Anomaly detection method for power dispatch monitoring data based on directional density ratio change rate

Abstract

This invention proposes an anomaly detection method for power dispatch monitoring data based on the rate of change of directed density ratios. The method includes: using historical power dispatch monitoring data as input data; calculating the local kernel density of samples through kernel density estimation and expanding the nearest neighbor set; combining the local kernel density with the distance vector between samples to calculate the sum of the directed density ratios of each sample and its nearest neighbors; calculating the rate of change of the directed density ratios of samples and their nearest neighbors using the L2 norm of the directed density ratios; summing the rates of change of directed density ratios under different parameters to obtain the final anomaly score for each sample, thereby identifying anomalous samples in the power dispatch monitoring data. The technical solution provided by this invention can improve the accuracy and stability of anomaly detection in power dispatch monitoring data.

Description

Technical field

[0001] This invention relates to a method for detecting anomalies in power dispatching and monitoring data, and more particularly to a method for detecting anomalies in power dispatching and monitoring data based on the directional density ratio change rate. [Background Technology]

[0002] With social development and the expansion of the power system, the construction of smart grids is also progressing rapidly. Smart grid systems aim to solve various problems throughout the entire process from power generation facilities through transmission and transformation to end users by actively utilizing communication and information technologies. As the smart grid advances, further demands are placed on the efficiency, safety, and stability of grid dispatching. The power dispatch automation system in the smart grid, as the command center of grid operation, plays a vital role in the power generation, transmission, transformation, and distribution stages of the power system, and is key to achieving the goals of grid reliability, safety, economy, efficiency, environmental friendliness, and safe operation.

[0003] While power dispatch automation systems can provide a wealth of power grid operation data and analysis results, facilitating accurate assessment of grid conditions by dispatch system operators, these systems possess massive amounts of data with numerous characteristic dimensions and complex interrelationships. As the command center of the power grid, any anomalies in the dispatch system's operations can significantly impact the grid's stable operation. Therefore, to ensure the safe and reliable operation of the smart grid, anomaly detection technology for power dispatch monitoring data has become increasingly important.

[0004] Power dispatch monitoring data is characterized by its large volume, numerous feature dimensions, and complex and diverse distribution patterns. Currently, most methods rely on expert experience to set thresholds for anomaly detection, which is inflexible and has limited accuracy. Therefore, it is necessary to introduce machine learning methods to achieve more accurate anomaly detection. However, with the large-scale application of the Internet of Things (IoT) in power grid systems, dispatch automation systems can acquire and record massive amounts of historical business data. Manually labeling all data is costly and time-consuming, making it difficult to use supervised machine learning methods for anomaly detection. In this situation, a better option is to use unlabeled data for unsupervised anomaly detection. Furthermore, in actual power dispatch monitoring data, due to the massive data volume and unknown number of anomalies, it is difficult to effectively specify parameters for unsupervised anomaly detection methods that require parameter setting and whose detection results are significantly affected by parameters, leading to unstable anomaly detection results. Therefore, this paper proposes an unsupervised anomaly detection algorithm that can adapt to multi-modal data distribution and is robust to parameter selection, aiming to improve the accuracy and stability of anomaly detection. This is of great significance for strengthening power grid status monitoring and ensuring power grid security. [Summary of the Invention]

[0005] In view of this, the present invention proposes a method for detecting anomalies in power dispatch monitoring data based on the directional density ratio change rate, so as to improve the performance of power dispatch monitoring data anomaly detection.

[0006] This invention proposes a method for detecting anomalies in power dispatch monitoring data based on the directional density ratio change rate, comprising the following steps:

[0007] (1) Using historical power dispatch monitoring data as the input dataset, the local kernel density of the samples is calculated through kernel density estimation and extended nearest neighbor set, specifically:

[0008] The real-time resource usage data of processes related to the power dispatching system's business, collected by the power dispatching and monitoring system, is used as the input to the historical dataset of power dispatching and monitoring, denoted as X={x1,x2,x3,...,x n}, where x n ∈R u R represents the set of real numbers, u = 6 represents the dimension of the data, specifically including process CPU utilization, memory utilization, disk I / O, network I / O, number of threads, and number of network connections; the total number of samples in the input dataset X is N;

[0009] Calculate the i-th sample x in X i Euclidean distance d(x) to all other samples i ,x j ), x j ∈X, and take the distance x i The k nearest samples are denoted as the nearest neighbor set kNN(x) i ); calculate kNN(x) for all samples i After that, iterate through all samples again, for sample x i Query whether it exists in another sample x m The nearest neighbor set kNN(x) m In ), x m ∈X, find all samples x that match. m Let RkNN(x) be the reverse nearest neighbor set. i ); kNN(x i ) and RkNN(x i By merging, we obtain sample x. i The extended nearest neighbor set EkNN(x) i )=kNN(x i )∪RkNN(x i );

[0010] Use x i The samples in the extended nearest neighbor set and the Gaussian kernel function on xi Perform kernel density estimation to obtain x i Local kernel density ρ(x) i ):

[0011]

[0012] Wherein, EkNN(x) i ) is the extended nearest neighbor set defined above; d(x) i ,x j ) represents sample x i To sample x j The Euclidean distance; i = 1, 2, ..., N, where N represents the total number of samples in the input dataset X; h is the bandwidth of the Gaussian kernel function, which is calculated by the following formula:

[0013]

[0014] Where d(x) i ,x j ) represents sample x i To sample x j The Euclidean distance; i = 1, 2, ..., N, where N represents the total number of samples in the input dataset X;

[0015] (2) Combining the local kernel density with the distance vector between samples, calculate the sum of the directed density ratios of each sample and its nearest neighbors, specifically:

[0016] Based on the local kernel density ρ(x) calculated in step (1) i By combining the distance vectors between samples, the distance vectors for each sample x in the input sample set X are calculated. i The sum of the directed density ratios of its nearest neighbor samples

[0017]

[0018] Where ρ(·) is the local kernel density calculated in step (1); Represents sample x in the feature space i To sample x j The vector; EkNN(x) i ) is the extended nearest neighbor set defined in step (1); i = 1, 2, ... N, N represents the total number of samples in the input dataset X;

[0019] (3) The rate of change of the directed density ratio between the sample and its nearest neighbors is calculated using the L2 norm of the directed density ratio, specifically:

[0020] Based on the sum of the directed density ratios of each sample and its nearest neighbors calculated in step (2), the rate of change of the directed density ratio is calculated to measure the degree of anomaly of the sample.i The directional density ratio change rate ΔDRD(i,k) is defined by the following equation:

[0021]

[0022] in This represents sample x. i The sum of the directional density ratios The second norm;

[0023] (4) Sum the rates of change of the directional density ratio under different parameters to obtain the final anomaly score for each sample, and use this to determine the abnormal samples in the power dispatch monitoring data, specifically:

[0024] Based on the directional density ratio change rate ΔDRD(i,k) calculated in step (3), the sample x is calculated by summing ΔDRD(i,k) for different k values. i The final abnormal score DCR(x) i ):

[0025]

[0026] Where K is a manually defined hyperparameter; based on the calculated anomaly score DCR(x) i The data samples are sorted, and the top t% of samples with the highest scores are identified as abnormal, thus realizing the detection of abnormal power dispatch and monitoring data.

[0027] In step (1) of the above method, the value of k is 6ln(N) rounded to the nearest integer, where N is the total number of samples in the input dataset X.

[0028] In step (4) of the above method, the value of K is 6ln(N) rounded to the nearest integer, where N is the total number of samples in the input dataset X.

[0029] In step (4) of the above method, the top t% of samples with the highest scores after sorting are judged as abnormal, and 5≤t≤20 is taken.

[0030] The proposed method for detecting anomalies in power dispatch and monitoring data improves the accuracy and stability of anomaly detection in power dispatch and monitoring data.

[0031] As can be seen from the above technical solutions, the present invention has the following beneficial effects:

[0032] In the technical solution implemented in this invention, by using extended nearest neighbor set and kernel density estimation, the local density of the sample can be determined more accurately. By using the directed density ratio that combines density, distance and direction to determine the degree of local anomaly of the sample, it can adapt to different data distribution patterns. By using the rate of change of directed density ratio under different parameters to calculate the final anomaly score, the robustness of the method can be improved, thereby improving the accuracy and stability of anomaly detection in power dispatch monitoring data. [Attached Image Description]

[0033] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of the framework of the power dispatch monitoring data anomaly detection method based on the directional density ratio change rate proposed in this invention.

[0035] Figure 2 This is a schematic diagram for calculating the directional density ratio;

[0036] Figure 3 This is a schematic diagram of the input data and output results of the algorithm of this invention.

Detailed Implementation Methods

[0037] To better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings.

[0038] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0039] This invention provides an anomaly detection method for power dispatch monitoring data based on the rate of change of directed density ratio. To meet the requirements for anomaly detection in power dispatch monitoring data, this invention calculates the sum of the directed density ratios of a sample and its nearest neighbors based on the local kernel density under extended nearest neighbor conditions, and determines anomalies by calculating the final anomaly score using the rate of change of directed density ratios under different parameters.

[0040] Figure 1 This is a schematic diagram of the framework of the power dispatching and monitoring data anomaly detection method based on the directional density ratio change rate proposed in this invention. The method includes the following steps:

[0041] Step 101: Using historical power dispatch monitoring data as the input dataset, the local kernel density of the samples is calculated through kernel density estimation and extended nearest neighbor set, specifically:

[0042] The real-time resource usage data of processes related to the power dispatching system's business, collected by the power dispatching and monitoring system, is used as the input to the historical dataset of power dispatching and monitoring, denoted as X={x1,x2,x3,...,x n}, where x n ∈R u R represents the set of real numbers, u = 6 represents the dimension of the data, specifically including process CPU utilization, memory utilization, disk I / O, network I / O, number of threads, and number of network connections; the total number of samples in the input dataset X is N.

[0043] Calculate the i-th sample x in X i Euclidean distance d(x) to all other samples i ,x j ), x j ∈X, and take the distance x i The k nearest samples are denoted as the nearest neighbor set kNN(x) i ); calculate kNN(x) for all samples i After that, iterate through all samples again, for sample x i Query whether it exists in another sample x m The nearest neighbor set kNN(x) m In ), x m ∈X, find all samples x that match. m Let RkNN(x) be the reverse nearest neighbor set. i ); kNN(x i ) and RkNN(x i By merging, we obtain sample x. i The extended nearest neighbor set EkNN(x) i )=kNN(x i )∪RkNN(x i ).

[0044] Use x i The samples in the extended nearest neighbor set and the Gaussian kernel function on x i Perform kernel density estimation to obtain x i Local kernel density ρ(x) i ):

[0045]

[0046] Wherein, EkNN(x) i ) is the extended nearest neighbor set defined above; d(x) i ,x j) represents sample x i To sample x j The Euclidean distance; i = 1, 2, ..., N, where N represents the total number of samples in the input dataset X; h is the bandwidth of the Gaussian kernel function, which is calculated by the following formula:

[0047]

[0048] Where d(x) i ,x j ) represents sample x i To sample x j The Euclidean distance; i = 1, 2, ..., N, where N represents the total number of samples in the input dataset X.

[0049] Step 102: Combining the local kernel density with the distance vector between samples, calculate the sum of the directed density ratios of each sample and its nearest neighbors, specifically:

[0050] Based on the local kernel density ρ(x) calculated in step 101 i By combining the distance vectors between samples, the distance vectors for each sample x in the input sample set X are calculated. i The sum of the directed density ratios of its nearest neighbor samples

[0051]

[0052] Where ρ(·) is the local kernel density calculated in step 101; Represents sample x in the feature space i To sample x j The vector; EkNN(x) i ) is the extended nearest neighbor set defined in step 101; i = 1, 2, ... N, N represents the total number of samples in the input dataset X.

[0053] Step 103: Calculate the rate of change of the directed density ratio between the sample and its nearest neighbors using the L2 norm of the directed density ratio. Specifically:

[0054] Based on the sum of the directed density ratios of each sample and its nearest neighbors calculated in step 102, the rate of change of the directed density ratio is calculated to measure the degree of anomaly of the sample. Sample x i The directional density ratio change rate ΔDRD(i,k) is defined by the following equation:

[0055]

[0056] in This represents sample x. i The sum of the directional density ratios The 2-norm.

[0057] Step 104: Sum the rates of change of the directional density ratio under different parameters to obtain the final anomaly score for each sample. Based on this score, abnormal samples in the power dispatch monitoring data are identified. Specifically:

[0058] Based on the directional density ratio change rate ΔDRD(i,k) calculated in step 103, the sample x is calculated by summing ΔDRD(i,k) for different k values. i The final abnormal score DCR(x) i ):

[0059]

[0060] Where K is a manually defined hyperparameter; based on the calculated anomaly score DCR(x) i The data samples are sorted, and the top t% of samples with the highest scores are identified as abnormal, thus realizing the detection of abnormal power dispatch and monitoring data.

[0061] Figure 3 This is a schematic diagram illustrating the input data and output results of the algorithm of this invention. The input of the algorithm is real-time resource usage data of processes related to the power dispatching system's business, collected by the power dispatching and monitoring system, including process CPU utilization, memory utilization, disk I / O, network I / O, number of threads, and number of network connections. The output of the algorithm is the anomaly score DCR(x). i The data samples with the highest scores after sorting the abnormal scores are identified as abnormal.

[0062] Algorithm 1 is the pseudocode for a method based on the anomaly detection of the directional density ratio change rate:

[0063]

[0064]

[0065] In this specific implementation, 22 publicly available datasets were used for testing. These datasets came from various domains, and all features were standardized using max-min scaling. Details of the datasets are shown in Table 1. To reduce the randomness of the results, all results are the average after 10-fold cross-validation.

[0066] Table 1 shows the datasets used in the specific embodiments.

[0067]

[0068]

[0069] To verify the effectiveness of the proposed algorithm, 12 typical unsupervised anomaly detection algorithms were used as comparison algorithms in the embodiments of the present invention, as shown in Table 2. In the table, the embodiments of the present invention are represented by DCROD.

[0070] Table 2 shows the algorithms compared in specific embodiments.

[0071]

[0072] This invention uses ROC curves to measure anomaly detection results. An ROC curve is a graph plotting accuracy versus false positive rate as a decision threshold changes. A perfect anomaly detector's curve should consist of a vertical line at a false positive rate of 0 and a horizontal line at an accuracy rate of 1. The ROC curve can be measured by a simple metric: the area under the curve (AUC). AUC ranges from 0 to 1; the higher the anomaly detection accuracy, the larger the AUC. A perfect anomaly detector achieves an AUC of 1, while the worst result results in an AUC of 0.

[0073] In this embodiment of the invention, t is set to 10 in the exception class marking method.

[0074] Table 3 shows the AUC results of the embodiments of the present invention and comparative methods based on k-nearest neighbors (including methods based on k-nearest neighbors (density) and methods based on k-nearest neighbors (distance)) on public datasets. It can be seen that the power dispatch monitoring data anomaly detection method based on the directed density ratio change rate of the present invention achieves higher AUCs than other methods on most public datasets, and obtains the highest average AUC and the lowest average rank.

[0075] Table 3 shows the AUC results compared to the k-nearest neighbor-based method on public datasets.

[0076] Dataset KNN LOF COF LDF RDOS LGOD CELOF MOD+ DCROD PenDigits 0.9743 0.8767 0.8818 0.9283 0.9289 0.9816 0.8901 0.9699 0.9587 Shuttle 0.7952 0.9015 0.8545 0.8352 0.8889 0.8144 0.8877 0.7389 0.9349 Glass 0.8049 0.7892 0.7966 0.8021 0.7856 0.8254 0.7712 0.8067 0.8545 Annthyroid 0.5736 0.6331 0.656 0.6013 0.6142 0.5482 0.6373 0.5669 0.641 Stamps 0.8885 0.7269 0.676 0.8179 0.7841 0.8577 0.7614 0.8863 0.8759 PageBlocks 0.8588 0.7628 0.7048 0.7681 0.6815 0.8185 0.7363 0.831 0.862 Cardiotoco 0.6201 0.5979 0.54 0.6056 0.6174 0.6661 0.6041 0.59 0.6506 WPBC 0.5086 0.5004 0.5139 0.5075 0.4913 0.5525 0.4949 0.545 0.5403 Pima 0.7276 0.6413 0.6289 0.6968 0.6539 0.7633 0.6394 0.7433 0.74 Ionosphere 0.8615 0.8883 0.8869 0.7865 0.8542 0.8835 0.897 0.6822 0.9166 SpamBase 0.5668 0.4349 0.4062 0.4392 0.512 0.6094 0.4551 0.5743 0.5759 Heart Disease 0.6633 0.5838 0.5749 0.6296 0.5863 0.8055 0.6085 0.7137 0.6356 Arrhythmia 0.7493 0.736 0.7387 0.7215 0.7448 0.7456 0.7545 0.7181 0.7606 WBC 0.9741 0.9307 0.9236 0.9693 0.989 0.9411 0.9885 0.9231 0.9778 WDBC 0.9565 0.9343 0.8996 0.9165 0.9406 0.9367 0.9504 0.9308 0.9398 Waveform 0.6938 0.7359 0.7132 0.748 0.7606 0.7504 0.7078 0.72 0.7301 Hepatitis 0.7008 0.6585 0.6367 0.7037 0.6977 0.6478 0.7149 0.7847 0.6599 Lymphogra 0.9987 0.9918 0.9916 0.9668 0.9974 0.9923 0.9884 0.9768 0.9985 Parkinson 0.3821 0.4142 0.6819 0.4114 0.3859 0.4944 0.4033 0.4406 0.4371 Wilt 0.3808 0.4561 0.4521 0.4209 0.4523 0.3951 0.369 0.387 0.4619 InternetAds 0.695 0.6718 0.6808 0.6712 0.6705 0.6338 0.6956 0.5995 0.7317 KDDCup99 0.965 0.654 0.6154 0.6151 0.5762 0.9637 0.6175 0.9489 0.9208 Average AUC 0.7427 0.7055 0.7025 0.7074 0.7097 0.7558 0.7079 0.7308 0.7638 Average rank 4.18 5.86 6.45 5.91 5.41 3.82 5.45 5.27 2.64

[0077] Table 4 shows the AUC results of the embodiments of this invention and other comparative methods on public datasets. It can be seen that the power dispatch monitoring data anomaly detection method based on the directed density ratio change rate of this invention achieves higher AUC values ​​than other methods on most public datasets, and obtains the highest average AUC and the lowest average rank. Combining the results in Tables 3 and 4, it is evident that the anomaly detection accuracy of the method in the embodiments of this invention is high.

[0078] Table 4 shows the AUC results compared with other methods on public datasets.

[0079]

[0080]

[0081] Similarly, the embodiments of the present invention are applied to three types of business anomalies in the smart grid dispatch and control system, namely data jump, application network outage, and telemetry meter not refreshing.

[0082] Table 5 shows the AUC results of the embodiments of the present invention and other comparative methods on these three anomalies.

[0083] Table 5 shows the AUC results compared with density-based methods on the three anomalies.

[0084] Data jump Application disconnected from the network Telemetry meter not refreshed CBLOF 0.9476 0.9478 0.9582 IFOREST 0.9493 0.9579 0.9557 HBOS 0.9474 0.9486 0.9605 COPOD 0.9479 0.9514 0.9526 KNN 0.9429 0.9611 0.9606 LOF 0.9463 0.9503 0.9614 COF 0.9418 0.9507 0.9645 LDF 0.9585 0.9516 0.9592 RDOS 0.9547 0.9575 0.9509 LGOD 0.9430 0.9469 0.9611 CELOF 0.9546 0.9561 0.9626 MOD+ 0.9404 0.9474 0.9597 DCROD 0.9619 0.9640 0.9691

[0085] As shown in Table 5, compared with all selected comparison methods, the power dispatch monitoring data anomaly detection method based on the directional density ratio change rate of this invention achieved better AUC results. The comparison results on three actual power dispatch monitoring data anomalies, combined with comparison results on a large number of publicly available datasets, demonstrate that this invention can effectively reduce the false alarm rate and improve stability when dealing with complex power dispatch monitoring data anomaly detection involving both local and clustered anomalies.

[0086] In summary, the embodiments of the present invention have the following beneficial effects:

[0087] In the technical solution of this invention, historical power dispatch monitoring data is used as the input dataset. The local kernel density of the samples is calculated by expanding the nearest neighbor set and kernel density estimation. Combining the local kernel density with the distance vector between samples, the sum of the directed density ratios of each sample and its nearest neighbors is calculated. The rate of change of the directed density ratio between a sample and its nearest neighbors is calculated using the L2 norm of the directed density ratio. The rates of change of the directed density ratios under different parameters are summed to obtain the final anomaly score, and anomaly samples in the data are identified. According to the technical solution provided by this embodiment, when facing complex power dispatch monitoring data with diverse distribution patterns and manifolds, and the presence of local anomalies, this method can improve the accuracy and stability of anomaly detection in power dispatch monitoring data compared to typical unsupervised anomaly detection methods.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, 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 detecting anomalies in power dispatching and monitoring data based on the directional density ratio change rate, characterized in that, Includes the following steps: (1) Using historical power dispatch monitoring data as the input dataset, the local kernel density of the samples is calculated through kernel density estimation and extended nearest neighbor set, specifically: The real-time resource usage data of processes related to the power dispatching system's business, collected by the power dispatching and monitoring system, is used as the input to the historical dataset of power dispatching and monitoring, denoted as X={x1,x2,x3,...,x n }, where x n ∈R u R represents the set of real numbers, u = 6 represents the dimension of the data, specifically including process CPU utilization, memory utilization, disk I / O, network I / O, number of threads, and number of network connections; the total number of samples in the input dataset X is N; Calculate the i-th sample x in X i Euclidean distance d(x) to all other samples i ,x j ), x j ∈X, and take the distance x i The k nearest samples are denoted as the nearest neighbor set kNN(x) i ); calculate kNN(x) for all samples i After that, iterate through all samples again, for sample x i Query whether it exists in another sample x m The nearest neighbor set kNN(x) m In ), x m ∈X, find all samples x that match. m Let RkNN(x) be the reverse nearest neighbor set. i ); kNN(x i ) and RkNN(x i By merging, we obtain sample x. i The extended nearest neighbor set EkNN(x) i )=kNN(x i )∪RkNN(x i ); Use x i The samples in the extended nearest neighbor set and the Gaussian kernel function on x i Perform kernel density estimation to obtain x i Local kernel density ρ(x) i ): Wherein, EkNN(x) i ) is the extended nearest neighbor set defined above; d(x) i ,x j ) represents sample x i To sample x j The Euclidean distance; i = 1, 2, ..., N, where N represents the total number of samples in the input dataset X; h is the bandwidth of the Gaussian kernel function, which is calculated by the following formula: Where d(x) i ,x j ) represents sample x i To sample x j The Euclidean distance; i = 1, 2, ..., N, where N represents the total number of samples in the input dataset X; (2) Combining the local kernel density with the distance vector between samples, calculate the sum of the directed density ratios of each sample and its nearest neighbors, specifically: Based on the local kernel density ρ(x) calculated in step (1) i By combining the distance vectors between samples, the distance vectors for each sample x in the input sample set X are calculated. i The sum of the directed density ratios of its nearest neighbor samples Where ρ(·) is the local kernel density calculated in step (1); Represents sample x in the feature space i To sample x j The vector; EkNN(x) i ) is the extended nearest neighbor set defined in step (1); i = 1, 2, ... N, N represents the total number of samples in the input dataset X; (3) The rate of change of the directed density ratio between the sample and its nearest neighbors is calculated using the L2 norm of the directed density ratio, specifically: Based on the sum of the directed density ratios of each sample and its nearest neighbors calculated in step (2), the rate of change of the directed density ratio is calculated to measure the degree of anomaly of the sample. i The directional density ratio change rate ΔDRD(i,k) is defined by the following equation: in This represents sample x. i The sum of the directional density ratios The second norm; (4) Sum the rates of change of the directional density ratio under different parameters to obtain the final anomaly score for each sample, and use this to determine the abnormal samples in the power dispatch monitoring data, specifically: Based on the directional density ratio change rate ΔDRD(i,k) calculated in step (3), the sample x is calculated by summing ΔDRD(i,k) for different k values. i The final abnormal score DCR(x) i ): Where K is a manually defined hyperparameter; based on the calculated anomaly score DCR(x) i The data samples are sorted, and the top t% of samples with the highest scores are identified as abnormal, thus realizing the detection of abnormal power dispatch and monitoring data.

2. The method for detecting anomalies in power dispatching and monitoring data based on the directional density ratio change rate as described in claim 1, characterized in that, In step (1), the value of k is 6ln(N) rounded to the nearest integer, where N is the total number of samples in the input dataset X.

3. The method for detecting anomalies in power dispatching and monitoring data based on the directional density ratio change rate as described in claim 1, characterized in that, In step (4), the value of K is 6ln(N) rounded to the nearest integer, where N is the total number of samples in the input dataset X.

4. The method for detecting anomalies in power dispatching and monitoring data based on the directional density ratio change rate as described in claim 1, characterized in that, In step (4), the top t% of samples with the highest scores after sorting are judged as abnormal, with 5≤t≤20.

Images