A cloud-edge collaborative power distribution network electric variable anomaly detection method and device

Abstract

The application discloses a power distribution network electric variable anomaly detection method and device based on cloud-edge cooperation, belongs to the technical field of intelligent operation and maintenance and safety control of power distribution networks, and is used for guaranteeing the safe, stable and high-quality operation of a power supply system. Firstly, based on multi-source data of the power distribution network on the cloud side, robust electric variable trajectory features are learned through unsupervised deep feature extraction and multi-state consistency analysis. The complex nonlinear dependence relationship between the trajectory features and abnormal modes is modeled by using a Copula theory, and high-precision abnormal density estimation is realized. An interpretable and deliverable abnormal detection rule template library is generated through modal self-organizing clustering. On the edge side, the optimized lightweight model and the rule library are deployed on intelligent terminals, online and rapid abnormal detection and classification early warning of real-time electric variable data are realized, and the defects that the traditional method only relies on electric variable threshold judgment and lacks multi-source information fusion and behavior inference capability are overcome.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance and safety control technology of distribution networks, and in particular to a cloud-edge collaborative method and device for detecting abnormal power variables in distribution networks. Background Technology

[0002] With the continuous improvement of the intelligence level of distribution networks, anomaly detection of electrical variables has become a key technology to ensure the safe and stable operation of the power grid. Currently, research and practice in this field are mainly based on electrical measurement technology, which involves real-time acquisition and threshold judgment of electrical quantities such as voltage, current, and power to detect obvious faults such as overload and short circuits. However, most of these methods focus only on the electrical energy information itself, failing to effectively integrate multi-source information such as distribution automation systems, power grid models, and user application data. This results in a single detection dimension, making it difficult to cope with increasingly complex distribution network anomaly scenarios.

[0003] Existing detection methods have significant limitations. On the one hand, they rely heavily on "outcome-based" indicators such as threshold exceedances of electrical quantities or waveform distortion, lacking the ability to perceive and warn of the evolution of anomalies, essentially acting as a reactive mechanism. On the other hand, because they fail to correlate electrical variable data with information such as grid topology, business expansion application capacity, and historical operation and maintenance records, the detection system cannot explain the root cause of anomalies. For example, it cannot distinguish between malicious electricity use by users and current anomalies caused by natural aging of equipment, lacking behavioral inference and interpretability, which severely restricts the efficiency and accuracy of operation and maintenance.

[0004] In recent years, although some studies have attempted to introduce unsupervised learning algorithms to improve the intelligence level of detection, these methods are often limited to feature mining within electrical variable data or simply splicing together multi-source data. These methods struggle to deeply characterize the complex nonlinear dependencies between electrical variables and factors such as business operations and the environment, resulting in high false alarm and false negative rates when facing new scenarios such as batch connection applications and distributed energy grid connection. Therefore, there is an urgent need for an innovative detection method that can integrate multi-source information from electrical measurement and distribution systems, possessing behavioral-level perception and early warning capabilities, to overcome current technological bottlenecks and promote the proactive and intelligent development of distribution network operation and maintenance. Summary of the Invention

[0005] To address the aforementioned technical issues, this application proposes a cloud-edge collaborative method and device for detecting abnormal power variables in power distribution networks.

[0006] The technical solution adopted in this application is: a cloud-edge collaborative method for detecting abnormal power variables in a distribution network, comprising the following steps:

[0007] S1: Integration and organization of multi-source operation and business information of distribution network: collect, organize and standardize multi-source information and data of distribution network on the cloud side, and build datasets for intelligent learning and training;

[0008] S2: Extraction and Polymorphic Consistency Analysis and Modeling of Electrical Variable Trajectory: Constructing a Rich Source Electrical Variable Trajectory Feature Extractor The dynamic change trajectory of electrical variables and key information signals in the multi-source information in step S1 are enhanced and fused to form rich source electrical variable trajectory features; then a multi-mode consistency discriminator is constructed. Consistency constraints are used to ensure the accuracy and stability of the trajectory characteristics of the Fuxin source power variables;

[0009] S3: Trajectory-Anomaly Pattern Density Estimation Based on Copula Information Theory: Constructing a Correlation Structure Decoupling and anomaly feature density estimator This paper uses Copula information theory to model the correlation between electrical variable trajectories and potential anomaly patterns in rich information source electrical variable trajectory features. First, a Copula correlation structure decoupling is used. Copula decoupling is performed on the dependencies between features; then an anomaly feature density estimator is used. The coupling relationship between the outcome anomalies represented by the electrical variable trajectories and the systemic mode anomalies of equipment or power grid represented by the cloud-side rich information source data is extracted, and the rich information source anomaly mode estimation density is finally output.

[0010] S4: Construction of a template library for self-organizing and detecting abnormal modes of power distribution network variables: Building a self-organizing mechanism Using modal self-organizing machines The system automatically identifies and classifies similarity relationships formed based on rich information source anomaly mode estimates, and generates a configurable and deployable anomaly detection rule template library based on expert knowledge. ;

[0011] S5: Real-time Intelligent Detection and Early Warning of Electrical Variable Anomalies: Real-time online detection, classification, and early warning of electrical variable data anomalies at the edge.

[0012] Furthermore, the collection, organization, and standardization of multi-source information and data from the cloud-side distribution network in step S1 is achieved through a distribution network information integrator. The specific steps to achieve this include:

[0013] S1.1: Multi-source information and data collection: Collect heterogeneous data from different sources and with different sampling frequencies from cloud-side data centers;

[0014] S1.2: Data quality improvement: including handling outliers and repairing missing data values;

[0015] S1.3: Rich Source Data Archive Modeling Around Detection Points: All source data for the same detection point are statistically aggregated and standardized to form a single detection point archive, also known as a matrix training dataset. Dataset Each row represents a measurable detection point, and each column represents a source of information.

[0016] Furthermore, step S2 specifically includes the following steps:

[0017] S2.1: Rich Source Electricity Variable Trajectory Feature Extraction: Extracting the dataset Input to RichSource Electric Variable Trajectory Feature Extractor In this process, a low-dimensional rich source electrical variable trajectory feature set is obtained. Among them, the files for each testing site Mapping to features The formula is: ,in For Rich Source Electricity Variable Trajectory Feature Extractor The set of learnable parameters;

[0018] S2.2: Multi-state detection point signal restoration: Reconstructing the feature set of the rich source electrical variable trajectory Input to polymorphic consistency discriminator In this process, the restored pseudo-detection point archive dataset containing polymorphic and multi-source information is obtained. Each feature Reconstructing the fake detection point files The formula is: ,in For polymorphic consistency discriminators The set of learnable parameters;

[0019] S2.3: Polymorphic Consistency Calculation: Calculate the polymorphic consistency error between the spurious detection point file and the detection point file. Then, based on this error, the rich source electrical variable trajectory feature extractor is updated using the gradient descent method. Learnable parameter set Polymorphic consistency discriminator Learnable parameter set ,in This is the total number of testing sites;

[0020] S2.4: Rich Source Electric Variable Trajectory Feature Optimization Training: Calculate and repeat steps S2.1-S2.3 at least 200 times to obtain the optimized Rich Source Electric Variable Trajectory Feature Set. And a fully trained Rich Source Electric Variable Trajectory Feature Extractor Polymorphic consistency discriminator .

[0021] Furthermore, step S3 specifically includes the following steps:

[0022] S3.1: Decoupling of the correlation structure between electrical variable trajectory features and anomalous behavior: using the Copula correlation structure decoupling device. The Fuxin source electrical variable trajectory feature set obtained in step S2.4 Each element is transformed into an independent Copula space to obtain a decoupled enhanced feature set. ,and and The dimensions are consistent;

[0023] S3.2: Rich Source Anomalous Mode Density Estimation: Based on Enhanced Feature Sets Hefuxin Source Electric Variable Trajectory Feature Set Using anomaly feature density estimator Abstracting key behavioral patterns between electrical variable trajectories and anomalous behavior modes to form a characteristic of rich source electrical variable trajectories. Mapping to the density of anomalous modes of rich information sources.

[0024] Furthermore, step S3 specifically includes the following steps: Step 3.2 specifically includes the following steps:

[0025] S3.2.1: From the feature set of the source electrical variable trajectory of Fuxin b feature data points are randomly sampled without replacement from the enhanced feature set. Also extract the corresponding b enhanced feature data and feed these data into the anomaly feature density estimator. In the middle, density estimates were obtained respectively. and ,in It is an anomaly feature density estimator The parameter set, and They refer to the first Single feature data and enhanced feature data;

[0026] S3.2.2: Calculate the density loss of the trajectory-abnormal pattern: ,in The function is defined as the rich source density evolution function, using algebraic... The formula for example is expressed as: ;

[0027] S3.2.3: Calculate the gradient of the trajectory-anomaly pattern density loss using the gradient ascent method, and update the parameters. ;

[0028] S3.2.4: Assume the parameters obtained from S3.2.3 are as follows. , The label representing the current iteration number is based on the formula: Update parameters ,in For a single configurable parameter, These are the parameters from the last update;

[0029] S3.2.5: Repeat steps S3.2.1-S3.2.4 until... Second-rate, Finally, the optimal feature density estimator is obtained. ;

[0030] S3.2.6: Define the rich source density correction function In algebra The formula for example is expressed as: Calculate the trajectory characteristics of all rich source electrical variables. Rich source anomalous mode estimation density matrix ,at this time Incorporated into the optimal feature density estimator middle.

[0031] Furthermore, step S4 specifically includes the following steps:

[0032] S4.1: Estimation density matrix of rich information source anomalous modes obtained in step S3 According to the formula: Calculate two detection points and The similarity measure between them, where It is a single adjustable parameter. and This represents the modality estimation density corresponding to two detection points, and then a similarity matrix is ​​constructed based on this formula. ;

[0033] S4.2: Use spectral clustering algorithm to... Perform unsupervised clustering to obtain There are self-organizing clusters, of which the _ ... Each cluster is labeled as Steps S4.1 and S4.2 are performed by the modal self-organizing machine. accomplish;

[0034] S4.3: Based on the self-organized cluster information obtained from the clustering results, trace back to the multi-source data information of the corresponding detection points, label each self-organized cluster with its category number, and then combine expert knowledge to define the abnormal behavior category to which it belongs;

[0035] S4.4: For each self-organizing cluster, find the cluster center and solve the... Cluster center The formula is: ,in Represents the density matrix of the abnormal modes of rich information sources. China belongs to The modality estimation density corresponding to the detection points of each category, each cluster center is called an anomaly detection rule template, and all templates together form an anomaly detection rule template library. .

[0036] Furthermore, step S5 specifically includes the following steps:

[0037] S5.1: Input the real-time electrical variable flow data of the edge-side detection point to be detected, as well as the multi-source information related to the detection point obtained from the cloud side, into the distribution network information integrator. In the process, the processed data is obtained;

[0038] S5.2: Input the processed data into the Fuxin Source Electric Variable Trajectory Feature Extractor. The corresponding features are obtained from the middle;

[0039] S5.3: Input features into the anomaly feature density estimator In this process, the estimated density value is obtained;

[0040] S5.4: Use the obtained estimated density values ​​and the anomaly detection rule template library All templates are matched using the similarity metric in step S4.1. The category corresponding to the most similar template is selected to label the data and obtain the category corresponding to the abnormal behavior.

[0041] S5.5: Based on the category of abnormal behavior and in conjunction with the security priority handling mechanism configured at this detection point, take corresponding response and control actions, and upload complete early warning information to the cloud side.

[0042] Furthermore, the RichSource Electricity Variable Trajectory Feature Extractor It consists of a basic feature extraction network, a multi-source feature extraction network, and an electrical variable feature trajectory generation network stacked sequentially, wherein:

[0043] Basic feature extraction network: constructed based on ResNet101 network;

[0044] The multi-source feature extraction network comprises multiple parallel subnetworks, specifically:

[0045] The pressure-current mutual inductance signal enhancement network is composed of two fully connected groups linearly spliced ​​together, with the last activation function changed to an exponential function.

[0046] Load characteristic sensing network: It is composed of three fully connected groups linearly spliced ​​together, with the last activation function changed to the tanh function;

[0047] Capacity ratio evaluation network: consists of one graph convolutional group;

[0048] Power factor regression network: It is composed of two convolutional groups and one fully connected group, which are linearly concatenated, with the last activation function changed to the softmax function;

[0049] The power parameter integration and extraction network is composed of three fully connected groups linearly spliced ​​together.

[0050] Electric variable feature trajectory generation network: composed of 4 fully connected groups linearly spliced ​​together;

[0051] The polymorphic consistency discriminator C is composed of an electrical variable trajectory backtracking network, a multi-source feature reconstruction network, and a state mapping network stacked sequentially, wherein:

[0052] Electric variable trajectory backtracking network: composed of 4 fully connected linear splices;

[0053] The multi-source feature reconstruction network comprises multiple parallel subnetworks, specifically:

[0054] The pressure-current mutual inductance signal restoration network is composed of two fully connected groups linearly spliced ​​together, with the last activation function removed.

[0055] Load characteristic restoration network: It is composed of three fully connected groups linearly spliced ​​together, with the last activation function removed;

[0056] Power grid capacity margin estimation network: consisting of a single graph convolutional group with the final activation function removed;

[0057] Active power and apparent power generation network: It is composed of two convolutional groups and one fully connected group linearly concatenated, with the last activation function removed;

[0058] The power parameter restoration network is composed of three fully connected groups linearly spliced ​​together, with the last activation function removed.

[0059] Each of the fully connected groups contains one fully connected layer, one ReLU activation function, and one batch normalization layer; each of the convolutional groups contains one 1D convolutional layer, one pooling layer, and one Leaky ReLU activation function; and each of the graph convolutional groups consists of one graph convolutional layer and one ReLU activation function.

[0060] Furthermore, the anomaly feature density estimator It consists of the following layers in sequence:

[0061] (1) Econometric fusion layer, used to perform weighted summation and nonlinear transformation on all features from the input layer, contains 4 fully connected layers followed by the GELU activation function;

[0062] (2) Protection logic layer: Through logical judgment, the features are screened and strengthened to protect information that contributes highly to the density features and suppress redundant or noisy information. It includes three convolutional layers with channel attention mechanism and self-attention mechanism.

[0063] (3) Flow control layer. This layer is used to control the flow and distribution of features. It includes two sets of normalization and residual connection layers. Each set has three fully connected layers with interlayer residual connections. A layer normalization mechanism is added after the third layer.

[0064] (4) Rich source density estimation layer, which outputs high-level density features after processing by the previous layers, consists of a fully connected layer with a sigmoid activation function.

[0065] A cloud-edge collaborative distribution network power variable anomaly detection device includes a Fuxinyuan distribution network power variable analysis and modeling instrument deployed on the cloud side and an intelligent edge detection terminal deployed on the edge side. The computer programs / instructions involved in steps S1-S4 of the cloud-edge collaborative distribution network power variable anomaly detection method are deployed and run in the Fuxinyuan distribution network power variable analysis and modeling instrument, and the computer programs / instructions involved in step S5 of the cloud-edge collaborative distribution network power variable anomaly detection method are deployed and run in the intelligent edge detection terminal.

[0066] The advantages of this application over the prior art are as follows:

[0067] Firstly, at the technical level, it has achieved a leap from single electrical variable measurement to multi-source information fusion detection. By introducing multi-dimensional information such as power grid model, business expansion application capacity and historical operation and maintenance data, and conducting deep coupling analysis with real-time electrical variable data, it overcomes the limitations of traditional methods that rely solely on electrical quantity threshold judgment, and significantly improves the perception capability and early warning accuracy of systemic anomalies and hidden faults in the distribution network.

[0068] Secondly, in terms of methodology, it innovatively applies Copula theory to model the dependency relationship between electrical variable trajectories and business behaviors, which can analyze complex nonlinear statistical correlations. This not only enables the detection of abnormal results, but also allows for tracing and inferring the root causes of user behavior or equipment status changes behind the anomalies, greatly enhancing the interpretability of detection results and the pertinence of operation and maintenance decisions.

[0069] Finally, in terms of system architecture, relying on the cloud-edge collaboration mechanism, the powerful modeling and analysis capabilities of the cloud are combined with the real-time response advantages of the edge side. This not only supports deep learning and model optimization of massive historical data, but also meets the low latency requirements of real-time on-site detection, providing an efficient, reliable and scalable comprehensive solution for power distribution network anomaly detection. Attached Figure Description

[0070] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0071] Figure 1 This is a schematic diagram of the entire process of steps S1-S5 in the method provided in the embodiments of this application.

[0072] Figure 2 The physical configuration diagram of the Fuxinyuan power distribution network variable analysis and modeling instrument provided in the embodiments of this application is shown.

[0073] Figure 3 This is a physical configuration diagram of the intelligent edge detection terminal provided in the embodiments of this application.

[0074] Figure 4 The logical structure diagram of the Fuxinyuan power distribution network variable analysis and modeling instrument provided in the embodiments of this application is shown.

[0075] Figure 5 This is a logical configuration diagram of the intelligent edge detection terminal provided in the embodiments of this application.

[0076] Figure 6 This is a schematic diagram of the structure of the Rich Source Electricity Variable Trajectory Feature Extractor provided in the embodiments of this application.

[0077] Figure 7 This is a schematic diagram of the structure of the polymorphic consistency discriminator provided in the embodiments of this application.

[0078] Figure 8 This is a schematic diagram of the structure of the anomaly feature density estimator provided in the embodiments of this application.

[0079] In the diagram: 1 is the control and logic unit, 2 is the intelligent analysis and modeling unit, 3 is the data storage and management unit, 4 is the high-speed communication and interface unit, 5 is the first host housing, 6 is the heat dissipation and power supply system, 7 is the first communication interface, 8 is the rack mounting and fixed structure, 9 is the embedded control unit, 10 is the intelligent detection unit, 11 is the embedded data storage and caching unit, 12 is the embedded communication and interface unit, 13 is the second host housing, 14 is the power supply system, 15 is the second communication interface, and 16 is the wall-mounted mounting and fixed structure. Detailed Implementation

[0080] like Figures 1 to 8As shown, this application provides a cloud-edge collaborative method for detecting abnormal power variables in a distribution network. Based on an intelligent distribution network system, this application not only relies on traditional power equipment and power grid model systems such as distribution terminals (DTU / FTU), transformer monitoring terminals (TTU), and smart meters, but also deeply integrates IoT sensing, cloud computing, and big data technologies to construct dedicated sensing devices including distributed electrical quantity sensors, line status indicators, and power quality monitoring units, as well as a full-scale power grid data warehouse that supports the access and management of multi-source heterogeneous data, forming a software and hardware collaborative infrastructure covering the three layers of the distribution network: cloud, edge, and terminal.

[0081] This method deploys tasks such as anomaly rule mining and modeling in distribution network anomaly detection on the cloud side (distribution network central dispatch and control center), as the cloud side aggregates the most comprehensive data from the entire distribution network. This data originates from various detection points, including distribution terminals (DTU / FTU), transformer monitoring terminals, concentrators, line monitoring indicators, and smart meters, encompassing multi-source information such as electrical variable data, power quality data, grid model data, business expansion application data, and dispatch and control data, laying a solid foundation for building a practical and scalable model. On the other hand, this method deploys the completed anomaly detection model on the edge side (including distribution transformer terminals / transformer areas) to achieve real-time anomaly judgment and rapid response. This cloud-edge collaborative structure can focus on the functional objectives of different sides and maintain local autonomy during network outages.

[0082] Based on the above cloud-edge deployment rules for devices, the implementation steps of the method in this application are as follows:

[0083] S1: Integration and Organization of Multi-Source Operation and Business Information in the Distribution Network. This step, based on cloud-side data infrastructure, integrates and organizes multi-source information and data from the distribution network, organized at the level of physical monitoring points within the network, ultimately forming a dataset suitable for subsequent analysis. The specific tasks of this step are handled by the distribution network information integrator. accomplish.

[0084] S2: Extraction and Polymorphic Consistency Analysis and Modeling of Electrical Variable Trajectory. This step utilizes the RichSource electrical variable trajectory feature extractor. The dynamic change trajectory of electrical variables and key information signals such as business indicators of detection points in the multi-source information of the cloud-side distribution network obtained in the previous step are enhanced and fused to form rich-source electrical variable trajectory features. Then, a polymorphic (including cloud-side state, edge-side state, and terminal state) consistency discriminator is used. The consistency constraints ensure the accuracy and stability of the trajectory characteristics of the rich source electrical variables.

[0085] S3: Trajectory-Anomaly Pattern Density Estimation Based on Copula Information Theory. This step uses Copula information theory to model the correlation between the electrical variable trajectories and potential anomaly patterns in the features obtained in the previous step. First, a correlation structure decoupling is used. Copula decoupling is applied to the dependencies between features to obtain a strengthened pattern of independent relationships. Then, an anomaly feature density estimator is used. The coupling relationship between the outcome anomalies represented by electrical variable trajectories and the systemic mode anomalies of equipment or power grid represented by rich source data on the cloud side is extracted, and the rich source anomaly mode estimation density is finally output.

[0086] S4: Construction of a template library for self-organizing and detecting abnormal modes of power distribution network variables. This step uses a modal self-organizing machine. The system automatically identifies and classifies similarity relationships formed based on rich information source anomaly mode estimates, and generates a configurable and deployable anomaly detection rule template library based on expert knowledge. .

[0087] S5: Real-time intelligent detection and early warning of electrical variable anomalies. This step integrates the distribution network information integrator obtained from the preceding steps. Optimized RichSource Electricity Variable Trajectory Feature Extractor Optimized anomaly feature density estimator and anomaly detection rule template library It enables real-time online anomaly detection, classification, and early warning of electrical variable data at the edge.

[0088] The main objective of step S1 is based on the distribution network information integrator. The process involves collecting, organizing, and standardizing multi-source information and data from the cloud-based power distribution network to construct a dataset suitable for intelligent learning and training. The specific processing steps are as follows:

[0089] S1.1: Multi-source information and data collection. Collecting heterogeneous data from different sources and sampling frequencies from cloud-side data centers;

[0090] S1.2: Data quality improvement. For obvious acquisition or transmission errors (such as negative power, over-range current, etc.), the quantile method is used to truncate abnormal power consumption data or construction parameter values. For missing data values, Gaussian filling or interpolation methods are used to repair them based on similar data from nearby power grid monitoring points.

[0091] S1.3: Modeling the Rich Source Data Archive Around the Detection Point. This step uses statistical aggregation and standardization (implemented using the Z-score method) to form a single detection point archive, also known as a matrix training dataset, where each row represents a measurable detection point and each column represents a type of source information. Specifically, the resulting dataset can be represented as follows: The first in the dataset Each testing site's file can be represented as: ,in It refers to the number of information entries in the files at each testing point. This is the total number of testing sites. It is the set of real numbers.

[0092] Step S2 specifically includes the following steps:

[0093] S2.1: Rich Source Electricity Variable Trajectory Feature Extraction. The dataset... Input to RichSource Electric Variable Trajectory Feature Extractor In this process, a low-dimensional rich source electrical variable trajectory feature set is obtained. Among them, the files for each testing site Mapping to features The formula is: ,in For Rich Source Electricity Variable Trajectory Feature Extractor The set of learnable parameters;

[0094] S2.2: Multi-state detection point signal restoration. This involves restoring the feature set of the RichSignal source electrical variable trajectory. Input to polymorphic consistency discriminator In this process, the restored pseudo-detection point archive dataset containing polymorphic and multi-source information is obtained. Each feature Reconstructing the fake detection point files The formula is: ,in For polymorphic consistency discriminators The set of learnable parameters;

[0095] S2.3: Polymorphic Consistency Calculation. Calculate the polymorphic consistency error between the spurious detection point file and the detection point file: Then, based on this error, the rich source electrical variable trajectory feature extractor is updated using the gradient descent method. Learnable parameter set Polymorphic consistency discriminator Learnable parameter set ;

[0096] S2.4: Optimization training of Rich Source Electric Variable Trajectory Features. The optimal Rich Source Electric Variable Trajectory Feature Set can be obtained by repeatedly executing steps S2.1-S2.3 at least 200 times. And a fully trained Rich Source Electric Variable Trajectory Feature Extractor Polymorphic consistency discriminator .

[0097] The rich source electrical variable trajectory feature extractor B consists of a basic feature extraction network, a multi-source feature extraction network, and an electrical variable feature trajectory generation network stacked sequentially, as follows: Figure 6 As shown. The multi-source feature extraction network and the electrical variable feature trajectory generation network are composed of one or more basic network groups, specifically including the following groups: fully connected group (each group contains 1 fully connected layer, 1 ReLU activation function and batch normalization layer), convolution group (each group contains 1 1D convolutional layer, 1 pooling layer and 1 LeakyReLU activation function) and graph convolution group (each group consists of 1 graph convolutional layer and 1 ReLU activation function).

[0098] The following section details the basic components of each network in the RichSource Electric Variable Trajectory Feature Extractor:

[0099] (1) The basic feature extraction network is constructed based on the Resnet101 network.

[0100] (2) The multi-source feature extraction network contains multiple parallel sub-networks used to amplify and enhance the multi-source signals of each power grid. The specific construction schemes of each sub-network are as follows: the voltage-current mutual inductance signal enhancement network is composed of two fully connected groups linearly spliced ​​together, with the last activation function changed to an exponential function; the load characteristic sensing network is composed of three fully connected groups linearly spliced ​​together, with the last activation function changed to a tanh function; the capacity-to-load ratio evaluation network is composed of one graph convolution group; the power factor regression network is composed of two convolution groups and one fully connected group linearly spliced ​​together, with the last activation function changed to a softmax function; and the power consumption parameter integration extraction network is composed of three fully connected groups linearly spliced ​​together.

[0101] (3) Electric variable characteristic trajectory generation network, which is composed of 4 fully connected groups linearly spliced ​​together.

[0102] The polymorphic consistency discriminator C is specifically composed of an electrical variable trajectory backtracking network, a multi-source feature reconstruction network, and a state mapping network stacked sequentially, as follows: Figure 7 As shown.

[0103] The electrical variable trajectory backtracking network and the multi-source feature reconstruction network are composed of one or more basic network components, specifically including the following groups: fully connected group (each group contains 1 fully connected layer, 1 ReLU activation function and batch normalization layer), convolution group (each group contains 1 1D convolutional layer, 1 pooling layer and 1 LeakyReLU activation function) and graph convolution group (each group consists of 1 graph convolutional layer and 1 ReLU activation function).

[0104] The following section details the basic network components of a polymorphic consistency discriminator:

[0105] (1) Electric variable trajectory backtracking network, which is composed of 4 fully connected groups linearly spliced ​​together.

[0106] (2) The multi-source feature restoration network contains multiple parallel sub-networks used to restore multi-source signals from various power grids. The specific construction schemes of each sub-network are as follows: the voltage-current mutual inductance signal restoration network is composed of two fully connected groups linearly spliced ​​together, with the last activation function removed; the load characteristic restoration network is composed of three fully connected groups linearly spliced ​​together, with the last activation function removed; the power grid capacity margin estimation network is composed of one graph convolution group with the last activation function removed; the active power and apparent power generation network is composed of two convolution groups and one fully connected group linearly spliced ​​together, with the last activation function removed; the power consumption parameter restoration network is composed of three fully connected groups linearly spliced ​​together, with the last activation function removed.

[0107] (3) The state mapping network is constructed based on the Resnet101 network.

[0108] Step S3 specifically includes the following steps:

[0109] S3.1: Decoupling of electrical variable trajectory characteristics and anomalous behavior correlation structure. Using the Copula correlation structure decoupling tool. The feature set of Fuxin source electrical variable trajectory obtained in S2.4 Each element is transformed into an independent Copula space to obtain a decoupled enhanced feature set. ,and and The dimensions are consistent. Specifically, in Lieutenant General The Middle Features The Elements of each dimension Convert to element at the corresponding position The formula is: ,in For feature set size, This is a mathematical symbol / function that indicates whether the conditions in the curly braces are met. If they are met, the function is calculated; otherwise, it is not calculated.

[0110] S3.2: Rich Source Anomalous Mode Density Estimation. This step is based on the enhanced feature set. Hefuxin Source Electric Variable Trajectory Feature Set Using anomaly feature density estimator Abstracting key behavioral patterns between electrical variable trajectories and anomalous behavior modes to form a characteristic of rich source electrical variable trajectories. The mapping to the estimation density of anomalous modes in rich information sources is as follows:

[0111] S3.2.1: From the feature set of the source electrical variable trajectory of Fuxin b feature data points are randomly sampled without replacement from the enhanced feature set. Also extract the corresponding b enhanced feature data. Feed these data into the anomaly feature density estimator. In the middle, density estimates were obtained respectively. and ,in It is an anomaly feature density estimator The parameter set, and They refer to the first Data on individual features and enhanced features;

[0112] S3.2.2: Calculate the density loss of the trajectory-abnormal pattern: ,in The function is defined as the rich source density evolution function, using algebraic... The formula for example is expressed as: ;

[0113] S3.2.3: Calculate the gradient of the trajectory-anomaly pattern density loss using the gradient ascent method, and update the parameters. ;

[0114] S3.2.4: Assume the parameters obtained from S3.2.3 are as follows. ( (Representing the label of the current iteration number), based on the formula: Update parameters ,in For a single configurable parameter, For the parameters updated last time (if) If =1, this step will not be performed);

[0115] S3.2.5: Repeat steps S3.2.1-S3.2.4 until... Second-rate( Finally, the optimal feature density estimator is obtained. ;

[0116] S3.2.6: Define the rich source density correction function In algebra The formula for example is expressed as: Calculate the trajectory characteristics of all rich source electrical variables. Rich source anomalous mode estimation density matrix ,at this time Incorporated into the optimal feature density estimator middle.

[0117] Among them, the anomaly feature density estimator Structure such as Figure 8 As shown, it consists of the following layers in sequence:

[0118] (1) Econometric fusion layer, used to perform weighted summation and nonlinear transformation on all features from the input layer. It contains 4 fully connected layers followed by the GELU activation function.

[0119] (2) Protection logic layer, whose function is analogous to that of a relay protection device, filters and strengthens features through its "logic" judgment, protects information that contributes highly to density features, and suppresses redundant or noisy information. It contains three convolutional layers with channel attention mechanism and self-attention mechanism.

[0120] (3) Flow control layer, which is used to control the “flow” and “distribution” of features. It contains two sets of normalization and residual connection layers, each set has three fully connected layers with interlayer residual connections, and a layer normalization mechanism is added after the third layer.

[0121] (4) Rich source density estimation layer, which outputs the high-level density features after processing by the previous layers. It consists of a fully connected layer with a sigmoid activation function.

[0122] The specific steps for constructing the distribution network power variable anomaly mode self-organization and anomaly detection rule template library in step S4 are as follows:

[0123] S4.1: Estimation density matrix of rich information source anomalous modes obtained from step S3 According to the formula: Calculate two detection points and The similarity measure between them, where It is a single adjustable parameter. and This represents the modality estimation density corresponding to two detection points. A similarity matrix is ​​then constructed based on this formula. (is) (A three-dimensional matrix preserves the similarity between all detection points).

[0124] S4.2: Use spectral clustering algorithm to... Perform unsupervised clustering to obtain There are self-organizing clusters, of which the _ ... Each cluster is labeled as S4.1 and S4.2 are derived from modal self-organizing machines. accomplish;

[0125] S4.3: The self-organized cluster information obtained from the clustering results can be traced back to the multi-source data information of the corresponding detection points. Each self-organized cluster is labeled with its category number, and then combined with expert knowledge, the abnormal behavior category to which it belongs is defined.

[0126] S4.4: For each self-organizing cluster, find the cluster center and solve the... Cluster center The formula is: ,in Represents the density matrix of the abnormal modes of rich information sources. China belongs to The modality estimation density corresponding to the detection points of each category, each cluster center is called an anomaly detection rule template, and all templates together form an anomaly detection rule template library. .

[0127] Step S5, which implements real-time intelligent detection and early warning of electrical variable anomalies, consists of the following steps:

[0128] S5.1: Input the real-time electrical variable flow data of the edge-side detection point to be detected, as well as the multi-source information related to the detection point obtained from the cloud side, into the distribution network information integrator. In the process, the processed data is obtained;

[0129] S5.2: Input the processed data into the Fuxin Source Electric Variable Trajectory Feature Extractor. The corresponding features are obtained from the middle;

[0130] S5.3: Input features into the anomaly feature density estimator In this process, the estimated density value is obtained;

[0131] S5.4: Use the obtained estimated density values ​​and the anomaly detection rule template library All templates are matched using the similarity metric in step S4.1. The category corresponding to the most similar template is selected to label the data and obtain the category corresponding to the abnormal behavior.

[0132] S5.5: Based on the category of abnormal behavior and in conjunction with the security priority handling mechanism configured at this detection point, take corresponding response and control actions, and upload complete early warning information to the cloud side.

[0133] Based on the above description of the method, this application also proposes a cloud-edge collaborative distribution network power variable anomaly detection device, including a Fuxinyuan distribution network power variable analysis and modeling instrument deployed on the cloud side (cloud platform / dispatch center of smart grid, etc.) and an intelligent edge detection terminal deployed on the edge side (station side, user side, controllable load, etc.). The core of steps S1-S4 in the above method is to learn and condense deployable anomaly detection rules from multi-source data, and deploy and run the computer programs / instructions involved therein in the Fuxinyuan distribution network power variable analysis and modeling instrument. Step S5 is the working step of real-time intelligent detection and early warning of power variable anomalies at the detection point, and deploy the computer programs / instructions involved therein in the intelligent edge detection terminal.

[0134] The structure of the Fuxinyuan distribution network electrical variable analysis and modeling instrument is as follows: Figure 2 As shown, the structure of the intelligent edge detection terminal is as follows: Figure 3 As shown, aside from differences in physical selection and performance, the two devices are essentially identical in structure and relationship; the specific hardware configuration can be determined based on requirements. The logic diagrams of the two devices are as follows: Figure 4 and Figure 5 As shown, the computer instructions / programs related to the methods described above are labeled, thus providing a clear visual understanding of the specific method steps included in the device. Both can be described from both physical implementation and logical function perspectives:

[0135] The Fuxinyuan power distribution network variable analysis and modeling instrument physically consists of the following two parts:

[0136] (1) Internal functional units. These include: control and logic unit 1, intelligent analysis and modeling unit 2, data storage and management unit 3, and high-speed communication and interface unit 4.

[0137] (2) External packaging and accessories. Including: first main unit housing 5, heat dissipation and power supply system 6, first communication interface 7, rack mounting and fixing structure 8.

[0138] Logically, the Fuxinyuan distribution network power variable analysis and modeling instrument consists of a distribution network information integrator. Fuxin Source Electric Variable Trajectory Feature Extractor Polymorphic consistency discriminator Copula related structure decoupling Anomaly feature density estimator Modal self-organizing machine and anomaly detection rule template library These components are stored in the data storage and management unit 3 of the Fuxinyuan distribution network power variable analysis and modeling instrument in the form of program code.

[0139] The control and logic unit 1 includes a high-performance central processing unit for managing and driving the method described in this application; the intelligent analysis and modeling unit 2 includes two high-performance tensor processors and a large-capacity memory for executing steps S1-S4; and the data storage and management unit 3 includes a high-speed solid-state drive for storing and caching data and files from step S1, as well as an anomaly detection rule template library. The computer program / instructions of the method of this application; the high-speed communication and interface unit 4 includes a fiber optic interface for data interaction with the cloud-side data warehouse / data lake.

[0140] The first host casing 5 can be made of industrial-grade metal or engineering plastic, and has the characteristics of dustproof, moisture-proof and electromagnetic interference-proof; the heat dissipation and power supply system 6 includes liquid cooling radiator, air cooling radiator, heat pipe and power management unit; the first communication interface 7 includes RJ45 network port, fiber optic interface, USB interface and RS485 interface.

[0141] The intelligent edge detection terminal physically consists of the following two parts:

[0142] (1) Internal functional units. These include: embedded control unit 9, intelligent detection unit 10, embedded data storage and caching unit 11, and embedded communication and interface unit 12.

[0143] (2) External encapsulation and accessories. Including: second main unit housing 13, power system 14, second communication interface 15, wall-mounted installation and fixing structure 16.

[0144] Logically, the intelligent edge detection terminal is composed of a power distribution network information integrator. Optimized RichSource Electricity Variable Trajectory Feature Extractor Optimized anomaly feature density estimator These components are stored in the embedded data storage unit 11 of the intelligent edge detection terminal in the form of program code and configuration files.

[0145] The embedded control unit 9 includes an embedded central processing unit for low-power operation; the intelligent detection unit 10 includes a lightweight computing accelerator for executing step S5; the embedded data storage and caching unit 11 includes an embedded storage module for storing and caching the computer program / instructions related to the method in step S5; and the second communication interface 15 includes a wired communication interface and a wireless communication interface for enabling data interaction with the cloud side and the nearby edge side.

[0146] The second host casing 13 also adopts an industrial-grade metal casing or engineering plastic casing, which has the characteristics of dustproof, moisture-proof and electromagnetic interference-proof; the second communication interface 15 includes an RJ45 network port and a USB interface.

[0147] This application significantly improves the ability to detect hidden anomalies and early faults in complex power distribution network environments and enhances the accuracy of early warning, while ensuring low latency and high reliability of real-time response at the edge.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A cloud-edge collaborative method for detecting anomalies in power distribution network variables, characterized in that: Includes the following steps: S1: Integration and organization of multi-source operation and business information of distribution network: collect, organize and standardize multi-source information and data of distribution network on the cloud side, and build datasets for intelligent learning and training; S2: Extraction and Polymorphic Consistency Analysis and Modeling of Electrical Variable Trajectory: Constructing a Rich Source Electrical Variable Trajectory Feature Extractor The dynamic change trajectory of electrical variables and key information signals in the multi-source information in step S1 are enhanced and fused to form rich source electrical variable trajectory features; Then, a polymorphic consistency discriminant is constructed. Consistency constraints are used to ensure the accuracy and stability of the trajectory characteristics of the Fuxin source power variables; Step S2 specifically includes the following steps: S2.1: Rich Source Electricity Variable Trajectory Feature Extraction: Extracting the dataset Input to RichSource Electric Variable Trajectory Feature Extractor In this process, a low-dimensional rich source electrical variable trajectory feature set is obtained. Among them, the files for each testing site Mapping to features The formula is: ,in For Rich Source Electricity Variable Trajectory Feature Extractor The set of learnable parameters; S2.2: Multi-state detection point signal restoration: Reconstructing the feature set of the rich source electrical variable trajectory Input to polymorphic consistency discriminator In this process, the restored pseudo-detection point archive dataset containing polymorphic and multi-source information is obtained. Each feature Reconstructing the fake detection point files The formula is: ,in For polymorphic consistency discriminators The set of learnable parameters; S2.3: Polymorphic Consistency Calculation: Calculate the polymorphic consistency error between the spurious detection point file and the detection point file. Then, based on this error, the rich source electrical variable trajectory feature extractor is updated using the gradient descent method. Learnable parameter set Polymorphic consistency discriminator Learnable parameter set ,in This is the total number of testing sites; S2.4: Rich Source Electric Variable Trajectory Feature Optimization Training: Calculate and repeat steps S2.1-S2.3 at least 200 times to obtain the optimized Rich Source Electric Variable Trajectory Feature Set. And a fully trained Rich Source Electric Variable Trajectory Feature Extractor Polymorphic consistency discriminator ; S3: Trajectory-Anomaly Pattern Density Estimation Based on Copula Information Theory: Constructing a Correlation Structure Decoupling and anomaly feature density estimator This paper uses Copula information theory to model the correlation between electrical variable trajectories and potential anomaly patterns in rich information source electrical variable trajectory features. First, a Copula correlation structure decoupling is used. Copula decoupling is performed on the dependencies between features; then an anomaly feature density estimator is used. The coupling relationship between the outcome anomalies represented by the electrical variable trajectories and the systemic mode anomalies of equipment or power grid represented by the cloud-side rich information source data is extracted, and the rich information source anomaly mode estimation density is finally output. Step S3 specifically includes the following steps: S3.1: Decoupling of the correlation structure between electrical variable trajectory features and anomalous behavior: using the Copula correlation structure decoupling device. The Fuxin source electrical variable trajectory feature set obtained in step S2.4 Each element is transformed into an independent Copula space to obtain a decoupled enhanced feature set. ,and and The dimensions are consistent; S3.2: Rich Source Anomalous Mode Density Estimation: Based on Enhanced Feature Sets Hefuxin Source Electric Variable Trajectory Feature Set Using anomaly feature density estimator Abstracting key behavioral patterns between electrical variable trajectories and anomalous behavior modes to form a characteristic of rich source electrical variable trajectories. Mapping to the density of anomalous modes estimation in rich information sources; Step 3.2 specifically includes the following steps: S3.2.1: From the feature set of the source electrical variable trajectory of Fuxin b feature data points are randomly sampled without replacement from the enhanced feature set. Also extract the corresponding b enhanced feature data and feed these data into the anomaly feature density estimator. In the middle, density estimates were obtained respectively. and ,in It is an anomaly feature density estimator The parameter set, and They refer to the first Single feature data and enhanced feature data; S3.2.2: Calculate the density loss of the trajectory-abnormal pattern: ,in The function is defined as the rich source density evolution function, using algebraic... The formula for example is expressed as: ; S3.2.3: Calculate the gradient of the trajectory-anomaly pattern density loss using the gradient ascent method, and update the parameters. ; S3.2.4: Assume the parameters obtained from S3.2.3 are as follows. , The label representing the current iteration number is based on the formula: Update parameters ,in For a single configurable parameter, These are the parameters from the last update; S3.2.5: Repeat steps S3.2.1-S3.2.4 until... Second-rate, Finally, the optimal feature density estimator is obtained. ; S3.2.6: Define the rich source density correction function In algebra The formula for example is expressed as: Calculate the trajectory characteristics of all rich source electrical variables. Rich source anomalous mode estimation density matrix ,at this time Incorporated into the optimal feature density estimator middle; S4: Construction of a template library for self-organizing and detecting abnormal modes of power distribution network variables: Building a self-organizing mechanism Using modal self-organizing machines The system automatically identifies and classifies similarity relationships formed based on rich information source anomaly mode estimates, and generates a configurable and deployable anomaly detection rule template library based on expert knowledge. ; Step S4 specifically includes the following steps: S4.1: Estimation density matrix of rich information source anomalous modes obtained in step S3 According to the formula: Calculate two detection points and The similarity measure between them, where It is a single adjustable parameter. and This represents the modality estimation density corresponding to two detection points, and then a similarity matrix is ​​constructed based on this formula. ; S4.2: Use spectral clustering algorithm to... Perform unsupervised clustering to obtain There are self-organizing clusters, of which the _ ... Each cluster is labeled as Steps S4.1 and S4.2 are performed by the modal self-organizing machine. accomplish; S4.3: Based on the self-organized cluster information obtained from the clustering results, trace back to the multi-source data information of the corresponding detection points, label each self-organized cluster with its category number, and then combine expert knowledge to define the abnormal behavior category to which it belongs; S4.4: For each self-organizing cluster, find the cluster center and solve the... Cluster center The formula is: ,in Represents the density matrix of the abnormal modes of rich information sources. China belongs to The modality estimation density corresponding to the detection points of each category, each cluster center is called an anomaly detection rule template, and all templates together form an anomaly detection rule template library. ; S5: Real-time Intelligent Detection and Early Warning of Electrical Variable Anomalies: Real-time online detection, classification, and early warning of electrical variable data anomalies at the edge.

2. The cloud-edge collaborative distribution network power variable anomaly detection method according to claim 1, characterized in that: In step S1, the collection, organization, and standardization of multi-source information and data from the cloud-side distribution network are achieved through the distribution network information integrator. The specific steps to achieve this include: S1.1: Multi-source information and data collection: Collect heterogeneous data from different sources and with different sampling frequencies from cloud-side data centers; S1.2: Data quality improvement: including handling outliers and repairing missing data values; S1.3: Rich Source Data Archive Modeling Around Detection Points: All source data for the same detection point are statistically aggregated and standardized to form a single detection point archive, also known as a matrix training dataset. Dataset Each row represents a measurable detection point, and each column represents a source of information.

3. The cloud-edge collaborative distribution network power variable anomaly detection method according to claim 1, characterized in that: Step S5 specifically includes the following steps: S5.1: Input the real-time electrical variable flow data of the edge-side detection point to be detected, as well as the multi-source information related to the detection point obtained from the cloud side, into the distribution network information integrator. In the process, the processed data is obtained; S5.2: Input the processed data into the Fuxin Source Electric Variable Trajectory Feature Extractor. The corresponding features are obtained from the middle; S5.3: Input features into the anomaly feature density estimator In this process, the estimated density value is obtained; S5.4: Use the obtained estimated density values ​​and the anomaly detection rule template library All templates are matched using the similarity metric in step S4.

1. The category corresponding to the most similar template is selected to label the data and obtain the category corresponding to the abnormal behavior. S5.5: Based on the category of abnormal behavior and in conjunction with the security priority handling mechanism configured at this detection point, take corresponding response and control actions, and upload complete early warning information to the cloud side.

4. The cloud-edge collaborative distribution network power variable anomaly detection method according to claim 1, characterized in that: Fuxin Source Electric Variable Trajectory Feature Extractor It consists of a basic feature extraction network, a multi-source feature extraction network, and an electrical variable feature trajectory generation network stacked sequentially, wherein: Basic feature extraction network: constructed based on ResNet101 network; The multi-source feature extraction network comprises multiple parallel subnetworks, specifically: The pressure-current mutual inductance signal enhancement network is composed of two fully connected groups linearly spliced ​​together, with the last activation function changed to an exponential function. Load characteristic sensing network: It is composed of three fully connected groups linearly spliced ​​together, with the last activation function changed to the tanh function; Capacity ratio evaluation network: consists of one graph convolutional group; Power factor regression network: It is composed of two convolutional groups and one fully connected group, which are linearly concatenated, with the last activation function changed to the softmax function; The power parameter integration and extraction network is composed of three fully connected groups linearly spliced ​​together. Electric variable feature trajectory generation network: composed of 4 fully connected groups linearly spliced ​​together; The polymorphic consistency discriminator C is composed of an electrical variable trajectory backtracking network, a multi-source feature reconstruction network, and a state mapping network stacked sequentially, wherein: Electric variable trajectory backtracking network: composed of 4 fully connected linear splices; The multi-source feature reconstruction network comprises multiple parallel subnetworks, specifically: The pressure-current mutual inductance signal restoration network is composed of two fully connected groups linearly spliced ​​together, with the last activation function removed. Load characteristic restoration network: It is composed of three fully connected groups linearly spliced ​​together, with the last activation function removed; Power grid capacity margin estimation network: consisting of a single graph convolutional group with the final activation function removed; Active power and apparent power generation network: It is composed of two convolutional groups and one fully connected group linearly concatenated, with the last activation function removed; The power parameter restoration network is composed of three fully connected groups linearly spliced ​​together, with the last activation function removed. Each of the fully connected groups contains one fully connected layer, one ReLU activation function, and a batch normalization layer; each of the convolutional groups contains one 1D convolutional layer, one pooling layer, and one LeakyReLU activation function; and each of the graph convolutional groups consists of one graph convolutional layer and one ReLU activation function. The state mapping network is built on the Resnet101 network.

5. The cloud-edge collaborative distribution network power variable anomaly detection method according to claim 1, characterized in that: Anomaly feature density estimator It consists of the following layers in sequence: (1) Econometric fusion layer, used to perform weighted summation and nonlinear transformation on all features from the input layer, contains 4 fully connected layers followed by the GELU activation function; (2) Protection logic layer: Through logical judgment, the features are screened and strengthened to protect information that contributes highly to the density features and suppress redundant or noisy information. It includes three convolutional layers with channel attention mechanism and self-attention mechanism. (3) Flow control layer. This layer is used to control the flow and distribution of features. It includes two sets of normalization and residual connection layers. Each set has three fully connected layers with interlayer residual connections. A layer normalization mechanism is added after the third layer. (4) Rich source density estimation layer, which outputs high-level density features after processing by the previous layers, consists of a fully connected layer with a sigmoid activation function.

6. A cloud-edge collaborative distribution network power variable anomaly detection device, characterized in that: The method includes a Fuxinyuan distribution network power variable analysis and modeling instrument deployed on the cloud side and an intelligent edge detection terminal deployed on the edge side. The computer programs / instructions involved in steps S1-S4 of the cloud-edge collaborative distribution network power variable anomaly detection method as described in any one of claims 1-5 are deployed and run in the Fuxinyuan distribution network power variable analysis and modeling instrument, and the computer programs / instructions involved in step S5 of the cloud-edge collaborative distribution network power variable anomaly detection method as described in any one of claims 1-5 are deployed and run in the intelligent edge detection terminal.

Images