Multi-modal abnormal data detection and restoration method and system for power business scene

Abstract

The invention discloses a multi-modal abnormal data detection and restoration method and system oriented to a power business scene. The method comprises the following steps: collecting multi-source heterogeneous power data, abstracting a power system into a weighted undirected graph, uniformly mapping the multi-source heterogeneous data into a graph signal, and preprocessing the collected data; extracting spatial features of nodes in a topological structure by adopting a graph convolutional network, and capturing a time dependency relationship in combination with a time sequence encoder; identifying various types of data abnormal points through an abnormal scoring function fusing the time sequence prediction error and the neighborhood consistency; a prediction-reconstruction combined repair strategy is adopted, time sequence prediction and neighborhood diffusion estimation are fused, and a preliminary repair value is generated; a lightweight parameter adapter is introduced, a scene feature vector is used as input, a repair weight and a regularization coefficient are dynamically generated, and a repair strategy is automatically adjusted; and performing physical consistency verification on a data result, wherein the physical consistency verification comprises power injection conservation constraint, voltage amplitude range constraint and time sequence continuity constraint.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems and smart grid, more particularly, to a multi-modal abnormal data detection and repair method and system for power business scenarios. BACKGROUND

[0002] With the development of new power systems, the operation environment of power grids is becoming increasingly complex, and the scale and modal number of operation data are showing a trend of continuous growth. Multi-source data from dispatching centers, substation monitoring systems, user-side smart meters, and meteorological systems play an important role in supporting power dispatching optimization, equipment state evaluation, and demand response. However, due to factors such as sensor failure, communication delay, data packet loss, equipment abnormality, and human input errors, power system operation data generally have problems of redundancy, missing, error, and abnormal values, which directly affect the accuracy of power load forecasting, state estimation, and operation decision-making, and reduce the application value of data in dispatching optimization and equipment operation and maintenance.

[0003] Traditional threshold detection methods based on statistical models can quickly identify isolated abnormal points, but have limited effect when dealing with high-dimensional, strong time-dependent, and multi-source coupled power data. Abnormal detection methods based on probability distribution and curve fitting correction methods can improve data consistency in some scenarios, but their modeling ability depends on prior distribution assumptions, making it difficult to adapt to cross-scenario and multi-modal power business data.

[0004] In recent years, some technical solutions have attempted to introduce machine learning and deep learning methods:

[0005] For example, using stacked autoencoders, variational autoencoders, or generative adversarial networks to reconstruct missing data can improve the repair effect to some extent. However, such methods are often sensitive to hyperparameters, have high training costs, and lack cross-scenario transferability. When faced with large-scale continuous missing or complex abnormal types, the stability of the repair results is poor. In addition, these methods usually ignore the physical topology constraints and business logic constraints of the power network, which may lead the repair results to deviate from the system operation rules, thereby limiting their direct application in power businesses.

[0006] For example, Chinese patent CN116401516A, a power load data abnormality detection and repair method based on deep learning, uses a combination of convolutional neural networks and recurrent neural networks to identify and repair data abnormalities. Although it improves the detection accuracy of load data to some extent, it mainly targets single load curve data and does not fully consider the heterogeneity and cross-scenario adaptability of multi-modal power business data. Moreover, it lacks the introduction of power physical constraints, making it difficult to meet the actual needs of complex power businesses.

[0007] Therefore, how to overcome the problems that the existing power system data governance method has the problems of being difficult to unify modeling of multi-modal and multi-source heterogeneous data, poor repair effect of continuous anomaly and large-scale missing, insufficient model cross-scene adaptability, and lack of power physical topology and business logic constraints is a problem that needs to be solved by those skilled in the art. SUMMARY

[0008] Therefore, the present application provides a multi-modal abnormal data detection and repair method and system for power business scenarios to solve some of the technical problems mentioned in the background art.

[0009] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0010] A multi-modal abnormal data detection and repair method for power business scenarios, comprising the following steps:

[0011] S1. Collecting multi-source heterogeneous power data, abstracting the power system as a weighted undirected graph, uniformly mapping the multi-source heterogeneous data to graph signals, and preprocessing the collected data;

[0012] S2. Extracting the spatial features of nodes in the topology structure using a graph convolution network, and capturing the time dependence relationship in combination with a time encoding encoder; identifying various types of data abnormal points by fusing the abnormal scoring function of the time prediction error and the neighborhood consistency;

[0013] S3. For the detected abnormal points, a prediction-reconstruction joint repair strategy is adopted and the time prediction and neighborhood diffusion estimation are fused to generate a preliminary repair value;

[0014] S4. Introducing a lightweight parameter adapter, taking a scene feature vector as input, dynamically generating repair weights and regularization coefficients, and automatically adjusting the repair strategy according to the needs of the dispatching station, substation or user side;

[0015] S5. After the repair is completed, the data results are checked for physical consistency, including power injection conservation constraints, voltage amplitude range constraints and time series continuity constraints.

[0016] Preferably, in step S1, the sources of multi-source heterogeneous data include dispatching centers, SCADA / PMU, smart meters and external systems, the nodes of the weighted undirected graph correspond to substation buses, devices or user load points, the edges represent power transmission lines or business logic connections, and the edge weights reflect electrical parameters or correlations.

[0017] Preferably, the data preprocessing includes normalization, interpolation and filtering of the collected data, and Laplacian constraint processing of the graph.

[0018] Preferably, the specific content of step S2 is:

[0019] The spatial features extracted by the graph convolution network are:

[0020]

[0021] wherein, is a normalized adjacency matrix after adding a self-loop, used to ensure that the features propagate within the neighborhood and avoid numerical divergence; D is a graph degree matrix, and the diagonal elements represent the total connection strength of the ith node; H (l) is the output feature matrix of the lth layer of the graph convolution network, W (l) is the weight matrix of the lth layer of the graph convolution network, and sigma is an activation function, and I is an identity matrix;

[0022] The representation after graph convolution is input into a time encoding encoder:

[0023]

[0024] wherein, is the time sequence feature vector of the ith power grid node v within the time window T, H t (v) is the graph convolution output feature of the ith power grid node v at time t;

[0025] Based on the reconstruction error and the neighborhood consistency, an anomaly score function is defined as:

[0026]

[0027] wherein, A(v, t) is the anomaly score of the ith power grid node v at time t; alpha is a time dimension anomaly weight coefficient, gamma is a spatial dimension anomaly weight coefficient, is the power data prediction value of the ith power grid node v at time t, x t (v) is the actual monitoring value of the ith power grid node v at time t, L is a graph Laplacian matrix, X t is an observation matrix of the power system at time t.

[0028] Preferably, in step S3, the optimization objective of the prediction-reconstruction joint repair strategy is:

[0029]

[0030] wherein, is the power data prediction value of the ith power grid node v at time t, x t (v) is the actual monitoring value of the ith power grid node v at time t, L repair is a repair loss function, is a data smoothing constraint function after repair, lambda is a spatial smoothing regularization coefficient, is a physical consistency verification function of the repaired data, mu is a physical consistency regularization coefficient;

[0031] Introducing time series prediction and neighborhood diffusion estimation:

[0032]

[0033] in, For time series prediction results, η is the estimated value after neighborhood diffusion, and η is the fusion coefficient that is dynamically adjusted according to the business scenario.

[0034] Preferably, in step S4, the lightweight parameter adapter specifically refers to:

[0035] η=σ(W s s+b s )

[0036] λ = softplus(W λ s+b λ )

[0037] Where s is the scene feature vector.

[0038] Preferably, in step S5, the power injection conservation constraint is:

[0039]

[0040] in, P is the power injection amount at node v. v gen Let v be the power generation of node v. Let B be the load power of node v, N(v) be the set of neighboring nodes of node v, and B be the load power of node v. vu Let θ be the susceptance between the two nodes. v θ u These are the voltage phase angles at nodes v and u, respectively.

[0041] The voltage amplitude range constraint and timing continuity constraint are as follows:

[0042]

[0043] in, Let V be the magnitude of the repair voltage at node v at time t. and κ represents the physical safety limits for voltage, and κ is the smoothness threshold. This represents the repair data for node v in the f-th mode at time t.

[0044] A multimodal anomaly data detection and repair system for power business scenarios, based on the aforementioned multimodal anomaly data detection and repair method for power business scenarios, includes: a data acquisition module, a graph construction module, a preprocessing module, an anomaly detection module, a repair decision module, a parameter adaptation module, and a physical verification module;

[0045] The data acquisition module is used to collect multi-source heterogeneous power data;

[0046] The graph construction module is used to abstract the power system into a weighted undirected graph and to map multi-source heterogeneous data into graph signals.

[0047] The preprocessing module is used to preprocess the acquired multi-source heterogeneous power data;

[0048] The anomaly detection module is used to extract the spatial features of nodes in the topology using a graph convolutional network and to capture temporal dependencies using a temporal encoder; it identifies various data anomalies by fusing an anomaly scoring function that integrates temporal prediction error and neighborhood consistency.

[0049] The repair decision module is used to generate preliminary repair values ​​by adopting a prediction-reconstruction joint repair strategy for detected anomalies and fusing time-series prediction and neighborhood diffusion estimation.

[0050] The parameter adaptation module is used to introduce a lightweight parameter adapter, which takes the scene feature vector as input, dynamically generates repair weights and regularization coefficients, and automatically adjusts the repair strategy according to the needs of the dispatching, substation or user side.

[0051] The physical verification module is used to perform physical consistency verification on the data results after the repair is completed, including power injection conservation constraints, voltage amplitude range constraints, and timing continuity constraints.

[0052] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for detecting and repairing multimodal abnormal data in power business scenarios.

[0053] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the aforementioned method for multimodal abnormal data detection and repair in power business scenarios.

[0054] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method and system for multimodal anomaly data detection and repair in power business scenarios. By constructing a power business graph structure and integrating graph convolution and temporal modeling, it achieves anomaly detection of multimodal data, and completes data repair under a prediction-reconstruction mechanism combined with physical consistency constraints. This approach balances data integrity, accuracy, and cross-scenario adaptability, thereby providing high-quality data support for power system scheduling optimization, equipment status assessment, and intelligent operation and maintenance. Specifically:

[0055] This invention achieves unified and accurate modeling of multi-source heterogeneous power data: it abstracts the power system into a weighted undirected graph, maps multimodal data into graph signals, and combines graph convolution and time-series coding to capture spatiotemporal features. It can identify isolated anomalies as well as capture multi-node, multimodal linkage anomalies, and significantly improves the adaptability to high-dimensional, strongly coupled power data.

[0056] This invention improves the robustness and cross-scenario adaptability of anomaly repair: by using a prediction-reconstruction joint strategy, combining temporal prediction and neighborhood diffusion to ensure the spatiotemporal consistency of repair values; at the same time, it introduces a parameter adapter to dynamically adjust parameters according to scenario characteristics, which can match the needs of different scenarios such as scheduling, substation, and user side, and the cross-scenario migration capability is significantly better than the existing technology.

[0057] This invention ensures that the repair results conform to the physical operation law of the power grid: after the repair, a physical consistency check is added, and the data conforms to the operation law of the power grid through power injection conservation, voltage amplitude constraints, etc., which solves the problem that the repaired data in the prior art can be used but not used. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0059] Figure 1 This invention provides a schematic diagram of a multimodal anomaly data detection and repair method for power business scenarios;

[0060] Figure 2 A schematic diagram of the multimodal anomaly data detection method provided by the present invention.

[0061] Figure 3 This is a schematic diagram of the multimodal abnormal data repair method provided by the present invention;

[0062] Figure 4This invention provides a schematic diagram of a multimodal anomaly data detection and repair system for power business scenarios. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] This invention discloses a method for detecting and repairing multimodal abnormal data in power business scenarios, such as... Figures 1-3 As shown, it includes the following steps:

[0065] S1. Collect multi-source heterogeneous power data, abstract the power system into a weighted undirected graph, map the multi-source heterogeneous data into graph signals, and preprocess the collected data;

[0066] S2. A graph convolutional network is used to extract the spatial features of nodes in the topology, and a temporal encoder is used to capture temporal dependencies. By fusing the anomaly scoring function of temporal prediction error and neighborhood consistency, various data anomalies are identified.

[0067] S3. For the detected anomalies, a prediction-reconstruction joint repair strategy is adopted and time-series prediction and neighborhood diffusion estimation are fused to generate preliminary repair values;

[0068] S4. Introduce a lightweight parameter adapter that takes the scene feature vector as input, dynamically generates repair weights and regularization coefficients, and automatically adjusts the repair strategy according to the needs of the dispatching, substation or user side.

[0069] S5. After the repair is completed, perform physical consistency verification on the data results, including power injection conservation constraints, voltage amplitude range constraints, and timing continuity constraints.

[0070] To further implement the above technical solution, in step S1, the sources of multi-source heterogeneous data include dispatch center, SCADA / PMU, smart meters and external systems. The nodes of the weighted undirected graph correspond to the substation bus, equipment or user load points. The edges represent transmission lines or business logic connections, and the edge weights reflect electrical parameters or correlations.

[0071] To further implement the above technical solution, data preprocessing includes normalizing, interpolating, and filtering the collected data, as well as processing the graph Laplace constraint.

[0072] In this embodiment, the power system is modeled as a weighted undirected graph:

[0073] G = (V, E, A)

[0074] Where V is the set of power grid nodes, corresponding to actual power equipment such as power station buses and transformers; E is the set of power grid edges, corresponding to business logic associations such as transmission lines and distribution lines; A is the power grid adjacency matrix, representing the electrical coupling strength or the degree of business association between nodes;

[0075] At time t, the observation matrix of the entire system is expressed as:

[0076]

[0077] Where, x t (v i ) represents the i-th power grid node v i Multimodal observation includes information such as voltage, current, power, temperature, and equipment status; N is the number of grid nodes; M is the number of power monitoring modes, corresponding to power-specific monitoring dimensions such as voltage, current, active power, and reactive power;

[0078] To ensure data consistency and stability, the graph Laplace balance constraint is introduced as follows:

[0079]

[0080] L=DA

[0081] This formula is a power grid data smoothing constraint function, designed to solve the problem of local noise in power sensors. Its function is to ensure the consistency of electrical data between adjacent nodes in the topology, where L is the Laplace matrix of the power grid graph.

[0082] By using graph structure modeling and graph Laplace constraints, multi-source heterogeneous power data has been transformed into standardized graph signals that combine spatial topological correlation and data consistency. This not only solves the problem of traditional preprocessing ignoring the physical connection of the power grid, but also provides high-quality data input for the next step of anomaly detection based on spatial-temporal features.

[0083] Based on the standardized power grid graph signal obtained in step S1, the spatial coupling characteristics and temporal evolution patterns of the data are further explored. The correlation information of nodes in the topology is extracted through graph convolutional networks, and the periodicity and continuity of power data are captured by a time-series encoder, thereby constructing an anomaly judgment standard that takes into account both spatial consistency and temporal rationality.

[0084] To further implement the above technical solution, the specific content of step S2 is as follows:

[0085] The spatial features extracted using graph convolutional networks are:

[0086]

[0087] in, The normalized adjacency matrix after adding self-loops is used to ensure that features propagate within the neighborhood and avoid numerical divergence; D is the grid degree matrix, and the diagonal elements represent the sum of the connection strengths of the i-th node; H (l) W is the output feature matrix of the l-th layer of the graph convolutional network, used to represent the spatial features of each node after l layers of convolution; (l) σ is the weight matrix of the l-th layer of the graph convolutional network, which is learned through model training and is used to adjust the contribution of different input features to the output features; σ is the activation function, which is used to introduce nonlinearity into the model, simulate the nonlinear relationship in power data, and avoid the model only learning linear laws; I is the identity matrix, representing the relationship between each node and itself.

[0088] The convolved representation of the graph is then input into the timing encoder:

[0089]

[0090] in, H represents the time-series feature vector of the i-th grid node v within the time window T, such as the morning peak electricity consumption characteristics of a user's load. t (v) represents the graph convolution output feature of the i-th node at time t; the temporal encoder concatenates the spatial features at T times to form a spatiotemporal fusion feature sequence, providing a comprehensive basis for subsequent anomaly scoring;

[0091] Based on reconstruction error and neighborhood consistency, the anomaly scoring function is defined as follows:

[0092]

[0093] Where A(v,t) is the anomaly score of the i-th power grid node v at time t; α is the anomaly weight coefficient in the time dimension, and γ is the anomaly weight coefficient in the spatial dimension. Let x be the predicted power data value of the i-th grid node v at time t. t (v) represents the actual monitored value of the i-th grid node v at time t, and L is the graph Laplace matrix used to calculate the spatial consistency deviation of the data across the entire system; X t Let be the observation matrix of the power system at time t, representing the actual monitoring data of the entire system at time t;

[0094] When A(v,t) exceeds the threshold δ, node v is judged as abnormal at time t;

[0095] The first term of the anomaly scoring function reflects the difference between temporal prediction and observation, while the second term reflects the inconsistency between the node and its neighborhood. The combination of the two ensures that the detection results take into account both temporal patterns and spatial coupling characteristics.

[0096] In this embodiment, by using spatial-temporal dual-dimensional anomaly scoring, missing, redundant, and erroneous anomalies in power grid data are accurately identified, and anomaly classification is completed. This not only avoids the limitations of traditional detection that only focuses on a single dimension, but also provides clear repair targets and classification basis for the next step of precise repair for different anomaly types.

[0097] For the various anomalies identified in step S2, a repair strategy that fits the operation law of the power grid is designed. Time series prediction is used to ensure that the repair value conforms to the time evolution trend of power data. Neighborhood diffusion is combined to use the power grid topology correlation to correct anomalies. At the same time, physical constraints are embedded to ensure that the repair results do not deviate from the core law of the power system.

[0098] To further implement the above technical solution, the optimization objective of the prediction-reconstruction joint repair strategy in step S3 is:

[0099]

[0100] in, Let x be the predicted power data value of the i-th grid node v at time t. t (v) represents the actual monitored value of the i-th power grid node v at time t, L repair The repair loss function is used to measure the quality of the repair results. The goal of model training is to minimize the value of this function to ensure that the repaired data simultaneously meets the three requirements of being close to the true value, spatial smoothness, and physical compliance. The data smoothing constraint function after repair, and R smooth (X t The form is consistent, only the input data is replaced with the repaired full system matrix; λ is the spatial smoothing regularization coefficient, which takes a value of 0-1 and is used to adjust the importance of spatial smoothing constraints. This is the physical consistency verification function for the repaired data. μ is the physical consistency regularization coefficient, which ranges from 0 to 1 and is used to adjust the importance of physical constraints.

[0101] The first term of the formula ensures that the repaired value is consistent with the actual observation, the second term ensures the smoothness of the repaired data in the graph structure, and the third term is a physical consistency check term.

[0102] To improve robustness, time-series prediction and neighborhood diffusion estimation are introduced as follows:

[0103]

[0104] in, For time series prediction results, η is the estimated value after neighborhood diffusion, and η is the fusion coefficient that is dynamically adjusted according to the business scenario; it can both ensure the rationality of the time trend and use neighborhood relationships to repair anomalies.

[0105] Considering that the S3 repair strategy has insufficient adaptability of parameters such as fusion coefficient and regularization coefficient to different power scenarios, the dispatch scenario needs to prioritize regional power balance, and the user-side scenario needs to retain personalized power consumption characteristics. Therefore, a parameter dynamic generation mechanism based on scenario characteristics is designed to realize scenario-based adjustment of the repair strategy.

[0106] To further implement the above technical solution, in step S4, the lightweight parameter adapter specifically refers to:

[0107] η=σ(W s s+b s )

[0108] λ = softplus(W λ s+b λ )

[0109] Where s is the scene feature vector;

[0110] When the scene requires smoothness, the λ value increases, and the repair process pays more attention to neighborhood consistency; when the scene allows local differences, the λ value decreases, thus preserving the individual characteristics of the nodes.

[0111] After completing the scenario-based repair in step S4, the repair results need to be physically verified specifically for the power industry. Through core power grid constraints such as power conservation, voltage range, and time sequence continuity, repair data that does not conform to physical laws is eliminated to ensure that the final output data can reliably support power system decision-making.

[0112] To further implement the above technical solution, in step S5, the final repair result must satisfy the physical constraints of the power system, and the power injection conservation constraint is as follows:

[0113]

[0114] in, The amount of power injected into node v. Let v be the power generation of node v. Let B be the load power of node v, N(v) be the set of neighboring nodes of node v, and B be the load power of node v. vu Let θ be the susceptance between the two nodes. v θ u These are the voltage phase angles of nodes v and u, respectively; this formula ensures the power balance of the nodes and conforms to the basic principles of power flow calculation.

[0115] The voltage amplitude range constraint and timing continuity constraint are as follows:

[0116]

[0117] in, Let V be the magnitude of the repair voltage at node v at time t. and κ represents the physical safety limits for voltage, and κ is the smoothness threshold. This represents the repair data for node v in the f-th mode at time t;

[0118] The first two terms of the formula limit the voltage within a safe range, while the last term limits the smoothness of data between adjacent time points. These physical constraints of the power system ensure that the repaired data is not only numerically reasonable but also conforms to the operating rules of the power system.

[0119] By screening out high-quality repair data that fully conforms to the operation rules of the power grid through physical consistency verification, a closed loop of the entire process from data collection to anomaly detection, scenario-based repair and result verification is completed. It also provides reliable data support for power dispatch optimization, equipment status assessment and intelligent operation and maintenance, and achieves the core goal of multimodal anomaly data governance for power business scenarios.

[0120] A multimodal anomaly data detection and repair system for power business scenarios, based on a multimodal anomaly data detection and repair method for power business scenarios, such as... Figure 4 As shown, it includes: a data acquisition module, a graph construction module, a preprocessing module, an anomaly detection module, a repair decision module, a parameter adaptation module, and a physical verification module;

[0121] The data acquisition module is used to collect multi-source heterogeneous power data;

[0122] The graph construction module is used to abstract the power system into a weighted undirected graph and to map multi-source heterogeneous data into graph signals.

[0123] The preprocessing module is used to preprocess the acquired multi-source heterogeneous power data;

[0124] The anomaly detection module is used to extract the spatial features of nodes in the topology using a graph convolutional network and to capture temporal dependencies using a temporal encoder; it identifies various data anomalies by fusing an anomaly scoring function that integrates temporal prediction error and neighborhood consistency.

[0125] The repair decision module is used to generate preliminary repair values ​​by adopting a prediction-reconstruction joint repair strategy for detected anomalies and fusing time-series prediction and neighborhood diffusion estimation.

[0126] The parameter adaptation module is used to introduce a lightweight parameter adapter, which takes the scene feature vector as input, dynamically generates repair weights and regularization coefficients, and automatically adjusts the repair strategy according to the needs of the dispatching, substation or user side.

[0127] The physical verification module is used to perform physical consistency verification on the data results after the repair is completed, including power injection conservation constraints, voltage amplitude range constraints, and timing continuity constraints.

[0128] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for detecting and repairing multimodal abnormal data in power business scenarios.

[0129] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements a method for detecting and repairing multimodal abnormal data for power business scenarios.

[0130] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0131] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting and repairing multimodal anomaly data in power business scenarios, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous power data, abstract the power system into a weighted undirected graph, map the multi-source heterogeneous data into graph signals, and preprocess the collected data; S2. A graph convolutional network is used to extract the spatial features of nodes in the topology, and a temporal encoder is combined to capture temporal dependencies; By integrating the anomaly scoring function of temporal prediction error and neighborhood consistency, various types of data anomalies are identified; S3. For the detected anomalies, a prediction-reconstruction joint repair strategy is adopted and time-series prediction and neighborhood diffusion estimation are fused to generate preliminary repair values; S4. Introduce a lightweight parameter adapter that takes the scene feature vector as input, dynamically generates repair weights and regularization coefficients, and automatically adjusts the repair strategy according to the needs of the dispatching, substation or user side. S5. After the repair is completed, perform physical consistency verification on the data results, including power injection conservation constraints, voltage amplitude range constraints, and timing continuity constraints.

2. The method for multimodal anomaly data detection and repair in power business scenarios according to claim 1, characterized in that, In step S1, the sources of multi-source heterogeneous data include the dispatch center, SCADA / PMU, smart meters and external systems. The nodes of the weighted undirected graph correspond to the substation bus, equipment or user load points. The edges represent transmission lines or business logic connections, and the edge weights reflect electrical parameters or correlations.

3. The method for multimodal anomaly data detection and repair in power business scenarios according to claim 1, characterized in that, Data preprocessing includes normalizing, interpolating, and filtering the collected data, as well as processing graph Laplace constraints.

4. The method for multimodal anomaly data detection and repair in power business scenarios according to claim 1, characterized in that, The specific content of step S2 is as follows: The spatial features extracted using graph convolutional networks are: in, The normalized adjacency matrix after adding self-loops is used to ensure that features propagate within the neighborhood and avoid numerical divergence; D is the grid degree matrix, and the diagonal elements represent the sum of the connection strengths of the i-th node; H (l) Let W be the output feature matrix of the l-th layer of the graph convolutional network. (l) Let be the weight matrix of the l-th layer of the graph convolutional network, σ be the activation function, and I be the identity matrix; The convolved representation of the graph is then input into the timing encoder: in, Let H be the time-series feature vector of the i-th grid node v within the time window T. t (v) represents the graph convolution output feature of the i-th power grid node v at time t; Based on reconstruction error and neighborhood consistency, the anomaly scoring function is defined as follows: Where A(v,t) is the anomaly score of the i-th power grid node v at time t; α is the time-dimensional anomaly fusion coefficient; and γ is the spatial-dimensional anomaly weight coefficient. Let x be the predicted power data value of the i-th grid node v at time t. t (v) represents the actual monitored value of the i-th power grid node v at time t, L is the graph Laplace matrix, and X t Let be the observation matrix of the power system at time t.

5. The method for multimodal anomaly data detection and repair in power business scenarios according to claim 1, characterized in that, In step S3, the optimization objective of the prediction-reconstruction joint repair strategy is: in, Let x be the predicted power data value of the i-th grid node v at time t. t (v) represents the actual monitored value of the i-th power grid node v at time t, L repair To repair the loss function, Here, λ is the smoothing constraint function for the repaired data, and λ is the spatial smoothing regularization coefficient. Here, μ is the physical consistency verification function for the repaired data, and μ is the physical consistency regularization coefficient. Introducing time series prediction and neighborhood diffusion estimation: in, For time series prediction results, η is the estimated value after neighborhood diffusion, and η is the fusion coefficient that is dynamically adjusted according to the business scenario.

6. The method for multimodal anomaly data detection and repair in power business scenarios according to claim 1, characterized in that, In step S4, the lightweight parameter adapter specifically refers to: η=σ(W s s+b s ) λ=softplus(W λ s+b λ ) Where s is the scene feature vector.

7. The method for multimodal anomaly data detection and repair in power business scenarios according to claim 1, characterized in that, In step S5, the power injection conservation constraint is: in, The amount of power injected into node v. Let v be the power generation of node v. Let B be the load power of node v, N(v) be the set of neighboring nodes of node v, and B be the load power of node v. vu θ is the susceptance between the two nodes. v θ u These are the voltage phase angles at nodes v and u, respectively; The voltage amplitude range constraint and timing continuity constraint are as follows: in, Let V be the magnitude of the repair voltage at node v at time t. and κ represents the physical safety limits for voltage, and κ is the smoothness threshold. This represents the repair data for node v in the f-th mode at time t.

8. A multimodal anomaly data detection and repair system for power business scenarios, characterized in that, A method for detecting and repairing multimodal abnormal data in power business scenarios according to any one of claims 1-7 includes: a data acquisition module, a graph construction module, a preprocessing module, an anomaly detection module, a repair decision module, a parameter adaptation module, and a physical verification module. The data acquisition module is used to collect multi-source heterogeneous power data; The graph construction module is used to abstract the power system into a weighted undirected graph and to map multi-source heterogeneous data into graph signals. The preprocessing module is used to preprocess the acquired multi-source heterogeneous power data; The anomaly detection module is used to extract the spatial features of nodes in the topology using a graph convolutional network and to capture temporal dependencies using a temporal encoder; it identifies various data anomalies by fusing an anomaly scoring function that integrates temporal prediction error and neighborhood consistency. The repair decision module is used to generate preliminary repair values ​​by adopting a prediction-reconstruction joint repair strategy for detected anomalies and fusing time-series prediction and neighborhood diffusion estimation. The parameter adaptation module is used to introduce a lightweight parameter adapter, which takes the scene feature vector as input, dynamically generates repair weights and regularization coefficients, and automatically adjusts the repair strategy according to the needs of the dispatching, substation or user side. The physical verification module is used to perform physical consistency verification on the data results after the repair is completed, including power injection conservation constraints, voltage amplitude range constraints, and timing continuity constraints.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the multimodal anomaly data detection and repair method for power business scenarios as described in any one of claims 1-7.

10. A processing terminal, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the multimodal abnormal data detection and repair method for power business scenarios as described in any one of claims 1-7.

Images