event, providing reliable transient feature support for subsequent islanding protection.

[0073] S105 performs pattern matching based on the abnormal transient characteristic vector and the load adjustment model constructed based on historical operation data, calculates the matching index, determines the islanding state through cluster analysis, and generates corresponding control instructions.

[0074] In this embodiment of the present invention, the abnormal transient characteristic vector reflects the dynamic characteristics of the power supply system in an islanded state. To accurately distinguish between normal load adjustment and islanding anomalies, it is necessary to perform pattern matching with the load adjustment model constructed from historical operating data, and to achieve state classification through cluster analysis. This step generates islanding isolation or preventive control instructions through multi-dimensional matching degree calculation and dynamic risk assessment, providing precise protection for the power supply system. The implementation process is described in detail below to ensure the accuracy of pattern matching and the reliability of classification results.

[0075] S1051 extracts features of the voltage and current feature sequences during the load adjustment process through a sliding window, calculates statistical indicators and performs principal component analysis to generate a first feature matrix.

[0076] In an embodiment of the present invention, a sliding window of 60 sampling points, covering a time span of 3 seconds, is used to intercept the voltage and current characteristic sequences in the historical operating data, calculate the mean, standard deviation, and volatility within the window, and generate the first load characteristic data. To ensure data quality, abnormal data segments where the voltage mean deviates from the rated value by ±10% or the standard deviation exceeds 3% are eliminated. Subsequently, the principal component analysis method is used to reduce the dimension of the valid data, and the top three principal components with a cumulative contribution rate of 95% are selected to reflect the basic characteristics, dynamic characteristics, and transient characteristics of the load, respectively, to construct the first characteristic matrix. The dimensionality reduction process reduces the computational complexity while retaining key characteristic information.

[0077] S1052 calculates a matching index between the abnormal transient characteristic vector and the load adjustment model based on the first characteristic matrix to generate first matching data.

[0078] In an embodiment of the present invention, the matching index includes Euclidean distance, cosine similarity and Pearson correlation coefficient. Euclidean distance measures the absolute difference of feature vectors in space, which is usually less than 0.25 under normal conditions; cosine similarity reflects the consistency of vector direction, and the normal value is greater than 0.94; Pearson correlation coefficient evaluates the correlation of change trends, and the normal range is above 0.88. For example, in a new energy power station, the Euclidean distance of the abnormal feature vector is 0.38, the cosine similarity is 0.86, and the correlation coefficient is 0.80, indicating a deviation from the normal mode. The calculation result generates the first matching data. The comprehensiveness of the matching is enhanced by multi-dimensional indicators.

[0079] S1053 performs nonlinear mapping and cluster analysis on the first matching degree data, and uses a density peak clustering algorithm to generate an island state classification result.

[0080] In an embodiment of the present invention, a support vector machine is used to perform nonlinear mapping with a radial basis kernel function, and the kernel function bandwidth is set to 0.9, and the three-dimensional matching data is mapped to a high-dimensional feature space to generate second matching data. Then, the density peak clustering algorithm is applied to calculate the local density of the sample and the distance to the high-density point. The local density threshold is set to 7 neighboring points, and the distance threshold is set to 0.6 standard deviations. The three categories of normal operation, mild islanding, and severe islanding are divided. For example, data analysis of a certain distribution system shows that the local density of the severe islanding category is 8, and the inter-class distance is 0.75, indicating that the classification result has a high degree of discrimination. This step achieves accurate division of the state through density clustering.

[0081] S1054 performs time series feature analysis on the clustering results, calculates the fluctuation characteristics and constructs risk assessment indicators, and generates control instructions.

[0082] In this embodiment of the present invention, a sliding window of 60 sampling points is used to cluster the results, calculating the fluctuation amplitude, frequency, and period. After normalization, the first fluctuation characteristic data is generated. Based on this, the fluctuation duration, amplitude mean, and frequency standard deviation are calculated, and combined with the Euclidean distance of samples within the cluster to generate the second fluctuation characteristic data. For example, in an islanding state, the fluctuation duration is 2.5 seconds, the amplitude mean is 10%, and the frequency standard deviation is 1.5 Hz. A risk assessment index is constructed, with weights of 0.35 for intra-cluster distance, 0.3 for inter-cluster distance, and 0.35 for fluctuation characteristics, with a risk score ranging from 0 to 1. If the score exceeds 0.82, an islanding isolation control code is generated, including circuit breaker tripping and load transfer. If the score is between 0.58 and 0.82, a preventive control code for frequency and voltage regulation is generated, and an early warning is triggered. For example, a photovoltaic power station with a risk score of 0.87 triggers an isolation command, successfully preventing the spread of islanding.

[0083] In this embodiment of the present invention, pattern matching and cluster analysis are used to effectively distinguish between normal load adjustments and islanding conditions. Compared to traditional single-threshold methods, this step, combined with principal component analysis, density clustering, and risk assessment, significantly improves classification accuracy and control reliability. For example, in an industrial park application, the system detected a mild islanding condition with a risk score of 0.73 and automatically issued a frequency increase instruction of 0.3 Hz, stabilizing system operation and providing efficient support for anti-islanding protection.

[0084] S106 performs weighted fusion based on the classification results of the islanding state to generate a final discrimination result, and uses the discrimination result to dynamically correct the power frequency offset to generate a stable frequency signal.

[0085] In this embodiment of the present invention, the islanding state classification results may be affected by various transient factors. To improve the classification accuracy, the classification results are weighted and fused. Based on the fused results, the power supply frequency offset is dynamically corrected to ensure the stability of the power supply system. This step generates a frequency stability signal containing state information through multidimensional feature fusion, neural network feature extraction, and dynamic frequency compensation. This provides reliable support for anti-islanding protection, ensuring the robustness of the classification results and the accuracy of frequency correction.

[0086] S1061 uses a sliding window to extract features from the island state classification results, calculates the confidence score and performs weighted fusion to generate the first discriminant data.

[0087] In this embodiment of the present invention, a sliding window of 60 sampling points, covering a 3-second time span, is selected to intercept the island state classification results, and three indicators within the window are calculated: fluctuation amplitude, duration, and state stability. The fluctuation amplitude reflects the dynamic changes in the classification results. For example, when it is less than 4%, it is assigned a weight of 0.85, and when it is greater than 15%, the weight is reduced to 0.25. Results with a duration exceeding 1.2 seconds or a state stability greater than 90% are assigned a high confidence score, such as 0.9. Fusion weights are assigned based on the confidence scores, and the first discriminant data is generated through weighted averaging and normalization.

[0088] S1062 verifies the first discrimination data, uses an adaptive neural network to extract features, and generates second discrimination data based on historical data training.

[0089] In this embodiment of the present invention, the first discriminant data is verified, and outlier data points with confidence scores below 0.65 are eliminated to ensure data quality. A five-layer adaptive neural network is used. The input layer receives the fused discriminant data, the three hidden layers contain 24, 48, and 24 neurons, respectively, and the output layer generates the state feature vector. The network is trained using 1500 sets of labeled data, including 1000 sets of normal operation data and 500 sets of island data. Parameters are adjusted using the Adam optimization algorithm, and the classification accuracy reaches 96% after training. The generated second discriminant data reflects the deep characteristics of the island state.

[0090] S1063 calculates the frequency offset based on the second discrimination data, dynamically updates it using a recursive least squares method, and generates first frequency compensation data.

[0091] In this embodiment of the present invention, the frequency offset is extracted based on the second discriminant data, and the offset sequence is dynamically updated using a recursive least squares method with a sampling interval of 50 milliseconds. The algorithm minimizes the prediction error through iterative optimization and calculates the frequency compensation increment. For example, in a substation scenario, if a frequency offset is detected from 50 Hz to 49.6 Hz, the algorithm estimates a compensation increment of 0.4 Hz with an update period of 150 milliseconds. The generated first frequency compensation data includes the offset and the increment sequence. This step achieves rapid tracking of frequency offset through dynamic estimation.

[0092] S1064 performs smoothing processing and dynamic correction on the first frequency compensation data to generate a frequency stabilization signal.

[0093] In this embodiment of the present invention, a low-pass filter is used to process the first frequency compensation data with a cutoff frequency set to 100 Hz to eliminate high-frequency interference and reduce the signal fluctuation amplitude from 0.25 Hz to 0.04 Hz. A correction coefficient is calculated based on the frequency change rate and deviation. For example, if the change rate exceeds 0.15 Hz per second and the deviation reaches 0.35 Hz, the correction coefficient is 1.15. The original frequency signal is then compensated to generate third frequency compensation data. Subsequently, based on the frequency stability range of 49.95 to 50.05 Hz, the data is segmented and analyzed to calculate the mean, variance, and volatility. This generates a frequency stability signal containing the frequency value, timestamp, and status indicator. For example, the output frequency of an industrial site is 50.02 Hz, the timestamp is 10:30:20:150 milliseconds, and the status indicator is stable. In this embodiment of the present invention, weighted fusion and dynamic correction can effectively improve the accuracy of islanding state detection and the real-time performance of frequency compensation. Compared with traditional fixed threshold methods, this step combines neural networks and recursive algorithms to improve the system's adaptability to complex operating conditions. For example, in a photovoltaic power station application, after detecting a frequency offset of 0.45 Hz, correction was completed within 150 milliseconds, stabilizing the frequency to 50.01 Hz, providing critical support for the reliable operation of the power system.

[0094] S107 analyzes the coupling relationship between the frequency stability signal and the voltage fluctuation through time series analysis, calculates the standard deviation of the voltage fluctuation, and corrects the protection setting of the anti-islanding protection device based on historical operating conditions and islanding status.

[0095] In this embodiment of the present invention, the coupling relationship between the frequency stability signal and voltage fluctuation reflects the dynamic behavior of the power system. To optimize the protection settings of the anti-islanding protection device, it is necessary to analyze the timing correlation between the two and evaluate voltage stability. This step generates accurate protection setting corrections through synchronous sampling, filtering, and neural network analysis, ensuring the reliability and adaptability of the protection device under complex operating conditions.

[0096] S1071 uses a sliding time window to synchronously sample the frequency stability signal and voltage fluctuation data, calculates the Pearson correlation coefficient of the frequency and voltage deviations, and generates first coupling characteristic data.

[0097] In this embodiment of the present invention, a sliding time window of 100 sampling points, covering a 5-second time span, is used to synchronously acquire frequency stability signals and voltage data. The Pearson correlation coefficient between the frequency deviation sequence and the voltage deviation sequence within the window is calculated to reflect the dynamic relationship between the two. For example, during normal operation, the correlation coefficient is typically below 0.35, but can rise to 0.7 during islanding. The correlation coefficient sequence is normalized to a range of 0 to 1 to generate the first coupling characteristic data. This step, through correlation analysis, provides the foundational data for the subsequent construction of the coupling matrix.

[0098] S1072 constructs a frequency-voltage coupling matrix based on the first coupling characteristic data, uses an adaptive Kalman filter to eliminate noise, and extracts the main characteristic components to generate second coupling characteristic data.

[0099] In an embodiment of the present invention, the first coupling feature data is organized into a frequency-voltage coupling matrix, where the matrix elements represent the correlation coefficients at different time points. An adaptive Kalman filter is used, with the noise covariance set to 0.01 and the measurement covariance set to 0.12, to filter out high-frequency noise and random interference. After filtering, the main feature components with eigenvalues ​​greater than 0.1 are extracted through singular value decomposition to generate the second coupling feature data. For example, in a photovoltaic power station, the proportion of the maximum eigenvalue after filtering increased from 60% to 82%, significantly enhancing the signal clarity. This step improves data quality through filtering and feature extraction.

[0100] S1073 performs time series segmentation on the second coupling characteristic data, calculates the voltage fluctuation standard deviation, marks the stable interval and generates a steady-state characteristic sequence.

[0101] In an embodiment of the present invention, the second coupling feature data is segmented into 1-second time windows, and the standard deviation of fluctuations in each segment of the voltage sampling data is calculated. Based on historical operating data, the mean standard deviation under normal operating conditions is 0.7%, and the variance is 0.15%. If the standard deviation is less than 1.1%, it is marked as a stable interval. For example, the standard deviation of 90% of the stable intervals of a certain distribution system is less than 1%, while it can reach 3.5% during disturbances. The frequency-voltage coupling characteristics within the stable interval are extracted to generate a steady-state feature sequence. This step accurately identifies the system's stable state through standard deviation analysis.

[0102] S1074 uses a long short-term memory network to extract features from steady-state feature sequences, combines the islanding state classification results to generate protection constant correction values ​​and encapsulate configuration signals.

[0103] In an embodiment of the present invention, a long short-term memory network is constructed. The input layer contains 48 neurons that receive steady-state feature sequences, the two hidden layers contain 96 and 48 neurons respectively, and the output layer generates a correction feature vector. The network is trained using 2,000 sets of historical data, including 1,500 sets of normal operating conditions and 500 sets of islanding data, with a classification accuracy of 97%. Combined with the islanding state classification results, dynamic protection constant correction rules are generated. For example, in a grid-connected photovoltaic power station, a standard deviation of 0.4% is detected and persists for 12 seconds. The frequency protection constant is adjusted from 49.4 Hz to 49.2 Hz, the voltage protection constant is adjusted from 89% to 87% of the rated value, and the delay is extended from 0.25 seconds to 0.35 seconds. The correction value is encapsulated as a configuration signal of 32-bit frequency constant, 32-bit voltage constant, and 16-bit delay constant, accurate to 3 decimal places. This step ensures the accuracy of constant correction through deep learning and rule generation.

[0104] In this embodiment of the present invention, through time series analysis and neural network processing, the coupling characteristics of frequency and voltage can be effectively captured and protection settings can be optimized. Compared with traditional fixed setting methods, this step, combined with Kalman filtering and long short-term memory networks, significantly improves the adaptability and reliability of correction. For example, in one industrial park, the system completed the setting adjustment when the voltage standard deviation dropped to 0.5%. The new setting avoided false operation in subsequent islanding events and ensured the stable operation of the power supply system.

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent anti-islanding method based on transient characteristics, characterized in that: The method comprises: Acquire real-time current data from UPS off-grid switching, analyze the high-frequency pulse noise spectrum in the current data, obtain spectrum characteristics, and extract the amplitude and time interval of the current spike. If the amplitude exceeds the preset amplitude threshold and the time interval is less than the preset interval threshold, it is determined that the node has experienced a momentary power outage; Identify the core switch link reconfiguration status when a transient power outage occurs, obtain the link reconfiguration timestamp, calculate the timing correlation between reconfiguration and power outage based on the moment the transient power outage occurs, and obtain the triggering characteristics of the link reconfiguration; If the rate of change of each operating parameter during the cooling system adjustment process exceeds a corresponding preset threshold, it is determined that the cooling system responds to a momentary power outage, and the operating parameters of the cooling system in response to the momentary power outage are analyzed to obtain the dynamic characteristics of the cooling system adjustment; When an islanding signal is detected that is temporally correlated with a link reconfiguration trigger feature or a cooling system adjustment dynamic feature, the abnormal electrical parameter values ​​of the anti-islanding protection device are obtained, and the time-frequency distribution of the power supply voltage and current characteristic signals are extracted to obtain an abnormal transient characteristic vector. Pattern matching is performed between the abnormal transient characteristic vector and the load adjustment model established based on statistical analysis of historical operating data to obtain a matching index. Cluster analysis is performed on the matching index to obtain the islanding state classification result. The classification results of the islanding state are weighted and fused to obtain the final discrimination result. The power supply frequency offset is dynamically corrected according to the final discrimination result to obtain a corrected frequency stability signal. By analyzing the corrected frequency stability signal and voltage fluctuations, the coupling relationship between frequency and voltage is obtained, and the voltage fluctuation standard deviation of the voltage sampling within the preset time window is calculated. If the standard deviation is less than the preset threshold, the protection setting of the anti-islanding protection device is corrected based on the historical operating condition data and the islanding state classification results.

2. The method according to claim 1, characterized in that The method of obtaining real-time current data of UPS off-grid switching, analyzing the high-frequency pulse noise spectrum in the current data to obtain spectrum characteristics, extracting the amplitude and time interval of the current spike, and determining that a momentary power outage has occurred at the node if the amplitude exceeds a preset amplitude threshold and the time interval is less than a preset interval threshold, includes: Sampling real-time current data of the uninterruptible power supply in an off-grid switching operation state, and performing median filtering on the sampled data to eliminate outliers; The filtered current data is band-limited by a band-pass filter and then decomposed at multiple scales. The amplitude distribution density function of the decomposed high-frequency pulse noise spectrum component is calculated to obtain the spectrum characteristic value. The spectral feature value is used to construct a feature vector, and a pre-trained neural network classifier is used to determine that the mean peak amplitude in the feature vector exceeds a preset amplitude threshold and the mean peak time interval is less than a preset time interval threshold to output a power-off status indicator.

3. The method according to claim 1, characterized in that The identifying of the core switch link reconfiguration state when the instantaneous power outage occurs, obtaining the link reconfiguration timestamp, and calculating the timing correlation between the reconfiguration and the power outage in combination with the instantaneous power outage occurrence time to obtain the triggering characteristics of the link reconfiguration include: The core switch link port status is polled and collected by a real-time communication agent program, and the collected data is normalized by a sliding window to obtain first link status data; Performing discrete wavelet transform on the first link state data to extract link state mutation characteristics including port state change rate and link topology change rate to obtain second link state data; Using a gradient clustering algorithm to classify the second link state data, identifying link reconfiguration events to obtain a first timestamp sequence; If the time interval between the reconfiguration occurrence time and the power off time is less than the preset time threshold and the reconfiguration duration is greater than the preset duration threshold, the reconfiguration event is determined to be a link reconfiguration event triggered by a momentary power off.

4. The method according to claim 1, wherein If the rate of change of each operating parameter during the cooling system adjustment process exceeds a corresponding preset threshold, it is determined that the cooling system responds to a momentary power outage, and the operating parameters of the cooling system in response to the momentary power outage are analyzed to obtain dynamic characteristics of the cooling system adjustment, including: After real-time acquisition and filtering and normalization processing of the operating parameters of the cooling system temperature, pressure, flow rate, and cooling power by a multi-channel collector, a differential algorithm is used to calculate the dynamic characteristic quantities of temperature change rate, pressure change rate, flow rate change rate, and cooling power change rate to obtain first dynamic characteristic data; Performing a change rate threshold determination on the first dynamic characteristic data, and if the dynamic characteristic exceeds the preset change rate threshold, determining that the cooling system responds to an instantaneous power outage to obtain first response characteristic data; A five-layer convolutional neural network is used to extract features from the first response feature data, wherein the convolutional neural network includes three hidden layers to obtain dynamic features of the cooling system regulation.

5. The method according to claim 1, wherein When an islanding signal temporally associated with a link reconfiguration trigger feature or a cooling system adjustment dynamic feature is detected, abnormal electrical parameter values ​​of the anti-islanding protection device are obtained, and the time-frequency distribution of the power supply voltage and current characteristic signals are extracted to obtain an abnormal transient characteristic vector, including: Sampling three sets of electrical parameters, namely, power supply frequency, voltage amplitude, and phase angle, and performing noise elimination on the electrical parameters to obtain a first electrical parameter sequence; Calculating a power frequency offset, a voltage fluctuation amplitude, and a phase jump amplitude based on a preset standard range according to the first electrical parameter sequence, and classifying the offsets to obtain first abnormal feature data; The envelope of the power supply voltage and current signals is extracted by Hilbert transform, and the transient component of the envelope curve is analyzed by wavelet decomposition to obtain the second abnormal characteristic data; A correlation analysis is performed based on the time coordinates of the first abnormal feature data and the second abnormal feature data, a recursive neural network is used to extract voltage and current features from the correlation analysis results, and a time spectrum distribution of the voltage and current features is calculated to obtain an abnormal transient feature vector.

6. The method according to claim 1, characterized in that The abnormal transient characteristic vector is pattern matched with the load adjustment model established based on the statistical analysis of historical operation data to obtain a matching index, and the matching index is clustered to obtain the island state classification result, including: intercepting the voltage and current characteristic sequence during the load adjustment process through a sliding window, and calculating first load characteristic data based on the voltage and current mean, standard deviation, and fluctuation rate within the window; Performing principal component analysis on the first load characteristic data, and constructing a first characteristic matrix using characteristic components whose cumulative contribution rates reach a preset threshold; Calculating a matching index between the first feature matrix and the load adjustment pattern using Euclidean distance, cosine similarity, and Pearson correlation coefficient to obtain first matching data; The local density of the sample and the distance to the nearest high-density point are calculated based on the first matching data. If the local density and the distance both meet the preset threshold conditions, the classification result is determined to be an island state classification result.

7. The method according to claim 6, characterized in that Also includes: Perform cluster analysis on the abnormal transient feature vectors to obtain initial clusters. Calculate the average matching index, volatility, and duration of the initial clusters and classify them. If the average matching index, volatility, and duration of the initial clusters are all above the corresponding preset risk threshold range, it is determined to be an island state and the island operation control instructions are activated. If it is between the preset risk ranges and is classified as a high-risk critical island, a warning signal is triggered and preventive control instructions are initiated, including: Adaptive fuzzy clustering is performed based on the Euclidean distance between abnormal transient feature vectors and the local density distribution, the silhouette coefficient is calculated using the ratio of the distance between samples to the density within the class, and the clustering result corresponding to the maximum silhouette coefficient is selected to obtain the first clustering result data; Calculating the fluctuation amplitude, fluctuation frequency, and fluctuation period of the first clustering result data using a sliding window with a length of 0.05, and normalizing the fluctuation amplitude, fluctuation frequency, and fluctuation period to obtain first fluctuation characteristic data; Calculate the fluctuation duration, fluctuation amplitude mean, and fluctuation frequency standard deviation based on the first fluctuation characteristic data, and calculate the dispersion degree of samples within the cluster using Euclidean distance to obtain the second fluctuation characteristic data; The average matching degree of each cluster is calculated for the second fluctuation characteristic data, and a risk assessment index is constructed based on the distance between samples within the cluster and the minimum distance between clusters. If the risk score is higher than the preset upper limit threshold, an island isolation control code is generated; if the risk score is between the upper and lower limit thresholds, a prevention control code is generated.

8. The method according to claim 1, characterized in that The weighted fusion of the classification results of the islanding state is performed to obtain a final discrimination result, and the power supply frequency offset is dynamically corrected according to the final discrimination result to obtain a corrected frequency stability signal, including: A sliding window is used to intercept the classification results of the island state, and the confidence score is calculated according to the fluctuation amplitude, duration and state stability of the classification results to obtain the first discriminant data; Verifying the first discriminant data, extracting features from the verified data using an adaptive neural network, and training network parameters based on the labeled historical data to obtain second discriminant data; Calculating a frequency offset according to the second discrimination data, and dynamically updating a frequency offset sequence using a recursive least squares method with a preset sampling interval to obtain first frequency compensation data; The frequency change rate and frequency deviation are calculated according to the first frequency compensation data, and the original frequency signal is compensated and corrected to obtain a corrected frequency stable signal.

9. The method according to claim 1, characterized in that The method includes: analyzing the corrected frequency stability signal and voltage fluctuation to obtain a coupling relationship between frequency and voltage, calculating the voltage fluctuation standard deviation of voltage samples within a preset time window, and correcting the protection setting of the anti-islanding protection device based on historical operating condition data and islanding state classification results if the standard deviation is less than a preset threshold. The frequency stability signal and the voltage fluctuation data are synchronously sampled using a sliding time window, and the Pearson correlation coefficient of the frequency deviation sequence and the voltage deviation sequence in the window is calculated based on the sampled data to obtain the first coupling characteristic data; constructing a frequency-voltage coupling matrix based on the first coupling characteristic data, performing noise elimination on the coupling matrix using an adaptive Kalman filter, and extracting main characteristic components to obtain second coupling characteristic data; Segmenting the voltage sampling data according to the second coupling characteristic data, and calculating the voltage fluctuation standard deviation of each data segment to obtain first standard deviation data; A long short-term memory network is used to extract features of the first standard deviation data, and a dynamic protection constant value correction rule is constructed to obtain a protection constant value correction amount.

Images