Power distribution network abnormal sample generation method based on dynamic balance between artificial intelligence model and mechanism model

Abstract

This invention discloses a method for generating abnormal samples in a distribution network based on a dynamic balance between an artificial intelligence model and a mechanistic model. The method includes: acquiring time-series data of the distribution network's operating status and a sample training dataset for the artificial intelligence model; processing the time-series data and the sample training dataset to obtain time-series correlation coefficients and accuracy correlation coefficients; calibrating continuous initial abnormal samples based on the accuracy correlation coefficients, the rate of change of real-time operating status mutation indicators, and response deviation values ​​to obtain a first calibrated abnormal sample set; and calibrating discontinuous initial abnormal samples based on time-series repair parameters, the rate of change of operating status mutation indicators, the accuracy correlation coefficients, and response deviation values ​​to obtain a second calibrated abnormal sample set. The generated samples conform to the physical operating logic of the distribution network and possess clear abnormal characteristic identification.

Description

Technical Field

[0001] This invention relates to the field of power distribution network technology, and more specifically, to a method for generating power distribution network anomaly samples based on the dynamic balance between artificial intelligence models and mechanistic models. Background Technology

[0002] With the expansion of distribution network scale and the diversification of load types, abnormal operating conditions of distribution networks are becoming increasingly complex. Artificial intelligence models have become the core tool for distribution network anomaly identification, but the training effect of the models is highly dependent on high-quality anomaly samples. Traditional methods of obtaining distribution network anomaly samples mostly rely on actual fault records. These samples suffer from problems such as small quantity and uneven type, with common voltage fluctuation anomaly samples accounting for a high proportion.

[0003] Traditional methods for generating outlier samples often deviate from the actual operating logic of distribution networks. Some methods generate outlier data solely through random disturbances, without considering the temporal correlation and physical constraints of the data. This results in samples that do not conform to the variation patterns of parameters such as voltage, current, and power in distribution networks. At the same time, actual distribution network operating data often suffers from temporal breaks due to communication interruptions and equipment failures. Traditional methods are unable to effectively repair such data or calibrate the impact of broken temporal sequences on the outlier characteristics of the samples. The generated samples have discontinuous temporal sequences and inaccurate features, further reducing the reliability of the model's identification.

[0004] Furthermore, the existing sample generation process lacks quantitative analysis of the relationship between sample features and model performance, and cannot clearly define the impact of sample time-series characteristics and abnormal mutation indicators on the accuracy of model recognition. This results in a lack of basis for sample calibration, low matching degree between generated samples and model training requirements, and ultimately affects the model's generalization ability in actual power distribution network scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method for generating abnormal power distribution network samples by dynamically balancing artificial intelligence models and mechanistic models.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for generating distribution network anomaly samples based on the dynamic balance between artificial intelligence models and mechanistic models includes the following steps:

[0008] Acquire time-series data of the operating status of the power distribution network and sample training datasets for artificial intelligence models;

[0009] The time-series data of the running status and the sample training dataset are processed to obtain the time-series correlation coefficient and the accuracy correlation coefficient;

[0010] The response bias value affected by the temporal characteristics of the sample is obtained by combining the temporal correlation coefficient and temporal correlation characteristics;

[0011] If the temporal continuity of the initial abnormal samples meets the preset requirements, the initial abnormal samples are calibrated according to the accuracy correlation coefficient, the rate of change of the real-time running status mutation index and the response deviation value to obtain the first calibration abnormal sample set.

[0012] If the temporal continuity of the initial abnormal samples does not meet the preset requirements, a temporal index coupling correction model is constructed based on the duration of temporal breaks, the location of breaks, and the accelerated change characteristics of abrupt changes in the operational status index of historical samples. Based on the temporal index coupling correction model, the temporal repair parameters and the rate of change of the operational status abrupt change index of the initial abnormal samples are obtained. The initial abnormal samples are then calibrated based on the temporal repair parameters, the rate of change of the operational status abrupt change index, the accuracy correlation coefficient, and the response deviation value to obtain the second calibrated abnormal sample set.

[0013] Preferably, the operating status time-series data includes line voltage values, equipment current values, active power output values, reactive power compensation values, and load connection data at different time points;

[0014] The sample training dataset includes historical normal operation time series samples, various abnormal operating condition time series samples, and corresponding model training performance data.

[0015] Preferably, the time-series data of the running status and the sample training dataset are processed to obtain the time-series correlation coefficient and the accuracy correlation coefficient, specifically including the following steps:

[0016] Temporal correlation features of sample data are extracted from the sample training dataset;

[0017] Operational state mutation indicators are extracted from operational state time-series data;

[0018] The temporal correlation coefficient between temporal correlation features and response parameters is extracted from historical data; the accuracy correlation coefficient between the rate of change of abrupt change indicators of distribution network operation status and the accuracy of abnormal sample identification is extracted from historical data.

[0019] Preferably, the temporal correlation features include data temporal correlation coefficient, temporal change smoothness, fluctuation range of data difference between adjacent time points, and temporal period stability;

[0020] The indicators of sudden changes in operating status include voltage change amplitude, current change rate, power fluctuation peak value, and load change frequency.

[0021] Preferably, the time-series correlation coefficient between time-series correlation features and response parameters is extracted based on historical data; the accuracy correlation coefficient between the rate of change of distribution network operation status abrupt change indicators and the accuracy of abnormal sample identification is extracted based on historical data, specifically including the following steps:

[0022] Collect historical sample time-series correlation features, response data, and abnormal sample feedback data;

[0023] Extract the historical actual values ​​of the temporal correlation features of sample data and the actual values ​​of the response parameters for the corresponding periods from historical correlation data;

[0024] The difference between the historical actual value and the optimal time series standard value of the temporal correlation feature of the sample data is used to obtain the temporal deviation amplitude, and the difference between the actual value and the ideal response standard value of the response parameter is used to obtain the performance response deviation amplitude.

[0025] The time-series correlation coefficient between the time-series correlation characteristics and the response parameters is obtained by taking the ratio of the performance response deviation magnitude and the time-series deviation magnitude.

[0026] Extract the historical rate of change of the distribution network operation status change index from the feedback data generated from historical anomaly samples, and the accuracy of anomaly sample identification corresponding to the rate of change;

[0027] The accuracy correlation coefficient is obtained by calculating the ratio between the accuracy of abnormal sample identification and the historical rate of change of the operational status mutation index.

[0028] Preferably, the response bias value affected by the temporal characteristics of the sample is obtained by combining the temporal correlation coefficient and temporal correlation features, specifically including the following steps:

[0029] Calculate the current deviation between the temporal correlation feature of the real-time sample data and the optimal temporal standard value;

[0030] The response deviation value affected by the temporal characteristics of the sample is obtained based on the temporal correlation coefficient and the current deviation value.

[0031] Preferably, the initial abnormal samples are calibrated based on the accuracy correlation coefficient, the rate of change of the real-time operating status mutation index, and the response deviation value to obtain the first calibration abnormal sample set, specifically including the following steps:

[0032] The standard identification accuracy of the initial abnormal sample is obtained by combining the accuracy correlation coefficient and the rate of change of the real-time running status mutation index. The actual identification accuracy is obtained by correcting the standard identification accuracy with the response deviation value.

[0033] Based on the actual recognition accuracy, the initial abnormal samples are subjected to temporal integrity completion and abnormal feature calibration to obtain the first calibrated abnormal sample set.

[0034] Preferably, a time-series index coupling correction model is constructed based on the duration of time-series fractures, fracture locations, and accelerated change characteristics of abrupt changes in operational status indicators of historical samples. Based on this model, the time-series repair parameters and the rate of change of operational status abrupt changes in the initial abnormal samples are obtained. Specifically, this includes the following steps:

[0035] Determine the temporal fracture parameters of the sample data, including the temporal fracture duration, fracture location, and correlation between the data before and after the fracture;

[0036] Capture the accelerated change characteristics of abrupt changes in operational status indicators, including the accelerated change trend, peak rate, and duration of the operational status abrupt changes indicators;

[0037] A time series index coupling correction model is constructed based on the historical sample time series fracture parameters, the accelerated change characteristics of the abrupt change index of the operating state, and the corresponding historical time series repair parameters and the change rate of the historical abrupt change index of the operating state.

[0038] The change rate of the initial abnormal sample time series break parameters and the accelerated change characteristics of the operational state mutation index are input into the time series index coupling correction model to obtain the change rate of the initial abnormal sample time series repair parameters and the operational state mutation index.

[0039] Preferably, the initial abnormal sample is calibrated based on the time-series repair parameters, the rate of change of the operational state mutation index, the accuracy correlation coefficient, and the response deviation value to obtain a second calibration abnormal sample set. This specifically includes the following steps:

[0040] The timing calibration weight coefficient of the initial abnormal sample is calculated based on the timing repair parameters, the rate of change of the operational status mutation index, the accuracy correlation coefficient, and the response deviation value of the initial abnormal sample.

[0041] The second calibration abnormal sample set is obtained by performing time-series repair, abnormal feature enhancement and quality screening on the initial abnormal samples according to the time-series calibration weight coefficient.

[0042] Preferably, it further includes:

[0043] A report is generated based on either the first or second set of calibration anomaly samples.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention integrates time-series data of distribution network operation status with sample training datasets, extracts time-series correlation features and operational status mutation indicators, and calculates time-series correlation coefficients and accuracy correlation coefficients. This ensures that sample generation is aligned with the actual operation patterns of the distribution network, guaranteeing that the generated samples conform to the physical operation logic of the distribution network while possessing clear anomaly identification characteristics, providing high-quality foundational data for model training. During sample generation, key information such as time-series repair parameters and the rate of change of operational status mutation indicators is retained, and a final output report including sample distribution, quality verification, and application suggestions is provided. This not only adapts to different distribution network operation scenarios such as peak residential loads and industrial load fluctuations, but also allows model trainers to clearly understand the source, characteristics, and applicable scope of the samples, facilitating flexible sample selection based on actual operation and maintenance needs. Furthermore, it provides a traceable basis for subsequent iterative optimization of samples.

[0046] This invention improves the efficiency and reliability of power distribution network anomaly identification, enabling artificial intelligence models to more accurately identify various abnormal operating conditions of the power distribution network in actual operation, reducing the impact range and duration of power distribution network faults, and indirectly improving the stability and reliability of power distribution network operation. Attached Figure Description

[0047] Figure 1 This is a schematic diagram illustrating a method for generating abnormal power distribution network samples by dynamically balancing artificial intelligence models and mechanistic models, as provided in an embodiment of the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0051] Reference Figure 1 As shown.

[0052] The embodiments further illustrate the method for generating abnormal power distribution network samples by dynamically balancing artificial intelligence models and mechanistic models proposed in this invention.

[0053] A method for generating distribution network anomaly samples based on the dynamic balance between artificial intelligence models and mechanistic models includes the following steps:

[0054] Acquire time-series data of the operating status of the power distribution network and sample training datasets for artificial intelligence models;

[0055] The time-series data of the running status and the sample training dataset are processed to obtain the time-series correlation coefficient and the accuracy correlation coefficient;

[0056] The response bias value affected by the temporal characteristics of the sample is obtained by combining the temporal correlation coefficient and temporal correlation characteristics;

[0057] If the temporal continuity of the initial abnormal samples meets the preset requirements, the initial abnormal samples are calibrated according to the accuracy correlation coefficient, the rate of change of the real-time running status mutation index and the response deviation value to obtain the first calibration abnormal sample set.

[0058] If the temporal continuity of the initial abnormal samples does not meet the preset requirements, a temporal index coupling correction model is constructed based on the duration of temporal breaks, the location of breaks, and the accelerated change characteristics of abrupt changes in the operational status index of historical samples. Based on the temporal index coupling correction model, the temporal repair parameters and the rate of change of the operational status abrupt change index of the initial abnormal samples are obtained. The initial abnormal samples are then calibrated based on the temporal repair parameters, the rate of change of the operational status abrupt change index, the accuracy correlation coefficient, and the response deviation value to obtain the second calibrated abnormal sample set. The methods for extracting fracture duration and fracture location are as follows: Using the time-series correlation characteristics of the normal operating state of the distribution network as a reference, including the time-series correlation coefficient, time-series change smoothness, and fluctuation range of data differences between adjacent times, the entire time-series of the sample is scanned point-by-point to identify numerical abrupt change intervals exceeding the benchmark threshold. The start time of this abrupt change interval is marked as the fracture start position, and the end time is marked as the fracture termination position; the time span between them is the fracture duration. During the extraction of fracture location and duration, the accelerated change characteristics of the operating state abrupt change indicators are used for anchoring. By calculating the instantaneous change rate and acceleration of operating indicators such as voltage, current, and power within the abrupt change interval, the moment with the largest abrupt change amplitude and the most drastic rate change is located, and this is taken as the core fracture location to enhance the representativeness of the abnormal characteristics. The complete time window from the initial acceleration to the rate decline within the abrupt change interval is taken as the final fracture duration.

[0059] The temporal continuity requirement means that the variation and correlation characteristics of the sample data in the time dimension conform to the physical logic and preset judgment criteria of the distribution network operation; the time-series data points are complete and without missing parts; the operating parameters (such as voltage, current, power, etc.) of adjacent time steps change smoothly without any jumps or gaps that exceed physical laws; the temporal correlation coefficient, change smoothness, and periodic stability all fall within the preset threshold range, reflecting the continuous evolution trend and correlation of the distribution network operating parameters over time. The fluctuation range of data differences conforms to the laws of normal operation or abnormal evolution, and there are no temporal breaks caused by communication interruptions or data loss. If the sample data meets the preset temporal continuity requirement, there is no temporal break phenomenon.

[0060] It should be noted that the initial anomaly samples are obtained through the following methods: Collecting various basic data from the historical operation of the distribution network, including time-series operating parameters such as line voltage, equipment current, active and reactive power, and load access, as well as historical fault records, abnormal operating condition monitoring data, and equipment maintenance logs; based on physical operating logic and equipment operating characteristics, simulating various preset abnormal scenarios through mechanism simulation, such as line short circuits, voltage drops, and abnormal load fluctuations, to generate basic anomaly samples that conform to the physical operating laws of the distribution network; combining generative adversarial networks and variational autoencoders to perform feature learning and expansion on historical anomaly samples, thereby obtaining the initial anomaly samples. The generation process of the initial anomaly samples is existing technology and is not the focus of this invention; therefore, it will not be described in detail here.

[0061] In this embodiment, the operating status time sequence data includes line voltage values, equipment current values, active power output values, reactive power compensation values, and load access data at different time points;

[0062] The sample training dataset includes historical normal operation time series samples, time series samples of various abnormal operating conditions, and corresponding model training performance data.

[0063] The model training performance data refers to a comprehensive set of performance feedback and operational metrics throughout the historical sample training process. Specifically, it includes recognition accuracy data for different types of sample inputs, such as accuracy, precision, recall, and F1 score for various abnormal operating conditions, with a particular focus on the recognition performance of low-probability, rare abnormal samples. This data is used to establish a quantitative correlation between operational state mutation indicators and recognition accuracy. It also covers response parameter data to the temporal characteristics of samples, including temporal feature extraction efficiency, response latency, and output result stability. Furthermore, it includes data on the model's generalization ability and robustness, such as performance degradation under different load scenarios, slight temporal fluctuations, or partial data loss. Finally, it includes feedback data from the model training iteration process, such as performance change curves after training different batches of samples, sample weight adjustment records, and gradient update information. These data collectively constitute the mapping relationship between sample quality and model training effectiveness.

[0064] In this embodiment, the time-series data of the running status and the sample training dataset are processed to obtain the time-series correlation coefficient and the accuracy correlation coefficient, specifically including the following steps:

[0065] Temporal correlation features of sample data are extracted from the sample training dataset;

[0066] Operational state mutation indicators are extracted from operational state time-series data;

[0067] The temporal correlation coefficient between temporal correlation features and response parameters is extracted from historical data; the accuracy correlation coefficient between the rate of change of abrupt change indicators of distribution network operation status and the accuracy of abnormal sample identification is extracted from historical data.

[0068] Temporal correlation characteristics include the temporal correlation coefficient of data, the smoothness of temporal changes, the fluctuation range of data differences between adjacent time points, and the stability of the time series period.

[0069] Operating status abrupt change indicators include voltage change amplitude, current change rate, power fluctuation peak value, and load change frequency.

[0070] This study extracts temporal correlation features from the training dataset and operational status mutation indicators from the operational status time-series data. Temporal correlation features are the core dimension characterizing the temporal dimension of the sample data. The temporal correlation coefficient measures the degree of correlation between data at different time points. For example, for voltage data of a certain line at 10 consecutive time points, the temporal correlation coefficient is obtained by calculating the linear correlation between voltage values ​​at adjacent time points. The temporal correlation coefficient is obtained by performing linear correlation analysis on operational parameter data at adjacent time points. Specifically, taking the voltage data of a distribution network line at consecutive time points as an example, firstly, voltage values ​​at adjacent time points in the sequence are extracted to form pairs of data. Then, the Pearson correlation coefficient formula is used for calculation. First, the mean of the two pairs of adjacent voltage data is calculated. Then, the sum of the products of the deviations of each pair of data from their respective means is calculated, as well as the square root of the sum of squares of the deviations of the two pairs of data. Finally, the sum of the products of deviations is divided by the square root of the sum of squares of the deviations to obtain a coefficient value between -1 and 1.

[0071] Time series smoothness reflects the degree of fluctuation in data over time. For example, if the change in current data of a certain device is consistently controlled within 0.3A over consecutive time intervals, it indicates that the time series smoothness of this data is high. Time series smoothness is obtained by calculating the statistical characteristics of the changes in operating parameters over consecutive time intervals. Taking distribution network equipment current data as an example, firstly, the current values ​​at adjacent time nodes in the sequence are extracted, and the absolute value of the current difference between every two adjacent time intervals is calculated to obtain a set of change sequences. Then, the mean, variance, or maximum value statistical indicators are calculated for this sequence to quantify the degree of fluctuation in the data over time. If the changes at all adjacent time intervals are controlled within a small threshold, and the statistical mean and variance are both at low levels, it indicates that the time series smoothness of the data is high, reflecting the stable transition characteristics of distribution network operating parameters over time. Conversely, if the changes fluctuate drastically or show large jumps, it indicates low smoothness, which may indicate abnormal disturbances or the risk of time series breaks.

[0072] The fluctuation range of data difference between adjacent time points is the difference between the maximum and minimum values ​​of data difference between adjacent time points. For example, if the minimum difference between adjacent time points in a certain active power data is 0.1kW and the maximum is 1.2kW, then the fluctuation range of its difference is 1.1kW.

[0073] Temporal periodicity is used to determine whether data exhibits stable periodic changes. For example, if the load access volume in a certain area peaks at fixed times each day, and the timing and duration of these peaks deviate relatively little, this indicates strong temporal periodicity. Temporal periodicity is characterized by periodic analysis and statistical deviation quantification, and measures whether sample data exhibits regular and robust periodic changes over time. First, select distribution network operation parameters with daily and weekly time periods, such as load access and voltage fluctuation patterns, and perform Fourier transform or autocorrelation analysis to extract the main periodic components of the data and clarify the inherent periodic characteristics of the data. Next, statistically analyze key node characteristic parameters within multiple periods, including the specific time of peak occurrence, peak duration, peak amplitude, and mean within the period. Calculate the deviations of these parameters across different periods, such as the standard deviation of time deviation and the range of amplitude deviation. If the time deviation, duration deviation, and amplitude fluctuation range of the peak occurrence are all controlled within preset thresholds, and the operating trend remains consistent across periods, it indicates strong temporal periodic stability of the data segment, reflecting a stable and predictable periodic pattern of distribution network operation parameters over time. Conversely, if key node parameters fluctuate drastically across different periods and the deviation exceeds the threshold, it indicates weak temporal periodic stability and unclear periodic characteristics of the data. Quantifying the periodic evolution pattern of distribution network operation data provides a quantitative basis for judging whether the data conforms to normal operating logic and identifying abnormal periodic fluctuations.

[0074] Operating status mutation indicators are key indicators for capturing abnormal changes in the distribution network. Voltage mutation amplitude refers to the magnitude of voltage change at a certain moment relative to the previous moment. For example, if the line voltage suddenly drops from 220V to 190V at a certain moment, the voltage mutation amplitude at that moment is 30V. Current mutation rate is the degree of change of current per unit time. For example, if the equipment current rises from 6A to 21A in 1 second, its current mutation rate is 15A / s. Power fluctuation peak value is the maximum deviation of power from the average value within a certain period of time. For example, if the average reactive power data of a certain period is 5kVar, and it reaches 12kVar at a certain moment, then the power fluctuation peak value of this data is 7kVar. Load mutation frequency refers to the number of times the load connection changes mutate per unit time. For example, if the load connection exceeds the preset threshold 4 times in 1 hour, its load mutation frequency is 4 times / hour.

[0075] Collect the time-series correlation feature data of historical samples and the response parameter data of the corresponding period. Then, calculate the difference between the historical actual value and the optimal time-series standard value of the time-series correlation feature of the sample data to obtain the time-series deviation amplitude. Calculate the difference between the actual value and the ideal response standard value of the response parameter to obtain the performance response deviation amplitude. Finally, compare the performance response deviation amplitude with the time-series deviation amplitude to obtain the time-series correlation coefficient. The time-series correlation coefficient = performance response deviation amplitude / time-series deviation amplitude.

[0076] The response parameter data refers to the set of performance feedback parameters output after inputting the time-series correlation features of the corresponding historical samples. It is the basis for quantifying the impact of time-series correlation features on model performance. Specifically, it includes the recognition response delay, feature extraction efficiency, output result stability, and generalization ability attenuation under the corresponding sample input, as well as the sensitivity feedback to features such as sample time-series integrity and smoothness.

[0077] The historical rate of change of the distribution network operation status change index is extracted from the historical abnormal sample generation feedback data, and the accuracy of abnormal sample identification under the corresponding rate of change is obtained by comparing the accuracy of abnormal sample identification with the historical rate of change value. The accuracy correlation coefficient is calculated as: Accuracy correlation coefficient = Accuracy of abnormal sample identification / Historical rate of change value.

[0078] Among them, the generated feedback data refers to the set of feedback information accumulated throughout the entire process of generating and calibrating historical abnormal samples, which is the data source for extracting the accuracy correlation coefficient. Specifically, it includes the rate of change of the abrupt change index of the initial generated abnormal sample's running state, the time-series repair parameters after sample calibration, the degree of abnormal feature enhancement, and the performance results of recognition accuracy, precision, and recall obtained after the corresponding sample is input into the model.

[0079] The time-series correlation coefficient between time-series correlation features and response parameters is extracted based on historical data; the accuracy correlation coefficient between the rate of change of abrupt change indicators of distribution network operation status and the accuracy of anomaly sample identification is extracted based on historical data, specifically including the following steps:

[0080] Collect historical sample time-series characteristics, response data, and abnormal sample feedback data;

[0081] Extract the historical actual values ​​of the temporal correlation features of sample data and the actual values ​​of the response parameters for the corresponding periods from historical correlation data;

[0082] The difference between the historical actual value and the optimal time series standard value of the temporal correlation feature of the sample data is used to obtain the temporal deviation amplitude, and the difference between the actual value and the ideal response standard value of the response parameter is used to obtain the performance response deviation amplitude.

[0083] The time-series correlation coefficient between the time-series correlation characteristics and the response parameters is obtained by taking the ratio of the performance response deviation magnitude and the time-series deviation magnitude.

[0084] Extract historical rate of change values ​​of power distribution network operation status change indicators from historical anomaly sample generation feedback data, and the accuracy of anomaly sample identification corresponding to the rate of change;

[0085] The accuracy correlation coefficient is obtained by calculating the ratio between the accuracy of abnormal sample identification and the historical rate of change.

[0086] Collect the temporal characteristics of historical samples, response data, and feedback data on the generation of abnormal samples. For example, the temporal characteristics of historical normal operation samples of a power distribution network over the past six months, the recognition accuracy data of the corresponding AI model within the same period, and the feedback data on the effects of generating abnormal samples in the past.

[0087] From these historical correlation data, the actual historical values ​​of the temporal correlation features of the sample data, as well as the actual values ​​of the response parameters for the corresponding periods, are extracted. The temporal correlation features include the temporal correlation coefficient and the smoothness of temporal changes. For example, the actual value of the temporal correlation coefficient between adjacent time points of the voltage data of a certain historical normal sample is 0.85. At the same time, the actual values ​​of the performance parameters of the recognition response speed of this type of sample within this period are also extracted.

[0088] The timing deviation amplitude is obtained by comparing the historical actual value and the optimal timing standard value of the timing correlation characteristics of the sample data. For example, if the optimal standard value of the timing correlation coefficient of voltage data is 0.9, then the timing deviation amplitude is 0.85 - 0.9 = -0.05. The performance response deviation amplitude is obtained by comparing the actual value and the ideal response standard value of the response parameter. For example, if the ideal standard value of the system identification response speed for the corresponding time period is 0.2 seconds and the actual value is 0.25 seconds, then the performance response deviation amplitude is 0.25 - 0.2 = 0.05.

[0089] Based on these two deviation magnitudes, the scheme calculates the time-series correlation coefficient, which is equal to the performance response deviation magnitude divided by the time-series deviation magnitude. Substituting the values ​​in the example above, we get a time-series correlation coefficient of -1. This coefficient quantifies the correlation strength between the time-series correlation features and the response parameters.

[0090] The historical change rate values ​​of power distribution network operation status change indicators are extracted from the feedback data generated from historical anomaly samples, along with the anomaly sample identification accuracy at the corresponding change rate. For example, in a historical voltage change event, the voltage change rate is 20V / s, and the anomaly sample identification accuracy at the corresponding voltage change rate is 90%.

[0091] The accuracy correlation coefficient is equal to the anomaly identification accuracy divided by the historical rate of change. Substituting the data, the accuracy correlation coefficient is 0.045s / V. The accuracy correlation coefficient reflects the correlation between the rate of change of the operational status mutation index and the anomaly identification accuracy.

[0092] The response bias value affected by the temporal characteristics of the samples is obtained by combining the temporal correlation coefficient and temporal correlation features, specifically including the following steps:

[0093] The current deviation between the time-series correlation characteristics of the real-time sample data and the optimal time-series standard value is calculated. The real-time sample data comes from the actual operation data of the distribution network that is currently undergoing time-series calibration and anomaly identification. It is the data object used for real-time verification and deviation calculation. It includes time-series operation parameters such as line voltage, equipment current, active and reactive power, and load access volume uploaded in real time by the distribution network field monitoring terminal, as well as the collected equipment status information and environmental monitoring data. This data is the real operation data of the current moment or the current batch, which is the direct basis for reflecting the current operation status of the distribution network and for determining the time-series deviation and deriving the response deviation value.

[0094] The response deviation value affected by the temporal characteristics of the sample is obtained based on the temporal correlation coefficient and the current deviation value.

[0095] Calculate the current deviation between the time-series correlation feature of the real-time sample data and the optimal time-series standard value. The time-series correlation feature includes dimensions such as the data time-series correlation coefficient, the smoothness of time-series changes, the fluctuation range of data differences between adjacent time points, and the stability of the time-series cycle. Each feature corresponds to a pre-set optimal time-series standard value, which is usually a reasonable threshold derived from historical data summarizing the normal and stable operation of the distribution network. Taking the smoothness of time-series changes as an example, assuming the optimal standard value for the smoothness of time-series changes of a certain line current data is 0.9 (the closer the value is to 1, the smoother the change), and the actual value of the smoothness of time-series changes of the current data of this line in the current real-time sample is 0.7, then the current deviation value is the actual real-time value minus the optimal time-series standard value, i.e., 0.7 - 0.9 = -0.2. Taking the fluctuation range of data differences between adjacent time points as an example, if the optimal standard value for this feature of a certain active power data is 0.8kW, and the actual value of the current real-time sample is 1.2kW, the corresponding current deviation value is 1.2 - 0.8 = 0.4.

[0096] The response bias value, influenced by the temporal correlation coefficient and the current bias value, is obtained. Response bias value = Temporal correlation coefficient × Current bias value.

[0097] Assuming the time-series correlation coefficient corresponding to the smoothness of the current time-series change is 2, it means that the deviation of this feature has a 2-fold impact, and the corresponding response deviation value of this feature is 2×(-0.2)=-0.4; while the time-series correlation coefficient corresponding to the fluctuation range of the difference between adjacent active power times is 1.5, and the corresponding response deviation value is 1.5×0.4=0.6.

[0098] By converting the temporal characteristic bias of real-time samples into response bias, the specific impact of sample temporal characteristics on model performance is clarified, providing a quantitative basis for subsequent adjustment of abnormal samples and optimization of model balance performance based on this bias.

[0099] The first calibration anomaly sample set is obtained by calibrating the initial anomaly samples based on the accuracy correlation coefficient, the rate of change of the real-time operating status mutation index, and the response deviation value. The specific steps include:

[0100] The standard identification accuracy of the initial abnormal sample is obtained by combining the accuracy correlation coefficient and the rate of change of the real-time running status mutation index. The actual identification accuracy is obtained by correcting the standard identification accuracy with the response deviation value.

[0101] Based on the actual recognition accuracy, the initial abnormal samples are subjected to temporal integrity completion and abnormal feature calibration to obtain the first calibrated abnormal sample set.

[0102] The standard identification accuracy of the initial anomaly sample is obtained based on the accuracy correlation coefficient and the rate of change of the real-time mutation index. The accuracy correlation coefficient is a coefficient calculated based on historical data, reflecting the correlation between the rate of change of the mutation index and the accuracy of anomaly sample identification. The rate of change of the real-time mutation index is the actual rate of change of the mutation index of the distribution network operation status in the current sample, such as the voltage mutation rate and the current mutation rate. Taking the voltage mutation index as an example, assuming that the voltage mutation rate (real-time mutation index change rate) of an initial anomaly sample is 15V / s, and the corresponding accuracy correlation coefficient is 0.045·s / V, then the formula for calculating the standard identification accuracy is: Standard identification accuracy = Accuracy correlation coefficient × Real-time mutation index change rate. Substituting the values, the standard identification accuracy of this sample is 0.045·s / V × 15V / s = 0.675.

[0103] The actual recognition accuracy is obtained by correcting the standard recognition accuracy based on the response deviation value. The response deviation value represents the degree of deviation affected by the temporal characteristics of the sample. During correction, it needs to be combined with the standard recognition accuracy. If the response deviation value is positive, it indicates that the temporal characteristic deviation of the sample will increase the recognition error, and this deviation value needs to be added to the standard recognition accuracy. If the response deviation value is negative, it indicates that the temporal characteristic deviation of the sample offsets part of the recognition error, and this deviation value needs to be subtracted from the standard recognition accuracy. Assuming the response deviation value for the above sample is -0.2, representing that the temporal characteristic deviation will reduce the recognition accuracy, the formula for calculating the actual recognition accuracy is: Actual Recognition Accuracy = Standard Recognition Accuracy + Response Deviation Value. Substituting the data, the actual recognition accuracy for this sample is 0.675 + (-0.2) = 0.475.

[0104] Based on the actual recognition accuracy, the initial abnormal samples are subjected to temporal integrity completion and abnormal feature calibration to obtain the first calibration abnormal sample set. Temporal integrity completion addresses the temporal missing issues present in the initial samples. For example, if an initial abnormal sample lacks current data for a certain period, the missing data needs to be completed based on the temporal correlation characteristics of adjacent periods (such as data temporal correlation coefficients and temporal change smoothness) to ensure the temporal continuity of the samples. The first calibration abnormal sample set corresponds to samples whose temporal continuity meets the preset requirements. However, meeting the preset requirements only means that the samples do not have large-scale temporal breaks or missing data, and does not equate to the complete absence of missing time series data or the absence of minor local discontinuities. In real-world scenarios, some initial abnormal samples may have local time series data missing data or data jumps at individual time points. Although these problems do not reach the threshold for determining temporal breaks, they still affect the integrity of the sample's temporal features and the model's recognition performance. Therefore, performing temporal integrity completion is to repair these local temporal defects, consolidate the temporal continuity of the samples, provide a more reliable data foundation for subsequent abnormal feature calibration, and ensure that the final generated calibration samples reach the optimal state in the temporal dimension. Abnormal feature calibration is to adjust the intensity of abnormal features of the samples according to the actual recognition accuracy. If the actual recognition accuracy is low, it means that the abnormal features of the samples are not obvious enough, and it is necessary to strengthen the voltage change amplitude and power fluctuation peak features. For example, the original voltage change amplitude can be increased from 20V to 25V to improve the model's recognition accuracy for the sample.

[0105] Anomaly feature enhancement has clear quantitative standards, based on the difference between the actual recognition accuracy and the preset accuracy threshold, as well as the intensity level of the current anomaly features in the sample. Specifically, firstly, a target recognition accuracy threshold is set for model training. When the actual recognition accuracy of the sample is lower than this threshold, the enhancement magnitude is determined according to the difference: the larger the difference, the weaker the anomaly feature identification, and the higher the enhancement magnitude; the smaller the difference, the more moderate the enhancement magnitude. Combining the amplitude and fluctuation range parameters of the current sample's anomaly features, the upper limit of enhancement is determined with the physical operation constraints of the distribution network as the boundary. This avoids over-enhancement that could cause the sample to deviate from the actual operating rules, ensuring that the actual recognition accuracy approaches the target threshold, thus guaranteeing that the anomaly features are sufficiently clear while ensuring that the sample conforms to the physical logic of distribution network operation.

[0106] This improves the ability of AI models to identify such abnormal samples while ensuring the rationality and usability of the samples.

[0107] Based on the duration, location, and accelerated change characteristics of abrupt change in historical sample time-series fractures, a time-series index coupling correction model is constructed. Based on this model, the time-series repair parameters and the rate of change of abrupt change in the initial abnormal sample are obtained. The specific steps include:

[0108] Determine the temporal fracture duration, fracture location, and correlation between data before and after the fracture in the sample data;

[0109] Capture the accelerated change trend, peak rate, and duration of mutation indicators;

[0110] A time series index coupling correction model is constructed based on the historical sample time series fracture parameters, the accelerated change characteristics of the abrupt change index of the operating state, and the corresponding historical time series repair parameters and the change rate of the historical abrupt change index of the operating state.

[0111] The initial abnormal sample time series breakage parameters and the accelerated change characteristics of mutation index are input into the time series index coupling correction model to obtain the initial abnormal sample time series repair parameters and the change rate of mutation index.

[0112] Determine the duration, location, and correlation of data before and after the time series breaks in the sample data. A time series break refers to a discontinuity in the sample data over time. For example, in the current time series data of a distribution network, 5 minutes of data within the interval [10:05, 10:10] are missing. In this case, the duration of the time series break is 5 minutes, and the break location is the period from 10:05 to 10:10. The correlation of data before and after the break refers to the degree of correlation between the current value at the last moment before the break (10:04) and the current value at the first moment after the break (10:10). For example, the current value at 10:04 is 8A, and the current value at 10:10 is 12A. The strength of the correlation is determined by calculating the time series correlation coefficient between the two values.

[0113] Capture the accelerating trend, peak rate, and duration of abrupt changes in indicators. Abrupt changes include voltage fluctuation amplitude and current fluctuation rate. An accelerating trend refers to the rate of change of the abrupt change indicator as it increases over time; for example, the current fluctuation rate gradually increases from 2 A / s to 8 A / s over a certain period, showing an accelerating upward trend. The peak rate refers to the maximum rate of change of the abrupt change indicator within that period; for example, the peak rate of the current fluctuation rate mentioned above is 8 A / s. The duration refers to the length of time the accelerated change process takes; for example, the process from 2 A / s to 8 A / s lasts for 3 minutes.

[0114] A time-series index coupling correction model is constructed based on historical sample time-series break data and accelerated change data of abrupt change indicators. The time-series index coupling correction model integrates similar time-series break case data from past distribution networks and the corresponding accelerated change patterns of abrupt change indicators to establish the coupling relationship between time-series break parameters and abrupt change indicator change characteristics. For example, in historical data, when the time-series break duration is 5 minutes and the current correlation before and after the break is 0.7, the corresponding peak current abrupt change rate usually increases by 20%. These patterns are incorporated into the model's calculation logic.

[0115] The initial abnormal sample time-series break parameters and accelerated change characteristics of abrupt change indices are input into the time-series index coupling correction model to obtain the initial abnormal sample time-series repair parameters and abrupt change rate of indices. The time-series repair parameters include the numerical range and change slope required to complete the broken data. For example, for a 5-minute current data break, the time-series repair parameters output by the time-series index coupling correction model are based on the 8A before the break, and the data is completed at a slope of 0.8A / min. The abrupt change rate of indices is the actual change rate after calibration. For example, the original current abrupt change rate is corrected from 8A / s to 9.6A / s, which conforms to the coupling relationship.

[0116] The initial abnormal samples are calibrated based on the time-series repair parameters, rate of change, accuracy correlation coefficient, and response deviation value to obtain the second calibration abnormal sample set. This process includes the following steps:

[0117] The temporal calibration weight coefficient of the initial abnormal sample is calculated based on the temporal repair parameters, mutation index change rate, accuracy correlation coefficient, and response deviation value of the initial abnormal sample.

[0118] The second calibration abnormal sample set is obtained by performing time-series repair, abnormal feature enhancement and quality screening on the initial abnormal samples according to the time-series calibration weight coefficient.

[0119] Information related to the coupling effect of time-series indicators is extracted from historical data, including historical time-series repair parameters, historical abrupt change rate, and the calibration effect data of outlier samples corresponding to these parameters. For example, in a certain historical data segment, the time-series repair parameter is to complete 5 minutes of broken data at a slope of 0.8 A / min, and the corresponding historical abrupt change rate is 9.6 A / s. The outlier sample calibration effect data of this parameter combination shows that the model recognition accuracy is improved by 15%.

[0120] The temporal calibration weight coefficient for the initial outlier sample is calculated based on the temporal repair parameters, mutation index change rate, accuracy correlation coefficient, and response deviation value. The temporal calibration weight coefficient is calculated as follows: (Temporal repair parameter fit × 0.3) + (mutation index change rate matching degree × 0.3) + (accuracy correlation coefficient × 0.2) + (response deviation value correction coefficient × 0.2). Here, the temporal repair parameter fit is the degree of agreement between the current temporal repair parameters and the historical best repair parameters. For example, the fit between the current slope of 0.8 A / min and the historical best of 0.7 A / min is 0.9. The mutation index change rate matching degree is the degree of agreement between the current rate and the historical effective rate. For example, the matching degree between the current rate of 9.6 A / s and the historical effective range is 0.85. Substituting the data, we get the time series calibration weight coefficient = (0.9×0.3) + (0.85×0.3) + (0.045×0.2) + (-0.2×0.2) = 0.27 + 0.255 + 0.009 - 0.04 = 0.494.

[0121] The initial abnormal samples are subjected to time-series repair, abnormal feature enhancement, and quality screening based on the time-series calibration weighting coefficient to obtain the second calibrated abnormal sample set. Time-series repair involves completing the broken time series of the samples according to the time-series repair parameters. For example, a 5-minute current data break is completed at a slope of 0.8A / minute to restore the continuity of the sample time. Abnormal feature enhancement involves adjusting the intensity of the mutation index according to the time-series calibration weighting coefficient. The time-series calibration weighting coefficient is an indicator that quantifies the degree of influence of time-series breaks on the abnormal features of the samples. It is calculated by combining the time-series repair parameters, the rate of change of the operating state mutation index, the accuracy correlation coefficient, and the response deviation value. The value directly reflects the degree to which the time-series break weakens the identification of abnormal features: the higher the weighting coefficient, the more serious the problem of abnormal feature ambiguity caused by the time-series break, and the more necessary it is to strengthen the mutation index; the lower the weighting coefficient, the smaller the impact of the time-series break, and the lower the intensity of the mutation index enhancement can be. When adjusting the intensity of the mutation index, the time-series calibration weighting coefficient is used as the basis, and the enhancement intensity is determined in combination with the physical operating constraints of the distribution network: the weighting coefficient is divided into different intervals, corresponding to different enhancement ratios. For example, when the weighting coefficient is in the range of 0.6-0.8, the amplitude of the sudden change index is increased by 20%; when the weighting coefficient is in the range of 0.8-1.0, the amplitude of the sudden change index is increased by 30%, ensuring that the enhancement magnitude matches the degree of impact. The enhanced amplitude of the sudden change index must meet the actual operating rules of the distribution network and must not exceed physical limits. For example, the voltage sudden change amplitude must not exceed 20% of the line's rated voltage to avoid the sample deviating from the real operating scenario. If the time-series calibration weighting coefficient of a sample is 0.7, the corresponding enhancement ratio is 20%, and the original voltage sudden change amplitude is 20V, then the enhanced amplitude is 20V × (1 + 20%) = 24V.

[0122] For example, increasing the voltage fluctuation amplitude from 20V to 24V enhances the anomaly identification of the samples; quality screening is based on calibration effect data to remove samples that still do not conform to physical laws or whose identification accuracy is not up to standard after repair. For example, samples with a timing calibration weight coefficient of less than 0.3 will be judged as unqualified and screened out. The samples that are finally retained form the second calibration anomaly sample set.

[0123] In some preferred embodiments, the method for generating distribution network anomaly samples based on the dynamic balance between artificial intelligence models and mechanistic models further includes:

[0124] A report is generated based on either the first or second set of calibration anomaly samples.

[0125] If the temporal continuity of the sample data meets the preset requirements, it corresponds to the first calibration abnormal sample set; if the sample has temporal breaks or the continuity does not meet the requirements, it corresponds to the second calibration abnormal sample set. These two types of sample sets are the core data foundation of the report. For example, the first calibration abnormal sample set contains samples with complete temporal sequence and calibrated abnormal features, while the second calibration abnormal sample set contains samples that have undergone temporal repair and feature enhancement.

[0126] Key information is extracted from the corresponding calibration anomaly sample set, covering the power distribution network operation scenario corresponding to the sample, the core characteristics and performance parameters of the sample. The information extracted for the first calibration anomaly sample set includes the time-series correlation characteristics of each sample (such as the data time-series correlation coefficient and the smoothness of time-series changes), the actual identification accuracy, and the anomaly type (such as voltage mutation and current mutation). For the second calibration anomaly sample set, additional time-series repair parameters (such as the completion slope and the break duration) and mutation index change rate and other repair-related parameters are added.

[0127] This information is integrated into an anomaly sample generation report. The basic information of the sample set includes the total number of samples, the distribution of anomaly types (e.g., voltage surge samples account for 30%, current surge samples account for 25%), and the corresponding distribution network operation scenario (e.g., peak residential load scenario, industrial load fluctuation scenario). The quality verification data includes the average actual identification accuracy of each sample, the average amplitude of the time series deviation between the time series correlation features and the optimal time series standard value, etc. For example, the average actual identification accuracy of the first calibration sample set is 0.5, and the average amplitude of the time series deviation is 0.1. For example, the application suggestions for the samples are which samples are suitable for improving the model's ability to identify anomalies with high mutation rates, and which samples are suitable for training distribution network models with specific grid structures.

[0128] Model trainers can quickly understand the quality, type distribution, and applicable scenarios of abnormal samples by generating reports, thereby selecting samples for targeted training and improving the model's balance performance and generalization ability in power distribution network anomaly identification.

[0129] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0130] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power distribution network abnormal sample generation method based on dynamic balance between artificial intelligence model and mechanism model, characterized in that, Includes the following steps: Acquire time-series data of the operating status of the power distribution network and sample training datasets for artificial intelligence models; The time-series data of the running status and the sample training dataset are processed to obtain the time-series correlation coefficient and the accuracy correlation coefficient; The response bias value affected by the temporal characteristics of the sample is obtained by combining the temporal correlation coefficient and temporal correlation characteristics; If the temporal continuity of the initial abnormal samples meets the preset requirements, the initial abnormal samples are calibrated according to the accuracy correlation coefficient, the rate of change of the real-time running status mutation index and the response deviation value to obtain the first calibration abnormal sample set. If the temporal continuity of the initial abnormal samples does not meet the preset requirements, a temporal index coupling correction model is constructed based on the duration of temporal breaks, the location of breaks, and the accelerated change characteristics of abrupt changes in the operational status index of historical samples. Based on the temporal index coupling correction model, the temporal repair parameters and the rate of change of the operational status abrupt change index of the initial abnormal samples are obtained. The initial abnormal samples are then calibrated based on the temporal repair parameters, the rate of change of the operational status abrupt change index, the accuracy correlation coefficient, and the response deviation value to obtain the second calibrated abnormal sample set.

2. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to claim 1, characterized in that, The operational status time-series data includes line voltage values, equipment current values, active power output values, reactive power compensation values, and load connection data at different time points. The sample training dataset includes historical normal operation time series samples, various abnormal operating condition time series samples, and corresponding model training performance data.

3. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to claim 1, characterized in that, The time-series data of the running status and the sample training dataset are processed to obtain the time-series correlation coefficient and the accuracy correlation coefficient. The specific steps include: Temporal correlation features of sample data are extracted from the sample training dataset; Operational state mutation indicators are extracted from operational state time-series data; The temporal correlation coefficient between temporal correlation features and response parameters is extracted from historical data; the accuracy correlation coefficient between the rate of change of abrupt change indicators of distribution network operation status and the accuracy of abnormal sample identification is extracted from historical data.

4. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to claim 3, characterized in that, The temporal correlation features include the data temporal correlation coefficient, temporal change smoothness, fluctuation range of data difference between adjacent time points, and temporal period stability. The indicators of sudden changes in operating status include voltage change amplitude, current change rate, power fluctuation peak value, and load change frequency.

5. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to claim 4, characterized in that, The time-series correlation coefficient between time-series correlation features and response parameters is extracted based on historical data; the accuracy correlation coefficient between the rate of change of abrupt change indicators of distribution network operation status and the accuracy of anomaly sample identification is extracted based on historical data, specifically including the following steps: Collect historical sample time-series correlation features, response data, and abnormal sample feedback data; Extract the historical actual values ​​of the temporal correlation features of sample data and the actual values ​​of the response parameters for the corresponding periods from historical correlation data; The difference between the historical actual value and the optimal time series standard value of the temporal correlation feature of the sample data is used to obtain the temporal deviation amplitude, and the difference between the actual value and the ideal response standard value of the response parameter is used to obtain the performance response deviation amplitude. The time-series correlation coefficient between the time-series correlation characteristics and the response parameters is obtained by taking the ratio of the performance response deviation magnitude and the time-series deviation magnitude. Extract the historical rate of change of the distribution network operation status change index from the feedback data generated from historical anomaly samples, and the accuracy of anomaly sample identification corresponding to the rate of change; The accuracy correlation coefficient is obtained by calculating the ratio between the accuracy of abnormal sample identification and the historical rate of change of the operational status mutation index.

6. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to claim 1, characterized in that, The response bias value affected by the temporal characteristics of the samples is obtained by combining the temporal correlation coefficient and temporal correlation features, specifically including the following steps: Calculate the current deviation between the temporal correlation feature of the real-time sample data and the optimal temporal standard value; The response deviation value affected by the temporal characteristics of the sample is obtained based on the temporal correlation coefficient and the current deviation value.

7. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to claim 1, characterized in that, The first calibration anomaly sample set is obtained by calibrating the initial anomaly samples based on the accuracy correlation coefficient, the rate of change of the real-time operating status mutation index, and the response deviation value. The specific steps include: The standard identification accuracy of the initial abnormal sample is obtained by combining the accuracy correlation coefficient and the rate of change of the real-time running status mutation index. The actual identification accuracy is obtained by correcting the standard identification accuracy with the response deviation value. Based on the actual recognition accuracy, the initial abnormal samples are subjected to temporal integrity completion and abnormal feature calibration to obtain the first calibrated abnormal sample set.

8. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to claim 1, characterized in that, Based on the duration, location, and accelerated change characteristics of abrupt changes in operational status indicators of historical samples, a time-series index coupling correction model is constructed. Based on this model, the time-series repair parameters and the rate of change of operational status abrupt changes in the initial abnormal samples are obtained. Specifically, the following steps are included: Determine the temporal fracture parameters of the sample data, including the temporal fracture duration, fracture location, and correlation between the data before and after the fracture; Capture the accelerated change characteristics of abrupt changes in operational status indicators, including the accelerated change trend, peak rate, and duration of the operational status abrupt changes indicators; A time series index coupling correction model is constructed based on the historical sample time series fracture parameters, the accelerated change characteristics of the abrupt change index of the operating state, and the corresponding historical time series repair parameters and the change rate of the historical abrupt change index of the operating state. The change rate of the initial abnormal sample time series break parameters and the accelerated change characteristics of the operational state mutation index are input into the time series index coupling correction model to obtain the change rate of the initial abnormal sample time series repair parameters and the operational state mutation index.

9. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to claim 8, characterized in that, The initial abnormal samples are calibrated based on the time-series repair parameters, the rate of change of the abrupt change index of the operating status, the accuracy correlation coefficient, and the response deviation value to obtain the second calibration abnormal sample set. The specific steps include: The timing calibration weight coefficient of the initial abnormal sample is calculated based on the timing repair parameters of the initial abnormal sample, the rate of change of the sudden change index of the running status, the accuracy correlation coefficient, and the response deviation value. The second calibration abnormal sample set is obtained by performing time-series repair, abnormal feature enhancement and quality screening on the initial abnormal samples according to the time-series calibration weight coefficient.

10. The method for generating distribution network anomaly samples based on the dynamic balance of artificial intelligence model and mechanism model according to any one of claims 1 to 9, characterized in that, Also includes: A report is generated based on either the first or second set of calibration anomaly samples.

Images