A method, program, device and storage medium for fast detection of key abnormal subsets of multi-source heterogeneous power data

Abstract

The application relates to a multi-source heterogeneous power data-oriented key abnormal subset fast detection method, program, equipment and storage medium. The application comprises the following steps: collecting historical multi-source time series data of power system associated parameters, pre-processing the multi-source time series data; calculating the time series cycle length of each multi-source time series data based on a fast Fourier transform method, and taking the mode of the total cycle length as the sliding window length of all time series; for the time series data subsequence of the power system associated parameters to be detected, a non-self-matching region is selected, the Euclidean distance between the to-be-detected subsequence and other subsequences in the non-self-matching region is calculated, and the minimum value is taken as the Matrix Profile of the time series data of the to-be-detected power system associated parameters; the Matrix Profiles of all the time series data of the power system associated parameters are summarized, all dimensions are traversed, the abnormal scores of all dimensions are calculated, and the subset with the maximum abnormal score is taken as the key abnormal subset.

Description

Technical Field

[0001] This invention belongs to the field of power system anomaly detection technology, specifically relating to a method, program, device, and storage medium for rapid detection of key anomaly subsets of multi-source heterogeneous power data. Background Technology

[0002] In recent years, with the continuous advancement of digital transformation in my country's power system, the dispatching cloud big data platform has achieved multi-disciplinary integration and inter-level dispatch sharing, unifying heterogeneous operational data from multiple sources such as meteorology, primary energy, generating units, and loads, providing strong support for intelligent power grid analysis. Data governance is the foundation for ensuring the efficient use of power grid data. As renewable energy gradually replaces traditional fossil fuels in my country's energy transition, the randomness of power output from new energy generation equipment, which is highly susceptible to weather changes, has increased, significantly impacting the supply-demand balance and stability of the power system. Any anomalies in dispatching system data can severely affect the stable operation of the power grid, even causing grid paralysis and resulting in serious economic losses.

[0003] In existing power system anomaly detection technologies, most are based on the analysis and modeling of low-dimensional data correlations to detect outliers in data within a specific time period. However, the power system has a wide variety of business operations and complex business logic interactions. The various business data are not completely independent, and there is strong coupling between different systems. The sampling periods and frequencies of different data are also different, resulting in power data with high dimensionality and complex spatial distribution. Therefore, joint modeling of high-dimensional data is very difficult.

[0004] Since anomalies often occur locally, affecting only a few key dimensions of data, identifying a critical subset of anomalies in high-dimensional data is crucial for targeted data monitoring and processing. Density-based algorithms, such as the local outlier factor, find local anomalies by calculating the density relationships between nodes. However, as data dimensionality increases, the distribution of sample points becomes sparser, making it difficult to distinguish the density of anomalous samples in high-dimensional data. Distance-based anomaly detection algorithms, like the k-nearest neighbors algorithm, primarily identify anomalous data based on the distance between neighboring points. However, these methods are susceptible to the "curse of dimensionality" in high-dimensional data, leading to decreased accuracy. Furthermore, the isolated forest method, based on random segmentation, analyzes the depth of nodes in the tree and the number of nodes in the segmented region to calculate anomaly scores. However, tree-based methods randomly select noisy features for segmentation, which is easily affected by noise in high-dimensional data, resulting in poor detection performance. Reconstruction-based deep learning methods can extract features and reduce data dimensionality through neural networks, and use reconstruction errors for anomaly detection. However, deep learning-based methods often require a large amount of data for training, are not suitable for small amounts of data and new scenarios, and have poor interpretability of results.

[0005] Therefore, there is an urgent need to develop a new multi-dimensional time-series critical anomaly subset rapid detection framework to make it applicable to various application scenarios containing multi-source heterogeneous power grid data, improve the efficiency, reliability and accuracy of online anomaly detection technology, and establish a safe and efficient intelligent anomaly detection and monitoring technology with autonomous learning capabilities, which is of great significance to further improve the reliability, security and stability of the power grid. Summary of the Invention

[0006] The purpose of this invention is to provide a rapid detection method for key anomaly subsets of multi-source heterogeneous power data, enhance the modeling capability of multi-dimensional data coupling and correlation in the power grid, improve the detection accuracy and real-time calculation speed of anomaly data in the transmission network, enhance the power grid data governance capability, ensure the quality of power grid data, and provide comprehensive, complete and accurate data support for typical power supply guarantee business scenarios such as high-precision load forecasting at multiple time scales and power supply and demand balance analysis of the entire network.

[0007] A rapid detection method for key anomaly subsets in multi-source heterogeneous power data includes the following steps:

[0008] Step 1: Collect historical multi-source time-series data of relevant parameters of the power system and preprocess the multi-source time-series data;

[0009] Step 2: Calculate the time series period length of each multi-source time series data using the Fast Fourier Transform method, and take the mode of the total period length as the sliding window length of all current time series.

[0010] Step 3: For the time series data subsequence of the power system associated parameters of the anomaly to be detected, select the non-self-matching region, calculate the Euclidean distance between the subsequence to be detected and other subsequences in the non-self-matching region, and take the minimum value as the Matrix Profile of the time series data of the power system associated parameters to be detected.

[0011] Step 4: Summarize the Matrix Profile of all time-series data of power system related parameters, traverse all dimensions, calculate the anomaly score for each dimension, and select the subset with the largest anomaly score as the key anomaly subset.

[0012] Furthermore, the historical multi-source time-series data of the power system-related parameters in step 1 includes millisecond-level PMU data, second-level SCADA data, and minute-level load and power generation-related data, specifically voltage amplitude, phase angle, active power, reactive power at nodes and branches, wind turbine and photovoltaic power generation, load power, and meteorological data-related parameters such as horizontal irradiance, wind speed, temperature, and humidity.

[0013] Furthermore, the preprocessing of multi-source time-series data in step 1 includes unifying data types, detecting and imputing missing values, removing duplicates, interpolating or sampling multi-source data with different sampling resolutions, and aligning timestamps.

[0014] Furthermore, the method for calculating the time series period length of each multi-source time series data based on the Fast Fourier Transform in step 2 is as follows: the time domain signal is decomposed into its frequency components by Fast Fourier Transform, the peak value in the frequency domain represents the dominant frequency in the signal, and the reciprocal of the dominant frequency is the time series period length of the corresponding time series data.

[0015] Furthermore, step 3 specifically includes:

[0016] Step 3.1: Take the time series data T of the power system associated parameter A at time position i and length m. A subsequence of For the current power subsequence that needs to be checked for abnormal conditions, 1≤i≤n-m+1, where m is the sliding window length calculated in step 2;

[0017] Step 3.2: Select non-self-matching regions, i.e., exclude regions that are not matched with the current subsequence. Other subsequences with intervals smaller than the subsequence length The distance calculation results ensure that the distance between non-overlapping subsequences is meaningful, i.e., |ij|≥m;

[0018] Step 3.3: Set the backtracking length parameter MTL, where MTL is a power of 2. k Configure it so that it is greater than the sliding window length m;

[0019] Step 3.4: Calculate the subsequence Other subsequences in non-self-matching regions The z-standardized Euclidean distance d between them i,j Take the minimum value as the subsequence The Matrix

[0020]

[0021]

[0022] in, and For subsequence The mean and standard deviation;

[0023] Step 3.5: Obtain the time-series data T of the current power system associated parameter A. A Matrix in the subsequence The maximum value is discord, which is the calculated value of the current least consistent subsequence;

[0024] Step 3.6: If the time series data T of the currently selected power system associated parameter A is... A If the discord value is not less than the historical discord value within the specified time period, then the backtracking length parameter MTL is increased to a larger power of 2, and steps 3.4 to 3.6 are repeated until MTL is increased to the initial position of the time series or the set maximum length; otherwise, the time series data T of the power system associated parameter A is output. A Matrix Profile (T) A ).

[0025] Furthermore, in step 3.3, when calculating the z-standardized Euclidean distance, the MASS algorithm based on Fast Fourier Transform is used to transform the time-series data of the power system-related parameters into the complex domain for rapid calculation.

[0026] Furthermore, step 4 specifically includes:

[0027] Step 4.1: Multi-source time-series data of power system related parameters All calculations are performed according to steps 2 and 3 within the same sliding window length m to obtain a Matrix Profile of all power system-related parameter time series data of the same length.

[0028] Step 4.2: Calculate the sum of the MatrixProfiles of all power system-related parameter time-series data according to the same time position i.

[0029] Step 4.3: From the power system associated parameter set [A1,A2,...,A] K Randomly select q items from [ ], that is Calculate Matrix Profile(T) i all The intersection of the minimum values ​​of each dimension in the q-dimensional subset of )

[0030]

[0031] Step 4.4: Iterate through q from 1 to the total dimension K, calculate the anomaly score (anomalyscore(i)) for all dimensions q at the current time i, and select the subset with the largest anomaly score. For the key anomaly subset at time i:

[0032]

[0033]

[0034] A computer device / apparatus / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method for rapid detection of key anomaly subsets of multi-source heterogeneous power data.

[0035] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described method for rapid detection of key anomaly subsets of multi-source heterogeneous power data.

[0036] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the aforementioned method for rapid detection of key anomaly subsets of multi-source heterogeneous power data.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention can quickly locate the location of anomalies and determine the subset with the greatest impact on the anomalies. It greatly reduces the problem of the accumulation of random noise overshadowing the anomalies in individual dimensions when the data dimensionality increases. It has strong robustness, avoids information overload caused by data surges, reduces the consumption of computing resources, and has the ability to detect anomalies in a timely and rapid manner. It is of great significance for practical applications.

[0039] This invention identifies anomalies by comparing the consistency of statistical patterns between current and historical data. It possesses strong interpretability and scalability, facilitating the analysis of potential anomalous behaviors influenced by multidimensional data. It can extract key dimensions of anomalies from high-dimensional data, making anomaly detection results more targeted. Compared to analyzing one-dimensional anomalies individually, it can obtain results showing the coupling and correlation of multiple factors.

[0040] This invention employs an unsupervised method, eliminating the need for prior labeling and providing anomaly data tags, significantly reducing the workload associated with labeling. Furthermore, it is applicable to new domains and novel, unknown anomaly patterns. Since power transmission networks involve multiple application scenarios such as power supply and demand balance and power load forecasting, the types of new anomaly patterns are uncertain when processing data from new scenarios. Therefore, unsupervised learning methods are more versatile. This method does not rely on prior knowledge; it calculates the current anomaly score based on newly emerging data and compares and judges it with previous values. Thus, it is suitable for anomaly detection in unknown domains and dynamic environments. Attached Figure Description

[0041] Figure 1 This is an example diagram of an abnormal sequence.

[0042] Figure 2This is the overall flowchart of the present invention.

[0043] Figure 3 This is a schematic diagram of the calculation process for a one-dimensional MatrixProfile.

[0044] Figure 4 A diagram showing the setup for the backtrack length MTL.

[0045] Figure 5 This is a topology diagram of the IEEE-38 system.

[0046] Figure 6 This is an example image showing the results of anomaly detection.

[0047] Figure 7 This is a data table comparing model accuracy.

[0048] Figure 8 This is a data table representing a subset of key anomalies. Detailed Implementation

[0049] The present invention will now be further described with reference to the accompanying drawings.

[0050] This invention provides a rapid detection method for key anomaly subsets in multi-source heterogeneous power data, enhancing the modeling capability of multi-dimensional data coupling and correlation in the power grid, improving the accuracy of anomaly detection and real-time calculation speed in the transmission network, improving power grid data governance capabilities, ensuring power grid data quality, and providing comprehensive, complete, and accurate data support for typical power supply guarantee business scenarios such as high-precision load forecasting at multiple time scales and power supply-demand balance analysis across the entire network. The anomaly detection model in this invention can not only be used for rapid location of anomalies in multi-dimensional time series of power systems, but also provide a set of key dimensions causing the anomalies and give an index value of the degree of anomaly.

[0051] Figure 2 The flowchart of a method for rapid detection of key anomaly subsets in multi-source heterogeneous power data provided by the present invention is shown.

[0052] The method provided by this invention includes the following steps:

[0053] Steps 1 through 4 belong to the initialization process, and each step of the initialization process is performed only once.

[0054] Step 1: Collect historical time-series data related to the power system. The data during the operation of the power system includes millisecond-level PMU data, second-level SCADA data, and minute-level load and power generation data. Specifically, this includes voltage amplitude, phase angle, active power, reactive power, etc. at nodes and branches, wind turbine and photovoltaic power generation, load power, and meteorological data related parameters such as horizontal irradiance, wind speed, temperature, and humidity.

[0055] Step 2: Preprocess the multi-source data. For meteorological and station data, unify the data type to float. Detect the percentage of missing values; if the percentage is greater than 50%, delete the variable. Fill in the missing values ​​using the nearest neighbor method. Remove duplicate timestamps for each variable. For data from sources with different sampling resolutions, unify the data to minute-level resolution and align the length and timestamps of each parameter.

[0056] Step 3: Calculate the time series period length. Use the Fast Fourier Transform (FFT) method to calculate parameters such as voltage amplitude, phase angle, power, horizontal irradiance, wind speed, and power generation. The period. The Fast Fourier Transform decomposes a time-domain signal into its frequency components. The peak value in the frequency domain represents the dominant frequency of the signal, and the reciprocal of the dominant frequency gives the period of the corresponding periodic component. For voltage amplitude data T of length n... A =[t1,t2,…,t n ], period m A The calculation formula is:

[0057] spectrum = |FFT(T) A )|

[0058] dominant frequency=argmax(spectrum)

[0059] m A =n / dominant frequency

[0060] Specifically, the node voltage amplitude and phase angle period length are 200 minutes, the horizontal irradiance is 524641 minutes, and the load power, wind turbine power, and photovoltaic power periods are 1440 minutes. After calculating the periods for all parameters, the total period length is obtained.

[0061] Step 4: Calculate the sliding window length. Take the total period length M. all The mode is used as the sliding window length m for all current time series. This invention avoids the problem that it is difficult to obtain a uniform window length for all series based on human experience when the time series dimension increases.

[0062] Steps 5 to 10 pertain to the one-dimensional matrix profile calculation process for various power system-related parameters. Figure 3 This is a detailed diagram of this section.

[0063] Step 5: Take the time series data T of the power system associated parameter A at time position i and length m. A subsequence of For the current power subsequence that needs to be checked for abnormal conditions, 1≤i≤n-m+1;

[0064] Step 6: Select non-self-matching regions, i.e., exclude regions that are not matched with the current subsequence. Other subsequences with intervals smaller than the subsequence length The distance calculation results ensure that the distance between non-overlapping subsequences is meaningful, i.e., |ij|≥m;

[0065] Step 7: Set the backtracking length parameter MTL, where MTL is a power of 2. k The settings are configured to be greater than the sliding window length *m* to enable rapid anomaly detection. This is because the historical data volume of power data is enormous, and the PMU collects data at a millisecond frequency; calculating the distance over the entire time period would require a significant computational load. Considering the dynamic behavior of the power grid, including changes in generation, load, and control, switches in the grid change states according to control needs, leading to dynamic changes in the grid topology. Therefore, most power data exhibits a data pattern similar to recent data, and the dependence between power data gradually weakens over time, while the patterns of more distant data may change. This invention only requires setting the calculation of subsequences within the current relatively recent time period's sequence length, rather than the entire sequence, effectively avoiding data model drift caused by long-term operation of power equipment.

[0066] Step 8: Calculate the subsequence Other subsequences in non-self-matching regions The z-standardized Euclidean distance d between them i,j Take the minimum value as the subsequence The Matrix

[0067]

[0068]

[0069] in, and For subsequence The mean and standard deviation;

[0070] When calculating the z-normalized Euclidean distance, the MASS (Mueen's ultra-fast Algorithm for Similarity Search) algorithm based on the fast Fourier transform is used to transform the time-series data of the power system's associated parameters into the complex domain for fast calculation.

[0071] Step 9: Obtain the time-series data T of the current power system associated parameter A. A Matrix in the subsequence The maximum value is discord, which is the calculated value of the current least consistent subsequence;

[0072] Step 10: If the time series data T of the currently selected power system associated parameter A is... A If the discord value within the specified time period is not less than the historical discord value, it means that there are no similar subsequence patterns in the recent time period. Therefore, it is necessary to further increase the backtracking length parameter MTL to a larger power of 2, and repeat step 7 until the MTL is increased to the initial position of the time series or the maximum length set manually. Figure 4 This step is illustrated in the diagram; otherwise, the time-series data T of the power system-related parameter A is output. A Matrix Profile (T) A ).

[0073] Steps 11 to 14 pertain to the calculation process of the multidimensional Matrix Profile and key anomaly subset of power data.

[0074] Step 11: Multi-source time-series data of power system related parameters All calculations are performed according to steps 2 and 3 within the same sliding window length m to obtain a Matrix Profile of all power system-related parameter time series data of the same length.

[0075] Step 12: Calculate the sum of the MatrixProfiles of all time-series data of all power system-related parameters at the same time position i.

[0076] Step 13: From the power system associated parameter set [A1,A2,...,A] K Randomly select q items from [ ], that is Calculate Matrix The intersection of the minimum values ​​of each dimension in the q-dimensional subset

[0077]

[0078] Step 14: Iterate through q from 1 to the total dimension K, calculate the anomaly score (anomalyscore(i)) for all dimensions q at the current time i, and select the subset with the largest anomaly score. For the key anomaly subset at time i:

[0079]

[0080]

[0081] Steps 15 to 20 pertain to the power simulation data generation and model verification process.

[0082] Step 15: Online Anomaly Detection Threshold Calculation. First, a pre-calculated Matrix Profile is performed based on historical data. As new data points arrive, a new Matrix Profile is calculated progressively. Finally, the dynamic threshold τ is used to determine the threshold. t Confirmed, in the following specific form:

[0083] τ t =μ t +kσ t

[0084] Where, μ t and σ t This represents the mean and standard deviation of the Matrix Profile for all subsequences in the historical electricity data to date. μ is dynamically updated based on the new Matrix Profile score after each new data point arrives. t and σ t τ will be updated later. t .

[0085] Step 16: Output the anomaly detection results. Determine whether the current subsequence is abnormal based on the threshold. If the calculated value is greater than the threshold, it is considered abnormal; otherwise, it is considered normal data.

[0086] Step 17: Generating Pseudo-Anomaly Data. Using the IEEE-38 node transmission network system as a benchmark, establish a cyber-physical system for the power grid that includes distributed energy resources. Figure 5 This is the topology diagram of the example system. SCADA and PMU measurement devices are configured at each node. Centralized renewable energy generators with an installed capacity of 1000MW are introduced at nodes 17 and 27. Three nodes in each region are selected to configure distributed renewable energy generators accounting for 20% of the daily average load of each node, along with their corresponding meteorological data. The base power per unit value in the example is 100MW, and the base voltage is 750KV.

[0087] Step 18: Node Load Setting. Actual grid load data is used. Node measurement data is obtained by superimposing normally distributed measurement noise with a mean of 0 onto the power flow calculation data. The standard deviation of the SCADA active and reactive power measurement errors is 0.02, and the standard deviations of the PMU voltage amplitude and phase angle measurement errors are 0.005 and 0.002, respectively. The data acquisition interval for each set of operating conditions is 15 minutes.

[0088] Step 19: Simulate the measurement status after an error or interference occurs in the power system information transmission network. Abnormal data will be generated by an increase or decrease of 50%-200% in the active or reactive power values ​​of randomly selected measurement nodes, and a random increase of 15%-25% in voltage amplitude and phase angle.

[0089] Step 20: Model Accuracy Validation. The present invention is compared with models OC-SVM, DAGMM, IFORS, and AutoEncoder. For model evaluation, the present invention uses three standard evaluation metrics: precision, recall, and F1 score, with the following formulas:

[0090]

[0091]

[0092]

[0093] In this model, TP represents the number of true positives (real anomalies detected), FP represents false positives (false positives), and FN represents false negatives (no detected anomalies). Precision is the proportion of samples with true anomalies to the total number of anomaly samples predicted by the model, while recall is the proportion of anomalies predicted by the model to the total number of anomaly samples. Therefore, a higher recall rate means more anomalies are detected; higher precision means fewer false positives.

[0094] The method proposed in this invention selects the first fifth of historical data to calculate a pre-conception Matrix Profile, and detects anomalies in the remaining data through real-time calculation. Figure 7 The results show the comparison of model accuracy. In many real-world scenarios, systems must detect as many actual attacks or anomalies as possible with minimal false positives, thus prioritizing recall and F1 score over precision. Therefore, our proposed multidimensional anomaly detection model exhibits higher accuracy.

[0095] Step 22: Verify the key anomaly subset. Taking a two-dimensional anomaly as an example, Figure 6 This shows the abnormal calculation process of some dimensions of power system data. Figure 8 The calculation results of the two-dimensional critical subset of anomalies are given. Combined with the visualized time series curve, it can be seen that the method proposed in this invention can effectively identify the time of anomaly occurrence and the key anomaly dimensions.

[0096] This invention can quickly locate the location of anomalies and determine the subset with the greatest impact on the anomalies. It greatly reduces the problem of the accumulation of random noise overshadowing the anomalies in individual dimensions when the data dimensionality increases. Therefore, the model is more robust to noise, avoids information inflation caused by data surges, reduces the consumption of computing resources, and has the ability to detect anomalies in a timely and rapid manner. It has great significance for practical applications.

[0097] This invention identifies anomalies by comparing the consistency of statistical patterns between current and historical data. It possesses strong interpretability and scalability, facilitating the analysis of potential anomalous behaviors influenced by multidimensional data. It can extract key dimensions of anomalies from high-dimensional data, making anomaly detection results more targeted. Compared to analyzing one-dimensional anomalies individually, it can obtain results showing the coupling and correlation of multiple factors. Figure 1 Example graph of an abnormal sequence.

[0098] This invention employs an unsupervised method, eliminating the need for prior labeling and providing anomaly data tags, significantly reducing the workload associated with labeling. It is also applicable to new domains and novel, unknown anomaly patterns. Since power transmission networks involve multiple application scenarios such as power supply and demand balance and power load forecasting, the types of new anomaly patterns are uncertain when processing data from new scenarios. Therefore, unsupervised learning methods are more versatile. This method does not rely on prior knowledge; it calculates the current anomaly score based on newly emerging data and compares it with previous values ​​for comprehensive judgment. Thus, it is suitable for anomaly detection in unknown domains and dynamic environments.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for rapid detection of key anomaly subsets in multi-source heterogeneous power data, characterized in that, Includes the following steps: Step 1: Collect historical multi-source time-series data of relevant parameters of the power system and preprocess the multi-source time-series data; Step 2: Calculate the time series period length of each multi-source time series data using the Fast Fourier Transform method, and take the mode of the total period length as the sliding window length of all current time series. Step 3: For the time series data subsequence of the power system associated parameters of the anomaly to be detected, select the non-self-matching region, calculate the Euclidean distance between the subsequence to be detected and other subsequences in the non-self-matching region, and take the minimum value as the Matrix Profile of the time series data of the power system associated parameters of the anomaly to be detected. Step 3.1: Take the time position at Location, length is Power system related parameters Time series data subsequence of For the current power subsequence that needs to be checked for anomalies, m is the sliding window length calculated in step 2; Step 3.2: Select non-self-matching regions, i.e., exclude regions that are not matched with the current subsequence. Other subsequences with intervals smaller than the subsequence length The distance calculation results ensure that the distances between non-overlapping subsequences are meaningful, i.e. ; Step 3.3: Set the backtracking length parameter , According to the powers of 2 k Configure it so that it is greater than the sliding window length m; Step 3.4: Calculate the subsequence Other subsequences in non-self-matching regions z-standard Euclidean distance between Take the minimum value as the subsequence of ; in, and For subsequence The mean and standard deviation; Step 3.5: Obtain relevant parameters of the current power system Time series data in the subsequence of The maximum value is discord, which is the calculated value of the current least consistent subsequence; Step 3.6: If the currently selected power system associated parameters Time series data If the discord value is not less than the historical discord value within the specified time period, then the backtrack length parameter is increased. To lengths that are larger powers of 2, repeat the steps. Up to step 3.6, until Expand to the initial position of the time series or the set maximum length; Otherwise, output the relevant parameters of the power system. Time series data of ; Step 4: Summarize the Matrix Profile of all time-series data of power system related parameters, traverse all dimensions, calculate the anomaly score for each dimension, and select the subset with the largest anomaly score as the key anomaly subset; Step 4.1: Multi-source time-series data of power system related parameters All steps 2 and 3 are performed within the same sliding window length. Calculations were performed to obtain time-series data of all power system-related parameters of the same length. ; Step 4.2: According to the same time position Calculate time series data of all power system associated parameters sum ; Step 4.3: From the set of power system related parameters Randomly select q items, that is ,calculate The intersection of the minimum values ​​of each dimension in the q-dimensional subset ; Step 4.4: Iterate through q from 1 to the total dimension K, and calculate the anomaly score for all dimensions q at the current time i. The subset with the largest outlier score For the key anomaly subset at time i: 。 2. The method for rapid detection of key anomaly subsets in multi-source heterogeneous power data according to claim 1, characterized in that: The historical multi-source time-series data of power system related parameters in step 1 includes millisecond-level PMU data, second-level SCADA data, and minute-level load and power generation related data, specifically voltage amplitude, phase angle, active power, reactive power at nodes and branches, wind turbine and photovoltaic power generation, load power, and meteorological data related parameters such as horizontal irradiance, wind speed, temperature, and humidity.

3. The method for rapid detection of key anomaly subsets in multi-source heterogeneous power data according to claim 1, characterized in that: Step 1, the preprocessing of multi-source time series data, includes unifying data types, detecting and filling missing values, removing duplicates, interpolating or sampling multi-source data with different sampling resolutions, and aligning timestamps.

4. The method for rapid detection of key anomaly subsets in multi-source heterogeneous power data according to claim 1, characterized in that: The method for calculating the time series period length of each multi-source time series data in step 2 based on the fast Fourier transform is as follows: the time domain signal is decomposed into its frequency components by fast Fourier transform, the peak value in the frequency domain represents the dominant frequency in the signal, and the reciprocal of the dominant frequency is the time series period length of the corresponding time series data.

5. The method for rapid detection of key anomaly subsets in multi-source heterogeneous power data according to claim 1, characterized in that: In step 3.3, when calculating the z-normalized Euclidean distance, an algorithm based on the Fast Fourier Transform is used. This transforms the time-series data of power system-related parameters into a complex domain for rapid calculation.

6. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 5.

8. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 5.

Images