Intelligent power dispatching data network fault prediction method and related device

Abstract

The invention discloses an intelligent power dispatching data network fault prediction method and a related device, and the method comprises the steps: segmenting a physical interval according to a network topology structure, obtaining power data, constructing an FR-HIN full-relation heterogeneous information network, and converting the FR-HIN full-relation heterogeneous information network into a low-dimensional vector to extract potential features and attributes. And constructing a context association model by using the fault update record, and performing attenuation accumulation analysis in combination with update time to generate aggregated context features. Implicit context features are identified in combination with equipment configuration information, position features are extracted by using a conditional adversarial space-time encoder, data distribution similarity is evaluated through a discriminator, and real-time space-time anomaly indexes are obtained. And inputting the potential structure features, the implicit context features and the real-time space-time anomaly indexes into a power grid fault prediction model to obtain prediction results of fault types, positions and time. The accuracy of power grid fault prediction is further improved.

Description

Technical Field

[0001] The present invention relates to a field, and in particular to a fault prediction method and device for an intelligent power dispatching data network, and computing equipment. Background Art

[0002] The safe and stable operation of the power dispatching data network is directly related to the reliability and economy of power supply. With the continuous expansion of the scale of power systems and the continuous improvement of their intelligence, the complexity of the power dispatching data network is also increasing, and failures are difficult to avoid.

[0003] At present, the following technologies are mainly used for fault prediction of power dispatching data network:

[0004] Methods based on statistical analysis: By analyzing historical failure data, statistical models are established to predict future failures (such as time series analysis, regression analysis, etc.), but it is difficult to capture nonlinear relationships in complex systems and the prediction accuracy is limited.

[0005] Machine learning-based methods: Using methods such as support vector machines (SVM), neural networks (NN), and decision trees to learn the relationship between historical fault data and fault features. Although this method improves prediction accuracy to a certain extent, it often requires a large amount of labeled data, and the model has poor interpretability, making it difficult to apply to actual scenarios.

[0006] Methods based on expert knowledge and rules require a large amount of expert knowledge accumulation, and the rules are difficult to fully cover all possible fault conditions, making it difficult to cope with the complex and changeable operation status of the power grid.

[0007] The above methods often only focus on the numerical characteristics of power data, while ignoring other types of information such as network topology, equipment association, fault description text, etc., resulting in insufficient information utilization, insufficient feature expression ability, and low prediction accuracy.

[0008] To solve the above problems, the present invention proposes a fault prediction method for an intelligent power dispatching data network, which inputs potential structural features, implicit context features and real-time spatiotemporal anomaly indicators into a power grid fault prediction model to further improve the accuracy of power grid fault prediction. Summary of the invention

[0009] In view of the above problems, the present invention provides a fault prediction method and device for an intelligent power dispatching data network, and a computing device.

[0010] According to one aspect of the present invention, there is provided a method for predicting faults in an intelligent power dispatching data network, comprising:

[0011] According to the topological structure data of the power dispatching data network to be predicted, the physical interval of the power dispatching data network is divided into multiple sections, and the power data of each section is obtained;

[0012] A FR-HIN full-relation heterogeneous information network is constructed based on the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, section and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include the interaction relationship between nodes;

[0013] The FR-HIN full-relation heterogeneous information network is converted into a low-dimensional real-valued vector representation to obtain the potential structural features and attribute information of different nodes and relationships; a context association model is constructed based on multiple fault update records of power data to obtain context association features between fault descriptions and maintenance record texts; attenuation accumulation analysis is performed based on the context association features and multiple update times to generate multiple aggregate context feature sets; implicit equipment association identification is performed on the aggregate context feature set based on the equipment configuration information of the power dispatching data network to obtain multiple implicit context feature sets;

[0014] The conditional adversarial spatiotemporal encoder is used to extract the location features of the segment or equipment, and the similarity between the generated power data distribution and the real data distribution is evaluated by the discriminator to obtain a set of real-time spatiotemporal anomaly indicators; the similarity between the generated power data distribution and the real data distribution is evaluated by the discriminator to obtain a set of real-time spatiotemporal anomaly indicators;

[0015] The potential structural features and attribute information of different nodes and relationships, the implicit context feature set and the real-time spatiotemporal anomaly indicator set are input into the power grid fault prediction model to obtain the fault prediction results of the power dispatching data network, wherein the fault prediction results include the fault type, fault location and fault time.

[0016] In an optional manner, the node types of the FR-HIN full-relation heterogeneous information network include device nodes, segment nodes, and power data type nodes; wherein the devices include transformers, circuit breakers, and lines; the segment nodes are physical areas formed after the power dispatching data network is segmented; the power data types include voltage, current, and power;

[0017] The edge types of the FR-HIN full-relation heterogeneous information network include device-device relationship, device-segment relationship, segment-segment relationship, device-power data type relationship, and power data type-segment relationship.

[0018] In an optional manner, performing device implicit association identification on the aggregated context feature set to obtain multiple implicit context feature sets further includes:

[0019] Construct an association graph between devices based on device configuration information, where nodes represent devices and edges represent connection relationships between devices;

[0020] Encode the association graph to extract implicit association information between devices;

[0021] The extracted implicit correlation information is combined with the aggregated context feature set, and implicit context features related to fault prediction are selected to form multiple implicit context feature sets.

[0022] In an optional manner, performing device implicit association identification on the aggregated context feature set to obtain multiple implicit context feature sets further includes:

[0023] The similarity between the generated power data distribution and the real data distribution is evaluated by the discriminator;

[0024] The output of the encoder is used as a real-time spatiotemporal anomaly indicator, wherein the spatiotemporal anomaly indicator reflects the spatiotemporal anomaly state of each section or device in the electric power dispatching data network;

[0025] The conditional adversarial spatiotemporal encoder includes a time series encoder and a space encoder, wherein the time series encoder is used to extract the time series features of the power data, and the space encoder is used to extract the location features of the segment or the equipment.

[0026] In an optional manner, converting the FR-HIN full-relation heterogeneous information network into a low-dimensional real-valued vector representation further includes:

[0027] The nodes and relationships in FR-HIN are mapped to a low-dimensional vector space using matrix decomposition, where the optimization objective function of matrix decomposition is:

[0028]

[0029] Where A is the adjacency matrix of FR-HIN; z i and z j are the low-dimensional vector representations of nodes i and j respectively; σ is the activation function; λ is the regularization parameter.

[0030] In an optional manner, encoding the association graph to extract implicit association information between devices further includes:

[0031] The representation of each device is iteratively updated through a multi-layer graph convolutional network, where the update formula of the device in each layer is:

[0032]

[0033] in, is the representation of node i at layer l+1; is the neighbor set of node i; W (l) and b (l) are the weight matrix and bias vector of the lth layer respectively; σ is the activation function; is the representation of node j at layer l.

[0034] In an optional manner, the power grid fault prediction model includes an input layer, a second-order wavelet scattering network module, a multi-scale invariant feature extraction module, a multi-scale invariant feature fusion module, a multi-scale invariant feature pooling module, a high-level semantic feature network layer, and an output layer;

[0035] Wherein, the second-order wavelet scattering network module includes a 0th-order scattering propagation path, a 1st-order scattering propagation path and a 2nd-order scattering propagation path.

[0036] In an optional manner, the conditional adversarial spatiotemporal encoder includes a generator, a discriminator, and a set condition;

[0037] Among them, the generator includes a temporal encoder and a spatial encoder; the temporal encoder is composed of multiple CNN and LSTM network layers; the spatial decoder is composed of FlowNet, and the fused features are decoded by the decoder; the decoder is composed of convolutional layers and deconvolutional layers; the discriminator is composed of multiple layers of fully convolutional network layers.

[0038] According to another aspect of the present invention, there is provided a fault prediction device for an intelligent power dispatching data network, comprising:

[0039] A section segmentation module is used to segment the physical interval of the power dispatching data network into multiple sections according to the topological structure data of the power dispatching data network to be predicted, and obtain power data of each section;

[0040] A FR-HIN construction module is used to construct a FR-HIN full-relation heterogeneous information network based on the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, section and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include the interaction relationship between nodes;

[0041] A preprocessing module is used to convert the FR-HIN full-relation heterogeneous information network into a low-dimensional real-valued vector representation to obtain potential structural features and attribute information of different nodes and relationships; construct a context association model based on multiple fault update records of power data to obtain context association features between fault descriptions and maintenance record texts; perform attenuation accumulation analysis based on the context association features and multiple update times to generate multiple aggregate context feature sets; perform implicit equipment association identification on the aggregate context feature set based on the equipment configuration information of the power dispatching data network to obtain multiple implicit context feature sets;

[0042] a similarity calculation module, for extracting the location features of the segment or device using the conditional adversarial spatiotemporal encoder, evaluating the similarity between the generated power data distribution and the real data distribution through the discriminator, and obtaining the real-time spatiotemporal anomaly indicator set; and evaluating the similarity between the generated power data distribution and the real data distribution through the discriminator, and obtaining the real-time spatiotemporal anomaly indicator set;

[0043] The fault prediction module is used to input the potential structural characteristics and attribute information of different nodes and relationships, the implicit context feature set and the real-time spatiotemporal anomaly indicator set into the power grid fault prediction model to obtain the fault prediction results of the power dispatching data network, wherein the fault prediction results include the fault type, fault location and fault time.

[0044] According to another aspect of the present invention, there is provided a computing device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0045] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the above-mentioned intelligent power dispatching data network fault prediction method.

[0046] According to the solution provided by the present invention, based on the topological structure data of the power dispatching data network to be predicted, the physical interval of the power dispatching data network is divided into multiple sections, and the power data of each section is obtained; a FR-HIN full-relation heterogeneous information network is constructed based on the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, section and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include interactive relationships between nodes; the FR-HIN full-relation heterogeneous information network is converted into a low-dimensional real-valued vector representation to obtain potential structural features and attribute information of different nodes and relationships; a context association model is constructed based on multiple fault update records of the power data to obtain context association features between the fault description and the maintenance record text; and multiple update records are constructed based on the context association features. The invention uses the topological structure of the power dispatching data network to divide the network into multiple sections and construct a FR-HIN full-relation heterogeneous information network, which effectively integrates the complex associations between equipment, sections and power data types. By analyzing the fault update records, building a context association model, and combining time decay and equipment configuration information, we deeply explore the context information in the fault description and maintenance records and the implicit association between devices. Using the conditional adversarial spatiotemporal encoder, we effectively extract the location features of the section or equipment, and use the discriminator to detect spatiotemporal anomalies in power data in real time, providing timely warning information for fault prediction. The structural features, attribute information, implicit context features and real-time spatiotemporal anomaly indicator sets are input into the power grid fault prediction model to achieve multi-dimensional information fusion, thereby significantly improving the accuracy of fault prediction.

[0047] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0049] Figure 1 A schematic diagram showing a flow chart of a method for predicting faults in an intelligent power dispatching data network according to an embodiment of the present invention;

[0050] Figure 2 A schematic diagram of a power grid fault prediction model according to an embodiment of the present invention is shown;

[0051] Figure 3 A schematic diagram of a second-order wavelet scattering network module according to an embodiment of the present invention is shown;

[0052] Figure 4 A schematic diagram of a conditional adversarial spatiotemporal encoder according to an embodiment of the present invention is shown;

[0053] Figure 5 A schematic diagram showing a framework of a fault prediction device for an intelligent power dispatching data network according to an embodiment of the present invention;

[0054] Figure 6 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0055] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0056] Figure 1 FIG. 1 is a flow chart showing a method for predicting faults in an intelligent power dispatching data network according to an embodiment of the present invention. Specifically, Figure 1 As shown, the following steps are included:

[0057] Step S101, according to the topological structure data of the power dispatching data network to be predicted, the physical interval of the power dispatching data network is divided into multiple sections, and the power data of each section is obtained.

[0058] In this embodiment, the topology and power data of the power grid are effectively integrated by dividing the physical intervals of the power dispatching data network and obtaining the power data of each section. The entire power dispatching data network is divided into multiple sections, and the power data of each section is analyzed in more detail, which effectively reduces the complexity of data processing and can discover local anomalies and potential fault phenomena.

[0059] Specifically, the topological structure data of the power dispatching data network is obtained, including equipment location, line connection status, etc. The voltage, current, and power data of each section are collected, and the physical interval is divided into multiple sections according to the topological structure data of the power dispatching data network using the minimum cut algorithm, and the power data of each section is extracted from the entire power dispatching data network according to the divided sections.

[0060] Step S102, constructing a FR-HIN full-relation heterogeneous information network based on the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, sections and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include interactive relationships between nodes.

[0061] In this embodiment, by constructing a FR-HIN full-relation heterogeneous information network, the complex relationships between equipment, sections, and power data types are effectively integrated. For example, the power dispatching data network contains three transformers (VT1, VT2, VT3) and a substation (VS1), and the node types of the FR-HIN full-relation heterogeneous information network are defined, including equipment (such as transformers, substations), sections, and power data types (such as voltage, current, power). Define the interactive relationships between nodes, such as the power transmission relationship between equipment, the power flow relationship between sections, etc. Based on the collected data, nodes of equipment, sections, and power data types are established, and edges between nodes are established based on the interactive relationships between nodes. Equipment nodes: VT1, VT2, VT3, VS1; Segment nodes: Segment 1 (VT1 to VS1), Segment 2 (VT2 to VS1), Segment 3 (VT3 to VS1); Power data type nodes: voltage, current, power; Edge relationships: Segment 1 from VT1 to VS1, Segment 2 from VT2 to VS1, Segment 3 from VT3 to VS1, and the edges of voltage, current, and power are marked. The construction of the FR-HIN full-relation heterogeneous information network is as follows:

[0062] Device nodes: VT1, VT2, VT3, VS1;

[0063] Segment nodes: Segment 1, Segment 2, Segment 3;

[0064] Power data type nodes: voltage, current, power;

[0065] Edge relationship: VT1-section 1-voltage, VT1-section 1-current, VT1-section 1-power.

[0066] In an optional manner, the node types of the FR-HIN full-relation heterogeneous information network include device nodes, segment nodes, and power data type nodes; wherein the devices include transformers, circuit breakers, and lines; the segment nodes are physical areas formed after the power dispatching data network is segmented; the power data types include voltage, current, and power;

[0067] The edge types of the FR-HIN full-relation heterogeneous information network include device-device relationship, device-segment relationship, segment-segment relationship, device-power data type relationship, and power data type-segment relationship.

[0068] In this embodiment, for example, the power dispatching data network includes: equipment: transformer T1, circuit breaker CB1, line L1, line L2; section: Zone A, Zone B; power data: voltage V1, current I1, power P1. Among them, the nodes include:

[0069] Device nodes: T1, CB1, L1, L2

[0070] Zone nodes: Zone A, Zone B

[0071] Power data type nodes: V1, I1, P1

[0072] Edge relations include:

[0073] Equipment-Equipment:

[0074] T1-CB1 (Transformer connection circuit breaker)

[0075] CB1-L1 (circuit breaker connection line)

[0076] Equipment-Section:

[0077] T1-Zone A (transformer is located in Zone A)

[0078] CB1-Zone A (circuit breaker is located in Zone A)

[0079] L1-Zone A (Line L1 is located in Zone A)

[0080] L2-Zone B (Line L2 is located in Zone B)

[0081] Section-Section:

[0082] Zone A-Zone B (Zone A is adjacent to Zone B)

[0083] Equipment-Electricity Data Type:

[0084] T1-V1 (transformer voltage)

[0085] T1-I1 (transformer current)

[0086] T1-P1 (transformer power)

[0087] L1-V1 (voltage of line L1)

[0088] L1-I1 (current of line L1)

[0089] L1-P1 (power of line L1)

[0090] Power Data Type-Section:

[0091] V1-Zone A (voltage of Zone A)

[0092] I1-Zone A (Current in Zone A)

[0093] P1-Zone A (Power of Zone A)

[0094] V1-Zone B (voltage of Zone B)

[0095] I1-Zone B (Current in Zone B)

[0096] P1-Zone B (Power of Zone B)

[0097] The above nodes and edges are connected to form a FR-HIN network.

[0098] Step S103, converting the FR-HIN full-relation heterogeneous information network into a low-dimensional real-valued vector representation to obtain potential structural features and attribute information of different nodes and relationships; constructing a context association model based on multiple fault update records of power data to obtain context association features between fault descriptions and maintenance record texts; performing attenuation cumulative analysis on the context association features and multiple update times to generate multiple aggregated context feature sets; performing implicit equipment association identification on the aggregated context feature set based on the equipment configuration information of the power dispatching data network to obtain multiple implicit context feature sets.

[0099] In this embodiment, the potential structure of nodes and relationships in the learning network is represented by embedding, rather than relying solely on explicit connection relationships, which enables the discovery of deeper association patterns. The meaning of fault descriptions and maintenance records is deeply understood through context association models and text analysis. The introduction of a time decay mechanism reduces the impact of early fault records on current analysis, and pays more attention to recent fault information. The context features are associated with the device configuration information to discover potential implicit connections between devices, providing more valuable information for fault diagnosis and prediction. Compared with the black box model, the feature vectors obtained by analyzing each step help to understand the reasoning process of the model and increase the interpretability of international power system decisions.

[0100] Specifically, FR-HIN is embedded through the Metapath2Vec graph embedding algorithm to generate a low-dimensional real-valued vector representation.

[0101] The word vector model is obtained by training the power field data through the GloVe algorithm, and the contextual association features between texts are learned through the BERT model. For example, the fault description text and the corresponding maintenance record text are spliced, and then the contextual association features are extracted through the model.

[0102] According to the fault update time, the context-related features are weighted averaged or accumulated using a time decay function to generate multiple aggregated context feature sets of different time ranges. For example, the fault context information within the past 1 day, 3 days, and 7 days can be aggregated respectively.

[0103] Obtain the equipment configuration information (equipment connection relationship, equipment type, etc.) of the power dispatching data network, and use the equipment configuration information to perform association analysis on the aggregated context feature set. For example, aggregate the context features of the equipment on the same line, or calculate the similarity of context features between different equipment to discover the implicit connection between the equipment. Generate a new implicit context feature set based on the results of the association analysis.

[0104] For example, device information includes:

[0105] Equipment A (transformer)-Line 1

[0106] Device B (switch) - Line 1

[0107] Device C (line)-belongs to line 2

[0108] The fault record is as follows:

[0109] 2023-10-25 10:00: Device A overheated, maintenance record: "Replace heat sink".

[0110] 2023-10-2612:00: Equipment B has insulation failure, maintenance record: "Replace insulation material".

[0111] 2023-10-28 14:00: Device A overheated again, maintenance record: "Check the cooling system".

[0112] The construction includes nodes such as "Device A", "Device B", and "Device C", as well as fault and maintenance record text nodes such as "Overheating Fault", "Insulation Fault", "Replace Heat Sink", "Replace Insulation Material", and "Check Cooling System". Relationships include: "Device A has an overheating fault", and "The maintenance record corresponding to the overheating fault is to replace the heat sink".

[0113] All nodes and relationships in FR-HIN are converted into low-dimensional real-valued vectors, and the context association model is trained to learn the association vectors of the texts "Device A overheating failure" and "Replace heat sink".

[0114] With a decay period of 1 day, when analyzing on 2023-10-28, the context feature of device A is the weighted sum of the decayed context vector of the fault record on 2023-10-25 and the context vector of the fault record on 2023-10-28. The context feature of device B only contains the decay result of the fault record vector on 2023-10-26.

[0115] Device A and device B both belong to line 1. The context feature vectors of device A and device B are aggregated to obtain the overall context feature vector of line 1.

[0116] By analyzing the aggregated contextual features and implicit association features, we can predict the equipment that may fail in the future, and prepare the corresponding maintenance materials in advance based on the historical maintenance records. For example, during the 10-28 analysis, it was found that equipment A had overheated several times in a short period of time and that line 1 had insulation failures in equipment B. It was inferred that line 1 might have more serious problems in the future. Based on the fault description and maintenance records, combined with implicit association features, the cause of the fault can be quickly located. For example, if multiple devices on the same line are found to have failed, it is necessary to focus on checking whether there is a problem with the line.

[0117] In an optional manner, performing device implicit association identification on the aggregated context feature set to obtain multiple implicit context feature sets further includes:

[0118] Construct an association graph between devices based on device configuration information, where nodes represent devices and edges represent connection relationships between devices;

[0119] Encode the association graph to extract implicit association information between devices;

[0120] The extracted implicit correlation information is combined with the aggregated context feature set, and implicit context features related to fault prediction are selected to form multiple implicit context feature sets.

[0121] In this embodiment, it is further clarified how to use device configuration information to mine implicit associations between devices and combine it with aggregated context features, so as to explicitly model the connection relationship and topological structure between devices, better understand the physical associations between devices, and more accurately select the implicit context features that are most critical for fault prediction, thereby better predicting power system faults.

[0122] In an optional manner, converting the FR-HIN full-relation heterogeneous information network into a low-dimensional real-valued vector representation further includes:

[0123] The nodes and relationships in FR-HIN are mapped to a low-dimensional vector space using matrix decomposition, where the optimization objective function of matrix decomposition is:

[0124]

[0125] Where A is the adjacency matrix of FR-HIN; z i and z j are the low-dimensional vector representations of nodes i and j respectively; σ is the activation function; λ is the regularization parameter.

[0126] In this embodiment, the matrix decomposition method is used to map the nodes and relationships in FR-HIN to a low-dimensional vector space. In essence, the structural information of FR-HIN is represented by an adjacency matrix, and then the low-dimensional vector representation of the nodes and relationships is learned through matrix decomposition. This can help understand the similarity between nodes and the strength of relationships, and has a certain degree of interpretability.

[0127] Specifically, all nodes in FR-HIN are assigned indexes, and all relationship types in FR-HIN are assigned indexes. Adjacency matrices of different relationship types are generated based on the node index and relationship index. If there is a certain relationship between node i and node j, the corresponding value is set in the corresponding adjacency matrix, such as using a binary value (0 / 1) to indicate whether there is a relationship, or using the relationship strength as the relationship value.

[0128] The adjacency matrices of different relationship types are concatenated or combined into a unified adjacency matrix, and the low-dimensional vectors of the initialized nodes are represented as random values.

[0129] The node vector representation is updated iteratively through stochastic gradient descent until the objective function converges or reaches a predetermined number of iterations. After the iterative process of matrix decomposition, a low-dimensional vector representation of each node is obtained. For example, the vector z of device A 1 , vector z of device B 2 , vector z of fault 4 4 wait.

[0130] In an optional manner, encoding the association graph to extract implicit association information between devices further includes:

[0131] The representation of each device is iteratively updated through a multi-layer graph convolutional network, where the update formula of the device in each layer is:

[0132]

[0133] in, is the representation of node i at layer l+1; is the neighbor set of node i; W (l) and b (l) are the weight matrix and bias vector of the lth layer respectively; σ is the activation function; is the representation of node j at layer l.

[0134] In this embodiment, traditional machine learning methods usually treat data as independent individuals and ignore the association between data. The multi-layer graph convolutional network iteratively updates the node representation so that the representation of each node not only contains its own information, but also incorporates the information of neighboring nodes. After multiple layers of iteration, the node can aggregate the information of more distant neighbors, thereby learning richer global context information. The multi-layer graph convolutional network operates directly on the graph structure, utilizes the connection relationship between nodes (devices), captures the mutual dependence and influence between devices, and more effectively extracts hidden related information.

[0135] Step S104, using the conditional adversarial spatiotemporal encoder to extract the location features of the segment or equipment, evaluating the similarity between the generated power data distribution and the real data distribution through the discriminator, and obtaining a real-time spatiotemporal anomaly indicator set; evaluating the similarity between the generated power data distribution and the real data distribution through the discriminator, and obtaining a real-time spatiotemporal anomaly indicator set.

[0136] In this embodiment, Figure 4 As shown in the figure, the conditional adversarial network (Conditional GAN) introduces location information as a condition, so that the generator can generate power data at a specific location, and the model can learn the specific pattern of each section or device, which helps to detect anomalies at that location. The discriminator evaluates the similarity between the generated data and the real data to achieve unsupervised anomaly detection, which effectively copes with the situation where abnormal samples are scarce or missing in actual scenarios, and can obtain abnormal indicators in real time to achieve rapid monitoring of the operating status of the power system.

[0137] In an optional manner, the evaluating the similarity between the generated power data distribution and the real data distribution by the discriminator further comprises:

[0138] The similarity between the generated power data distribution and the real data distribution is evaluated by the discriminator;

[0139] The output of the encoder is used as a real-time spatiotemporal anomaly indicator, wherein the spatiotemporal anomaly indicator reflects the spatiotemporal anomaly state of each section or device in the electric power dispatching data network;

[0140] The conditional adversarial spatiotemporal encoder includes a time series encoder and a space encoder, wherein the time series encoder is used to extract the time series features of the power data, and the space encoder is used to extract the location features of the segment or the equipment.

[0141] In this embodiment, the time encoder and the space encoder are separated, the time encoder learns the time dependency of the power data, and the space encoder learns the location characteristics of the device.

[0142] In an optional manner, the conditional adversarial spatiotemporal encoder includes a generator, a discriminator, and a set condition;

[0143] Among them, the generator includes a temporal encoder and a spatial encoder; the temporal encoder is composed of multiple CNN and LSTM network layers; the spatial decoder is composed of FlowNet, and the fused features are decoded by the decoder; the decoder is composed of convolutional layers and deconvolutional layers; the discriminator is composed of multiple layers of fully convolutional network layers.

[0144] In this embodiment, the conditional adversarial network not only generates data, but also generates data that meets the conditions according to the set conditions. Among them, although FlowNet was originally used for optical flow estimation, the present invention applies it to power fault prediction, and regards the changes in spatially adjacent sensor data as "movement" to capture the propagation mode of faults in the power grid. Instead of directly predicting future current, voltage, temperature, etc., it predicts the "flow" (changing trend) of current, voltage, and temperature in space to assist in predicting future fault conditions.

[0145] Step S105, input the potential structural features and attribute information of different nodes and relationships, the implicit context feature set and the real-time spatiotemporal anomaly indicator set into the power grid fault prediction model to obtain the fault prediction results of the power dispatching data network, wherein the fault prediction results include the fault type, fault location and fault time.

[0146] In this embodiment, the power grid fault prediction model takes into account the structure, attributes, context and real-time indicators, and more accurately predicts the type, location and time of the fault, finds early signs of the fault earlier, and provides more warning time. It can give the potential cause of the fault, improve the interpretability of the model, and help operation and maintenance personnel understand the operation status of the power grid.

[0147] In an optional manner, the power grid fault prediction model includes an input layer, a second-order wavelet scattering network module, a multi-scale invariant feature extraction module, a multi-scale invariant feature fusion module, a multi-scale invariant feature pooling module, a high-level semantic feature network layer, and an output layer;

[0148] Wherein, the second-order wavelet scattering network module includes a 0th-order scattering propagation path, a 1st-order scattering propagation path and a 2nd-order scattering propagation path.

[0149] In this embodiment, Figure 2 , Figure 3 As shown in the figure, the second-order wavelet scattering network extracts features of different scales from the power grid data. The 0th order captures the overall trend of the data, the first order captures the mid-frequency details, and the second order captures the high-frequency details. By cooperating with the wavelet scattering network, it has a certain invariance to the slight translation and rotation of the input signal, and further extracts the invariant features at different scales. After multi-scale feature fusion and pooling, the high-level semantic feature network layer further extracts abstract features, which improves the discriminative ability of the features to capture key information related to the fault.

[0150] Specifically, three scattering propagation paths are constructed according to the Daubechies wavelet basis function, wherein the 0th order scattering propagation path directly convolves the input data to output low-frequency components. The 1st order scattering propagation path convolves and nonlinearly activates the input data, and then performs wavelet transform to output intermediate-frequency components. The 2nd order scattering propagation path convolves and nonlinearly activates the output of the 1st order, and then performs wavelet transform to output high-frequency components.

[0151] The feature maps output by the three paths are connected together as the output of the second-order wavelet scattering network module.

[0152] The multi-scale invariant feature extraction module uses methods such as maximum pooling or average pooling to extract invariant features at different scales. The output of the second-order wavelet scattering network is pooled using pooling windows of different sizes to form a multi-scale feature branch.

[0153] The multi-scale invariant feature fusion module splices the feature maps output by feature branches of different scales together. The spliced ​​features are fused using convolutional layers or fully connected layers to form a unified multi-scale invariant feature representation.

[0154] The multi-scale invariant feature pooling module uses global average pooling or global maximum pooling to convert the feature map into a fixed-length feature vector.

[0155] The high-level semantic feature network layer extracts spatial features through convolutional neural networks (CNN), processes temporal features through recurrent neural networks (RNN), and captures long-distance dependencies through the Transformer model. It uses linear activation functions or Sigmoid activation functions to output the predicted fault risk value.

[0156] According to the solution provided by the present invention, based on the topological structure data of the power dispatching data network to be predicted, the physical interval of the power dispatching data network is divided into multiple sections, and the power data of each section is obtained; a FR-HIN full-relation heterogeneous information network is constructed based on the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, section and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include interactive relationships between nodes; the FR-HIN full-relation heterogeneous information network is converted into a low-dimensional real-valued vector representation to obtain potential structural features and attribute information of different nodes and relationships; a context association model is constructed based on multiple fault update records of the power data to obtain context association features between the fault description and the maintenance record text; and multiple update records are constructed based on the context association features. The invention uses the topological structure of the power dispatching data network to divide the network into multiple sections and construct a FR-HIN full-relation heterogeneous information network, which effectively integrates the complex associations between equipment, sections and power data types. By analyzing the fault update records, building a context association model, and combining time decay and equipment configuration information, we deeply explore the context information in the fault description and maintenance records and the implicit association between devices. Using the conditional adversarial spatiotemporal encoder, we effectively extract the location features of the section or equipment, and use the discriminator to detect spatiotemporal anomalies in power data in real time, providing timely warning information for fault prediction. The structural features, attribute information, implicit context features and real-time spatiotemporal anomaly indicator sets are input into the power grid fault prediction model to achieve multi-dimensional information fusion, thereby significantly improving the accuracy of fault prediction.

[0157] Figure 5The schematic diagram of the framework of the intelligent power dispatching data network fault prediction device according to the embodiment of the present invention is shown. The intelligent power dispatching data network fault prediction device comprises:

[0158] The section segmentation module 510 is used to segment the physical interval of the power dispatching data network into multiple sections according to the topological structure data of the power dispatching data network to be predicted, and obtain the power data of each section;

[0159] A FR-HIN construction module 520, configured to construct a FR-HIN full-relation heterogeneous information network according to the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, section and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include interactive relationships between nodes;

[0160] The preprocessing module 530 is used to convert the FR-HIN full-relation heterogeneous information network into a low-dimensional real-valued vector representation to obtain potential structural features and attribute information of different nodes and relationships; construct a context association model based on multiple fault update records of power data to obtain context association features between fault descriptions and maintenance record texts; perform attenuation accumulation analysis based on the context association features and multiple update times to generate multiple aggregate context feature sets; perform implicit equipment association identification on the aggregate context feature set based on the equipment configuration information of the power dispatching data network to obtain multiple implicit context feature sets;

[0161] A similarity calculation module 540 is used to extract the location features of the segment or device using the conditional adversarial spatiotemporal encoder, evaluate the similarity between the generated power data distribution and the real data distribution through the discriminator, and obtain the real-time spatiotemporal anomaly indicator set; evaluate the similarity between the generated power data distribution and the real data distribution through the discriminator, and obtain the real-time spatiotemporal anomaly indicator set;

[0162] The fault prediction module 550 is used to input the potential structural characteristics and attribute information of different nodes and relationships, the implicit context feature set and the real-time spatiotemporal anomaly indicator set into the power grid fault prediction model to obtain the fault prediction results of the power dispatching data network, wherein the fault prediction results include the fault type, fault location and fault time.

[0163] Figure 6 The schematic diagram of the structure of the computing device embodiment of the present invention is shown. The specific embodiment of the present invention does not limit the specific implementation of the computing device.

[0164] like Figure 6As shown, the computing device may include: a processor (processor) 602 , a communications interface (Communications Interface) 64 , a memory (memory) 606 , and a communication bus 608 .

[0165] The processor 602, the communication interface 604, and the memory 606 communicate with each other via the communication bus 608. The communication interface 604 is used to communicate with other devices such as a client or other server network elements. The processor 602 is used to execute the program 610, which can specifically execute the relevant steps in the above-mentioned intelligent power dispatching data network fault prediction method embodiment.

[0166] Specifically, the program 610 may include program codes, which include computer operation instructions.

[0167] The processor 602 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention. The one or more processors included in the computing device may be processors of the same type, such as one or more CPUs; or may be processors of different types, such as one or more CPUs and one or more ASICs.

[0168] The memory 606 is used to store the program 610. The memory 606 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0169] According to the solution provided by the present invention, based on the topological structure data of the power dispatching data network to be predicted, the physical interval of the power dispatching data network is divided into multiple sections, and the power data of each section is obtained; a FR-HIN full-relation heterogeneous information network is constructed based on the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, section and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include interactive relationships between nodes; the FR-HIN full-relation heterogeneous information network is converted into a low-dimensional real-valued vector representation to obtain potential structural features and attribute information of different nodes and relationships; a context association model is constructed based on multiple fault update records of the power data to obtain context association features between the fault description and the maintenance record text; and multiple update records are constructed based on the context association features. The invention uses the topological structure of the power dispatching data network to divide the network into multiple sections and construct a FR-HIN full-relation heterogeneous information network, which effectively integrates the complex associations between equipment, sections and power data types. By analyzing the fault update records, building a context association model, and combining time decay and equipment configuration information, we deeply explore the context information in the fault description and maintenance records and the implicit association between devices. Using the conditional adversarial spatiotemporal encoder, we effectively extract the location features of the section or equipment, and use the discriminator to detect spatiotemporal anomalies in power data in real time, providing timely warning information for fault prediction. The structural features, attribute information, implicit context features and real-time spatiotemporal anomaly indicator sets are input into the power grid fault prediction model to achieve multi-dimensional information fusion, thereby significantly improving the accuracy of fault prediction.

[0170] Those skilled in the art will appreciate that the modules in the devices in the embodiments can be adaptively changed and set in one or more devices different from the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and in addition they can be divided into multiple submodules or subunits or subcomponents. Except that at least some of such features and / or processes or units are mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device disclosed in this manner can be combined in any combination. Unless otherwise explicitly stated, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) can be replaced by an alternative feature that provides the same, equivalent or similar purpose. In addition, those skilled in the art will appreciate that although some embodiments herein include certain features included in other embodiments but not other features, the combination of features of different embodiments means being within the scope of the present invention and forming different embodiments. For example, in the following claims, any one of the claimed embodiments can be used in any combination. The present invention can be implemented by hardware including several different elements and by appropriately programmed computers. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The steps in the above embodiments should not be understood as limiting the execution order unless otherwise specified.

Claims

1. A fault prediction method for an intelligent power dispatching data network, characterized in that: include: According to the topological structure data of the power dispatching data network to be predicted, the physical interval of the power dispatching data network is divided into multiple sections, and the power data of each section is obtained; A FR-HIN full-relation heterogeneous information network is constructed based on the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, section and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include the interaction relationship between nodes; The FR-HIN full-relation heterogeneous information network is converted into a low-dimensional real-valued vector representation to obtain the potential structural features and attribute information of different nodes and relationships; a context association model is constructed based on multiple fault update records of power data to obtain context association features between fault descriptions and maintenance record texts; attenuation accumulation analysis is performed based on the context association features and multiple update times to generate multiple aggregate context feature sets; implicit equipment association identification is performed on the aggregate context feature set based on the equipment configuration information of the power dispatching data network to obtain multiple implicit context feature sets; The conditional adversarial spatiotemporal encoder is used to extract the location features of the segment or equipment, and the similarity between the generated power data distribution and the real data distribution is evaluated by the discriminator to obtain a set of real-time spatiotemporal anomaly indicators; the similarity between the generated power data distribution and the real data distribution is evaluated by the discriminator to obtain a set of real-time spatiotemporal anomaly indicators; The potential structural features and attribute information of different nodes and relationships, the implicit context feature set and the real-time spatiotemporal anomaly indicator set are input into the power grid fault prediction model to obtain the fault prediction results of the power dispatching data network, wherein the fault prediction results include the fault type, fault location and fault time.

2. The fault prediction method of the intelligent power dispatching data network according to claim 1 is characterized in that: The node types of the FR-HIN full-relation heterogeneous information network include device nodes, segment nodes, and power data type nodes; wherein the devices include transformers, circuit breakers, and lines; the segment nodes are physical areas formed after the power dispatching data network is segmented; the power data types include voltage, current, and power; The edge types of the FR-HIN full-relation heterogeneous information network include device-device relationship, device-segment relationship, segment-segment relationship, device-power data type relationship, and power data type-segment relationship.

3. The fault prediction method of the intelligent power dispatching data network according to claim 1 is characterized in that: The step of performing device implicit association identification on the aggregated context feature set to obtain multiple implicit context feature sets further includes: Construct an association graph between devices based on device configuration information, where nodes represent devices and edges represent connection relationships between devices; Encode the association graph to extract implicit association information between devices; The extracted implicit correlation information is combined with the aggregated context feature set, and implicit context features related to fault prediction are selected to form multiple implicit context feature sets.

4. The fault prediction method of the intelligent power dispatching data network according to claim 1 is characterized in that: The evaluating the similarity between the generated power data distribution and the real data distribution by the discriminator further comprises: The similarity between the generated power data distribution and the real data distribution is evaluated by the discriminator; The output of the encoder is used as a real-time spatiotemporal anomaly indicator, wherein the spatiotemporal anomaly indicator reflects the spatiotemporal anomaly state of each section or device in the electric power dispatching data network; The conditional adversarial spatiotemporal encoder includes a time series encoder and a space encoder, wherein the time series encoder is used to extract the time series features of the power data, and the space encoder is used to extract the location features of the segment or the equipment.

5. The fault prediction method of the intelligent power dispatching data network according to claim 1 is characterized in that: The converting the FR-HIN full-relation heterogeneous information network into a low-dimensional real-valued vector representation further includes: The nodes and relationships in FR-HIN are mapped to a low-dimensional vector space using matrix decomposition, where the optimization objective function of matrix decomposition is: Where A is the adjacency matrix of FR-HIN; z i and z j are the low-dimensional vector representations of nodes i and j respectively; σ is the activation function; λ is the regularization parameter.

6. The fault prediction method of the intelligent power dispatching data network according to claim 3 is characterized in that: The step of encoding the association graph to extract implicit association information between devices further includes: The representation of each device is iteratively updated through a multi-layer graph convolutional network, where the update formula of the device in each layer is: in, is the representation of node i at layer l+1; is the neighbor set of node i; W (l) and b (l) are the weight matrix and bias vector of the lth layer respectively; σ is the activation function; is the representation of node j at layer l.

7. The fault prediction method of the intelligent power dispatching data network according to claim 1 is characterized in that: The power grid fault prediction model includes an input layer, a second-order wavelet scattering network module, a multi-scale invariant feature extraction module, a multi-scale invariant feature fusion module, a multi-scale invariant feature pooling module, a high-level semantic feature network layer, and an output layer; Wherein, the second-order wavelet scattering network module includes a 0th-order scattering propagation path, a 1st-order scattering propagation path and a 2nd-order scattering propagation path.

8. The fault prediction method of the intelligent power dispatching data network according to claim 1, characterized in that: The conditional adversarial spatiotemporal encoder includes a generator, a discriminator, and a set condition; Among them, the generator includes a temporal encoder and a spatial encoder; the temporal encoder is composed of multiple CNN and LSTM network layers; the spatial decoder is composed of FlowNet, and the fused features are decoded by the decoder; the decoder is composed of convolutional layers and deconvolutional layers; the discriminator is composed of multiple layers of fully convolutional network layers.

9. A fault prediction device for an intelligent power dispatching data network, characterized in that: include: A section segmentation module is used to segment the physical interval of the power dispatching data network into multiple sections according to the topological structure data of the power dispatching data network to be predicted, and obtain power data of each section; A FR-HIN construction module is used to construct a FR-HIN full-relation heterogeneous information network based on the power data of each section, wherein the node types of the FR-HIN full-relation heterogeneous information network include equipment, section and power data types, and the edge types of the FR-HIN full-relation heterogeneous information network include the interaction relationship between nodes; A preprocessing module is used to convert the FR-HIN full-relation heterogeneous information network into a low-dimensional real-valued vector representation to obtain potential structural features and attribute information of different nodes and relationships; construct a context association model based on multiple fault update records of power data to obtain context association features between fault descriptions and maintenance record texts; perform attenuation accumulation analysis based on the context association features and multiple update times to generate multiple aggregate context feature sets; perform implicit equipment association identification on the aggregate context feature set based on the equipment configuration information of the power dispatching data network to obtain multiple implicit context feature sets; a similarity calculation module, for extracting the location features of the segment or device using the conditional adversarial spatiotemporal encoder, evaluating the similarity between the generated power data distribution and the real data distribution through the discriminator, and obtaining the real-time spatiotemporal anomaly indicator set; and evaluating the similarity between the generated power data distribution and the real data distribution through the discriminator, and obtaining the real-time spatiotemporal anomaly indicator set; The fault prediction module is used to input the potential structural characteristics and attribute information of different nodes and relationships, the implicit context feature set and the real-time spatiotemporal anomaly indicator set into the power grid fault prediction model to obtain the fault prediction results of the power dispatching data network, wherein the fault prediction results include the fault type, fault location and fault time.

10. A computing device comprising: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the above-mentioned intelligent power dispatching data network fault prediction method.

Images