Optical fiber differential protection decision-making method and system based on adaptive multiple transmission lines

Abstract

The invention provides an optical fiber differential protection decision-making method and system based on adaptive multiple transmission lines, and the method comprises the steps: collecting the real-time optical signal waveform data of each line in the multiple transmission lines, and extracting a differential current feature set corresponding to each line, generating a joint feature sequence, inputting the joint feature sequence into a pre-trained adaptive multi-transmission line model, and obtaining protection priority parameters corresponding to each line; based on the matched differential protection threshold corresponding to each line, generating a self-adaptive trigger condition of each line; the actual differential current value of each line is monitored in real time, and if it is detected that the actual differential current value of the target line reaches the corresponding self-adaptive triggering condition, the differential protection action for the target line is activated; and according to priority parameters output by the adaptive multi-transmission line model, monitoring frequencies of other lines which do not trigger protection actions are adjusted. According to the invention, when the optical fiber differential protection is carried out, the fault positioning precision and the system operation stability can be improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a fiber optic differential protection decision-making method and system based on adaptive multiple transmission lines. Background Art

[0002] Fiber-optic differential protection technology, a core method for fault detection in critical power system lines, primarily detects fault occurrence by comparing the amplitude and phase differences of current signals at both ends of a transmission line. Traditional solutions rely on preset fixed differential thresholds, triggering protection action when the real-time differential current exceeds the threshold. These protection criteria are based on isolated analysis of the current characteristics of a single line and lack the ability to globally assess the dynamic coupling effects of multiple lines operating in a coordinated manner. Existing methods typically employ static parameter configuration, such as setting a unified protection threshold based on historical fault data. These methods fail to consider the impact of factors such as line load fluctuations and shifts in fiber transmission characteristics caused by ambient temperature variations on the accuracy of differential currents. This results in high-load lines being susceptible to false tripping due to transient current fluctuations, while low-load lines may mask true fault characteristics due to signal attenuation. Particularly in complex networking scenarios with multiple transmission lines, existing technologies are unable to effectively identify the propagation paths of differential current anomalies caused by inter-line energy transfer. Furthermore, the rigid monitoring resource allocation mechanism, with all lines collecting data at the same sampling frequency, results in missing transient anomalies on high-risk lines due to long sampling intervals, while low-risk lines suffer from hardware resource waste due to continuous high-frequency monitoring. In addition, traditional differential protection models lack the ability to dynamically filter interference components such as harmonic distortion and transient pulses in the signal waveform, making it difficult to distinguish between real fault currents and noise interference, further exacerbating the contradiction between protection sensitivity and anti-interference performance. There is an urgent need for a differential protection decision-making method that can integrate the dynamic characteristics of multiple lines, adaptively optimize protection thresholds, and realize intelligent allocation of monitoring resources. Summary of the Invention

[0003] The present invention aims to provide a method and system for making decisions for optical fiber differential protection based on adaptive multiple transmission lines. The present invention is achieved as follows: In a first aspect, an embodiment of the present invention provides a method for making decisions for optical fiber differential protection based on adaptive multiple transmission lines, the method comprising: collecting real-time optical signal waveform data for each line in the multiple transmission lines and extracting a differential current feature set corresponding to each line; generating a joint feature sequence based on the differential current feature sets of all lines, and inputting the joint feature sequence into a pre-trained adaptive multiple transmission line model to obtain a protection priority parameter corresponding to each line; dynamically matching the differential protection threshold corresponding to each line based on the protection priority parameter to generate an adaptive trigger condition for each line; monitoring the actual differential current value of each line in real time, and activating the differential protection action for the target line if the actual differential current value of the target line reaches its corresponding adaptive trigger condition; and adjusting the monitoring frequency of the remaining lines for which the protection action has not been triggered based on the priority parameter output by the adaptive multiple transmission line model.

[0004] On the other hand, the present invention provides a fiber differential protection decision system, comprising: one or more processors; a memory; and one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method described above is implemented.

[0005] The present invention provides an adaptive multi-transmission line-based optical fiber differential protection decision-making method. This method extracts differential current feature sets and generates a joint feature sequence by collecting real-time optical signal waveform data from multiple transmission lines. A pre-trained adaptive multi-transmission line model dynamically outputs protection priority parameters for each line. Based on the priority parameters, the method dynamically matches differential protection thresholds to generate adaptive triggering conditions. The method monitors differential current values ​​in real time and triggers protection actions. Simultaneously, the monitoring frequency of untriggered lines is adjusted based on the priority parameters to form a closed-loop feedback loop. By synergizing multi-dimensional data fusion (optical signal waveforms, differential current features, and signal transmission delay parameters) with dynamic model decision-making, the method accurately captures abnormal line fluctuations and the risk of cross-line cascading failures, effectively improving the response sensitivity and anti-interference capability of differential protection and avoiding the problems of false triggering or protection delays caused by traditional fixed thresholds. A priority parameter-driven threshold dynamic compensation mechanism and a closed-loop monitoring frequency adjustment strategy achieve an adaptive balance between efficient protection resource allocation and anomaly detection sensitivity. This method prioritizes transient fault characteristics in highly loaded lines while reducing resource consumption caused by redundant monitoring in stable lines, enhancing the overall robustness of differential protection systems in complex power grid environments. Based on the dynamic weight superposition calculation of the joint feature sequence and real-time load parameters, it ensures that the protection priority parameters can synchronously reflect the correlation between the current operating status of the line and the historical load trend. Combined with the multi-level protection trigger logic of harmonic distortion rate detection and buffer zone verification, it further suppresses the interference of environmental noise on the differential current judgment criteria, forming a three-dimensional protection decision-making system that is coordinated by data drive and physical rules, and ultimately achieves a dual improvement in fault location accuracy and system operation stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is a flow chart of a fiber differential protection decision method based on adaptive multiple transmission lines provided by an embodiment of the present invention; Figure 2 The diagram is a composition diagram of a fiber optic differential protection decision system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0007] The following describes the embodiments of the present invention in conjunction with the accompanying drawings. The terms used in the implementation methods of the embodiments of the present invention are only used to explain the specific embodiments of the present invention, and are not intended to limit the present invention.

[0008] The execution subject of the optical fiber differential protection decision method based on adaptive multiple transmission lines in the embodiment of the present invention is an optical fiber differential protection decision system, including but not limited to a server, a personal computer, a laptop computer, a tablet computer, etc. Figure 1 As shown, the method includes: Step S100: collecting real-time optical signal waveform data of each line in multiple transmission lines, and extracting a differential current feature set corresponding to each line.

[0009] Real-time optical signal waveform data refers to the waveform information of the optical signal transmitted by each line in the multiple transmission lines at the current moment, which changes with time, and reflects the real-time state of the optical signal. The differential current feature set is a collection of a series of features related to the differential current extracted from the real-time optical signal waveform data. These features can be used for subsequent analysis and judgment of the line status. In an embodiment of the present application, the real-time optical signal waveform data of each line in the multiple transmission lines can be collected by a special collection device, and then the signal processing and analysis algorithm is used to extract the differential current feature set corresponding to each line from the collected real-time optical signal waveform data. For example, in a system containing three transmission lines, the real-time optical signal waveform data of the three lines are collected separately using an optical signal collector, and then the differential current feature set of each line is extracted by algorithms such as Fourier transform, including current amplitude, frequency, phase and other features.

[0010] Step S200: generating a joint feature sequence based on the differential current feature sets of all lines, and inputting the joint feature sequence into a pre-trained adaptive multi-transmission line model to obtain a protection priority parameter corresponding to each line.

[0011] The joint feature sequence is a sequence containing the feature information of each line generated by integrating and processing the differential current feature sets of all lines, which can comprehensively reflect the overall characteristics of multiple transmission lines. The pre-trained adaptive multi-transmission line model is obtained through training with a large amount of historical data, and can output the protection priority parameters corresponding to each line based on the input joint feature sequence. The protection priority parameters are used to measure the importance of each line in the protection decision. In the embodiment of the present application, a joint feature sequence is first generated based on the differential current feature sets of all lines. The specific generation method can refer to the subsequent implementation method; then the generated joint feature sequence is input into the pre-trained adaptive multi-transmission line model, and the model outputs the protection priority parameters corresponding to each line after calculation and analysis. For example, in a scenario with five transmission lines, a joint feature sequence is generated based on the differential current feature sets of these five lines, and the sequence is input into the pre-trained adaptive multi-transmission line model. The model outputs the protection priority parameters corresponding to each of the five lines to determine the order of each line in the protection decision.

[0012] As an implementation method, in step S200, a joint feature sequence is generated based on the differential current feature sets of all lines, which may specifically include: step S210: filtering out feature subsets whose amplitude fluctuation rate exceeds a preset fluctuation threshold from the differential current feature set of each line, and aligning each feature subset by timestamp.

[0013] Amplitude fluctuation refers to the degree of time-varying fluctuation in the amplitude of a feature value in a differential current feature set. The preset fluctuation threshold is a pre-set standard value used to determine whether the amplitude fluctuation is excessive. A feature subset is a collection of features selected from the differential current feature set that meet specific conditions (e.g., amplitude fluctuation exceeding the preset fluctuation threshold). A timestamp is information used to mark the time when the feature data was generated. In this embodiment of the present application, the differential current feature set for each line is screened to identify the feature subset whose amplitude fluctuation exceeds the preset fluctuation threshold. The feature subsets for each line are then aligned based on the timestamps, ensuring temporal correspondence between the feature subsets for different lines. For example, in a system comprising four transmission lines, the differential current feature set for each line is screened to identify the feature subset whose amplitude fluctuation exceeds the preset fluctuation threshold (e.g., 5%). The feature subsets for these four lines are then aligned based on the timestamps to facilitate subsequent analysis and processing.

[0014] Step S220: Calculate the energy distribution parameters of the feature subset of each line within a preset time window, and extract the maximum energy gradient value in the energy distribution parameters of each line.

[0015] The preset time window is a pre-set time interval used for segmented analysis of feature subsets. The energy distribution parameter is a parameter that describes the energy distribution of the feature subset within the preset time window, which reflects the energy change characteristics of the feature subset. The maximum energy gradient value is the gradient value of the place where the energy changes fastest in the energy distribution parameter, which can reflect the severity of the energy change of the feature subset. In an embodiment of the present application, for the feature subset of each line, its energy distribution parameters within the preset time window are calculated, and then the maximum energy gradient value is extracted from these energy distribution parameters. For example, for the feature subset of a transmission line, the preset time window is set to 10 seconds, the energy distribution parameters of the feature subset within the 10 seconds are calculated, and then the maximum energy gradient value is determined to understand the severity of the energy change of the line within this time period.

[0016] In one embodiment, step S220 calculates the energy distribution parameters of the feature subset of each line within a preset time window and extracts the maximum energy gradient value in the energy distribution parameters of each line, which may specifically include: step S221: dividing the feature subset into multiple continuous time segments according to the length of the preset time window, each time segment contains amplitude fluctuation rate data of a fixed duration.

[0017] A time segment is a small time interval obtained by segmenting a feature subset according to the length of a preset time window. Each time segment contains amplitude fluctuation rate data of a fixed duration, which can reflect the amplitude fluctuation of the differential current characteristics within that time period. In an embodiment of the present application, the feature subset is segmented according to the length of the preset time window to obtain multiple consecutive time segments, each of which has a fixed duration and contains corresponding amplitude fluctuation rate data. For example, if the preset time window is 5 seconds, the feature subset of a line is segmented according to the length of 5 seconds to obtain multiple consecutive 5-second time segments, each of which contains amplitude fluctuation rate data within that 5-second period.

[0018] Step S222: Accumulate the differential current values ​​corresponding to all amplitude fluctuation rates in each time segment to generate an energy accumulation value corresponding to each time segment.

[0019] The energy accumulation value is the result of summing the differential current values ​​corresponding to all amplitude fluctuations within each time segment, reflecting the energy accumulation of the differential current within that time segment. In the embodiment of the present application, the differential current values ​​corresponding to the amplitude fluctuations within each time segment are accumulated to obtain the energy accumulation value corresponding to each time segment. For example, within a time segment, there are multiple differential current values ​​corresponding to the amplitude fluctuations. These values ​​are added together to obtain the energy accumulation value for that time segment, which is used to measure the degree of energy accumulation within that time segment.

[0020] Step S223: Calculate the boundary energy change rate of each time segment according to the difference in energy accumulation values ​​of adjacent time segments, and arrange all boundary energy change rates into an energy difference sequence in chronological order.

[0021] The boundary energy change rate refers to the difference in the energy cumulative values ​​of adjacent time segments, which reflects the rate of energy change between adjacent time segments. The energy difference sequence is a sequence obtained by arranging the boundary energy change rates of all time segments in chronological order, which can show the energy changes of the feature subset over the entire time range. In an embodiment of the present application, the difference in the energy cumulative values ​​of adjacent time segments is calculated to obtain the boundary energy change rate of each time segment, and then these boundary energy change rates are arranged in chronological order into an energy difference sequence. For example, there are three consecutive time segments, and the difference between the energy cumulative values ​​of the second time segment and the first time segment, and the third time segment and the second time segment are calculated respectively to obtain two boundary energy change rates. These two boundary energy change rates are arranged in chronological order to form part of the energy difference sequence.

[0022] Step S224: Filter out the difference values ​​of all forward transitions from the energy difference sequence, and perform weighted summation on the filtered difference values ​​according to a preset weight coefficient to generate a dynamic energy distribution parameter within a preset time window.

[0023] A positive transition difference value refers to a difference value in the energy difference sequence that is positive and has a clear growth trend, indicating a dramatic positive change in energy within that time period. A preset weight coefficient is a pre-set coefficient used to weight the selected positive transition difference values. Different positive transition difference values ​​may correspond to different weight coefficients. The dynamic energy distribution parameter is a parameter obtained by weighting and summing the selected positive transition difference values ​​according to the preset weight coefficient. It can more accurately reflect the dynamic distribution of energy within a preset time window. In an embodiment of the present application, the difference values ​​of all positive transitions are determined from the energy difference sequence, and then these difference values ​​are weighted and summed according to the preset weight coefficient to obtain the dynamic energy distribution parameter within the preset time window. For example, the difference values ​​of several positive transitions are selected from the energy difference sequence, and these difference values ​​are weighted and summed according to the preset weight coefficient (e.g., the first difference value has a weight of 0.3, the second difference value has a weight of 0.5, etc.) to obtain the dynamic energy distribution parameter.

[0024] Step S225: traverse the gradient change trend corresponding to each difference value in the dynamic energy distribution parameter, identify the interval segment with continuous increasing difference values, and extract the maximum rising slope of the difference value in the interval segment as the maximum energy gradient value.

[0025] The gradient change trend refers to the change slope of each difference value in the dynamic energy distribution parameter, which reflects the speed and direction of the energy change. The interval segment with continuous difference value increase refers to the time period in which the difference value in the dynamic energy distribution parameter increases continuously, and the energy shows an upward trend in this interval segment. The maximum rising slope refers to the slope of the place where the difference value rises the fastest in the interval segment with continuous difference value increase. It is used as the maximum energy gradient value and can reflect the drastic degree of energy change in this interval segment. In an embodiment of the present application, the gradient change trend of each difference value in the dynamic energy distribution parameter is traversed to determine the interval segment with continuous difference value increase, and then the maximum rising slope of the difference value is extracted from these interval segments as the maximum energy gradient value. For example, in the dynamic energy distribution parameter, it is found that there is a continuous difference value that increases, and the rising slope of the difference value in this interval segment is calculated, and the largest rising slope among them is determined as the maximum energy gradient value.

[0026] Step S226: superimpose the maximum energy gradient value and the average change rate in the dynamic energy distribution parameter to generate a corrected energy gradient parameter, and output the corrected energy gradient parameter as the maximum energy gradient value in the energy distribution parameter.

[0027] The average rate of change is the average rate of change of all the difference values ​​in the dynamic energy distribution parameters, which reflects the average change in energy over the entire time range. The corrected energy gradient parameter is the parameter obtained by superimposing the maximum energy gradient value with the average rate of change in the dynamic energy distribution parameters, which comprehensively considers the severity of the energy change and the average change. In the embodiment of the present application, the maximum energy gradient value is added to the average rate of change in the dynamic energy distribution parameters to obtain the corrected energy gradient parameter, which is output as the maximum energy gradient value in the energy distribution parameters.

[0028] Step S230: The maximum energy gradient values ​​of each line are concatenated into an initial joint sequence according to the line number sequence, and the initial joint sequence is normalized.

[0029] The initial joint sequence is a sequence formed by arranging the maximum energy gradient values ​​of each line in sequence according to the line number order, which contains the key information of the energy change of each line. Normalization processing is to process the data in the initial joint sequence so that its value range is within a preset interval (usually [0, 1]). This can eliminate the dimensional differences between the maximum energy gradient values ​​of different lines and facilitate subsequent analysis and processing. In an embodiment of the present application, the maximum energy gradient values ​​of each line are spliced ​​into an initial joint sequence in the order of the line number, and then the sequence is normalized. For example, there are three lines with maximum energy gradient values ​​of 5, 8, and 3 respectively. They are spliced ​​into the initial joint sequence [5, 8, 3] in the order of the line number, and then the sequence is normalized so that its value range is between [0, 1].

[0030] Step S240: matching the normalized initial joint sequence with the standard pattern in the historical joint feature library, and removing abnormal feature segments in the initial joint sequence whose overlap with the historical abnormal pattern exceeds a preset overlap.

[0031] The historical joint feature library is a database that stores historical joint feature sequences of multiple transmission lines, containing various normal and abnormal standard patterns. Standard patterns are representative feature patterns in the historical joint feature library and are used to match the current normalized initial joint sequence. Historical abnormal patterns are feature patterns marked as abnormal in the historical joint feature library. The preset overlap is a pre-set standard value used to determine the degree of similarity between the current sequence and the historical abnormal pattern. Abnormal feature segments are portions of the normalized initial joint sequence whose overlap with the historical abnormal pattern exceeds the preset overlap. In this embodiment of the present application, the normalized initial joint sequence is compared with the standard patterns in the historical joint feature library to identify abnormal feature segments whose overlap with the historical abnormal pattern exceeds the preset overlap, and these segments are removed from the initial joint sequence. For example, if the preset overlap is 80%, the normalized initial joint sequence is matched with the standard patterns in the historical joint feature library, and a segment is determined to have an overlap of 85% with a historical abnormal pattern. This segment is then removed from the initial joint sequence as an abnormal feature segment.

[0032] Step S250: rearrange the remaining feature segments in chronological order to generate a final joint feature sequence.

[0033] The remaining feature segments are the portion remaining in the initial joint sequence after the abnormal feature segments are removed. The final joint feature sequence is the sequence obtained by rearranging the remaining feature segments in chronological order. It more accurately reflects the normal feature information of multiple transmission lines. In this embodiment of the present application, the remaining feature segments after removing the abnormal feature segments are rearranged in chronological order to generate the final joint feature sequence. For example, if three feature segments remain after removing the abnormal feature segments from the initial joint sequence, these three feature segments are rearranged in chronological order to generate the final joint feature sequence.

[0034] As an implementation method, in the above step S200, the joint feature sequence is input into the pre-trained adaptive multi-transmission line model to obtain the protection priority parameters corresponding to each line, which can specifically include the following steps: Step S260: Divide the final joint feature sequence into multiple continuous time segments, each time segment contains a fixed number of feature points.

[0035] A feature point is each data point in the final joint feature sequence, representing feature information at a specific moment. In an embodiment of the present application, the final joint feature sequence is divided into multiple consecutive time segments according to a certain rule, with each time segment containing a fixed number of feature points. For example, the final joint feature sequence is divided into time segments each containing 10 feature points, which facilitates subsequent analysis and processing of each time segment.

[0036] Step S270: Input each time segment into the feature fusion layer of the adaptive multi-transmission line model, and output the line state weight coefficient corresponding to each time segment.

[0037] The feature fusion layer is a network layer in the adaptive multi-transmission line model, which can perform feature fusion processing on the input time segment and comprehensively consider the relationship between each feature point. The line state weight coefficient is a coefficient output by the feature fusion layer based on the input time segment for measuring the line state, which reflects the importance and state information of the line within the time segment. In the embodiment of the present application, each time segment is input into the feature fusion layer of the adaptive multi-transmission line model in turn, and the feature fusion layer outputs the line state weight coefficient corresponding to each time segment after calculation and analysis. For example, a time segment containing 10 feature points is input into the feature fusion layer, and the feature fusion layer outputs a line state weight coefficient corresponding to the time segment of 0.8, indicating that the state of the line in the time segment is more important.

[0038] Step S280: Calculate the cumulative weight value of each line in the entire monitoring period according to the line state weight coefficient of each time segment, and sort the lines from high to low according to the cumulative weight value.

[0039] The cumulative weight value is the result of adding up the line state weight coefficients of each line in each time segment, which comprehensively reflects the importance of the line in the entire monitoring period. In the embodiment of the present application, the line state weight coefficients of each line in each time segment are added up to obtain the cumulative weight value of each line, and then the lines are sorted from high to low according to the cumulative weight value. For example, the line state weight coefficients of a line in three time segments are 0.7, 0.8, and 0.6 respectively. They are added together to obtain a cumulative weight value of 2.1, and then all lines are sorted from high to low according to the cumulative weight value.

[0040] Step S290: Map the sorted line numbers to preset priority intervals to generate dynamic protection priority parameters corresponding to each line.

[0041] The preset priority interval is a pre-set range used to prioritize lines. The dynamic protection priority parameter is a parameter obtained by mapping the sorted line numbers to the preset priority interval. It can dynamically reflect the priority of each line in the protection decision. In the embodiment of the present application, the sorted line numbers are mapped to the preset priority interval to obtain the dynamic protection priority parameter corresponding to each line. For example, if the preset priority interval is [1, 5] and the sorted line number is 3, the dynamic protection priority parameter of the line obtained by mapping it to the priority interval is 3.

[0042] Step S2100: Perform weighted superposition on the dynamic protection priority parameter and the real-time load parameter to obtain a final protection priority parameter.

[0043] The real-time load parameter represents the load borne by the line at the current moment, reflecting the actual working status of the line. Weighted superposition is the operation of adding the dynamic protection priority parameter and the real-time load parameter according to certain weights. The final protection priority parameter is the parameter obtained through weighted superposition, which comprehensively considers the priority and real-time load conditions of the line and more accurately determines the importance of each line in the protection decision. In the embodiment of the present application, the dynamic protection priority parameter and the real-time load parameter are weighted and superimposed according to preset weights to obtain the final protection priority parameter.

[0044] In one embodiment, the training process of the pre-trained adaptive multi-transmission line model may include the following steps: Step S10: collecting multiple groups of optical signal waveform sample data of historical transmission lines, and extracting the differential current feature vector corresponding to each group of sample data.

[0045] The optical signal waveform sample data of the historical transmission line refers to the collection of waveform information of the optical signals of multiple transmission lines that change with time over a period of time in the past, which contains rich line status information. The differential current characteristic vector is a vector composed of features related to the differential current extracted from the optical signal waveform sample data, which can be used to describe the current characteristics of the line. In an embodiment of the present application, multiple groups of optical signal waveform sample data of historical transmission lines are collected by a data acquisition device, and then the corresponding differential current characteristic vector is extracted from each group of sample data using a preset signal processing and analysis algorithm. For example, 100 groups of optical signal waveform sample data of historical transmission lines are collected, and each group of data is processed to extract the corresponding differential current characteristic vector, each vector containing features such as current amplitude, frequency, and phase.

[0046] Step S20: performing a joint analysis of the differential current characteristic vector in the time domain and the frequency domain to generate a characteristic signature sequence for each set of sample data.

[0047] Joint analysis of the time domain and frequency domain is an analysis method that comprehensively considers the characteristics of the signal in both time and frequency dimensions, which can provide a more comprehensive understanding of the characteristics of the signal. The feature marker sequence is a sequence containing line feature marker information generated by performing a joint analysis of the differential current feature vector in the time domain and frequency domain, which can more accurately describe the state of the line. In the embodiment of the present application, the extracted differential current feature vector is subjected to a joint analysis of the time domain and frequency domain. The specific analysis method can refer to the subsequent implementation method, and then a feature marker sequence for each group of sample data is generated based on the analysis results. For example, a set of differential current feature vectors is subjected to a joint analysis of the time domain and frequency domain to obtain a feature marker sequence for the set of sample data, and each marker in the sequence represents a preset line feature.

[0048] As an implementation method, step S20 performs a joint analysis of the differential current characteristic vector in the time domain and frequency domain to generate a characteristic marker sequence for each group of sample data, which may specifically include: step S21: dividing the differential current characteristic vector into multiple time domain segments, and calculating the average amplitude and phase offset of each time domain segment.

[0049] A time domain segment is a small time period into which the differential current characteristic vector is divided in chronological order. The average amplitude refers to the average value of the differential current amplitude within each time domain segment, and the phase offset refers to the offset of the differential current phase relative to the reference phase within each time domain segment. In an embodiment of the present application, the differential current characteristic vector is divided into multiple time domain segments at certain time intervals, and then the differential current amplitude and phase within each time domain segment are calculated to obtain the average amplitude and phase offset. For example, a differential current characteristic vector is divided into 10 time domain segments at a time interval of 1 second, and the average amplitude and phase offset of each time domain segment are calculated.

[0050] Step S22: Perform fast Fourier transform on each time domain segment to extract the dominant frequency component in the frequency domain energy distribution of each time domain segment.

[0051] Fast Fourier transform is used to convert time domain signals into frequency domain signals. Each time domain segment can be converted to the frequency domain through fast Fourier transform to obtain its frequency domain energy distribution. The dominant frequency component is the frequency component with the largest energy share in the frequency domain energy distribution, which reflects the main frequency characteristics of the time domain segment. In an embodiment of the present application, a fast Fourier transform is performed on each time domain segment to convert it from the time domain to the frequency domain, and then the dominant frequency component is determined from the frequency domain energy distribution. For example, a time domain segment is fast Fourier transformed to obtain its frequency domain energy distribution, and it is found that the component with a frequency of 50Hz has the largest energy share, and 50Hz is taken as the dominant frequency component of the time domain segment.

[0052] Step S23: Match the dominant frequency component with a preset typical fault frequency library, and mark the potential fault type corresponding to each time domain segment.

[0053] The preset typical fault frequency library is a pre-established database containing frequency information corresponding to various typical faults. By matching the dominant frequency component with the frequency information in the library, the potential fault type corresponding to each time domain segment can be determined. In an embodiment of the present application, the dominant frequency component of each time domain segment is compared with the preset typical fault frequency library to determine the fault type that matches it, and the potential fault type corresponding to each time domain segment is marked. For example, the preset typical fault frequency library records that a frequency of 50Hz may correspond to a line short circuit fault. The dominant frequency component of a time domain segment is 50Hz, and the time domain segment is marked as a possible line short circuit fault.

[0054] Step S24: performing weighted fusion on the average amplitude and phase offset of the time domain segments according to the potential fault type to generate time-frequency joint feature parameters.

[0055] It can be understood that weighted fusion is a process of assigning different weights to the average amplitude and phase offset according to the potential fault type, and then calculating them comprehensively. The time-frequency joint feature parameter is a parameter obtained by weighted fusion that comprehensively considers the time domain and frequency domain characteristics, which can more comprehensively describe the state of the line. In an embodiment of the present application, different weights are assigned to the average amplitude and phase offset of each time domain segment according to the potential fault type corresponding to the segment, and then they are weighted fused to generate the time-frequency joint feature parameter. For example, a time domain segment is marked as a possible line short circuit fault, and the average amplitude of the time domain segment is assigned a weight of 0.6, and the phase offset is assigned a weight of 0.4. They are weighted fused to obtain the time-frequency joint feature parameter.

[0056] Step S25: Arrange all the time-frequency joint feature parameters in the same set of sample data in chronological order to form a feature tag sequence.

[0057] In this embodiment of the present application, the time-frequency joint feature parameters of each time domain segment in the same set of sample data are arranged in chronological order to form a feature tag sequence. This sequence contains the comprehensive feature information of the sample data in the time and frequency domains, providing more accurate data for subsequent model training. For example, if a set of sample data contains 10 time domain segments, the time-frequency joint feature parameters of these 10 time domain segments are arranged in chronological order to form a feature tag sequence containing 10 elements.

[0058] Step S30: inputting the signature sequence into the convolution kernel alignment layer in the initial adaptive multi-transmission line model, and outputting the alignment weight of each signature in the time dimension.

[0059] The convolution kernel alignment layer is a network layer in the initial adaptive multi-transmission line model, which can process the input feature marker sequence and determine the alignment relationship of each feature marker in the time dimension. The alignment weight is the weight value output by the convolution kernel alignment layer based on the feature marker sequence to represent the degree of alignment of each feature marker in the time dimension, which can make the feature marker sequence more orderly and unified in time. In an embodiment of the present application, the generated feature marker sequence is input into the convolution kernel alignment layer of the initial adaptive multi-transmission line model, and the convolution kernel alignment layer outputs the alignment weight of each feature marker in the time dimension after calculation and analysis. For example, a feature marker sequence containing 10 feature markers is input into the convolution kernel alignment layer to obtain the alignment weight of each feature marker in the time dimension, such as the alignment weight of the first feature marker is 0.8, the alignment weight of the second feature marker is 0.7, and so on.

[0060] As an implementation method, step S30, inputting the feature marker sequence into the convolution kernel alignment layer in the initial adaptive multi-transmission line model, and outputting the alignment weight of each feature marker in the time dimension, can specifically include: step S31: setting multiple sliding windows of different scales in the convolution kernel alignment layer, each sliding window is used to capture local patterns in the feature marker sequence.

[0061] The sliding window is a window in the convolution kernel alignment layer used to perform local analysis on the feature marker sequence. Sliding windows of different scales can capture local patterns of different sizes in the feature marker sequence. Local patterns are local fragments with certain regularities and characteristics in the feature marker sequence. By capturing these local patterns, the structure and characteristics of the feature marker sequence can be better understood. In an embodiment of the present application, multiple sliding windows of different scales are set in the convolution kernel alignment layer. For example, sliding windows of scales 3, 5, and 7 are set, and these sliding windows are allowed to slide on the feature marker sequence to capture the local patterns therein.

[0062] Step S32: Calculate the similarity score between the feature markers covered by each sliding window and the preset reference pattern.

[0063] The preset benchmark pattern is a representative feature pattern that is pre-set. The similarity score is a score used to measure the degree of similarity between the feature markers covered by each sliding window and the preset benchmark pattern, with a higher score indicating a higher similarity. In the embodiment of the present application, for each feature marker covered by the sliding window, a similarity score is calculated between it and the preset benchmark pattern. For example, a cosine similarity algorithm is used to calculate the similarity score between the feature markers covered by the sliding window and the preset benchmark pattern, resulting in a score between 0 and 1.

[0064] Step S33: Dynamically adjust the step size and coverage of each sliding window according to the similarity score to generate multi-scale alignment parameters.

[0065] The step size refers to the distance the sliding window slides each time, and the coverage refers to the length that the sliding window covers on the feature marker sequence. The multi-scale alignment parameter is a parameter obtained by dynamically adjusting the step size and coverage of the sliding window according to the similarity score, which can enable the sliding window to more accurately capture the local pattern in the feature marker sequence. In an embodiment of the present application, the step size and coverage of each sliding window are dynamically adjusted according to the similarity score between the feature marker covered by each sliding window and the preset reference pattern to generate the multi-scale alignment parameter. For example, if the similarity score between the feature marker covered by a certain sliding window and the preset reference pattern is low, the step size is appropriately reduced and the coverage is expanded to improve the accuracy of capturing the local pattern.

[0066] Step S34: Input the multi-scale alignment parameters into the weight distribution network, and output the position weight of each feature marker on the time axis.

[0067] The weight distribution network is a pre-trained network that can assign a position weight on the time axis to each feature marker based on the input multi-scale alignment parameters. The position weight is used to represent the importance and relative position relationship of each feature marker on the time axis. In an embodiment of the present application, the generated multi-scale alignment parameters are input into the weight distribution network, and the network outputs the position weight of each feature marker on the time axis after calculation. For example, the input multi-scale alignment parameters contain adjustment information for different sliding windows, and the weight distribution network assigns position weights to each feature marker in the feature marker sequence based on this information, such as the position weight of the first feature marker is 0.6, the position weight of the second feature marker is 0.8, etc. These weights reflect the difference in importance of each feature marker on the time axis.

[0068] For example, the weight distribution network can adopt a fully connected neural network architecture, consisting of an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimensionality of the multi-scale alignment parameters. Each element in the multi-scale alignment parameters corresponds to a neuron in the input layer, which is responsible for receiving the input multi-scale alignment parameter data. The number of neurons in the hidden layer can be flexibly adjusted to meet the needs of tasks of varying complexity. In each hidden layer, neurons perform a weighted sum of the outputs of the previous layer and perform a nonlinear transformation using an activation function (such as the ReLU function). The ReLU function can address the vanishing gradient problem, enhance the network's expressiveness, and enable the network to learn more complex features and patterns.

[0069] The number of neurons in the output layer is equal to the number of feature labels in the feature label sequence, and the output value of each neuron is the position weight of the corresponding feature label on the time axis. During the training phase, the weight allocation network can be trained using a large amount of historical data. The training process uses a common backpropagation algorithm to continuously adjust the network weights and biases to minimize the error between the network's output position weights and the true expected weights. Specifically, the multi-scale alignment parameters are first input into the network to obtain the network output. A loss function (such as the mean squared error loss function) is then calculated between the output and the true label. The partial derivatives of the loss function with respect to the network weights and biases are then taken. Finally, an optimization algorithm (such as stochastic gradient descent) is used to update the network weights and biases based on the partial derivatives. After multiple iterations of training, the weight allocation network gradually learns the mapping between the multi-scale alignment parameters and the position weights, thereby accurately outputting reasonable position weights based on the input multi-scale alignment parameters. Position weights can provide a more accurate basis for subsequent nonlinear interpolation of feature marker sequences, improving the accuracy and effectiveness of model training. This allows the model to better capture the temporal information of feature markers when processing data related to multiple transmission lines, providing more reliable support for tasks such as line protection priority determination.

[0070] Step S35: Perform nonlinear interpolation on the feature tag sequence according to the position weight to generate a normalized feature sequence aligned in the time dimension.

[0071] Nonlinear interpolation is a method of processing a feature marker sequence according to position weights to make the sequence smoother and more orderly in the time dimension. The standardized feature sequence is a feature sequence aligned in the time dimension after nonlinear interpolation, which eliminates the inconsistency of the feature markers in time and facilitates subsequent model training. In an embodiment of the present application, a nonlinear interpolation operation is performed on the feature marker sequence based on the position weight of each feature marker. For example, for two adjacent feature markers in the feature marker sequence, the interpolation point in the middle is calculated according to their position weights, so that the sequence is more continuous and regular in the time dimension, and finally a standardized feature sequence after alignment in the time dimension is generated.

[0072] Step S40: Dynamically sample the feature tag sequence based on the alignment weight to generate a standardized feature set for training.

[0073] Dynamic sampling is a targeted sampling of feature marker sequences based on alignment weights to obtain more representative and effective data. The standardized feature set is a feature set generated after dynamic sampling for training the model, which contains key feature information filtered out from the feature marker sequence. In an embodiment of the present application, the feature marker sequence is dynamically sampled according to the alignment weights output by the convolution kernel alignment layer. The specific sampling method can refer to the subsequent implementation method, and finally a standardized feature set for training is generated. For example, the higher-weighted parts of the feature marker sequence are sampled according to the alignment weights to form a standardized feature set to improve the efficiency and accuracy of model training.

[0074] As an implementation method, step S40 dynamically samples the feature marker sequence based on the alignment weight to generate a standardized feature set for training, which may specifically include: step S41: filtering out feature markers with weight values ​​higher than a preset weight threshold according to the position weight as key feature points.

[0075] The preset weight threshold is a pre-set standard value for judging the importance of feature markers. The key feature point is a feature marker whose position weight in the feature marker sequence is higher than the preset weight threshold, representing important information in the sequence. In the embodiment of the present application, the position weight of each feature marker in the feature marker sequence is compared, and feature markers with weight values ​​higher than the preset weight threshold are screened out and used as key feature points. For example, the preset weight threshold is 0.7, and there are 10 feature markers in the feature marker sequence, of which the position weights of 3 feature markers are 0.8, 0.9, and 0.8, respectively, which are higher than the preset weight threshold. These 3 feature markers are used as key feature points.

[0076] Step S42: performing mean filling on the interval areas between the key feature points to generate a continuous feature trajectory curve.

[0077] Mean filling refers to filling the interval area between key feature points with the mean between them to form a continuous curve. The characteristic trajectory curve is a continuous curve obtained after mean filling that reflects the changing trend of the characteristic mark, which can make the characteristic information more complete and coherent. In an embodiment of the present application, for the selected key feature points, the mean of the interval area between them is calculated, and the mean is filled into the interval area to generate a continuous characteristic trajectory curve. For example, there are two key feature points, and the interval area between them contains 5 data points. The mean of the two key feature points is calculated, and the mean is filled into the position of these 5 data points to form a continuous characteristic trajectory curve.

[0078] Step S43: resampling is performed on the characteristic trajectory curve at equal time intervals to obtain uniformly distributed characteristic sampling points.

[0079] Equal time interval resampling refers to selecting data points at fixed time intervals on the feature trajectory curve to obtain evenly distributed feature sampling points. These feature sampling points can more accurately reflect the characteristic information of the feature marker sequence, facilitating subsequent model training. In the embodiment of the present application, on the generated feature trajectory curve, resampling is performed according to a pre-set equal time interval to obtain evenly distributed feature sampling points. For example, the equal time interval is set to 0.1 seconds, and a data point is selected every 0.1 seconds on the feature trajectory curve to obtain a series of evenly distributed feature sampling points.

[0080] Step S44: Bind the feature sampling points with the corresponding potential fault type labels to generate standardized training samples.

[0081] The potential fault type label is obtained in the aforementioned steps based on the matching of the dominant frequency component with the preset typical fault frequency library, and is used to indicate the possible fault type corresponding to each feature sampling point. The standardized training sample is a sample obtained by binding the feature sampling point with the corresponding potential fault type label. It contains feature information and corresponding fault type information and is the basic data unit for model training. In an embodiment of the present application, the resampled feature sampling points are bound one-to-one with the corresponding potential fault type labels to generate standardized training samples. For example, if the potential fault type corresponding to a feature sampling point is a line short circuit fault, the feature sampling point is bound to the label "line short circuit fault" to form a standardized training sample.

[0082] Step S45: Randomly shuffle and batch-divide the standardized training samples of all historical transmission lines to generate a standardized feature set for final training.

[0083] During random shuffling, the order of all standardized training samples is randomly disrupted to prevent the model from being affected by the order of samples during training. Batch division is to group the standardized training samples after random shuffling according to a certain number, and each group is called a batch, which is convenient for batch training of the model. The standardized feature set used for final training is the feature set for model training obtained after random shuffling and batch division. In an embodiment of the present application, the standardized training samples of all historical transmission lines are randomly shuffled, and then batch divided according to a preset batch size to generate a standardized feature set for final training. For example, there are 1,000 standardized training samples, and the preset batch size is 100. These samples are randomly shuffled and divided into 10 batches to form a standardized feature set for final training.

[0084] Step S50: iteratively train the initial adaptive multi-transmission line model using the standardized feature set until the error rate between the protection priority parameter output by the initial adaptive multi-transmission line model and the preset reference parameter is lower than a preset error threshold, thereby obtaining the adaptive multi-transmission line model.

[0085] Iterative training refers to the process of repeatedly inputting a standardized feature set into an initial adaptive multi-transmission line model, continuously adjusting the model's parameters so that the model's output gradually approaches preset benchmark parameters. The error rate refers to the degree of difference between the protection priority parameters output by the model and the preset benchmark parameters. The preset error threshold is a pre-set standard value used to determine whether the model training is qualified. An adaptive multi-transmission line model, obtained through iterative training, is a model that more accurately determines line protection priority by ensuring that the error rate between its output protection priority parameters and the preset benchmark parameters is below the preset error threshold. The preset benchmark parameters are standard values ​​determined based on extensive historical data and actual operational experience, and can be adaptively adjusted based on actual conditions. They represent reasonable protection priority parameters for each line under different line conditions. For example, they can be values ​​determined based on statistical analysis and expert evaluation of actual fault identification results, or empirical values ​​derived from industry standards and long-term operations. In this embodiment, the final standardized feature set used for training is input into the initial adaptive multi-transmission line model, and multiple iterative training is performed. During each iteration, a predicted protection priority parameter is calculated and output based on the input standardized feature set. The error rate between the predicted parameter and the preset benchmark parameter is then calculated. Based on the error rate, a preset optimization algorithm (such as stochastic gradient descent) is used to update the model parameters so that the model output gradually approaches the preset benchmark parameters. For example, for a set of standardized feature sets, the initial adaptive multi-transmission line model outputs protection priority parameters for each line. These parameters are compared with the preset benchmark parameters, and the error rate is calculated. If the error rate is high, it indicates that the model output deviates significantly from the expected result, and the model parameters need to be adjusted. After multiple iterations of training, when the error rate falls below a preset error threshold, the model training is considered qualified, and the adaptive multi-transmission line model is obtained. For example, the initial adaptive multi-transmission line model can adopt a multi-layer perceptron (MLP) architecture, consisting of an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer matches the dimensionality of the standardized feature set and is responsible for receiving data from the standardized feature set. The number of neurons in the hidden layer can be adjusted based on actual conditions. Each hidden layer performs a nonlinear transformation on the input using an activation function (such as a ReLU function) to enhance the model's expressiveness. The number of neurons in the output layer matches the number of lines, and the output of each neuron is the protection priority parameter for the corresponding line. During the training process, a standardized feature set is used as input, and preset benchmark parameters are used as training labels. The error is calculated through the back-propagation algorithm and the weights and biases of the model are updated. This enables the model to learn the mapping relationship between the standardized feature set and the protection priority parameters, ultimately achieving accurate line protection priority judgment and improving the accuracy and reliability of line protection decisions.

[0086] As an implementation manner, step S50, iteratively training the initial adaptive multi-transmission line model using the standardized feature set, may specifically include: step S51: in each iteration process, inputting the current batch of standardized training samples into the initial adaptive multi-transmission line model, and outputting the predicted protection priority parameters.

[0087] During each training iteration, a batch of standardized training samples is selected from the standardized feature set used for final training and fed into the initial adaptive multi-transmission line model. The model performs calculations and analysis based on the input sample data and outputs predicted protection priority parameters. For example, if the current batch contains 100 standardized training samples, these samples are fed into the initial adaptive multi-transmission line model, and the model outputs the predicted protection priority parameters corresponding to these 100 samples.

[0088] Step S52: Calculate the mean square error between the predicted protection priority parameter and the corresponding reference parameter, and update the weight of the initial adaptive multi-transmission line model according to the mean square error.

[0089] The mean square error is an indicator that measures the degree of difference between a predicted value and a true value. It is obtained by calculating the average of the squares of the differences between the predicted protection priority parameters and the corresponding reference parameters. The weights of the model are parameters in the model, and the performance of the model can be adjusted by updating the weights. In an embodiment of the present application, the mean square error between the protection priority parameters predicted by the model and the corresponding reference parameters is calculated, and then the weights of the initial adaptive multi-transmission line model are updated based on this mean square error using a preset optimization algorithm (such as a stochastic gradient descent algorithm) to make the output of the model closer to the reference parameters. For example, the mean square error between the predicted protection priority parameters and the reference parameters is calculated to be 0.05. The weights of the model are updated based on this error value so that the model can output more accurate results in the next iteration.

[0090] Step S53: After each iteration, randomly select samples from the validation set to evaluate the model performance and calculate its fault detection accuracy and false trigger rate.

[0091] The validation set is a pre-partitioned set of data used to evaluate model performance during training. Fault detection accuracy refers to the proportion of faulty lines correctly detected by the model, while the false trigger rate refers to the proportion of protection actions incorrectly triggered by the model. After each training iteration, a certain number of samples are randomly selected from the validation set and fed into the currently trained model. Based on the model's output, the fault detection accuracy and false trigger rate are calculated to evaluate model performance. For example, if 200 samples are randomly selected from the validation set, the model correctly detects 180 faulty lines and falsely triggers 10 protection actions, resulting in a calculated fault detection accuracy of 90% and a false trigger rate of 5%.

[0092] Step S54: If the fault detection accuracy does not improve and the false trigger rate does not decrease after multiple consecutive iterations, the early stopping mechanism is triggered and the current optimal model parameters are saved.

[0093] The early stopping mechanism is a strategy used to prevent model overfitting. If the model's fault detection accuracy does not improve and the false trigger rate does not decrease over multiple consecutive iterations, it indicates that the model may have reached its optimal state or is beginning to overfit. At this time, the early stopping mechanism is triggered, the training process is stopped, and the current optimal model parameters are saved. For example, if the judgment criteria are set as five consecutive iterations, and the model's fault detection accuracy remains at 90% and the false trigger rate remains at 5% over these five consecutive iterations, the early stopping mechanism is triggered and the current model parameters are saved.

[0094] Step S55: loading the optimal model parameters into the initial adaptive multi-transmission line model to generate a pre-trained adaptive multi-transmission line model.

[0095] After triggering the early stopping mechanism and saving the optimal model parameters, these parameters are loaded into the initial adaptive multi-transmission line model, replacing the original parameters to generate a pre-trained adaptive multi-transmission line model. This trained and optimized model can more accurately output protection priority parameters for each line based on the input joint feature sequence, providing more reliable support for subsequent fiber differential protection decisions. For example, by loading the saved optimal model parameters into the initial adaptive multi-transmission line model, parameter replacement is completed, generating a pre-trained adaptive multi-transmission line model that can be used in actual line protection decision-making.

[0096] Step S300: Dynamically matching the differential protection threshold corresponding to each line according to the protection priority parameter to generate an adaptive trigger condition for each line.

[0097] The differential protection threshold is a current value standard used to determine whether the line needs to trigger the differential protection action. Different lines may correspond to different differential protection thresholds. The adaptive triggering condition is generated based on the differential protection threshold obtained by dynamic matching of the protection priority parameters, which can adaptively adjust the conditions for triggering the protection action according to the actual situation of the line. In an embodiment of the present application, according to the protection priority parameters of each line, the corresponding differential protection threshold is dynamically matched from the pre-set threshold range, and then the adaptive triggering conditions for each line are generated based on these thresholds. For example, for lines with higher protection priorities, the matched differential protection threshold may be lower, and the corresponding adaptive triggering conditions are also stricter; while for lines with lower protection priorities, the matched differential protection threshold may be higher, and the adaptive triggering conditions are relatively loose.

[0098] As an implementation method, step S300 dynamically matches the differential protection threshold corresponding to each line according to the protection priority parameters to generate an adaptive trigger condition for each line, which may specifically include: step S310: obtaining the historical differential current peak data of each line, and extracting the reference peak range in the historical differential current peak data that matches the current environmental parameters.

[0099] The historical differential current peak data refers to the maximum value record of the differential current that appeared in each line in the past period of time, which reflects the current peak situation of the line under different circumstances. The current environmental parameters refer to the environmental conditions in which the line is currently located, such as temperature, humidity, voltage, etc. Different environmental parameters may affect the current characteristics of the line. The reference peak range is the current peak range that matches the current environmental parameters screened out from the historical differential current peak data, which provides a reference basis for the subsequent determination of the differential protection threshold. In an embodiment of the present application, the historical differential current peak data of each line is obtained through a data storage system, and then a reference peak range that matches the current environmental parameters is screened out from these data. For example, the current ambient temperature is 25°C, and the current peak range when the temperature is around 25°C is screened out from the historical differential current peak data as the reference peak range.

[0100] Step S320: adjusting the upper limit of the reference peak range according to the final protection priority parameter to generate an initial differential protection threshold.

[0101] The initial differential protection threshold is a threshold value obtained by adjusting the upper limit value of the reference peak range according to the final protection priority parameter, which preliminarily determines the current standard for the line to trigger the differential protection action. In the embodiment of the present application, the upper limit value of the reference peak range is adjusted according to the final protection priority parameter of each line to obtain the initial differential protection threshold value. For example, for lines with a higher final protection priority, the upper limit value of the reference peak range is appropriately lowered, and the generated initial differential protection threshold value is also lower; for lines with a lower final protection priority, the upper limit value of the reference peak range is appropriately increased, and the generated initial differential protection threshold value is also higher.

[0102] Step S330: collecting signal transmission delay parameters between adjacent nodes of each line in real time, and dynamically compensating the initial differential protection threshold according to the signal transmission delay parameters.

[0103] The signal transmission delay parameter refers to the time required for signal transmission between adjacent nodes in a line, which reflects the delay of the signal during transmission. Since the signal transmission delay may cause errors in current measurement, it is necessary to dynamically compensate the initial differential protection threshold to improve the accuracy of protection. In an embodiment of the present application, the signal transmission delay parameters between adjacent nodes of each line are collected in real time by a special measuring device, and then the initial differential protection threshold is adjusted according to these parameters to obtain the compensated differential protection threshold. For example, when the signal transmission delay is large, the initial differential protection threshold is appropriately increased to avoid false triggering due to delay; when the signal transmission delay is small, the initial differential protection threshold can be appropriately lowered to improve the sensitivity of protection.

[0104] Step S340: If it is detected that the compensated differential protection threshold of the target line is lower than the preset minimum protection threshold, the preset minimum protection threshold is used as the final trigger condition of the target line; otherwise, the compensated differential protection threshold is multiplied by the deviation coefficient of the real-time differential current value to generate the adaptive trigger condition of the target line.

[0105] The preset minimum protection threshold is a pre-set minimum differential protection threshold that ensures basic line safety. When the compensated differential protection threshold is lower than this value, the preset minimum protection threshold is used as the final trigger condition to ensure line safety. The deviation coefficient is a coefficient calculated based on the difference between the real-time differential current value and the compensated differential protection threshold. It is used to adjust the compensated differential protection threshold to generate an adaptive trigger condition that is more in line with the actual situation. In an embodiment of the present application, the compensated differential protection threshold of the target line is judged. If it is lower than the preset minimum protection threshold, the preset minimum protection threshold is used as the final trigger condition; otherwise, the compensated differential protection threshold is multiplied by the deviation coefficient to obtain the adaptive trigger condition of the target line. For example, the compensated differential protection threshold of the target line is 5A, the preset minimum protection threshold is 3A, and the deviation coefficient between the real-time differential current value and the compensated differential protection threshold is 0.8. Since 5A is greater than 3A, 5A is multiplied by 0.8 to obtain an adaptive trigger condition of 4A.

[0106] Step S400: monitor the actual differential current value of each line in real time. If it is detected that the actual differential current value of the target line reaches its corresponding adaptive trigger condition, activate the differential protection action for the target line.

[0107] The actual differential current value refers to the actual differential current value of the line at the current moment, which is obtained in real time by the current measuring device. The differential protection action is a series of protection measures that are activated to protect the safety of the line when the actual differential current value of the line reaches the adaptive trigger condition, such as tripping, alarming, etc. In the embodiment of the present application, the actual differential current value of each line is monitored in real time by a current monitoring device. When it is detected that the actual differential current value of the target line reaches its corresponding adaptive trigger condition, the differential protection action for the target line is immediately activated. For example, the adaptive trigger condition of the target line is 4A. When the actual differential current value of the line is monitored in real time to reach 4A, the differential protection action is activated.

[0108] As an implementation mode, in step S400, if it is detected that the actual differential current value of the target line reaches its corresponding adaptive trigger condition, the differential protection action for the target line is activated, which may specifically include: step S410: when the actual differential current value of the target line exceeds its adaptive trigger condition for the first time, the protection action delay counter is started, and the harmonic distortion rate of the target line is continuously collected during the delay period.

[0109] The protection action delay counter is a device used to record the delay time from the first time the actual differential current value exceeds the adaptive trigger condition to the time the protection action is triggered, which can avoid false triggering due to instantaneous current fluctuations. Harmonic distortion rate refers to the ratio of the harmonic content to the fundamental content of the current in the target line, which reflects the degree of distortion of the current waveform. Excessively high harmonic distortion rate may indicate that there is a fault in the line. In an embodiment of the present application, when the actual differential current value of the target line exceeds its adaptive trigger condition for the first time, the protection action delay counter is immediately started, and the harmonic distortion rate of the target line is continuously collected during the delay period through a special measuring device. For example, the adaptive trigger condition of the target line is 5A. When the actual differential current value reaches 5.1A, the protection action delay counter is started, and the harmonic distortion rate of the line is continuously collected during the delay of 10 seconds.

[0110] Step S420: If the harmonic distortion rate continues to exceed the preset distortion threshold during the delay period, a differential protection trip command is triggered.

[0111] The preset distortion threshold is a pre-set standard value used to determine whether the harmonic distortion rate is too high. When the harmonic distortion rate exceeds the threshold, it indicates that there may be a serious fault in the line and protective measures need to be taken immediately. The differential protection tripping instruction is an instruction for cutting off the power supply of the line, which can prevent the fault from further expanding. In an embodiment of the present application, during the protection action delay period, if it is detected that the harmonic distortion rate of the target line continues to exceed the preset distortion threshold, the differential protection tripping instruction is immediately triggered. For example, the preset distortion threshold is 10%. Within the 10-second delay, if the harmonic distortion rate of the target line remains above 12%, the differential protection tripping instruction is triggered to cut off the power supply of the line.

[0112] Step S430: If the harmonic distortion rate does not exceed the preset distortion threshold, recheck the deviation between the actual differential current value of the target line and the adaptive trigger condition after the delay ends.

[0113] The deviation refers to the difference between the actual differential current value of the target line and the adaptive trigger condition, which reflects the degree to which the actual current value deviates from the trigger condition. In this embodiment of the present application, if the harmonic distortion rate of the target line does not exceed the preset distortion threshold during the protection action delay period, then after the delay period, the actual differential current value of the target line is measured again, and the deviation from the adaptive trigger condition is calculated. For example, if the adaptive trigger condition of the target line is 6A, and the actual differential current value measured after the delay period is 6.2A, the calculated deviation is 0.2A.

[0114] Step S440: When the deviation is greater than the preset buffer range, a differential protection alarm signal is triggered and a backup line switching protocol is started.

[0115] The preset buffer zone is a pre-set range that allows the actual differential current value to fluctuate around the adaptive trigger condition. When the deviation exceeds this range, it indicates that there may be a potential fault in the line. The differential protection alarm signal is a signal used to remind the operation and maintenance personnel that there may be a problem with the line. The backup line switching protocol is a series of operating rules for switching the load to the backup line when a fault occurs in the main line. In an embodiment of the present application, when the deviation between the actual differential current value of the target line and the adaptive trigger condition is greater than the preset buffer zone, the differential protection alarm signal is triggered and the backup line switching protocol is started at the same time. For example, if the preset buffer zone is 0.1A and the deviation of the target line is 0.2A, which is greater than the preset buffer zone, the differential protection alarm signal is triggered, and the backup line switching protocol is started to switch the load to the backup line.

[0116] Step S450: When the deviation is less than or equal to the preset buffer period, the protection action delay counter is reset and the actual differential current value of the target line is continuously monitored.

[0117] In this embodiment of the present application, if the deviation between the actual differential current value of the target line and the adaptive trigger condition is less than or equal to the preset buffer range, indicating that the current fluctuation of the line is within the acceptable range, the protection action delay counter is reset so that the next time a current anomaly is detected, the timing will be restarted and the actual differential current value of the target line will continue to be monitored in real time. For example, if the preset buffer range is 0.1A and the deviation of the target line is 0.05A, which is less than the preset buffer range, the protection action delay counter is reset and the actual differential current value of the line will continue to be monitored.

[0118] Step S500: adjusting the monitoring frequencies of the remaining lines that have not triggered protection actions according to the priority parameters output by the adaptive multi-transmission line model.

[0119] The monitoring frequency refers to the time interval for data collection and status monitoring of the line. Adjusting the monitoring frequency can reasonably allocate monitoring resources according to the importance and status of the line. In the embodiment of the present application, the monitoring frequency of the remaining lines that have not triggered the protection action is adjusted according to the priority parameters of each line output by the adaptive multi-transmission line model. For lines with higher priorities, the monitoring frequency is increased to detect potential problems in a timely manner; for lines with lower priorities, the monitoring frequency is appropriately reduced to reduce the waste of monitoring resources. For example, if the adaptive multi-transmission line model outputs a higher priority for a certain line, the monitoring frequency of the line will be increased from once per hour to once every half hour.

[0120] As an implementation method, step S500 adjusts the monitoring frequency of the remaining lines that have not triggered protection actions based on the priority parameters output by the adaptive multi-transmission line model, which may specifically include: step S510: obtaining the final protection priority parameter corresponding to each line output by the adaptive multi-transmission line model, and dividing the final protection priority parameter into a high priority interval, a medium priority interval, and a low priority interval.

[0121] The final protection priority parameter is a parameter used to measure the line protection priority after comprehensively considering various factors of the line. The high priority interval, medium priority interval, and low priority interval are pre-set ranges for dividing line priorities. In an embodiment of the present application, the final protection priority parameter corresponding to each line is obtained from the adaptive multi-transmission line model, and then these parameters are divided into a high priority interval, a medium priority interval, and a low priority interval according to the pre-set range. For example, the final protection priority parameter range [1, 10] is divided into a high priority interval [7, 10], a medium priority interval [3, 6], and a low priority interval [1, 2], and each line is divided into the corresponding interval according to its final protection priority parameter.

[0122] Step S520: Generate corresponding initial monitoring frequency adjustment factors according to the priority intervals of each line, wherein the high priority interval corresponds to the first adjustment factor, the medium priority interval corresponds to the second adjustment factor, and the low priority interval corresponds to the third adjustment factor.

[0123] The initial monitoring frequency adjustment factor is a coefficient used to adjust the initial monitoring frequency of the line, and different priority intervals correspond to different adjustment factors. The first adjustment factor, the second adjustment factor, and the third adjustment factor are adjustment factors corresponding to the high priority interval, the medium priority interval, and the low priority interval, respectively. They have different values ​​and are used to achieve differentiated adjustments to the monitoring frequencies of lines with different priorities. In the embodiment of the present application, corresponding initial monitoring frequency adjustment factors are generated according to the priority interval to which each line belongs. For example, the first adjustment factor corresponding to the high priority interval is 1.5, the second adjustment factor corresponding to the medium priority interval is 1, and the third adjustment factor corresponding to the low priority interval is 0.5.

[0124] Step S530: In real time, the signal interference strength and line load fluctuation rate between adjacent nodes of the line that has not triggered the protection action are collected, and the signal interference strength and load fluctuation rate are input into the dynamic frequency compensation model to output the real-time dynamic adjustment parameters of each line.

[0125] Signal interference intensity refers to the degree of interference to the signal between adjacent nodes of a line, which affects the signal transmission quality and the accuracy of current measurement. Line load fluctuation rate refers to the degree of fluctuation of the line load over time, which reflects the stability of the line load. The dynamic frequency compensation model is a pre-trained model that can output real-time dynamic adjustment parameters for each line based on the input signal interference intensity and line load fluctuation rate, which are used to further adjust the monitoring frequency of the line. In an embodiment of the present application, the signal interference intensity and line load fluctuation rate between adjacent nodes of the line that has not triggered the protection action are collected in real time by a special measuring device. For example, a signal strength tester is used to measure the signal interference intensity between adjacent nodes, and the line load fluctuation rate is calculated by statistically analyzing the line load data. The collected signal interference intensity and load fluctuation rate are then used as input data and input into the dynamic frequency compensation model. After internal calculation and analysis, the model outputs the real-time dynamic adjustment parameters corresponding to each line.

[0126] For example, the dynamic frequency compensation model can employ a long short-term memory (LSTM) network, consisting of an input layer, an LSTM layer, and an output layer. The input layer receives two data series: signal interference intensity and line load fluctuation, and converts them into a format suitable for processing by the LSTM layer. The LSTM layer consists of multiple LSTM units, each of which includes an input gate, a forget gate, and an output gate. These gating mechanisms control the flow and memory of information, thereby learning the patterns and regularities of signal interference intensity and line load fluctuation over time. The output layer, based on the LSTM layer outputs, performs linear transformations and other operations to output the real-time dynamic adjustment parameters for each line. During the training phase, a large amount of historical signal interference intensity and line load fluctuation data, along with the corresponding actual monitored frequency adjustment results, can be used as training samples. The model parameters are continuously adjusted through a backpropagation algorithm, enabling the model to accurately predict appropriate real-time dynamic adjustment parameters based on the input signal interference intensity and line load fluctuation. This improves the accuracy and effectiveness of line monitoring frequency adjustment and better adapts to the real-time operating conditions of the line.

[0127] Step S540: superimpose the initial monitoring frequency adjustment factor and the corresponding real-time dynamic adjustment parameter to generate the final monitoring frequency coefficient of each line that has not triggered the protection action, and adjust the sampling interval of its data acquisition device according to the final monitoring frequency coefficient.

[0128] The final monitoring frequency coefficient is the coefficient obtained by superimposing the initial monitoring frequency adjustment factor and the real-time dynamic adjustment parameter. It comprehensively considers the priority and real-time status of the line and is used to determine the final monitoring frequency of the line. The sampling interval of the data acquisition device refers to the time interval for the data acquisition device to collect data once. The line monitoring frequency can be adjusted by adjusting the sampling interval. In an embodiment of the present application, the initial monitoring frequency adjustment factor of each line that has not triggered the protection action is added to the corresponding real-time dynamic adjustment parameter to obtain the final monitoring frequency coefficient, and then the sampling interval of the data acquisition device is adjusted according to the coefficient. For example, the initial monitoring frequency adjustment factor of a line is 1.2, and the real-time dynamic adjustment parameter is 0.8. The final monitoring frequency coefficient obtained by adding them together is 2. According to this coefficient, the sampling interval of the data acquisition device is shortened from the original 10 minutes to 5 minutes.

[0129] Step S550: Continue to obtain the differential current feature set of each line at the adjusted sampling interval. If it is detected that the differential current feature set of the same line does not contain a feature subset exceeding the preset fluctuation threshold in multiple consecutive sampling periods, reduce its final monitoring frequency coefficient according to the preset attenuation ratio.

[0130] The preset fluctuation threshold is a pre-set standard value used to determine whether the differential current feature set has abnormal fluctuations, and the preset attenuation ratio is a pre-set ratio used to reduce the final monitoring frequency coefficient. In the embodiment of the present application, the differential current feature set of each line is continuously collected at the adjusted sampling interval. If it is found that the differential current feature set of the same line does not have a feature subset that exceeds the preset fluctuation threshold in multiple consecutive sampling periods, it indicates that the state of the line is relatively stable, and its final monitoring frequency coefficient is reduced according to the preset attenuation ratio. For example, if the preset fluctuation threshold is 5% and the preset attenuation ratio is 0.1, and the amplitude fluctuation rate of the differential current feature set of a certain line does not exceed 5% in 5 consecutive sampling periods, then the final monitoring frequency coefficient of the line is reduced from 2 to 1.8.

[0131] Step S560: Feedback the reduced final monitoring frequency coefficient to the dynamic frequency compensation model, update the weight distribution relationship of the real-time dynamic adjustment parameters, and form a closed-loop monitoring frequency control mechanism based on the linkage between priority parameters and line status.

[0132] The weight distribution relationship is the weight ratio of each factor used to calculate the real-time dynamic adjustment parameters in the dynamic frequency compensation model. In an embodiment of the present application, the final monitoring frequency coefficient after the reduction is fed back to the dynamic frequency compensation model, and the model updates the weight distribution relationship of the real-time dynamic adjustment parameters according to this new coefficient, thereby forming a closed-loop monitoring frequency control mechanism based on the priority parameter and the line status linkage. This mechanism can automatically adjust the monitoring frequency according to the priority and real-time status of the line to achieve optimal allocation of monitoring resources. For example, the final monitoring frequency coefficient after the reduction is 1.8 and fed back to the dynamic frequency compensation model. The model updates the weight distribution relationship of the signal interference intensity and the line load fluctuation rate, so that the subsequently calculated real-time dynamic adjustment parameters are more in line with the actual status of the line.

[0133] As an embodiment, after forming a closed-loop monitoring frequency control mechanism based on the linkage between priority parameters and line status in step S560, the method provided by the present invention may also include: step S570: based on the final monitoring frequency coefficient of each line in the closed-loop monitoring frequency control mechanism, generating a dynamic monitoring priority queue for each line, and allocating the hardware resource occupancy ratio of the data acquisition equipment according to the dynamic monitoring priority queue.

[0134] The dynamic monitoring priority queue is a queue generated based on the final monitoring frequency coefficient of each line, and the lines in the queue are arranged from high to low according to the monitoring priority. The hardware resource occupancy ratio refers to the ratio of hardware resources allocated by the data acquisition device when monitoring different lines. Reasonable allocation of the hardware resource occupancy ratio can improve monitoring efficiency. In the embodiment of the present application, a dynamic monitoring priority queue for each line is generated based on the final monitoring frequency coefficient of each line in the closed-loop monitoring frequency control mechanism, and then the hardware resource occupancy ratio of the data acquisition device is allocated based on this queue. For example, a line with a higher final monitoring frequency coefficient ranks higher in the dynamic monitoring priority queue, and the hardware resource occupancy ratio of the data acquisition device allocated to the line is also higher.

[0135] Step S580: acquiring in real time the differential current feature set collected from each line at the adjusted sampling interval, and extracting potential abnormal segments in the differential current feature set that match the historical abnormal waveform.

[0136] Potential abnormal segments refer to portions of the differential current feature set that are similar to historical abnormal waveforms. These segments may indicate potential faults in the line. In the embodiment of the present application, the differential current feature set of each line is continuously collected at an adjusted sampling interval, and then the collected differential current feature set is compared with the historical abnormal waveform to extract potential abnormal segments that match the historical abnormal waveform. For example, if the historical abnormal waveform exhibits a sudden increase in current amplitude, and a segment of the waveform in the collected differential current feature set also exhibits a sudden increase in current amplitude, this segment of the waveform is extracted as a potential abnormal segment.

[0137] Step S590: input the potential abnormal segment into the abnormality prediction module in the adaptive multi-transmission line model, and output the real-time abnormality confidence parameter of each line.

[0138] The anomaly prediction module is a module within the adaptive multi-transmission line model that analyzes and determines potential anomaly segments. The real-time anomaly confidence parameter is a parameter output by the anomaly prediction module based on the input potential anomaly segments, indicating the likelihood of an anomaly on the line. In this embodiment of the present application, the extracted potential anomaly segments are input into the anomaly prediction module of the adaptive multi-transmission line model. After analysis and calculation, the module outputs the real-time anomaly confidence parameter for each line. For example, if a potential anomaly segment for a particular line is input into the anomaly prediction module, the module outputs a real-time anomaly confidence parameter of 0.8 for that line, indicating a high probability of an anomaly on that line.

[0139] For example, the anomaly prediction module can adopt a convolutional neural network (CNN) architecture, consisting of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer receives data on potential anomaly segments and converts it into a format suitable for processing by the convolutional layer. The convolutional layer, composed of multiple convolution kernels, extracts local features from potential anomaly segments through convolution operations. Each convolution kernel learns different characteristic patterns, such as the changing trend of current amplitude or frequency fluctuations. The pooling layer downsamples the output of the convolutional layer to reduce the data dimension while retaining important feature information and enhancing the robustness of the model. The fully connected layer integrates the features output by the pooling layer and establishes global connections between the features. The output layer, after undergoing linear transformation and activation function processing (such as the sigmoid function), outputs real-time anomaly confidence parameters for each line. The sigmoid function maps the output value to a range between 0 and 1 to indicate the probability of an anomaly on the line. During the training phase, historical potential abnormal fragment data and the corresponding actual abnormal situations can be used as training samples. The parameters of the model can be continuously adjusted through the general back-propagation algorithm, so that the abnormality prediction module can accurately output reasonable real-time abnormality confidence parameters based on the input potential abnormal fragments, thereby improving the accuracy of predicting line abnormalities.

[0140] Step S5100: dynamically adjust the sorting weight of the dynamic monitoring priority queue according to the real-time anomaly confidence parameter, and reallocate the hardware resource occupancy ratio to give priority to processing the lines whose real-time anomaly confidence parameter is higher than the preset threshold.

[0141] The sorting weight is a weight value used to determine the order of lines in the dynamic monitoring priority queue, and the preset threshold is a pre-set standard value used to determine whether the possibility of line abnormality is high. In an embodiment of the present application, the sorting weight of the dynamic monitoring priority queue is dynamically adjusted according to the real-time abnormality confidence parameter of each line, and the lines with real-time abnormality confidence parameters higher than the preset threshold are placed in front, and then the hardware resource occupancy ratio of the data acquisition equipment is reallocated according to the new queue, and these lines are monitored and processed first. For example, the preset threshold is 0.7, and the real-time abnormality confidence parameter of a certain line is 0.8. The sorting weight of the line in the dynamic monitoring priority queue is increased, and the hardware resource occupancy ratio of the data acquisition equipment allocated to the line is increased accordingly.

[0142] Step S5110: The reallocated hardware resource occupancy ratio is coupled with the final monitoring frequency coefficient to generate an optimized monitoring strategy for each line, and the frequency adjustment rules in the closed-loop monitoring frequency control mechanism are updated according to the optimized monitoring strategy.

[0143] Coupling is an operation that comprehensively considers and combines the proportion of hardware resources occupied after reallocation with the final monitoring frequency coefficient. The optimized monitoring strategy is a more reasonable and effective monitoring strategy obtained through coupling, which comprehensively considers the priority of the line, the possibility of abnormality and the allocation of hardware resources. The frequency adjustment rule is a rule used to adjust the line monitoring frequency in the closed-loop monitoring frequency control mechanism. Updating it according to the optimized monitoring strategy can make the closed-loop monitoring frequency control mechanism more accurate and effective. In an embodiment of the present application, the proportion of hardware resources occupied after reallocation is coupled with the final monitoring frequency coefficient to generate an optimized monitoring strategy for each line, and then the frequency adjustment rules in the closed-loop monitoring frequency control mechanism are updated according to this strategy. For example, the proportion of hardware resources occupied after the reallocation of a certain line increases, and the final monitoring frequency coefficient is also higher. The optimized monitoring strategy generated after coupling is to further increase the monitoring frequency of the line. The frequency adjustment rules in the closed-loop monitoring frequency control mechanism are updated according to this strategy, so that the line can receive more timely attention in subsequent monitoring.

[0144] As an implementation method, in step S5110, after generating the optimized monitoring strategy for each line, the method provided by the present invention may further include: step S5120: constructing a cross-line abnormality association network based on the real-time abnormality confidence parameters of each line in the optimized monitoring strategy, wherein each node represents a line, and the connection weight between the nodes represents the synchronous change rate of the abnormality confidence parameters of the two lines.

[0145] The abnormal association network is a network used to represent the abnormal association relationship between each line. The nodes represent the lines, and the connection weights represent the synchronous change rate of the abnormal confidence parameters of the two lines. The synchronous change rate reflects the correlation between the abnormal conditions of the two lines. In an embodiment of the present application, an abnormal association network across lines is constructed based on the real-time abnormal confidence parameters of each line in the optimized monitoring strategy. For example, there are three lines A, B, and C. The synchronous change rate of the abnormal confidence parameters of line A and line B is calculated to be 0.6, the synchronous change rate of line A and line C is 0.3, and the synchronous change rate of line B and line C is 0.4. Based on these data, an abnormal association network is constructed, and the connection weights between nodes A, B, and C are 0.6, 0.3, and 0.4, respectively.

[0146] Step S5130: traverse the node pairs whose connection weights exceed the preset synchronization threshold in the abnormal association network, and mark the node pairs as an association abnormality group.

[0147] The preset synchronization threshold is a pre-set standard value used to determine whether the abnormal conditions of two lines are highly correlated. The associated abnormal group is a group composed of node pairs in the abnormal associated network whose connection weight exceeds the preset synchronization threshold. The abnormal conditions of these lines may affect each other. In an embodiment of the present application, the connection weights of all node pairs in the abnormal associated network are traversed to determine the node pairs whose connection weight exceeds the preset synchronization threshold, and these node pairs are marked as associated abnormal groups. For example, the preset synchronization threshold is 0.5. In the abnormal associated network, the connection weight of line A and line B is 0.6, which exceeds the preset synchronization threshold. Line A and line B are marked as an associated abnormal group.

[0148] Step S5140: For each line in the associated anomaly group, the final monitoring frequency coefficient is synchronously adjusted so that the lines in the same associated anomaly group use the same sampling interval for data collection.

[0149] In this embodiment of the present application, for each line group marked as an associated anomaly group, the final monitoring frequency coefficients of each line within the group are synchronously adjusted so that these lines use the same sampling interval for data collection. This ensures consistent monitoring of lines within the associated anomaly group and facilitates the detection and handling of cross-line anomalies. For example, if an associated anomaly group contains lines A and B, their final monitoring frequency coefficients are adjusted to the same value, ensuring that their data collection devices use the same sampling interval for data collection.

[0150] Step S5150: At the synchronously adjusted sampling interval, the differential current feature set of all lines in the associated abnormal group is collected, and whether there is a chain abnormal waveform pattern across the lines.

[0151] A chained abnormal waveform pattern refers to a correlated abnormal waveform pattern that appears simultaneously on multiple lines within an associated abnormal group, which may indicate the presence of a cross-line fault. In an embodiment of the present application, the differential current feature set of all lines within the associated abnormal group is collected at a synchronously adjusted sampling interval, and then the collected feature set is analyzed to detect whether there is a chained abnormal waveform pattern across lines. For example, if a waveform pattern with a sudden increase in current amplitude and similar phase changes is detected simultaneously on two lines within an associated abnormal group, it is considered a chained abnormal waveform pattern across lines.

[0152] Step S5160: If a chain abnormal waveform pattern is detected, the global protection coordination protocol is triggered, the final monitoring frequency coefficient of all lines in the associated abnormal group is forcibly increased to the preset maximum monitoring frequency, and the effective time of the global protection coordination protocol is locked.

[0153] The global protection collaborative protocol is a protocol for coordinating the handling of cross-line abnormalities. The preset maximum monitoring frequency is a pre-set maximum monitoring frequency that ensures that abnormalities can be discovered and handled in a timely manner. The effective duration is the effective time range of the global protection collaborative protocol. In an embodiment of the present application, if it is detected that there is a chain abnormal waveform pattern across lines in the associated abnormal group, the global protection collaborative protocol is immediately triggered, and the final monitoring frequency coefficients of all lines in the associated abnormal group are forced to be increased to the preset maximum monitoring frequency, and the effective duration of the global protection collaborative protocol is locked. For example, the preset maximum monitoring frequency is once per minute. After the chain abnormal waveform pattern is detected, the final monitoring frequency coefficients of all lines in the associated abnormal group are adjusted to the corresponding monitoring frequency of once per minute, and the effective duration of the global protection collaborative protocol is locked to 30 minutes.

[0154] Step S5170: After the validity period ends, the locking state of the global protection coordination protocol is released, and the final monitoring frequency coefficient of each line based on the closed-loop monitoring frequency control mechanism is restored.

[0155] After the global protection coordination agreement expires, the agreement is unlocked and the preset maximum monitoring frequency is no longer enforced. Instead, the final monitoring frequency coefficient of each line based on the closed-loop monitoring frequency control mechanism is restored, returning the line's monitoring frequency to its normal control state. For example, if the global protection coordination agreement expires after 30 minutes, the agreement is unlocked and the final monitoring frequency coefficient of each line is restored to the value determined by the closed-loop monitoring frequency control mechanism.

[0156] As an embodiment, in step S5170, after restoring the final monitoring frequency coefficient of each line based on the closed-loop monitoring frequency control mechanism, the method provided by the present invention may also include: step S5180: obtaining the differential current feature set of all lines in the associated abnormal group collected during the effectiveness of the global protection coordination protocol, and extracting the abnormal waveform recovery parameters in the differential current feature set.

[0157] Abnormal waveform recovery parameters refer to parameters used to describe the abnormal waveform recovery in the differential current feature set, which can reflect the degree of recovery of the abnormal state of the line. In an embodiment of the present application, the differential current feature sets of all lines in the associated abnormal group collected during the period when the global protection coordination protocol is in effect are obtained from the data storage system, and then these feature sets are analyzed to extract the abnormal waveform recovery parameters therein. For example, by analyzing the changes in parameters such as current amplitude, frequency, and phase, the abnormal waveform recovery parameters are extracted, such as the time it takes for the current amplitude to return to the normal range, the degree of convergence of frequency fluctuations, etc.

[0158] Step S5190: Determine whether the abnormal state of each line has been eliminated based on the abnormal waveform recovery parameter. If eliminated, generate a corresponding abnormal recovery flag; otherwise, generate an abnormality continuation flag.

[0159] The abnormal recovery mark is a mark used to indicate that the abnormal state of the line has been eliminated, and the abnormal persistence mark is a mark used to indicate that the abnormal state of the line still exists. In the embodiment of the present application, the abnormal state of each line is judged based on the extracted abnormal waveform recovery parameters. If the abnormal state has been eliminated, a corresponding abnormal recovery mark is generated; if the abnormal state still exists, an abnormal persistence mark is generated. For example, the abnormal waveform recovery parameters of a certain line show that the current amplitude and frequency have returned to the normal range. It is judged that the abnormal state of the line has been eliminated and an abnormal recovery mark is generated; the abnormal waveform recovery parameters of another line show that the current amplitude still fluctuates greatly. It is judged that the abnormal state of the line still exists and an abnormal persistence mark is generated.

[0160] Step S5200: Input the abnormal recovery mark into the closed-loop monitoring frequency control mechanism, trigger the gradient attenuation mechanism of the final monitoring frequency coefficient of the corresponding line, and gradually reduce its sampling interval until it reaches the initial setting value in the closed-loop monitoring frequency control mechanism.

[0161] The gradient attenuation mechanism is a mechanism used in the closed-loop monitoring frequency control mechanism to gradually reduce the final monitoring frequency coefficient, which can gradually restore the monitoring frequency of the line to a normal level after the abnormal state is eliminated. In an embodiment of the present application, the abnormal recovery mark is input into the closed-loop monitoring frequency control mechanism, triggering the gradient attenuation mechanism for the final monitoring frequency coefficient of the corresponding line, and gradually reducing the final monitoring frequency coefficient of the line according to a certain gradient, thereby gradually increasing its sampling interval until the initial setting value in the closed-loop monitoring frequency control mechanism is reached. For example, after the abnormal state of a certain line is eliminated, the gradient attenuation mechanism is triggered, and the final monitoring frequency coefficient is reduced by 0.1 at regular intervals until the initial setting value is reached.

[0162] Step S5210: input the abnormal persistence flag into the priority re-evaluation module in the adaptive multi-transmission line model, recalculate the final protection priority parameter of the corresponding line, and update its final monitoring frequency coefficient according to the recalculated parameter.

[0163] The priority reassessment module is a module within the adaptive multi-transmission line model used to reassess line priorities. It can recalculate the final protection priority parameters based on the line's latest status. In this embodiment of the present application, the abnormality persistence flag is input into the priority reassessment module of the adaptive multi-transmission line model. This module uses the abnormality persistence flag as input and, in conjunction with the line's historical operating data, previous protection priority parameters, and other relevant line status information, recalculates the final protection priority parameters for the corresponding line. For example, it analyzes factors such as the duration of the abnormality and its severity, and comprehensively considers and calculates these factors based on pre-set weights and rules.

[0164] Exemplarily, the priority reassessment module can be implemented using a rule-based architecture that combines machine learning. The rule component contains a series of pre-defined rules. For example, if an anomaly persists for longer than a preset duration, the final protection priority parameter for the line is appropriately increased. If the severity of the anomaly reaches a certain level, the priority parameter is adjusted accordingly. Machine learning can employ a decision tree algorithm. The decision tree uses the anomaly persistence flag and other relevant line status information as input features and constructs a decision tree model by learning from historical data. Each internal node in the decision tree model is a test on a feature, each branch is a test output, and each leaf node is a category (which can be understood here as a different level of the final protection priority parameter). When a new anomaly persistence flag and related information are input, the decision tree makes a judgment based on its internal decision rules and outputs the corresponding final protection priority parameter. After obtaining the recalculated final protection priority parameter, the priority reassessment module updates the final monitoring frequency coefficient for the line based on the pre-defined correspondence. For example, if the final protection priority parameter increases, the final monitoring frequency coefficient is increased accordingly, thereby increasing the line's monitoring frequency to more promptly detect further abnormalities. Conversely, if the final protection priority parameter decreases, the final monitoring frequency coefficient is appropriately reduced to reduce unnecessary monitoring resource consumption. Based on this, the priority reassessment module can dynamically adjust the line's priority and monitoring frequency when a line abnormality persists, making the entire line monitoring system more flexible and efficient, and enhancing its ability to respond to line abnormalities.

[0165] Step S5220: Synchronize the updated final monitoring frequency coefficient to the connection weight calculation of the dynamic monitoring priority queue and the associated anomaly group, forming a dynamic frequency backtracking mechanism based on the anomaly recovery state to adaptively balance resource allocation and anomaly detection sensitivity in subsequent monitoring cycles.

[0166] The dynamic frequency backtracking mechanism is a mechanism that dynamically adjusts the monitoring frequency according to the abnormal recovery state of the line, which can adaptively balance the allocation of monitoring resources and the sensitivity of abnormal detection in subsequent monitoring cycles. In an embodiment of the present application, the updated final monitoring frequency coefficient is synchronized to the connection weight calculation of the dynamic monitoring priority queue and the associated abnormal group, so that these mechanisms can be adjusted according to the latest state of the line, forming a dynamic frequency backtracking mechanism based on the abnormal recovery state. For example, after the final monitoring frequency coefficient of a certain line is updated, it is synchronized to the dynamic monitoring priority queue, and the position of the line in the queue is adjusted. At the same time, the new state of the line is taken into account in the connection weight calculation of the associated abnormal group, so as to more reasonably allocate resources and improve the sensitivity of abnormal detection in subsequent monitoring.

[0167] The embodiment of the present invention provides a fiber differential protection decision system, such as Figure 2 As shown, the optical fiber differential protection decision system 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the optical fiber differential protection decision system 100 may further include a transceiver 104. It should be noted that in practical applications, the number of transceivers 104 is not limited to one, and the structure of the optical fiber differential protection decision system 100 does not constitute a limitation on the embodiments of the present invention.

[0168] Processor 101 may be a CPU, a general-purpose processor, a GPU, a DSP, an ASIC, an FPGA, or other programmable logic device, a transistor logic device, a hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the present disclosure. Processor 101 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0169] The bus 102 may include a path for transmitting information between the above components. The bus 102 may be a PCI bus or an EISA bus, etc. The bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 2 Only one thick line is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0170] The memory 103 may be a ROM or other type of static storage device that can store static information and instructions, a RAM or other type of dynamic storage device that can store information and instructions, or an EEPROM, a CD-ROM or other optical disk storage, an optical disc storage (including a compact disc, a laser disc, an optical disc, a digital versatile disc, a Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto.

[0171] The memory 103 is used to store application code for executing the solution of the present invention, and is controlled by the processor 101. The processor 101 is used to execute the application code stored in the memory 103 to implement the content shown in any of the above method embodiments.

[0172] An embodiment of the present invention provides a fiber differential protection decision system. The fiber differential protection decision system in the embodiment of the present invention includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors. When the one or more programs are executed by the processors, the above method is implemented.

[0173] An embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program runs on a processor, the processor can execute the corresponding contents of the aforementioned method embodiment.

[0174] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0175] The above descriptions are only partial embodiments of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and modifications without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A fiber differential protection decision method based on adaptive multi-transmission lines, characterized in that: The method includes: collecting real-time optical signal waveform data of each line in multiple transmission lines and extracting a differential current feature set corresponding to each line; generating a joint feature sequence based on the differential current feature sets of all lines, and inputting the joint feature sequence into a pre-trained adaptive multi-transmission line model to obtain a protection priority parameter corresponding to each line; dynamically matching the differential protection threshold corresponding to each line according to the protection priority parameter to generate an adaptive trigger condition for each line; monitoring the actual differential current value of each line in real time, and activating the differential protection action for the target line if it is detected that the actual differential current value of the target line reaches its corresponding adaptive trigger condition; and adjusting the monitoring frequency of the remaining lines for which the protection action has not been triggered according to the priority parameter output by the adaptive multi-transmission line model.

2. The method according to claim 1, characterized in that The method of generating a joint feature sequence based on the differential current feature sets of all lines includes: selecting a feature subset whose amplitude fluctuation rate exceeds a preset fluctuation threshold from the differential current feature set of each line, and aligning each feature subset by time stamp; calculating the energy distribution parameter of the feature subset of each line within a preset time window, and extracting the maximum energy gradient value from the energy distribution parameter of each line; splicing the maximum energy gradient values ​​of each line into an initial joint sequence in line number order, and normalizing the initial joint sequence; matching the normalized initial joint sequence with a standard pattern in a historical joint feature library, and eliminating abnormal feature segments in the initial joint sequence whose overlap with historical abnormal patterns exceeds a preset overlap; and rearranging the remaining feature segments in chronological order to generate a final joint feature sequence.

3. The method according to claim 2, characterized in that Inputting the joint feature sequence into a pre-trained adaptive multi-transmission line model to obtain a protection priority parameter corresponding to each line includes: dividing the final joint feature sequence into multiple continuous time segments, each time segment containing a fixed number of feature points; inputting each time segment into a feature fusion layer in the adaptive multi-transmission line model, and outputting a line state weight coefficient corresponding to each time segment; calculating the cumulative weight value of each line within the overall monitoring period based on the line state weight coefficient of each time segment, and sorting the lines from high to low according to the cumulative weight value; mapping the sorted line sequence number to a preset priority interval to generate a dynamic protection priority parameter corresponding to each line; and weighted superposition of the dynamic protection priority parameter and the real-time load parameter to obtain a final protection priority parameter.

4. The method according to claim 3, characterized in that The method of dynamically matching the differential protection threshold corresponding to each line according to the protection priority parameter to generate an adaptive trigger condition for each line includes: obtaining historical differential current peak data of each line and extracting a reference peak range that matches the current environmental parameters in the historical differential current peak data; adjusting the upper limit of the reference peak range according to the final protection priority parameter to generate an initial differential protection threshold; collecting signal transmission delay parameters between adjacent nodes of each line in real time, and dynamically compensating the initial differential protection threshold according to the signal transmission delay parameters; if it is detected that the compensated differential protection threshold of the target line is lower than the preset minimum protection threshold, then using the preset minimum protection threshold as the final trigger condition of the target line; otherwise, multiplying the compensated differential protection threshold by the deviation coefficient of the real-time differential current value to generate the adaptive trigger condition of the target line.

5. The method according to claim 4, characterized in that If it is detected that the actual differential current value of the target line reaches its corresponding adaptive trigger condition, the differential protection action for the target line is activated, including: when the actual differential current value of the target line exceeds its adaptive trigger condition for the first time, the protection action delay counter is started, and the harmonic distortion rate of the target line is continuously collected during the delay period; if the harmonic distortion rate continues to exceed the preset distortion threshold during the delay period, the differential protection trip command is triggered; if the harmonic distortion rate does not exceed the preset distortion threshold, the deviation between the actual differential current value of the target line and the adaptive trigger condition is rechecked after the delay ends; when the deviation is greater than the preset buffer range, a differential protection alarm signal is triggered and a backup line switching protocol is started; when the deviation is less than or equal to the preset buffer range, the protection action delay counter is reset and the actual differential current value of the target line is continuously monitored.

6. The method according to claim 1, characterized in that The training process of the pre-trained adaptive multi-transmission line model includes: collecting multiple groups of optical signal waveform sample data of historical transmission lines, and extracting the differential current feature vector corresponding to each group of sample data; performing a joint time domain and frequency domain analysis on the differential current feature vector to generate a feature marker sequence for each group of sample data; inputting the feature marker sequence into the convolution kernel alignment layer in the initial adaptive multi-transmission line model, and outputting the alignment weight of each feature marker in the time dimension; dynamically sampling the feature marker sequence based on the alignment weight to generate a standardized feature set for training; and iteratively training the initial adaptive multi-transmission line model using the standardized feature set until the error rate between the protection priority parameter output by the initial adaptive multi-transmission line model and the preset reference parameter is lower than a preset error threshold, thereby obtaining the adaptive multi-transmission line model.

7. The method according to claim 6, characterized in that The method of performing a joint time domain and frequency domain analysis on the differential current characteristic vector to generate a characteristic marker sequence for each group of sample data includes: dividing the differential current characteristic vector into multiple time domain segments and calculating the average amplitude and phase offset of each time domain segment; performing a fast Fourier transform on each time domain segment to extract the dominant frequency component in the frequency domain energy distribution of each time domain segment; matching the dominant frequency component with a preset typical fault frequency library to mark the potential fault type corresponding to each time domain segment; performing a weighted fusion on the average amplitude and phase offset of the time domain segment according to the potential fault type to generate a time-frequency joint characteristic parameter; and arranging all the time-frequency joint characteristic parameters in the same group of sample data in chronological order to form the characteristic marker sequence.

8. The method according to claim 7, characterized in that The method of inputting the feature marker sequence into a convolution kernel alignment layer in an initial adaptive multi-transmission line model and outputting the alignment weight of each feature marker in the time dimension includes: setting a plurality of sliding windows of different scales in the convolution kernel alignment layer, each sliding window being used to capture a local pattern in the feature marker sequence; calculating a similarity score between the feature marker covered by each sliding window and a preset reference pattern; dynamically adjusting the step size and coverage of each sliding window according to the similarity score to generate multi-scale alignment parameters; inputting the multi-scale alignment parameters into a weight distribution network to output the position weight of each feature marker on the time axis; and performing nonlinear interpolation on the feature marker sequence according to the position weight to generate a standardized feature sequence after time dimension alignment.

9. The method according to claim 8, characterized in that The method dynamically samples the feature marker sequence based on the alignment weight to generate a standardized feature set for training, including: screening feature markers with weight values ​​higher than a preset weight threshold as key feature points based on the position weight; performing mean filling on the interval areas between the key feature points to generate a continuous feature trajectory curve; resampling at equal time intervals on the feature trajectory curve to obtain uniformly distributed feature sampling points; binding the feature sampling points to corresponding potential fault type labels to generate standardized training samples; randomly shuffling and batch dividing the standardized training samples of all historical transmission lines to generate a standardized feature set for final training; and iterating the initial adaptive multi-transmission line model using the standardized feature set. The method includes: during each iteration, inputting the standardized training samples of the current batch into the initial adaptive multi-transmission line model and outputting a predicted protection priority parameter; calculating the mean square error between the predicted protection priority parameter and the corresponding reference parameter, and updating the weight of the initial adaptive multi-transmission line model according to the mean square error; after each iteration, randomly selecting samples from the validation set to evaluate the model performance and calculating the fault detection accuracy and false trigger rate; if the fault detection accuracy does not improve and the false trigger rate does not decrease after multiple consecutive iterations, triggering an early stopping mechanism and saving the current optimal model parameters; and loading the optimal model parameters into the initial adaptive multi-transmission line model to generate the pre-trained adaptive multi-transmission line model.

10. An optical fiber differential protection decision system, characterized in that: include: one or more processors; Memory; One or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, implement the method according to any one of claims 1 to 9.

Images