Search

CN-121524897-B - Cloud-edge cooperative power distribution network electrical variable anomaly detection method and device

CN121524897BCN 121524897 BCN121524897 BCN 121524897BCN-121524897-B

Abstract

The application discloses a cloud-edge cooperative power distribution network electrical variable anomaly detection method and device, which belong to the technical field of intelligent operation and maintenance and safety control of power distribution networks and are used for guaranteeing safe, stable and high-quality operation of a power supply system; the method comprises the steps of firstly, extracting multi-source data of a power distribution network on a cloud side, analyzing consistency of the multi-source data through unsupervised depth features, learning to obtain robust electric variable track features, modeling complex nonlinear dependency between track features and abnormal modes through a Copula theory to realize high-precision abnormal density estimation, generating an interpretable and downloadable abnormal detection rule template library through modal self-organizing clustering, deploying an optimized lightweight model and rule library on an intelligent terminal on an edge side to realize online and rapid abnormal detection and classification early warning of real-time electric variable data, and overcoming the defects that a traditional method only depends on electric variable threshold judgment and lacks multi-source information fusion and behavior inference capability.

Inventors

  • LIU CHAOYANG
  • LI TAOCHEN
  • WU XIAOBIN
  • ZHANG RUI
  • HAO JIN
  • LI XI
  • YANG BIN
  • LU YUJIE
  • WANG YING
  • LI HUIQI

Assignees

  • 国网山西省电力有限公司信息通信分公司

Dates

Publication Date
20260512
Application Date
20260116

Claims (6)

  1. 1. A cloud-edge cooperative power distribution network electrical variable anomaly detection method is characterized by comprising the following steps: s1, integrating and arranging multisource operation and business information of a power distribution network, namely collecting, arranging and standardizing multisource information and data of the power distribution network on a cloud side, and constructing a data set for intelligent learning training; S2, electric variable running track extraction and polymorphic consistency analysis and modeling, namely constructing a rich information source electric variable track feature extractor Enhancing and fusing the dynamic change track of the electric variable in the multisource information and the key information signals in the step S1 to form the characteristics of the electric variable track of the rich information source, and then constructing a polymorphic consistency discriminator Ensuring the accuracy and stability of the rich source electric variable track characteristics through consistency constraint; The step S2 specifically comprises the following steps: s2.1, extracting the trace characteristics of the rich source electric variable, namely, collecting the data set Input to rich source electricity variable track feature extractor Obtaining a low-dimensional rich source electric variable track feature set Wherein each detection point profile Mapping to features The formula of (2) is: wherein Electrical variable track feature extractor for rich information source Is a set of learnable parameters; S2.2, restoring the signals of the multi-state detection points, and gathering the characteristics of the rich information source electric variable track Input to polymorphic consistency discriminator Obtaining the restored pseudo detection point file data set containing multi-state multi-information source Each feature is provided with Restoring the pseudo-detection point file The formula of (2) is: wherein Is polymorphic consistency discriminator Is a set of learnable parameters; S2.3, calculating polymorphism consistency discrimination errors between the pseudo detection point files and the detection point files: then updating the rich information source electric variable track characteristic extractor by using a gradient descent method according to the error Is a set of learnable parameters of (a) And polymorphic identity discriminator Is a set of learnable parameters of (a) Wherein Is the total number of detection points; s2.4, optimizing and training the rich source electric variable track characteristics, namely calculating and repeatedly executing the steps S2.1-S2.3 for at least 200 times to obtain an optimized rich source electric variable track characteristic set Complete rich source electricity variable track feature extractor of training And polymorphic identity discriminator ; S3, track-abnormal mode density estimation based on Copula information theory, namely, constructing a correlation structure decoupler And outlier feature density estimator Modeling the association relation between the electric variable track in the rich information source electric variable track characteristic and the potential abnormal mode by using Copula information theory, firstly using a Copula association structure decoupler Copula decoupling the dependency relationship between features, and then using an outlier feature density estimator Extracting a coupling relation between a resultant abnormality represented by an electric variable track and equipment or grid system mode abnormality represented by cloud side rich source data, and finally outputting rich source abnormality modal estimated density; The step S3 specifically comprises the following steps: s3.1, decoupling the electrical variable track characteristics from the abnormal behavior association structure by using a Copula association structure decoupler The rich information source electric variable track feature set obtained in the step S2.4 Each element of (a) is converted to an independent Copula space to obtain a decoupled enhanced feature set And (2) and And Is consistent in dimension; s3.2, estimating abnormal modal density of rich source based on the enhanced feature set And rich source electric variable track feature set Using outlier feature density estimators Abstracting key behavior modes between the electric variable track and the abnormal behavior mode to form electric variable track characteristics from rich information sources Mapping to the abnormal modal estimation density of the rich information source; Step 3.2 specifically comprises the following steps: S3.2.1 Electrical variable trajectory feature set from rich sources Taking b pieces of feature data without replacement from the random sampling, correspondingly from the enhanced feature set Corresponding b pieces of enhanced feature data are also taken out and put into an abnormal feature density estimator Respectively obtaining density estimated values And Wherein Is an outlier feature density estimator Is used for the parameter set of (a), And Respectively refer to the first Bar characteristic data and enhanced characteristic data; S3.2.2 calculating the trace-anomaly pattern density loss: wherein The function is defined as a rich source density evolution function algebraically The formula for the example is expressed as: ; s3.2.3 track-anomaly mode density loss gradients are calculated using gradient ascent and parameters are updated ; S3.2.4 assume that the parameters currently obtained according to S3.2.3 are , The number of times tag representing the current iteration is according to the formula: Updating parameters Wherein For a single one of the parameters that can be set, The last updated parameter; s3.2.5 repeating the steps S3.2.1-S3.2.4 until In a second time, the first time, Finally, an optimal characteristic density estimator is obtained ; S3.2.6 defining a rich source density correction function Algebraic by algebra The formula for the example is expressed as: Calculating all rich source electric variable track characteristics Is used for estimating density matrix of abnormal modes of rich source At this time Is incorporated into an optimal feature density estimator In (a) and (b); s4, constructing a power distribution network electric variable abnormal modal self-organization and abnormal detection rule template library, namely constructing a modal self-organizing loom Using modal self-organizing looms The similarity relation formed based on the abnormal modal estimation value of the rich information source is automatically identified and classified, and a configurable and issued abnormal detection rule template library is generated based on expert knowledge ; The step S4 specifically comprises the following steps: s4.1, estimating a density matrix according to the abnormal mode of the rich information source obtained in the step S3 According to the formula: Calculating two detection points And A measure of similarity between, wherein Is a single parameter that can be adjusted, And Representing the corresponding mode estimated density of the two detection points, and constructing a similarity matrix according to the formula ; S4.2 using spectral clustering algorithm pairs Unsupervised clustering is performed to obtain Self-organizing clusters, wherein The clusters are marked as Steps S4.1 and S4.2 are performed by a modal self-organizing loom Realizing; S4.3, tracing the self-organizing cluster information obtained according to the clustering result to the multi-source data information of the corresponding detection point, marking each self-organizing cluster with a class number, and then defining the abnormal behavior class to which each self-organizing cluster belongs by combining expert knowledge; s4.4, solving a cluster center for each self-organizing cluster, and solving the first Center of each cluster The formula of (2) is: wherein Density matrix representing abnormal mode estimation of rich information source Belongs to the field of The mode estimation density corresponding to the detection points of the category is called an abnormal detection rule template at the center of each cluster, and all templates jointly form an abnormal detection rule template library ; And S5, real-time intelligent detection and early warning of the abnormal electric variable, namely realizing real-time online abnormal detection, classification and early warning of the electric variable data at the edge side.
  2. 2. The cloud-edge cooperative power distribution network electrical variable anomaly detection method of claim 1, wherein in step S1, the collection, arrangement and standardization of cloud-side power distribution network multi-source information and data are achieved through a power distribution network information integration arrangement device The implementation method comprises the following specific steps: s1.1, multi-source information and data collection, namely collecting heterogeneous data from different sources and different sampling frequencies from a cloud-side data center; S1.2, improving the data quality, wherein the data quality comprises the processing of abnormal values and the restoration of data missing values; S1.3 modeling the rich source data files around the detection points, namely forming a single detection point file by means of statistical aggregation and standardization processing of all source data of the same detection point, which is also called matrix training data set Data set Each row represents a measurable detection point and each column represents a source of information.
  3. 3. The cloud-edge cooperative power distribution network electrical variable anomaly detection method of claim 1, wherein the step S5 specifically comprises the following steps: s5.1, inputting the electric variable flow data of the edge side real-time detection point to be detected and the related multi-source information of the detection point obtained from the cloud side into the information integration collator of the power distribution network Obtaining the data after being tidied; s5.2, inputting the processed data into a rich source electric variable track feature extractor Corresponding to the obtained characteristics; S5.3 inputting the features to the outlier feature density estimator Obtaining an estimated density value; S5.4, using the obtained estimated density value and the abnormal detection rule template library Matching all templates in the step S4.1 by using the similarity measure, and selecting the category corresponding to the most similar template to mark the data so as to obtain the category corresponding to the abnormal behavior; And S5.5, according to the abnormal behavior category, making corresponding response and control actions by combining the security priority processing mechanism configured by the detection point, and uploading complete early warning information to the cloud side.
  4. 4. The cloud-edge collaborative power distribution network electrical variable anomaly detection method according to claim 1, wherein the rich information source electrical variable track feature extractor is characterized by comprising the following steps of The system consists of a basic feature extraction network, a multi-source feature extraction network and an electric variable feature track generation network which are stacked in sequence, wherein: A basic feature extraction network is constructed based on Resnet network 101; the multi-information source characteristic extraction network comprises a plurality of sub-networks which are connected in parallel, and specifically comprises the following steps: the voltage-current mutual inductance signal enhancement network is formed by linearly splicing 2 full-connection groups in sequence, wherein the last activation function is changed into an exponential function; The load characteristic sensing network is formed by linearly splicing 3 full-connection groups in sequence, wherein the last activation function is changed into a tanh function; The capacity-to-load ratio evaluation network consists of 1 graph roll-up group; the power factor regression network is formed by linearly splicing 2 convolution groups and 1 full connection group in sequence, wherein the last activation function is changed into a softmax function; the electric parameter integrated extraction network is formed by linearly splicing 3 full-connection groups in sequence; the electric variable characteristic track generation network is formed by linearly splicing 4 groups of full-connection groups; the polymorphic consistency discriminator C is formed by stacking an electric variable track backtracking network, a multi-source characteristic restoring network and a state mapping network in sequence, wherein: The electric variable track backtracking network is formed by linearly splicing 4 groups of full-connection groups; The multi-information source characteristic restoring network comprises a plurality of sub-networks which are connected in parallel, and specifically comprises the following steps: The voltage-current mutual inductance signal restoring network is formed by linearly splicing 2 full-connection groups in sequence, wherein the last activation function is removed; the load characteristic restoring network is formed by linearly splicing 3 full-connection groups in sequence, wherein the last activation function is removed; the power grid capacity margin estimation network consists of 1 graph roll group and removes the last activation function; The active power and apparent power generating network is formed by linearly splicing 2 convolution groups and 1 full connection group in sequence, wherein the last activation function is removed; an electrical parameter restoring network is formed by linearly splicing 3 full-connection groups in sequence, wherein the last activation function is removed; wherein each of the fully connected groups comprises 1 fully connected layer, 1 ReLU activation function and a batch normalization layer, each of the convolved groups comprises 1-dimensional convolved layer, 1 pooled layer and 1 LeakyReLU activation function, and each of the graph convolved groups consists of 1-layer graph convolution layer and 1 ReLU activation function; the state mapping network is built based on Resnet network.
  5. 5. The cloud-edge cooperative power distribution network electrical variable anomaly detection method of claim 1, wherein the anomaly characteristic density estimator is characterized by comprising the following steps of The method comprises the following steps of: (1) The metering fusion layer is used for carrying out weighted summarization and nonlinear transformation on all the characteristics from the input layer, and comprises 4 full-connection layers, and then is connected with GELU activation functions; (2) The protection logic layer is used for screening and strengthening the characteristics through logic judgment, protecting the information with high contribution degree to the density characteristics and inhibiting redundant or noise information, and comprises a 3-layer convolution layer with a channel attention mechanism and a self-attention mechanism; (3) The tide regulation and control layer is used for regulating and controlling the flow and distribution of the characteristics and comprises 2 groups of normalization and residual error connection layers, each group is provided with 3 layers of full connection layers which are connected by interlayer residual errors, and 1 layer of normalization mechanism is arranged behind a third layer; (4) The rich source density estimation layer outputs the advanced density characteristics processed by the previous layers and consists of 1 fully connected layer with a sigmoid activation function.
  6. 6. The cloud-edge cooperative power distribution network electric variable anomaly detection device is characterized by comprising a rich information source power distribution network electric variable analysis modeling instrument deployed on a cloud side and an intelligent edge detection terminal deployed on an edge side, wherein a computer program/instruction related to the steps S1-S4 in the cloud-edge cooperative power distribution network electric variable anomaly detection method according to any one of claims 1-5 is deployed and operated in Yu Fuxin source power distribution network electric variable analysis modeling instrument, and a computer program/instruction related to the step S5 in the cloud-edge cooperative power distribution network electric variable anomaly detection method according to any one of claims 1-5 is deployed and operated in the intelligent edge detection terminal.

Description

Cloud-edge cooperative power distribution network electrical variable anomaly detection method and device Technical Field The application relates to the technical field of intelligent operation and maintenance and safety control of power distribution networks, in particular to a cloud-edge cooperative power distribution network electric variable anomaly detection method and device. Background Along with the continuous promotion of the intelligent level of the power distribution network, the abnormal detection of the electric variable becomes a key technology for guaranteeing the safe and stable operation of the power grid. At present, research and practice in the field are mainly based on an electrical measurement technology, and through real-time acquisition and threshold judgment of electrical quantities such as voltage, current, power and the like, obvious faults such as overload and short circuit are detected. However, most of the methods focus on the electric energy information, and cannot effectively fuse multi-source information such as a power distribution automation system, a power grid model, user report data and the like, so that the detection dimension is single, and increasingly complex abnormal scenes of the power distribution network are difficult to deal with. The existing detection means have significant limitations. On the one hand, the method is mainly based on the judgment of 'resultance' indexes such as threshold value out-of-limit of electric quantity or waveform distortion, and the like, and lacks the sensing and early warning capability of an abnormal evolution process, so that the method is essentially a post-hoc response mechanism. On the other hand, because the electric variable data cannot be associated and analyzed with the information such as the power grid topology, the business expansion capacity, the historical operation and maintenance records and the like, the detection system cannot explain the root cause of the abnormality, for example, cannot distinguish whether the user side maliciously uses electricity or the current caused by the natural aging of equipment is abnormal, and the operation and maintenance efficiency and the accuracy are seriously restricted due to lack of behavior deduction and interpretability. In recent years, research attempts have been made to introduce an unsupervised learning algorithm to promote the level of detection intelligence, but the method is still limited to feature mining inside the electrical variable data or simply splicing the multi-source data. The method is difficult to deeply describe complex nonlinear dependency relationships among electric variables, business, environment and other factors, so that false alarm rate and false alarm rate are high when facing new scenes such as batch installation access, distributed energy grid connection and the like. Therefore, an innovative detection method capable of integrating multi-source information of an electric measurement and distribution system and having behavior level sensing and early warning capabilities is urgently needed, so that the current technical bottleneck is broken through, and the operation and maintenance of the distribution network are promoted to develop towards the initiative and the intellectualization directions. Disclosure of Invention In order to solve the technical problems, the application provides a cloud-edge cooperative power distribution network electrical variable anomaly detection method and device. The technical scheme adopted by the application is that the cloud-edge cooperative power distribution network electrical variable anomaly detection method comprises the following steps: s1, integrating and arranging multisource operation and business information of a power distribution network, namely collecting, arranging and standardizing multisource information and data of the power distribution network on a cloud side, and constructing a data set for intelligent learning training; S2, electric variable running track extraction and polymorphic consistency analysis and modeling, namely constructing a rich information source electric variable track feature extractor Enhancing and fusing the dynamic change track of the electric variable in the multisource information and the key information signals in the step S1 to form the characteristics of the electric variable track of the rich information source, and then constructing a polymorphic consistency discriminatorEnsuring the accuracy and stability of the rich source electric variable track characteristics through consistency constraint; S3, track-abnormal mode density estimation based on Copula information theory, namely, constructing a correlation structure decoupler And outlier feature density estimatorModeling the association relation between the electric variable track in the rich information source electric variable track characteristic and the potential abnormal mode by using Copula information theory, fir