Electronic current transformer operating data calculation system

Abstract

This invention provides a data calculation system for electronic current transformer operation, relating to the field of data processing technology. The system includes: a system for scanning the device-dependent adjacency matrix using a sliding time window; calculating the local reachability density under dynamic nearest neighbor topological distance using a local outlier factor algorithm; and extracting a multidimensional anomaly feature vector set when the total harmonic distortion rate gradient enters the statistical hypothesis test rejection region or the winding temperature rise rate exceeds the material's thermal aging critical threshold. The system then uses this multidimensional anomaly feature vector set to generate a risk quantification index set through a pre-trained fault propagation probability model. Based on the risk quantification index set, a multi-objective optimization function is established, and the NSGA-II genetic algorithm is used to perform non-dominated sorting evolution on the disposal instruction chromosomes. A Pareto front solution set is used to screen dynamically optimized instruction sequences that match the real-time load fluctuation coefficient, generating a structured disposal instruction set. This invention improves the reliability of power system operation.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a system for calculating operating data of an electronic current transformer. Background Technology

[0002] Electronic current transformers have become core measurement devices in new energy power generation grid connection and flexible DC transmission scenarios. By collecting high-frequency current signals, they enable real-time sensing of critical conditions such as short-circuit current, harmonic distortion, and equipment temperature rise. However, traditional monitoring methods based on single data thresholds are insufficient to meet the condition assessment needs of complex scenarios.

[0003] Traditional anomaly detection methods, some of which rely on single thresholds, are ill-suited to the complex load fluctuations in power systems. Traditional algorithms also suffer from computational errors due to duplicate data points and loss of temporal features, and fail to consider grid topology. Furthermore, existing fault propagation models are largely based on offline simulations, making it difficult to handle real-time load fluctuations and generate timely and accurate effective response strategies. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an electronic current transformer operation data calculation system, which improves the reliability of power system operation.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] In a first aspect, an electronic current transformer operation data calculation system includes:

[0007] The acquisition module is used to acquire the raw high-frequency sampling data stream of the electronic current transformer, perform signal preprocessing and compression storage, and generate a standardized multi-dimensional time series dataset.

[0008] The building module is used to construct a device dependency adjacency matrix based on a standardized multidimensional time-series dataset and the power grid SCADA topology.

[0009] The calculation module is used to perform a sliding time window scan of the device-dependent adjacency matrix, and to calculate the local reachability density under the dynamic nearest neighbor topological distance using the local outlier factor algorithm. If the total harmonic distortion rate gradient is detected to enter the statistical hypothesis test rejection region or the winding temperature rise rate exceeds the material thermal aging critical threshold, a multidimensional abnormal feature vector set is extracted.

[0010] The training module is used to generate a set of risk quantification indicators from a multidimensional set of anomaly feature vectors through a pre-trained fault propagation probability model.

[0011] The optimization module is used to establish a multi-objective optimization function based on the risk quantification index group, and to perform non-dominated sorting evolution of the disposal instruction chromosomes using the NSGA-II genetic algorithm. The dynamic optimization instruction sequence that matches the real-time load fluctuation coefficient is selected through the Pareto front solution set to generate a structured disposal instruction set.

[0012] Furthermore, the raw high-frequency sampling data stream from the electronic current transformer is acquired, preprocessed, compressed, and stored to generate a standardized multi-dimensional time-series dataset, including:

[0013] The high-speed sampling data stream generated by the electronic current transformer is analyzed to extract the original structured dataset containing the complex harmonic spectrum sequence, the distributed temperature gradient matrix, and the ADC sampling status word.

[0014] Based on the original structured dataset, the harmonic spectrum complex sequence is scaled, the layer-related dynamic threshold is calculated based on the risk estimation principle, and soft thresholding is performed on the high-frequency detail coefficients. The time-domain waveform is transformed and reconstructed, and after verification by mean square error, a set of time-domain waveforms for noise reduction parameters is generated.

[0015] The time-domain waveform set of noise reduction parameters is encoded and compressed into a timestamp sequence. This is combined with energy threshold truncation to achieve lossy compression of the harmonic spectrum. The sampling rate of the temperature gradient data is adaptively adjusted to generate a standardized multidimensional time-series dataset with time synchronization identifier.

[0016] Furthermore, based on a standardized multidimensional time-series dataset, a device dependency adjacency matrix is ​​constructed based on the power grid SCADA topology, including:

[0017] The device geographic labels and electrical connection relationships in the standardized multidimensional time-series dataset are analyzed, and an initial device adjacency matrix is ​​constructed based on the topology file exported from the power grid SCADA system.

[0018] Based on the electrical connection relationships in the initial device adjacency matrix, the distance attenuation weight coefficient is calculated and generated; the real-time device health status data in the standardized multidimensional time series dataset is fused to calculate the health status coupling factor; the distance attenuation weight coefficient and the health status coupling factor are multiplied and fused, and the elements of the initial adjacency matrix are dynamically weighted and updated to generate a dynamic weighted device adjacency matrix with probabilistic fault propagation intensity.

[0019] By integrating the dynamic weighted device adjacency matrix with real-time operating parameters from a standardized multidimensional time-series dataset, the fault propagation weights between devices are iteratively updated to construct a device dependency adjacency matrix.

[0020] Furthermore, the device relies on an adjacency matrix for sliding time window scanning, and employs a local outlier factor algorithm to calculate the local reachability density under dynamic nearest neighbor topological distance. If the total harmonic distortion rate gradient is detected to enter the statistical hypothesis test rejection region or the winding temperature rise rate exceeds the material's thermal aging critical threshold, a multidimensional abnormal feature vector set is extracted, including:

[0021] The sliding time window width is set as an integer multiple of the power system frequency cycle, and the device-dependent adjacency matrix is ​​time-series segmented to generate a window matrix sequence. Based on the sparsity distribution characteristics of the window matrix sequence, the nearest neighbor topology search radius of each device is dynamically calculated. Based on the dynamic nearest neighbor search radius, the local outlier factor algorithm is executed to calculate the local reachability density of device nodes and generate a real-time outlier factor matrix.

[0022] The real-time outlier factor matrix is ​​analyzed. When the total harmonic distortion rate gradient enters the statistical anomaly confidence interval or the winding temperature rise rate exceeds the critical threshold of insulation thermal aging, a spatiotemporal coordinate set of anomaly events is generated, and a multidimensional anomaly feature vector set is extracted.

[0023] Furthermore, the multidimensional anomaly feature vector set is used to generate a risk quantification index set through a pre-trained fault propagation probability model, including:

[0024] The harmonic distortion acceleration component of the multidimensional anomaly feature vector set is analyzed and matched with the acceleration threshold boundary of the insulation dielectric breakdown failure mode in the equipment's historical fault database to generate the insulation breakdown risk probability; the thermal stress mutation coefficient component of the multidimensional anomaly feature vector set is analyzed and matched with the temperature rise gradient critical curve of the thermal fatigue cumulative failure mode of the winding material to generate the thermal fatigue risk probability; the data packet loss Shannon entropy component of the multidimensional anomaly feature vector set is analyzed and the relay protection maloperation risk probability value is calculated.

[0025] By integrating the probability of insulation breakdown risk, the probability of thermal fatigue risk, and the probability of malfunction, an initial probability distribution vector of equipment cascade failure is generated.

[0026] Based on the initial cascade failure probability distribution vector, the real-time stress aging coefficient of the equipment is injected for correction, generating dynamically updated equipment cascade failure probability values; based on the equipment cascade failure probability values, the remaining effective operating time window of the equipment is calculated; the equipment cascade failure probability values ​​and the remaining effective operating time window are integrated to generate a risk quantification index group.

[0027] Furthermore, based on the risk quantification index set, a multi-objective optimization function is established. The NSGA-II genetic algorithm is used to perform non-dominated sorting evolution on the disposal instruction chromosomes. A dynamically optimized instruction sequence matching the real-time load fluctuation coefficient is selected through the Pareto front solution set, generating a structured disposal instruction set, including:

[0028] A set of multi-objective optimization functions is established based on a risk quantification index group. Constraints are generated by analyzing the real-time topological connection status of the power grid. The elements of the disposal instructions are encoded into chromosome gene sequences to generate an initial genetic population.

[0029] Based on a set of multi-objective optimization functions, the NSGA-II algorithm is used to perform non-dominated sorting and crowding calculation on the initial population, and a Pareto front solution set is generated through genetic evolution. The real-time load fluctuation coefficient is fused, and a dynamic optimization instruction sequence with timeliness matching is selected from the Pareto front solution set. The instruction sequence is then decoded to generate a structured disposal instruction set.

[0030] Furthermore, a multi-objective optimization function set is established based on the risk quantification index set, the constraints for generating the real-time topology connection status of the power grid are analyzed, and the elements of the disposal instructions are encoded into chromosome gene sequences to generate an initial genetic population, including:

[0031] Based on the probability value of cascading failure of equipment and the remaining effective operating time window in the risk quantification index group as input parameters, a set of multi-objective optimization functions is constructed with the objectives of minimizing the probability of cascading failure propagation, maximizing the operating time margin of key equipment, and minimizing the load shedding amount.

[0032] Analyze the real-time topology connection status of the power grid, and based on the operation time margin constraints in the multi-objective optimization function set, extract the circuit breaker operation blocking logic relationship and the available capacity of the reactive power compensation node to generate equipment operation sequence constraints.

[0033] Based on the equipment operation sequence constraints, the circuit breaker opening and closing status commands, reactive power compensation device adjustment commands, and load transfer path commands are encoded into chromosome gene fragments; based on the remaining effective operation time window, the relay protection action time window constraints are set, and the chromosome gene fragments are combined under the time window constraints to generate the initial genetic population.

[0034] Furthermore, based on a multi-objective optimization function set, the NSGA-II algorithm is used to perform non-dominated sorting and crowding calculation on the initial population, and a Pareto front solution set is generated through genetic evolution; real-time load fluctuation coefficients are fused, and time-matched dynamic optimization instruction sequences are screened from the Pareto front solution set; the instruction sequences are decoded to generate a structured disposal instruction set, including:

[0035] Based on a set of multi-objective optimization functions, non-dominated sorting stratification is performed on the initial genetic population, and crowding distance is calculated for individuals at each stratum; parent individuals are selected based on the non-dominated sorting stratification and crowding distance.

[0036] The parent individuals are manipulated to generate the offspring population; the parent individuals and the offspring population are merged, and the non-dominated sorting and elite preservation strategies are re-executed. The evolution is iterated until the convergence condition is met, and a Pareto front solution set is generated; the execution time window of each instruction sequence in the Pareto front solution set is analyzed, and the instruction execution time matching degree is calculated based on the real-time load fluctuation coefficient.

[0037] Based on the matching degree of instruction execution time, the timeliness matching and dynamic optimization of instruction sequences are selected, and the structured disposal instruction set is generated by decoding chromosome gene fragments.

[0038] In a second aspect, a computing device includes:

[0039] One or more processors;

[0040] A storage device for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the system.

[0041] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the system.

[0042] The above-described solution of the present invention has at least the following beneficial effects:

[0043] By employing a dynamic detection mechanism combining sliding time window scanning and a local outlier factor algorithm, the shortcomings of the single threshold method in adapting to load fluctuations are addressed. Through a dual-threshold triggering mechanism using the total harmonic distortion gradient and winding temperature rise rate, the response speed for anomaly feature extraction is improved, the false positive rate of harmonic distortion is reduced, and the missed detection problem caused by the loss of time dimension features in traditional algorithms is avoided.

[0044] A pre-trained fault propagation probability model generates a set of quantitative indicators including fault rate and load loss. Combined with the multi-objective optimization strategy of the NSGA-II genetic algorithm, intelligent evolution of response commands is achieved. By matching the Pareto front solution set with the real-time load fluctuation coefficient, the generation speed of dynamically optimized command sequences is accelerated, effectively compressing fault propagation time in practical applications and improving response efficiency in emergency situations. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of an electronic current transformer operation data calculation system provided by an embodiment of the present invention.

[0046] Figure 2 This is a schematic diagram of the process for constructing the device-dependent adjacency matrix according to the present invention. Detailed Implementation

[0047] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0048] like Figure 1 As shown, an embodiment of the present invention proposes an electronic current transformer operation data calculation system, comprising:

[0049] The acquisition module is used to acquire the raw high-frequency sampling data stream of the electronic current transformer, perform signal preprocessing and compression storage, and generate a standardized multi-dimensional time series dataset.

[0050] The building module is used to construct a device dependency adjacency matrix based on a standardized multidimensional time-series dataset and the power grid SCADA topology.

[0051] The calculation module is used to perform a sliding time window scan of the device-dependent adjacency matrix, and to calculate the local reachability density under the dynamic nearest neighbor topological distance using the local outlier factor algorithm. If the total harmonic distortion rate gradient is detected to enter the statistical hypothesis test rejection region or the winding temperature rise rate exceeds the material thermal aging critical threshold, a multidimensional abnormal feature vector set is extracted.

[0052] The training module is used to generate a set of risk quantification indicators from a multidimensional set of anomaly feature vectors through a pre-trained fault propagation probability model.

[0053] The optimization module is used to establish a multi-objective optimization function based on the risk quantification index group, and to perform non-dominated sorting evolution of the disposal instruction chromosomes using the NSGA-II genetic algorithm. The dynamic optimization instruction sequence that matches the real-time load fluctuation coefficient is selected through the Pareto front solution set to generate a structured disposal instruction set.

[0054] In this embodiment of the invention, a dynamic detection mechanism combining sliding time window scanning and local outlier factor algorithm is adopted to solve the problem that the single threshold method cannot adapt to load fluctuations. By using a dual threshold triggering mechanism of total harmonic distortion gradient and winding temperature rise rate, the response speed of abnormal feature extraction is improved, the false positive rate of harmonic distortion is reduced, and the missed detection problem caused by the loss of time dimension features in traditional algorithms is avoided.

[0055] A pre-trained fault propagation probability model generates a set of quantitative indicators including fault rate and load loss. Combined with the multi-objective optimization strategy of the NSGA-II genetic algorithm, intelligent evolution of response commands is achieved. By matching the Pareto front solution set with the real-time load fluctuation coefficient, the generation speed of dynamically optimized command sequences is accelerated, effectively compressing fault propagation time in practical applications and improving response efficiency in emergency situations.

[0056] In a preferred embodiment of the present invention, the raw high-frequency sampling data stream of the electronic current transformer is acquired, preprocessed and compressed for storage, and a standardized multi-dimensional time-series dataset is generated, including:

[0057] This process involves analyzing the high-speed sampling data stream generated by the electronic current transformer to extract the original structured dataset containing the harmonic spectrum complex sequence, distributed temperature gradient matrix, and ADC sampling status word. Specifically, this includes: acquiring the original sampling data stream output by the electronic current transformer through a dedicated data receiving interface; parsing the data frame structure according to standard communication protocols to extract metadata such as frame header identifier, sampling time, and data length; and splitting the data frame payload to separate the harmonic spectrum complex sequence (containing amplitude and phase information from the fundamental to the 50th harmonic) and the distributed temperature gradient matrix (distributed along the transformer winding axis with sampling point spacing of 5mm-20mm). The original structured dataset is formed by combining the ADC sampling status word (including gain coefficient, parity bit, and other status identifiers) with the sampled data (mm). The sampling frequency is set to 1kHz-2MHz, which can be adaptively adjusted according to the power grid operating conditions. Verification operations are performed on the extracted original data: the integrity of the data frame transmission is verified by CRC cyclic redundancy check, and abnormal frames that fail the verification are removed; boundary value checks are performed on the temperature gradient matrix to remove invalid values ​​that exceed the normal operating range of -40℃ to 125℃; and continuity checks are performed on the harmonic spectrum complex sequence to ensure that the amplitude change rate of adjacent sampling points does not exceed 30% (to avoid jump data caused by sudden interference).

[0058] Based on the original structured dataset, the harmonic spectrum complex sequence is scaled and decomposed. The layer-related dynamic threshold is calculated based on the risk estimation principle, and soft thresholding is applied to the high-frequency detail coefficients. The time-domain waveform is reconstructed by transformation and verified by mean square error to generate a set of time-domain waveforms for noise reduction parameters. Specifically, the harmonic spectrum complex sequence is decomposed by wavelet fast decomposition. The number of decomposition layers is adjusted in the range of 2-8 layers according to the harmonic complexity of the power grid (a higher number of decomposition layers is taken when there are more harmonic components). During each layer decomposition, the current layer signal is divided into low-frequency approximate components and high-frequency detail components. The original sampling interval of the time series is kept unchanged during the decomposition process to ensure that no time information is lost.

[0059] Model Construction: Over 1000 sets of harmonic spectrum data from the past 12 months of normal transformer operation were selected as the sample set. Each sample set includes the coefficient matrix of each decomposition layer and the corresponding noise level label. The standard deviation, mean, and correlation of coefficients from adjacent layers were used as input features, and the manually labeled optimal noise reduction threshold was used as the output label. A risk estimation model based on random forest was constructed, containing 100-200 decision trees, each with a depth of 5-15 layers. Model Training: The sample set was divided into training and validation sets in a 7:3 ratio. In each iteration, 80% of the training samples and 50% of the feature subset were randomly selected to construct the decision trees. Mean squared error was used as the loss function, and a grid search method was used to optimize hyperparameters such as the number and depth of decision trees. The number of training iterations was set to 500-200. Training stops when the validation set accuracy stabilizes above 90% for 10 consecutive iterations. Model implementation: The trained model is embedded in an embedded processor (such as ARM Cortex-A9), which receives the coefficient features of each decomposition layer in real time and outputs dynamic thresholds: the threshold for low-frequency layers is 1.2-2.0 times the standard deviation of the coefficients of that layer, and the threshold for high-frequency layers is 0.8-1.5 times the standard deviation of the coefficients of that layer. Soft thresholding is performed on high-frequency detail coefficients. Coefficients with absolute values ​​less than the threshold are set to zero directly, and coefficients with absolute values ​​greater than the threshold are retained as the difference between the coefficient and the threshold (i.e., the coefficient value minus the threshold) as the corrected coefficient.

[0060] The inverse wavelet transform algorithm is used to reconstruct the processed low-frequency approximation coefficients and high-frequency detail coefficients to restore the time-domain waveform. The mean square error between the reconstructed waveform and the original waveform before denoising is calculated. This is done by traversing all sampling points of the two waveforms, calculating the sum of squared differences at each corresponding point, and then dividing by the total number of sampling points to obtain the error value. When the error accounts for less than 5% of the energy of the original waveform, it is considered a valid reconstruction, and a time-domain waveform set of denoising parameters is generated. If the error exceeds the limit, the decomposition layer number (±1 layer) and threshold coefficient (±0.2 times) are automatically adjusted, and the decomposition and threshold processing process is re-executed.

[0061] Select wavelet basis functions consistent with the decomposition process, and determine the reconstruction scale parameters based on the original signal sampling frequency and the number of decomposition levels; where the number of decomposition levels is 2-8, the corresponding reconstruction scale parameters are set at a ratio of 1:2. nThe scaling factor (n being the number of decomposition levels) is set to ensure that the temporal resolution of the reconstructed signal matches that of the original signal. The low-frequency approximation coefficients after soft-thresholding are dimensionally expanded by zero-padding to lengthen the coefficient sequence. Padding is done at both ends of the coefficient sequence, with the amount determined by the difference between the original signal length and the approximation coefficient length (5%-20% of the original signal length), ensuring the expanded approximation coefficient length matches the number of sampling points in the original signal. The coefficients are processed sequentially from highest to lowest decomposition level, and the coefficient length is adjusted using repeated interpolation to ensure that the length of the detail coefficients at each level is the same as the expanded low-frequency approximation coefficient length. During interpolation, the extreme points of the coefficients are kept constant, and linear padding is only applied to smooth regions. Basic weights (weight values ​​ranging from 1.0 to 1) are assigned to the low-frequency approximation coefficients. .5) Assign decreasing weights to the high-frequency detail coefficients according to the number of layers (weight of the highest layer is 0.3-0.5, and weight of the lowest layer is 0.8-1.0); The extended low-frequency approximation coefficients and the aligned high-frequency detail coefficients of each layer are superimposed point by point in time series. During superposition, the value of each sampling point is the sum of the product of the corresponding point value of each component and the weight, generating a preliminary reconstructed time-domain waveform; The preliminary reconstructed waveform is smoothed and corrected by using a sliding window (window size is 3-11 sampling points) to mean the abrupt change points in the waveform, and the correction amplitude does not exceed 10% of the value of that point; After correction, the mean square error between the reconstructed waveform and the original waveform before noise reduction is calculated. When the error ratio is less than 5%, the reconstruction is considered valid, and the time-domain waveform set of noise reduction parameters is output; If the error exceeds the standard, the weight coefficients are adjusted and superposition is performed again.

[0062] The time-domain waveform set of denoising parameters is compressed into a time-stamp sequence through encoding, and lossy harmonic spectrum compression is achieved by combining energy threshold truncation. The sampling rate of temperature gradient data is adaptively adjusted to generate a standardized multidimensional time-series dataset with time synchronization identifiers. Specifically, this includes: extracting spectral features from the denoised time-domain waveform set to identify and retain the fundamental wave and higher harmonic components with an energy percentage ≥1%; compressing the time-stamp sequence using differential encoding: for consecutively sampled timestamps, only the original value of the first timestamp is stored, and subsequent timestamps store the difference from the previous time (the difference range is limited to 1ms-100ms; if the difference exceeds this range, the original value is re-stored); combining an energy threshold truncation strategy, key harmonic spectrum data with a total energy percentage of 90%-95% is retained, and low-energy noise components are removed to complete harmonic data compression; calculating the temperature difference between adjacent sampling periods, and setting the sampling rate to 1 when the rate of change is ≤0.5℃ / s. The sampling rate is dynamically selected based on the stability of the ambient temperature (with the lower limit used when the temperature difference fluctuation is <0.2℃). When the rate of change is >0.5℃ / s, the sampling rate is automatically increased to 20Hz-100Hz (the greater the temperature difference fluctuation, the closer the sampling rate is to the upper limit). Linear interpolation is performed on the adjusted data to ensure that the time axis is continuous without breaks. The compressed harmonic data and temperature data are time-synchronized and calibrated using a GPS precise time protocol, with synchronization accuracy controlled within the range of 1μs-1ms. The data is packaged according to a preset standardized format: each data record includes a data type identifier (1 byte), a sampling time synchronization identifier (8 bytes), multi-dimensional parameter values ​​(harmonic amplitude / phase, temperature value, etc.), and a check code (2 bytes). The packaged data is stored in a circular buffer with a buffer capacity set to 10min-60min of sampled data to achieve efficient data caching and retrieval.

[0063] In the scenario of increasing the temperature gradient data sampling rate (i.e., when the original sampling interval is greater than the target sampling interval), the timestamps and temperature values ​​of two adjacent valid original sampling points are first located. Let the time of original sampling point A be t1 and the temperature be T1, and the time of original sampling point B be t2 and the temperature be T2, where the time interval between t2 and t1 ranges from 10ms to 1000ms, and both T1 and T2 are within the valid range of -40℃ to 125℃. Based on the target sampling rate, the interpolated data interval Δt (the time interval corresponding to the target sampling rate, ranging from 10ms to 50ms) is determined, and the number of interpolation points N that need to be added within the original interval is calculated. Specifically: N = float or[(t2-t1) / Δt]-1, where N ranges from 2 to 20 (when the original interval is too large, the number of interpolation points in a single segment is controlled to not exceed 20 through piecewise interpolation to avoid error accumulation); based on the linear change assumption, the temperature change slope k between adjacent original points is calculated, that is, the temperature change rate per unit time is determined by the ratio of the temperature difference between two points to the time difference; when calculating the slope, it is necessary to ensure that t2≠t1. If the timestamp is repeated, the redundant point is removed and a new valid interval is selected; starting from the starting point t1, the timestamps t1+Δt, t1+2Δt, ..., t2-Δt of the interpolation points are generated sequentially according to the target interval Δt; for each interpolation point t i (i = 1, 2, ..., N), its temperature value Tᵢ is determined by the following method: taking the initial temperature T1 as a reference, the slope k is accumulated and the time increment (t) is increased. i The product of -t1) is used to obtain the temperature values ​​at each interpolation point, ensuring that T... i A linear transition is performed between T1 and T2; the calculated N interpolation points are inserted in chronological order between the original sampling points A and B to form a continuous temperature gradient data sequence; after interpolation, a rationality check is performed: check whether the temperature difference between adjacent interpolation points is uniform (the deviation does not exceed 0.1℃), and the temperature deviation between the final interpolation point and the endpoint B does not exceed 0.5℃. If the deviation range is exceeded, the interpolation interval is readjusted or the number of segments is increased.

[0064] In this embodiment of the invention, multi-level data verification and dynamic threshold noise reduction effectively eliminate invalid data caused by electromagnetic interference and transmission anomalies, ensuring the quality of the original data; energy-sensing compression and adaptive sampling rate adjustment are adopted to reduce data storage volume and transmission bandwidth occupation while retaining key features, balancing data accuracy and storage efficiency; soft threshold processing and interpolation completion technology reduce data distortion and improve waveform reconstruction accuracy.

[0065] like Figure 2 As shown, based on a standardized multidimensional time-series dataset, a device dependency adjacency matrix is ​​constructed based on the power grid SCADA topology, including:

[0066] This paper analyzes the geographical labels and electrical connections of equipment in a standardized multidimensional time-series dataset. Based on the topology file exported from the power grid SCADA system, an initial equipment adjacency matrix is ​​constructed. Specifically, this includes: parsing the metadata fields in the standardized multidimensional time-series dataset to extract the geographical labels (including installation location coordinates and bay number, with an accuracy controlled within ±0.5m) and electrical connection relationships (such as busbar connections, outgoing line associations, and protection circuit links) of each electronic current transformer; importing the XML format topology file exported from the power grid SCADA system to extract the substation primary equipment topology structure (including the connections of circuit breakers, disconnectors, busbars, and other components). First, establish a one-to-one mapping relationship between device IDs and topology nodes, ensuring a mapping error rate ≤ 0.1%. Then, determine the matrix dimension based on the number of mapped devices. Assuming the substation contains N instrument transformers and associated devices, the initial device adjacency matrix is ​​an N×N square matrix. Matrix elements are defined as binary values: if device q and device j have a direct electrical connection (e.g., physical connection via cable or busbar), the matrix element (q, j) is assigned a value of 1; otherwise, it is assigned a value of 0. Perform symmetry verification on the matrix (electrical connections are bidirectional), remove invalid isolated nodes (isolated node percentage ≤ 5%), and generate the initial device adjacency matrix.

[0067] Based on the electrical connection relationships in the initial device adjacency matrix, distance attenuation weight coefficients are calculated and generated. Real-time device health status data from a standardized multidimensional time-series dataset are integrated to calculate the health status coupling factor. The distance attenuation weight coefficients and the health status coupling factor are multiplied and fused, and the elements of the initial adjacency matrix are dynamically weighted and updated to generate a dynamic weighted device adjacency matrix with probabilistic fault propagation strength. Specifically, this includes: extracting physical distance parameters (cable length or busbar span) between devices based on the electrical connection relationships in the initial adjacency matrix, and using a piecewise attenuation function... The distance attenuation weighting coefficient is calculated as follows: when the distance between two devices is ≤50m, the coefficient is 0.8-1.0; when the distance is between 50m and 200m, the coefficient decreases linearly to 0.4-0.7; when the distance is >200m, the coefficient is 0.1-0.3. After the coefficient is calculated, it is normalized to ensure that all coefficients fall within the range of [0, 1]. Real-time health status data of the equipment is extracted from the standardized multidimensional time series dataset, including parameters such as winding temperature rise rate (℃ / min), total harmonic distortion rate (%), and insulation resistance (MΩ).

[0068] The piecewise decay function is used to dynamically adjust the crossover and mutation probabilities in genetic operations. It is divided into three stages based on the number of generations: initial exploration stage (generations 1-20), local optimization stage (generations 21-60), and convergence fine-tuning stage (generation 61 to the final generation). This division is based on the population diversity variation patterns in historical training data. In the initial stage, a higher probability is maintained to enhance exploration capabilities, while the probability is gradually reduced in the later stages to stabilize convergence. The base probability settings are: an initial crossover probability of 0.85 and an initial mutation probability of 0.04. The decay coefficient is calculated linearly, with the coefficient value being 0.8 + 0.01 × (current generation - 1), ranging from 0.8 to 0.99. For example, the coefficient for generation 1 is 0.8. The coefficient for generation 10 is 0.89, and the coefficient for generation 20 is 0.99. Actual probability calculation: Current crossover probability = base crossover probability × decay coefficient, current mutation probability = base mutation probability × decay coefficient, with the result rounded to two decimal places (e.g., for generation 5, crossover probability = 0.85 × 0.84 = 0.714, rounded to 0.71). Base probability adjustment: The base crossover probability is reduced to 0.75, and the base mutation probability is reduced to 0.03. Decay coefficient calculation: A step-wise decay is used, with the coefficient decreasing by 0.05 every 10 generations. The initial coefficient is 0.95 (generations 21-30), 0.90 for generations 31-40, 0.85 for generations 41-50, and 0.80 for generations 51-60. Actual probability calculation: Current probability = base probability × The decay coefficient is set at a minimum, with a crossover probability of at least 0.6 and a mutation probability of at least 0.02 (e.g., in generation 45, the crossover probability = 0.75 × 0.85 = 0.6375, so we take 0.64). The base probability is fixed: the crossover probability remains at 0.6, and the mutation probability remains at 0.02. The decay coefficient is calculated using exponential decay: coefficient = 0.8 × e(-0.01 × (current generation - 60)), with a range of 0.5-0.8. For example, the coefficient for generation 61 is approximately 0.792, for generation 80 it is approximately 0.67, and for generation 100 it is approximately 0.56. The actual probability is calculated as: current probability = base probability × decay coefficient, rounded to two decimal places, with a crossover probability ≥ 0.5 and a mutation probability ≥ 0.01 (e.g., in generation 70, the mutation probability = 0). .02 × 0.73 = 0.0146, take 0.01); Probability truncation: If the calculation result exceeds the preset range (crossover probability 0.5-0.9, mutation probability 0.01-0.05), then take the nearest boundary value. For example, if the calculated crossover probability of a certain generation is 0.92, it is automatically truncated to 0.9; When the actual number of generations exceeds the preset maximum number of generations, the crossover probability is forcibly set to 0.5 and the mutation probability is set to 0.01 to ensure stable termination of the algorithm; In the genetic operation module of the NSGA-II algorithm, the piecewise decay function is called through conditional judgment statements, and the current probability value is automatically calculated and updated to the algorithm parameter library before each generation of evolution; Through 5 sets of different power grid scenarios, it is ensured that the population diversity index (Shannon entropy) is maintained at 1 after probability adjustment at each stage.Between 2 and 2.0, it improves diversity by more than 30% compared to the fixed probability strategy, and accelerates convergence speed by 20%.

[0069] A health status assessment model is constructed, and the model training process is as follows:

[0070] Training data preparation: Collect normal operation data (≥1000 sets) and pre-fault data (≥200 sets) from the past 3 years, and divide them into training and validation sets in a 7:3 ratio; determine the weights of each parameter using the analytic hierarchy process (AHP), with temperature rise rate accounting for 35%-45%, harmonic distortion rate accounting for 25%-35%, and insulation resistance accounting for 20%-30%; when the health condition is normal, the coupling factor is 0.8-1.0; when a single parameter exceeds the limit, the factor is reduced to 0.5-0.7; when multiple parameters exceed the limit, the factor is reduced to 0.2-0.4; the model validation accuracy must be ≥90% before it can be put into use; multiply the distance attenuation weight coefficient with the health condition coupling factor element by element to obtain dynamic weight values ​​(range 0-1); replace the elements with a value of 1 in the initial adjacency matrix with the dynamic weight values; leave the elements with a value of 0 unchanged, generating a dynamic weighted device adjacency matrix containing probabilistic fault propagation intensity.

[0071] This method integrates a dynamic weighted device adjacency matrix with real-time operating parameters from a standardized multidimensional time-series dataset. It iteratively updates the fault propagation weights between devices and constructs a device dependency adjacency matrix. Specifically, this involves: extracting real-time operating parameters from the standardized multidimensional time-series dataset, including load current amplitude, power factor, and voltage deviation; constructing parameter influence factors: when load current > 80% of rated value, the influence factor is 1.1-1.3; when power factor < 0.9, it is 1.05-1.2; when voltage deviation > ±5%, it is 1.05-1.2; and when there are no abnormalities, the influence factor is 1.0. The dynamic weighted device adjacency matrix and the real-time operating parameter influence factors are then multiplied element-wise to obtain the initial fault propagation weights. Iterative update rules are then set. In each iteration, if the anomaly of the real-time parameters of device q changes by more than 10% compared to the previous cycle, its weight value with that of associated device j is adjusted by ±5%-15% (increasing the weight when the anomaly worsens and decreasing it when it eases). The number of iterations is set to 10-50 times, stopping when the weight change rate between two adjacent iterations is less than 2%, ensuring that the weight convergence is stable. The matrix after iteration is standardized so that the sum of the elements in each row is 1 (representing the probability distribution of the transmission of a device's fault to other devices). The matrix is ​​verified through a historical fault case library: ≥50 typical fault cases are selected to verify that the matching rate between the high-risk transmission path predicted by the matrix and the actual fault propagation path is ≥85%. If the condition is not met, the iteration parameters are adjusted and recalculated, and finally, a device dependency adjacency matrix is ​​generated.

[0072] In this embodiment of the invention, a dynamic weight matrix is ​​constructed by combining geographical topology and real-time health data, solving the problem that traditional static topology matrices cannot reflect changes in equipment status and improving the accuracy of fault propagation path prediction. The hierarchical weighting and iterative update mechanism ensures that the matrix can respond to changes in operating parameters in real time, shortening the weight adjustment delay time and meeting the dynamic monitoring needs of power systems.

[0073] In a preferred embodiment of the present invention, the device is scanned using a sliding time window based on the adjacency matrix, and the local reachability density under dynamic nearest neighbor topological distance is calculated using the local outlier factor algorithm. If the total harmonic distortion rate gradient is detected to enter the statistical hypothesis test rejection region or the winding temperature rise rate exceeds the material thermal aging critical threshold, a multidimensional abnormal feature vector set is extracted, including:

[0074] A sliding time window width is set as an integer multiple of the power system frequency cycle. The device dependency adjacency matrix is ​​time-series segmented to generate a window matrix sequence. Based on the sparsity distribution characteristics of the window matrix sequence, the nearest neighbor topology search radius for each device is dynamically calculated. A local outlier factor algorithm is executed based on the dynamic nearest neighbor search radius to calculate the local reachability density of device nodes and generate a real-time outlier factor matrix. Specifically, this includes: setting the sliding time window width as an integer multiple of the power system frequency cycle, specifically n times the power system frequency cycle (n is an integer from 1 to 5, corresponding to a time range of 0.02s-0.1s, with the power system frequency cycle set to 0.02s according to a 50Hz standard); dynamically adjusting the value of n according to the grid operating conditions: when the system load fluctuation frequency is high, n takes a value of 1-2; when the load is stable, n takes a value of 3-5; and segmenting the device dependency adjacency matrix according to the set time window width. Time-series segmentation is performed, with the sliding step size set to 1 / 3-1 / 2 of the window width (i.e., when the window width is 0.02s, the step size is 0.0067s-0.01s). After each sliding, matrix data within the current time window is extracted to generate a continuous window matrix sequence. The sequence length dynamically expands according to the monitoring duration, and the storage size of a single sequence is controlled within the range of 10-50 window matrices. A local outlier factor algorithm is executed based on the dynamic nearest neighbor search radius: for each device node, neighboring nodes within the search radius are selected in the current window matrix, and the number of neighboring nodes is counted; the relative distance between nodes is calculated through the weight distribution between neighboring nodes, thereby obtaining the local reachability density of each device node, with the density value range mapped to the 0-1 interval; the local reachability densities of each device node are arranged according to the original matrix topology to generate a real-time outlier factor matrix.

[0075] The system analyzes the real-time outlier matrix. When the total harmonic distortion (THD) gradient enters the statistical anomaly confidence interval or the winding temperature rise rate exceeds the critical threshold for insulation thermal aging, it generates a spatiotemporal coordinate set of anomaly events and extracts a multidimensional anomaly feature vector set. Specifically, for the THD gradient, a statistical anomaly confidence interval is set at a 95% confidence level, and hypothesis testing is performed using the average gradient of three consecutive window matrices. For the winding temperature rise rate, the material thermal aging critical threshold is set to the maximum allowable temperature rise rate of the insulation material. When the THD gradient is detected to enter the statistical anomaly confidence interval or the winding temperature rise rate exceeds the critical threshold, the timestamp of the anomaly and the corresponding equipment number are recorded, generating a spatiotemporal coordinate set of anomaly events. Feature parameters corresponding to the event are extracted from the standardized multidimensional time-series dataset, including the harmonic component amplitude, phase deviation, temperature distribution gradient, outlier value, and changes in the correlation weight of adjacent equipment at the time of the anomaly, forming a multidimensional anomaly feature vector set with the vector dimension controlled within the range of 8-15 dimensions.

[0076] In this embodiment of the invention, the dynamic time window and adaptive search radius design address the insufficient adaptability of fixed windows under complex operating conditions, improving the real-time performance and accuracy of anomaly detection and reducing misjudgments caused by load fluctuations. The anomaly determination mechanism, integrating topological relationships and multi-dimensional features, upgrades from single-parameter monitoring to multi-factor correlation analysis, enhancing the ability to identify latent faults. The structured anomaly feature vector set provides accurate input for subsequent fault propagation modeling, ensuring the reliability of risk quantification assessment.

[0077] In a preferred embodiment of the present invention, a risk quantification index set is generated from the multidimensional anomaly feature vector set through a pre-trained fault propagation probability model, including:

[0078] The following steps are taken to generate the insulation breakdown risk probability: First, analyze the harmonic distortion acceleration component of the multidimensional anomaly feature vector set and match it with the acceleration threshold boundary of the insulation breakdown failure mode in the equipment's historical fault database. Second, analyze the thermal stress mutation coefficient component of the multidimensional anomaly feature vector set and match it with the temperature rise gradient critical curve of the thermal fatigue cumulative failure mode of the winding material to generate the thermal fatigue risk probability. Third, analyze the data loss Shannon entropy component of the multidimensional anomaly feature vector set to calculate the relay protection maloperation risk probability value. Specifically, this includes analyzing the harmonic distortion acceleration component of the multidimensional anomaly feature vector set and extracting the trend data of this component over three consecutive sampling periods. The extracted data was compared with sample data of insulation dielectric breakdown failure modes in the equipment's historical fault database. This database covers over 200 insulation breakdown cases of similar equipment over the past 5 years, including harmonic distortion acceleration characteristics under different operating conditions. Ten threshold intervals were divided at 5% intervals, each corresponding to a different insulation breakdown risk level. When a harmonic distortion acceleration component falls into a certain threshold interval, the corresponding risk probability benchmark value for that interval was matched, and then corrected based on the equipment's operating years to generate an insulation breakdown risk probability, with the probability value controlled between 0 and 1. The thermal characteristics of the multidimensional anomaly feature vector set were analyzed. The stress mutation coefficient component is analyzed, and its maximum and average values ​​are obtained within the sliding time window. The critical temperature rise gradient curve of the cumulative failure mode of thermal fatigue for the winding material is retrieved. This curve is plotted based on material handbook parameters and data from over 1000 thermal cycles, including critical values ​​for thermal fatigue in different temperature ranges. The maximum value of the thermal stress mutation coefficient component is matched with the corresponding temperature point on the critical curve. When the maximum value exceeds the critical value, a base probability is calculated proportionally to the excess value. This probability is then adjusted based on the real-time ambient temperature (the probability increases by 3%-6% for every 2°C increase in ambient temperature above the standard temperature) to generate a thermal fatigue risk probability. The probability value range is specified in the original text. The range is 0-1; the Shannon entropy component of the data packet loss in the multidimensional abnormal feature vector set is analyzed, and the packet loss rate and Shannon entropy change rate within 10 consecutive data frames are statistically analyzed; according to the communication reliability standard of relay protection device, the normal fluctuation range of Shannon entropy is set to [0.2, 0.8]; when the Shannon entropy component exceeds this range, the excess amplitude is calculated. For every 0.1 increase in the excess amplitude, the basic false trip probability increases by 8%-12%. The corresponding relationship between packet loss rate and false trip probability in historical false trip cases is referenced (for every 1% increase in packet loss rate, the false trip probability increases by 5%-8%). The relay protection false trip risk probability value is generated comprehensively, and the probability value range is 0-1.

[0079] The initial probability distribution vector of cascaded equipment faults is generated by integrating the probabilities of insulation breakdown risk, thermal fatigue risk, and relay protection maloperation risk. Specifically, this involves: integrating the previously generated probabilities of insulation breakdown risk, thermal fatigue risk, and relay protection maloperation risk, and using a weighted summation method to generate the initial probability distribution vector of cascaded equipment faults; the weighting coefficients are set according to the degree of influence of the three risks on cascaded equipment faults, with the weight of insulation breakdown risk being 0.4-0.5, thermal fatigue risk being 0.3-0.4, and relay protection maloperation risk being 0.2-0.3. The weights can be fine-tuned according to the equipment type (such as busbar transformers and outgoing line transformers); and arranging the probability values ​​in order of equipment number to form the initial probability distribution vector.

[0080] Based on the initial cascade failure probability distribution vector, the real-time stress aging coefficient of the equipment is injected for correction, generating dynamically updated equipment cascade failure probability values. The remaining effective operating time window of the equipment is calculated based on these cascade failure probability values. The cascade failure probability values ​​and the remaining effective operating time window are integrated to generate a risk quantification index group, specifically including: the real-time stress aging coefficient of the equipment is injected for correction based on the initial cascade failure probability distribution vector; the real-time stress aging coefficient is calculated by analyzing parameters such as equipment runtime, cumulative load, and ambient humidity in a standardized multi-dimensional time-series dataset. For every 1000-hour increase in runtime, the coefficient increases by 0.02-0.05; for every 10% increase in cumulative load beyond the rated load, the coefficient increases by 0.03-0.06; and for every 60% increase in ambient humidity, the coefficient increases by 0.01-0.03. The revised formula is: Dynamic probability value = Initial probability value × (1 + Real-time stress aging coefficient), generating dynamically updated equipment cascade failure probability values, with the probability value range remaining between 0 and 1. Based on the equipment cascade failure probability values, a probability threshold range is set: when the probability value ≤ 0.3, the remaining effective operating time window is 24-48 hours; when 0.3 < probability value ≤ 0.6, the time window is 6-24 hours; when 0.6 < probability value ≤ 0.8, the time window is 1-6 hours; when the probability value > 0.8, the time window is 0-1 hour. This is then adjusted in conjunction with historical equipment failure handling time data to generate the final remaining effective operating time window. The equipment cascade failure probability value and the remaining effective operating time window are integrated, supplemented with equipment number, abnormal feature vector extraction timestamp, current values ​​of key abnormal parameters, etc., and encapsulated according to a preset data format (including indicator name, value, unit, and confidence level) to generate a risk quantification indicator group. Each indicator in the indicator group is accompanied by a ±5% error range indicator to ensure data reliability.

[0081] A fault propagation probability model is constructed. The input layer consists of key components of a multidimensional anomaly feature vector set (harmonic distortion acceleration, thermal stress mutation coefficient, data loss Shannon entropy, etc.). The hidden layer includes equipment status assessment nodes and fault type identification nodes. The output layer consists of various risk probabilities and cascading fault probabilities. The connection strength between network nodes is initialized based on the physical connection relationship of the equipment and historical fault propagation path data, with a connection strength range of 0-1. The training data consists of equipment operation data and fault records from the past 3 years, totaling more than 5000 valid samples, which are divided into training and validation sets in a 7:3 ratio. During training, the anomaly feature vectors of historical samples are first input into the model. The conditional probability table between nodes is adjusted using the maximum likelihood estimation method. The model prediction error is calculated every 100 iterations. Training is stopped when the error is less than 5% for 3 consecutive iterations. After training, the model is tested using validation set data to correct parameters and ensure that the prediction accuracy of the model under different operating conditions is ≥90%. The trained model is deployed on the local edge computing node of the substation and integrated into the risk assessment module of the data processing system. The model receives multidimensional anomaly feature vector set data in real time, calls the pre-stored conditional probability table for inference calculation, and the time for a single inference is controlled within 100ms to meet the real-time requirements; a model periodic update mechanism is set up, and new fault data is imported every quarter for fine-tuning to ensure the model's adaptability.

[0082] In this embodiment of the invention, the multi-dimensional risk probability generation mechanism combines historical fault data with real-time features to achieve refined quantification of risk, avoiding the limitations of single-indicator assessment and improving the accuracy of risk assessment. The dynamic correction mechanism introduces a real-time stress aging coefficient, enabling fault probability assessment to adapt to changes in equipment operating status and solving the problem that static models struggle to cope with complex operating conditions. The calculation of the remaining effective operating time window provides clear time constraints for fault handling, shortening the fault response decision time.

[0083] In a preferred embodiment of the present invention, a multi-objective optimization function is established based on a risk quantification index set. The NSGA-II genetic algorithm is used to perform non-dominated sorting evolution on the disposal instruction chromosomes. A dynamically optimized instruction sequence matching the real-time load fluctuation coefficient is selected through the Pareto front solution set to generate a structured disposal instruction set, including:

[0084] A multi-objective optimization function set is established based on a risk quantification index set. Constraints are generated by analyzing the real-time topology connection status of the power grid. The elements of the disposal instructions are encoded into chromosome gene sequences to generate an initial genetic population. Specifically, this includes: extracting the cascade failure probability values ​​(range 0-1.0) of each device from the risk quantification index set, assigning weights according to device importance with an average value of 0.6-0.8; and assigning weights of 0.2-0.4 to ordinary devices (such as branch line transformers). When calculating the function value, the failure probability of a single device is multiplied by its corresponding weight and then summed. For example, if hub device A has a probability of 0.7 and a weight of 0.7, and ordinary device B has a probability of 0.5 and a weight of 0.3, then the function value = 0.7 × 0.7 + 0.5 × 0.3 = 0.64. The target control value is ≤0.5 (low-risk threshold). Real-time load data from a standardized multi-dimensional time-series dataset is collected every 5 seconds, and the average load within 1 minute is used as the baseline value. Predict the load transfer amount after the execution of the disposal instruction, and calculate the deviation rate = (absolute value of transferred load / rated load of equipment) × 100%, with a constraint deviation rate ≤ 20%. For example, for equipment with a rated load of 1000A, the transferred load ≤ 200A, and the function value is this deviation rate, with a target deviation rate ≤ 10%. Assign time cost coefficients according to the urgency of the operation: 1.0-1.2 for routine operations, and 1.5-2.0 for emergency operations (risk probability > 0.6). Function value = number of operation steps × coefficient, for example, the function value of 3-step emergency operation = 3 × 1.8 = 5.4, with a target control function value ≤ 20. Extract the equipment connection relationship table from the SCADA topology file, mark directly connected equipment pairs, and only allow load transfer between marked pairs. In the equipment capacity constraint, the rated capacity is taken from the equipment parameter ledger, and the transfer amount must be ≤ ledger value × 80%, for example, the maximum transfer amount of a 1000A equipment ≤ 800A. Operation sequence constraints are implemented through timing logic. The implementation of the verification process requires manual input of the timestamp difference between disconnect and connect commands, which must be ≥5s. The winding temperature rise rate is calculated using a temperature gradient matrix, with the temperature difference value taken every 10s. The rate is calculated as (current temperature - temperature 10s ago) / 10s, with a mandatory constraint of ≤0.5℃ / min. The total harmonic distortion rate is extracted from the noise-reduced time-domain waveform and calculated according to standards, with a constraint of ≤5%. Commands exceeding this limit are automatically marked as invalid. The operation type gene uses a fixed 2-bit binary code (00 / 01 / 10), corresponding to the three types of operations in the preset operation list. The equipment number gene length is determined by the maximum number of devices in the substation (e.g., 50 devices require a 6-bit binary code, ranging from 000001 to 110010). The execution time gene uses a 3-bit decimal code (001-010), corresponding to 1-10min. The entire chromosome length is 2+6+3=11 bits.

[0085] A random generation and historical filtering method is employed: First, 80 chromosomes are randomly generated, then filtered through two layers of checks: operation type and device type matching check (e.g., "parameter adjustment" is only allowed to be combined with the gene of the current transformer number); execution time conflict check (the difference in execution time genes between any two chromosomes is ≥2 minutes). 20% of the instructions are extracted from the historical valid instruction library to supplement the population, ensuring that the proportion of valid genes is ≥60% (valid genes refer to gene positions that have passed the verification), thus generating the initial population.

[0086] Based on a multi-objective optimization function set, the NSGA-II algorithm is used to perform non-dominated sorting and crowding calculation on the initial population, and a Pareto front solution set is generated through genetic evolution. Real-time load fluctuation coefficients are fused, and time-matched dynamic optimization instruction sequences are screened from the Pareto front solution set. The instruction sequences are decoded to generate a structured disposal instruction set, specifically including: pairwise comparison of 80 chromosomes in the population, for example, comparing chromosomes X and Y: if the risk function value of X ≤ Y, the load function value ≤ Y, the cost function value ≤ Y, and at least one function value < Y, Then X dominates Y; classify all chromosomes according to their dominance relationship, with undominated chromosomes classified as level 1 (approximately 10-15 chromosomes), chromosomes dominated by level 1 classified as level 2 (approximately 20-25 chromosomes), and so on down to levels 3-5; for level 1 chromosomes, arrange them in ascending order of function value along the risk function dimension, taking the first and last chromosomes as boundaries, and the crowding distance of intermediate chromosomes = (value of subsequent chromosomes - value of preceding chromosomes) / (maximum value - minimum value); similarly calculate the distances along the load and cost dimensions, with crowding degree being the mean of the three-dimensional distances, ranging from 0 to 1.0, for example, a certain The three-dimensional distances of the chromosomes are 0.3, 0.5, and 0.4, respectively, and the crowding degree is (0.3 + 0.5 + 0.4) / 3 = 0.4. A tournament selection method is used, with three chromosomes randomly selected from the population, prioritizing chromosomes with lower rank. If rank is the same, chromosomes with higher crowding degree are selected. The probability of selecting a rank 1 chromosome is 0.4, rank 2 is 0.25, and rank 3 and below is ≤0.15, ensuring that high-quality chromosomes enter the mating pool first. Chromosomes in the mating pool are paired with a crossover probability of 0.8, starting at the 5th position of the gene sequence (operation type and setting). Single-point crossover is performed at the gene boundary (within the specified number). For example, parent 1: 00010101005, parent 2: 10110010008, after crossover, offspring 1: 00110010005, offspring 2: 10010101008. After crossover, the device type matching needs to be re-verified. Gene positions are randomly selected with a mutation probability of 0.03. When the operation type gene mutates, it only switches between 00 / 01 / 10. The mutation range of the device number gene is ≤±5, and the execution time gene mutation is ±1min (still maintaining the range of 1-10min). For example, the original gene 005 mutates to 006, and 001 mutates to 002.

[0087] Each generation of evolution retains all non-dominated solutions (approximately 15-20) from both parents and offspring. The remaining slots are selected from other chromosomes based on "rank priority + crowding density," maintaining a population size of 80 chromosomes. After 50 iterations, if the objective function value of the rank 1 chromosome fluctuates by ≤5% for five consecutive generations (e.g., the risk function value stabilizes between 0.38 and 0.40), iteration stops. The average crowding density of the rank 1 chromosome is calculated. When the average crowding density is ≥0.6 (e.g., the average crowding density of 10 chromosomes is 0.65), the solution set is considered uniformly distributed, and a set containing 15 chromosomes is generated. The Pareto front solution set of non-dominated solutions is obtained; the real-time load fluctuation coefficient (current value 1.1) is extracted from the standardized dataset, and the load impact prediction value of each instruction in the Pareto solution set is calculated (e.g., the prediction coefficient of a certain instruction is 1.08). The deviation rate is approximately 1.8% ≤ 10% = |1.08-1.1| / 1.1×100% ≤ 1.8%, and is marked as a candidate solution; a total of 5 candidate solutions are selected; the risk reduction rate of the candidate solutions is calculated. For example, if the initial risk value is 0.6 and the optimized risk value is 0.3, the reduction rate is (0.6-0.3) / 0.6 = 50% ≥ 20%. Sort by risk reduction rate in descending order and cost function value in ascending order, and select the top 3 as optimal candidate solutions; each instruction generates a unique ID (e.g., "CMD20250818153001"), and parameter values ​​are checked according to constraints (e.g., load transfer 700A≤800A). A checksum (4-digit hexadecimal) is calculated for the instruction content (ID+time+object+type+parameter), and the final instruction format is "ID|time|object|type|parameter|checksum"; the population size is dynamically adjusted based on the number of power grid equipment and instruction complexity, with 80-120 chromosomes for single substation scenarios and 120-150 chromosomes for regional power grid scenarios; this range is verified through 200 sets of simulation experiments to ensure that the solution set convergence is achieved within 30-50 generations, and the computational resource utilization rate is controlled at 60%-80%.

[0088] An adaptive adjustment mechanism is adopted, with the initial value set at 0.75-0.85. When the population diversity index (mean crowding) is <0.4, it is automatically increased to 0.85-0.9 to increase the probability of gene recombination; when the diversity index is >0.8, it is reduced to 0.6-0.75 to stabilize high-quality genes. The base value is set at 0.02-0.03, reducing the mutation probability of chromosomes with dominance level 1 (optimal solution) to 0.01-0.015, and increasing the mutation probability of chromosomes with dominance level ≥3 to 0.04-0.05. According to the severity of the fault, the high-risk scenario (cascading fault probability >0.6) is set to 30-50 generations, and the medium-low risk scenario is set to 50-80 generations, ensuring that the decision response time of the high-risk scenario is ≤30s.

[0089] For any two chromosomes A and B, if the risk minimization function value of A is less than or equal to B, the load effect function value is less than or equal to B, and the operation cost function value is less than or equal to B, and at least one function value is strictly less than B, then A is determined to dominate B. This rule is implemented using a logical decision tree to avoid numerical calculation errors, with the decision time controlled to 0.01-0.05ms per chromosome. All chromosomes are initialized to an unsorted state. Chromosomes without dominance are traversed and marked as level 1. After removing level 1 chromosomes, chromosomes without dominance are repeatedly marked as level 2 in the remaining population. This process continues until all chromosomes are classified, with the number of levels controlled to 3-5 (lower-level chromosomes are merged when there are more than 5 levels). Chromosomes within each dominance level are sorted in ascending order of their objective function values. Boundary chromosomes (minimum / maximum) are assigned infinite crowding, and the crowding of intermediate chromosomes is the sum of the distances between adjacent chromosomes in each objective dimension. The distance calculation range is limited to the actual interval of the objective function values ​​within that level. The crowding values ​​are normalized to the 0-1.0 range based on the maximum value within the level for easier cross-level comparison. Chromosomes with a crowding ≥0.6 after normalization are marked as high-diversity solutions and are preferentially retained during the selection operation.

[0090] We collected power grid fault handling cases from the past three years, covering substations of different voltage levels (220kV, 110kV, and 35kV), including three scenarios: normal operating conditions, minor faults, and severe faults, with a total sample size of over 1000 groups. Each sample group includes original risk indicators, constraints, historical best command sequences, and execution effect evaluation data. Outlier cleaning (removing invalid cases with operation costs > 50 or risk reduction rate < 10%) and standardization (mapping risk values ​​and load impact values ​​to the 0-1 range) were performed on the original data. The data was then divided into a training set (700+ sets) and a validation set (300+ sets) in a 7:3 ratio. The convergence speed and quality of the validation set's solution set were used as optimization metrics, and core parameters were adjusted using a controlled variable method. With a fixed population size of 100, combinations with crossover probabilities of 0.6-0.9 and mutation probabilities of 0.01-0.05 were tested, and validation set metrics for each combination were recorded to determine the initial parameter range. An adaptive parameter mechanism was introduced, and the training model dynamically adjusted parameters based on population diversity. Thresholds (e.g., diversity thresholds of 0.4 and 0.8) were optimized through 50 iterations. The optimized model was tested on the validation set, requiring a convergence speed improvement of ≥ 20% and a reduction in the mean of the optimal solution's objective function of ≥ 15%; otherwise, the model was returned to the parameter readjustment range. The process involves: performing 10 independent evolutions on each validation set sample, calculating the standard deviation of the convergent generation (≤5); adding 5%-10% random noise to the input data, testing the effectiveness of the model's output instructions (the proportion of risk reduction ≥20% after execution), requiring an effectiveness rate ≥90%; otherwise, strengthening the noise tolerance of the mutation operation; developing the core algorithm module in C++ (execution efficiency ≥1 million times / second), implementing data preprocessing and result visualization in Python, and deploying it on an industrial-grade server (CPU ≥8 cores, memory ≥16GB), supporting multi-threaded parallel computing (simultaneously processing 3-5 substation scenarios); receiving standardized multi-dimensional time-series datasets and risk quantification indicator groups, with data transmission latency ≤100ms; sending structured disposal instruction sets to the SCADA system, including instruction ID, execution time, checksum, etc., with interface compatibility conforming to the IEC61850 standard.

[0091] Computational acceleration strategies: Bitmap index optimization is used for dominance sorting (storing chromosome comparison results as binary bits), improving query efficiency by 50%; a pre-sorting caching mechanism is used for congestion calculation to reduce redundant computation time; a "priority calculation" mode is triggered in high-risk scenarios, automatically allocating ≥50% of CPU resources to ensure the evolution process takes ≤30 seconds; a "power-saving mode" is used in low-risk scenarios, keeping resource utilization below 30%; dual verification is performed on the generated disposal instruction set, including syntax verification (checking device number and time format validity); automatic re-evolution is triggered when verification fails; when the model calculation is abnormal (e.g., failing to converge 3 times consecutively), it automatically switches to the historical best instruction library matching mode, retrieving the instruction sequence of the most similar case based on the current scenario characteristics to ensure the continuity of disposal instructions.

[0092] In this embodiment of the invention, a multi-objective function weighting mechanism achieves a quantitative balance between risk, load, and cost, improving overall benefits compared to single-objective optimization and avoiding excessive load fluctuations caused by overemphasis on risk reduction. The parameter constraints and verification mechanism of the genetic operation accelerates algorithm convergence, ensuring timely fault handling. Real-time load factor matching and priority ranking enhance the stability of power grid operation.

[0093] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0094] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the system as described above. All implementations in the above system embodiments are applicable to this embodiment and can achieve the same technical effects.

[0095] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A system for calculating operating data of an electronic current transformer, characterized in that, include: The acquisition module is used to acquire the raw high-frequency sampling data stream of the electronic current transformer, perform signal preprocessing and compression storage, and generate a standardized multi-dimensional time series dataset. The building module is used to construct a device dependency adjacency matrix based on a standardized multidimensional time-series dataset and the power grid SCADA topology. The calculation module is used to perform a sliding time window scan of the device-dependent adjacency matrix, and to calculate the local reachability density under the dynamic nearest neighbor topological distance using the local outlier factor algorithm. If the total harmonic distortion rate gradient is detected to enter the statistical hypothesis test rejection region or the winding temperature rise rate exceeds the material thermal aging critical threshold, a multidimensional abnormal feature vector set is extracted. The training module is used to generate a set of risk quantification indicators from a multidimensional set of anomaly feature vectors through a pre-trained fault propagation probability model. The optimization module is used to establish a multi-objective optimization function based on the risk quantification index group, and to perform non-dominated sorting evolution of the disposal instruction chromosomes using the NSGA-II genetic algorithm. The dynamic optimization instruction sequence that matches the real-time load fluctuation coefficient is selected through the Pareto front solution set to generate a structured disposal instruction set.

2. The electronic current transformer operation data calculation system according to claim 1, characterized in that, The raw high-frequency sampling data stream from the electronic current transformer is acquired, preprocessed, compressed, and stored to generate a standardized multi-dimensional time-series dataset, including: The high-speed sampling data stream generated by the electronic current transformer is analyzed to extract the original structured dataset containing the complex harmonic spectrum sequence, the distributed temperature gradient matrix, and the ADC sampling status word. Based on the original structured dataset, the harmonic spectrum complex sequence is scaled, the layer-related dynamic threshold is calculated based on the risk estimation principle, and soft thresholding is performed on the high-frequency detail coefficients. The time-domain waveform is transformed and reconstructed, and after verification by mean square error, a set of time-domain waveforms for noise reduction parameters is generated. The time-domain waveform set of noise reduction parameters is encoded and compressed into a timestamp sequence. This is combined with energy threshold truncation to achieve lossy compression of the harmonic spectrum. The sampling rate of the temperature gradient data is adaptively adjusted to generate a standardized multidimensional time-series dataset with time synchronization identifier.

3. The electronic current transformer operation data calculation system according to claim 2, characterized in that, Based on a standardized multidimensional time-series dataset, a device dependency adjacency matrix is ​​constructed based on the power grid SCADA topology, including: The device geographic labels and electrical connection relationships in the standardized multidimensional time-series dataset are analyzed, and an initial device adjacency matrix is ​​constructed based on the topology file exported from the power grid SCADA system. Based on the electrical connection relationships in the initial device adjacency matrix, the distance attenuation weight coefficient is calculated and generated; the real-time device health status data in the standardized multidimensional time series dataset is fused to calculate the health status coupling factor; the distance attenuation weight coefficient and the health status coupling factor are multiplied and fused, and the elements of the initial adjacency matrix are dynamically weighted and updated to generate a dynamic weighted device adjacency matrix with probabilistic fault propagation intensity. By integrating the dynamic weighted device adjacency matrix with real-time operating parameters from a standardized multidimensional time-series dataset, the fault propagation weights between devices are iteratively updated to construct a device dependency adjacency matrix.

4. The electronic current transformer operation data calculation system according to claim 3, characterized in that, The device relies on an adjacency matrix for sliding time window scanning. A local outlier factor algorithm is used to calculate the local reachability density under dynamic nearest-neighbor topological distance. If the total harmonic distortion rate gradient enters the statistical hypothesis rejection region or the winding temperature rise rate exceeds the material's thermal aging critical threshold, a multidimensional abnormal feature vector set is extracted, including: The sliding time window width is set as an integer multiple of the power system frequency cycle. The device-dependent adjacency matrix is ​​time-series segmented to generate a window matrix sequence. Based on the sparsity distribution characteristics of the window matrix sequence, the nearest neighbor topology search radius of each device is dynamically calculated. Based on the dynamic nearest neighbor search radius, the local outlier factor algorithm is executed to calculate the local reachability density of device nodes and generate a real-time outlier factor matrix. The real-time outlier factor matrix is ​​analyzed. When the total harmonic distortion rate gradient enters the statistical anomaly confidence interval or the winding temperature rise rate exceeds the critical threshold of insulation thermal aging, a spatiotemporal coordinate set of anomaly events is generated, and a multidimensional anomaly feature vector set is extracted.

5. The electronic current transformer operation data calculation system according to claim 4, characterized in that, The multidimensional anomaly feature vector set is used to generate a risk quantification index set through a pre-trained fault propagation probability model, including: The harmonic distortion acceleration component of the multidimensional anomaly feature vector set is analyzed and matched with the acceleration threshold boundary of the insulation dielectric breakdown failure mode in the equipment's historical fault database to generate the insulation breakdown risk probability; the thermal stress mutation coefficient component of the multidimensional anomaly feature vector set is analyzed and matched with the temperature rise gradient critical curve of the thermal fatigue cumulative failure mode of the winding material to generate the thermal fatigue risk probability; the data packet loss Shannon entropy component of the multidimensional anomaly feature vector set is analyzed and the relay protection maloperation risk probability value is calculated. By integrating the probability of insulation breakdown risk, the probability of thermal fatigue risk, and the probability of malfunction, an initial probability distribution vector of equipment cascade failure is generated. Based on the initial cascade failure probability distribution vector, the real-time stress aging coefficient of the equipment is injected for correction, generating dynamically updated equipment cascade failure probability values; based on the equipment cascade failure probability values, the remaining effective operating time window of the equipment is calculated; the equipment cascade failure probability values ​​and the remaining effective operating time window are integrated to generate a risk quantification index group.

6. The electronic current transformer operation data calculation system according to claim 5, characterized in that, Based on a risk quantification index set, a multi-objective optimization function is established. The NSGA-II genetic algorithm is used to perform non-dominated sorting evolution on the disposal instruction chromosomes. A dynamically optimized instruction sequence matching the real-time load fluctuation coefficient is selected through Pareto front solution set screening, generating a structured disposal instruction set, including: A set of multi-objective optimization functions is established based on a risk quantification index group. Constraints are generated by analyzing the real-time topological connection status of the power grid. The elements of the disposal instructions are encoded into chromosome gene sequences to generate an initial genetic population. Based on a set of multi-objective optimization functions, the NSGA-II algorithm is used to perform non-dominated sorting and crowding calculation on the initial population, and a Pareto front solution set is generated through genetic evolution. The real-time load fluctuation coefficient is fused, and a dynamic optimization instruction sequence with timeliness matching is selected from the Pareto front solution set. The instruction sequence is then decoded to generate a structured disposal instruction set.

7. The electronic current transformer operation data calculation system according to claim 6, characterized in that, A multi-objective optimization function set is established based on a risk quantification index group. Constraints are generated by analyzing the real-time topology connection status of the power grid. The elements of the disposal instructions are encoded into chromosome gene sequences to generate an initial genetic population, including: Based on the probability value of cascading failure of equipment and the remaining effective operating time window in the risk quantification index group as input parameters, a set of multi-objective optimization functions is constructed with the objectives of minimizing the probability of cascading failure propagation, maximizing the operating time margin of key equipment, and minimizing the load shedding amount. Analyze the real-time topology connection status of the power grid, and based on the operation time margin constraints in the multi-objective optimization function set, extract the circuit breaker operation blocking logic relationship and the available capacity of the reactive power compensation node to generate equipment operation sequence constraints. Based on the equipment operation sequence constraints, the circuit breaker opening and closing status commands, reactive power compensation device adjustment commands, and load transfer path commands are encoded into chromosome gene fragments; based on the remaining effective operation time window, the relay protection action time window constraints are set, and the chromosome gene fragments are combined under the time window constraints to generate the initial genetic population.

8. The electronic current transformer operation data calculation system according to claim 7, characterized in that, Based on a set of multi-objective optimization functions, the NSGA-II algorithm is used to perform non-dominated sorting and crowding calculation on the initial population, and a Pareto front solution set is generated through genetic evolution; real-time load fluctuation coefficients are fused, and dynamic optimization instruction sequences with timeliness matching are selected from the Pareto front solution set. Decode the instruction sequence to generate a structured processing instruction set, including: Based on a set of multi-objective optimization functions, non-dominated sorting stratification is performed on the initial genetic population, and crowding distance is calculated for individuals at each stratum; parent individuals are selected based on the non-dominated sorting stratification and crowding distance. The parent individuals are manipulated to generate the offspring population; the parent individuals and the offspring population are merged, and the non-dominated sorting and elite preservation strategies are re-executed. The evolution is iterated until the convergence condition is met, and a Pareto front solution set is generated; the execution time window of each instruction sequence in the Pareto front solution set is analyzed, and the instruction execution time matching degree is calculated based on the real-time load fluctuation coefficient. Based on the matching degree of instruction execution time, the timeliness matching and dynamic optimization of instruction sequences are selected, and the structured disposal instruction set is generated by decoding chromosome gene fragments.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the system as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the system as described in any one of claims 1 to 8.

Images