Power grid equipment status prediction method based on edge intelligence and multi-source data fusion

Abstract

The present invention relates to the field of electric power technology, and discloses a method for predicting the state of power grid equipment based on edge intelligence and multi-source data fusion, the method comprising collecting and processing the state data of a plurality of power grid equipment through edge computing technology to obtain multi-source state data, and extracting characterization data from the multi-source state data; the characterization data has characteristic information and attribute information of the multi-source state data; constructing an objective function based on the characterization data, and fusing the multi-source state data through the objective function to obtain a multi-source state data fusion matrix; inputting the multi-source state data fusion matrix into a lightweight neural network model for processing, and outputting the state prediction result of the power grid equipment. The present invention combines edge computing and deep learning technology to achieve accurate state prediction of power grid equipment.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a method for predicting the state of power grid equipment based on edge intelligence and multi-source data fusion. Background Art

[0002] With the continuous development of data processing technology, deep learning technology has been applied to the state prediction process of a large number of power grid equipment.

[0003] There are many types of existing power grid equipment, and the relative independence of equipment in different areas makes it difficult to collect heterogeneous data. The diversity of the data also makes the existing neural network model's prediction results on the status of power grid equipment inaccurate.

[0004] It can be seen that how to solve the problem of low accuracy in predicting the status of power grid equipment has become a technical problem that needs to be urgently solved by those skilled in the art. Summary of the invention

[0005] The present invention provides a method for predicting the state of power grid equipment based on edge intelligence and multi-source data fusion, which solves the problem of low accuracy in predicting the state of power grid equipment.

[0006] In order to solve the above technical problems, the first aspect of the present invention provides a method for predicting the state of power grid equipment based on edge intelligence and multi-source data fusion, comprising:

[0007] The state data of the plurality of power grid devices are collected and processed by edge computing technology to obtain multi-source state data, and characterization data is extracted from the multi-source state data; the characterization data has characteristic information and attribute information of the multi-source state data;

[0008] Constructing an objective function based on the representation data, and fusing the multi-source state data through the objective function to obtain a multi-source state data fusion matrix;

[0009] The multi-source state data fusion matrix is ​​input into a lightweight neural network model for processing, and a state prediction result of the power grid equipment is output.

[0010] As one of the preferred solutions, the state data of the plurality of power grid devices are collected and processed by edge computing technology to obtain multi-source state data, including:

[0011] The deployed first regression model and the first deep neural network model are trained on the terminal mobile device, and the deployed second regression model and the second deep neural network model are trained on the edge server to obtain the trained first regression model, the trained first deep neural network model, the trained second regression model and the trained second deep neural network model; the first regression model and the second regression model have different numbers of deep neural network layers; the first deep neural network model and the second deep neural network model both have multiple exit points;

[0012] Based on the measured bandwidth and delay requirements between the terminal mobile device and the edge server, searching for a first optimal branch point of the first deep neural network model that has been trained through a first regression model that has been trained, searching for a second optimal branch point of the second deep neural network model that has been trained through a second regression model that has been trained, and deploying corresponding branch deep neural network models on the terminal mobile device and the edge server according to the first optimal branch point and the second optimal branch point, so that the terminal mobile device that has been deployed collects status data of a plurality of the power grid devices and transmits the data to the edge server that has been deployed;

[0013] The state data is processed by the edge server that has been deployed to obtain multi-source state data, so that the edge server that has been deployed can store the multi-source state data to each edge node through a content distribution network based on user traffic requirements.

[0014] As one of the preferred solutions, based on the measured bandwidth and delay requirements between the terminal mobile device and the edge server, searching for the first optimal branch point of the trained first deep neural network model through the trained first regression model includes:

[0015] Select any branch layer of the first deep neural network model that has completed training, and predict the running delay of the branch layer on the terminal mobile device and the edge server respectively through the first regression model that has completed training;

[0016] Based on the measured bandwidth between the terminal mobile device and the edge server, respectively calculate the communication delay of the terminal mobile device and the edge server, and combine the running delay to obtain a total delay;

[0017] The branch layer is iteratively updated to calculate the total delay through the updated branch layer, and the branch layer with the smallest total delay is used as the first optimal branch point of the first deep neural network model that has completed training.

[0018] As one of the preferred solutions, the edge server that has completed deployment stores the multi-source state data to each edge node through a content distribution network based on user traffic requirements, including:

[0019] The user traffic is monitored in real time by the deployed edge server, and the user traffic is analyzed by a machine learning model to predict the user traffic demand;

[0020] Building a content distribution network, and configuring each edge node in the content distribution network according to the user traffic demand;

[0021] The multi-source status data are classified according to access frequencies, so as to store the multi-source status data in corresponding edge nodes according to the classification results.

[0022] As one preferred solution, the extracting the representation data from the multi-source state data includes:

[0023] Integrating the multi-source status data to obtain multi-source status integrated data with a unified format and structure;

[0024] Setting the data mode and entity relationship of the multi-source state data to construct the multi-source state integration data into a state data graph;

[0025] Based on the Laplace matrix of the state data map, characterization data is extracted from the multi-source state data.

[0026] As one preferred solution, the objective function is expressed by the following formula:

[0027]

[0028] In the formula, is the multi-source state data fusion matrix; U v is the data matrix; F is the label matrix; m and n are integers; is the norm operation; x v is multi-source state data; u v To characterize the data; w v is the weight; is the regularization parameter; is the Laplace matrix of the multi-source state data fusion matrix; is a hyperparameter; Tr is the trace number; T is the transpose operation; I is the unit matrix; , are the data of the i-th and j-th columns of the data matrix respectively; s ij is the data in the i-th row and j-th column of the multi-source state data fusion matrix.

[0029] As one of the preferred solutions, fusing the multi-source state data through the objective function to obtain a multi-source state data fusion matrix includes:

[0030] Under the constraint conditions, the parameters in the objective function are updated by an alternating iteration method until the objective function reaches convergence, thereby obtaining an optimized weight and an optimized matrix representing data; the parameters include weights and a matrix representing data, and the expression of the weights is represented by a Lagrangian function of the objective function;

[0031] Based on the optimization weights, the individual representation data optimization matrices are fused to obtain the multi-source state data fusion matrix, so as to achieve fusion of the multi-source state data.

[0032] As one of the preferred solutions, the parameters also include a label matrix and a multi-source state data fusion matrix; wherein,

[0033] The updating of the parameters in the objective function by the alternating iteration method comprises:

[0034] The characterization data matrix is ​​updated, and the updating process is expressed by the following formula:

[0035]

[0036] Where: is a multi-source state data matrix;

[0037] The weight is updated, and the updating process is expressed by the following formula:

[0038]

[0039] The label matrix is ​​updated, and the updating process is expressed by the following formula:

[0040]

[0041] In the formula, k is the label category;

[0042] The multi-source state data fusion matrix is ​​updated, and the updating process is expressed by the following formula:

[0043]

[0044]

[0045] Where g is the multi-source vector; f is the data in the label matrix.

[0046] As one of the preferred solutions, the lightweight neural network model includes a deep separable convolutional network and a recursive neural network;

[0047] The step of inputting the multi-source state data fusion matrix into a lightweight neural network model for processing and outputting a state prediction result of the power grid equipment includes:

[0048] Extracting features from the multi-source state data fusion matrix through the deep separable convolutional network to obtain fusion state features;

[0049] The fusion state feature is modeled in time series by the recursive neural network to obtain a state prediction result of the power grid equipment.

[0050] As one preferred solution, the recursive neural network includes a first update gate, a reset gate and a second update gate;

[0051] The step of performing time series modeling on the fusion state feature through the recursive neural network to obtain a state prediction result of the power grid device includes:

[0052] The fusion state feature and its weight at the current time step and the hidden feature and its weight corresponding to the fusion state feature at the previous time step are processed respectively through the first update gate to obtain a first processing result;

[0053] The fusion state feature and its weight at the current time step and the hidden feature and its weight corresponding to the fusion state feature at the previous time step are processed respectively through the reset gate to obtain a second processing result;

[0054] The first processing result and the second processing result are processed by the second update gate to obtain a state prediction result of the power grid device.

[0055] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0056] (1) Edge computing technology is used to extract and process the status data of various power grid equipment, which improves the timeliness and accuracy of the data. By obtaining multi-source status data of power grid equipment, the operating status of the equipment can be more comprehensively reflected, avoiding the information loss or deviation that may be caused by a single data source;

[0057] (2) By replacing the original data with the representation data for modeling, the initial value sensitivity problem caused by the original state data is effectively avoided. The representation data that retains the key feature information and attribute information of the original state data has a high information density and representativeness, which helps to reduce the impact of redundant data on the prediction results and provides a solid foundation for subsequent data fusion and state prediction.

[0058] (3) By constructing an objective function to fuse multi-source status data, it is helpful to discover the correlation and complementarity between different data sources, thereby more accurately reflecting the actual operating status of power grid equipment; using a lightweight neural network model to process the fused data reduces the complexity and computing requirements of the model, and improves the prediction speed and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the implementation mode will be briefly introduced below. Obviously, the drawings described below are only some implementation modes of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0060] Figure 1 It is a flow chart of a method for predicting power grid equipment status based on edge intelligence and multi-source data fusion provided by a certain embodiment of the present invention;

[0061] Figure 2 It is a schematic diagram of the objective function construction process provided by a certain embodiment of the present invention;

[0062] Figure 3 is a schematic diagram of a multi-source state data fusion process provided by an embodiment of the present invention;

[0063] Figure 4 It is a schematic diagram of the structure of a lightweight neural network model provided by a certain embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] In the description of the present invention, the terms "first", "second", "third", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first", "second", "third", etc. may explicitly or implicitly include one or more of the feature. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0066] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two components. The terms "vertical", "horizontal", "left", "right", "upper", "lower" and similar expressions used herein are only for illustrative purposes, and do not indicate or imply that the system or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0067] In the description of the present invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0068] In one embodiment, if Figure 1 As shown, the first aspect of the present invention provides a method for predicting power grid equipment status based on edge intelligence and multi-source data fusion, comprising:

[0069] S1. Collect and process the status data of the plurality of power grid devices through edge computing technology to obtain multi-source status data, and extract characterization data from the multi-source status data; the characterization data has feature information and attribute information of the multi-source status data;

[0070] Specifically, power grid equipment is a key component of the power system. They are used for the transmission, distribution, conversion and control of electricity to ensure that electric energy is stably and reliably supplied to users. There are many types of power grid equipment, including power distribution equipment, switchgear, primary equipment, secondary equipment, power cables, transformers, etc., and the status data for monitoring power grid equipment is also diverse. Taking cables as an example, their status data includes cable temperature measurement data, cable fault location data, cable circulation monitoring data, etc. These multi-source heterogeneous data are used to monitor the status of cables, and then judge the performance of the cables or whether they have failed. The present invention analyzes the real-time status data of power grid equipment to predict its multi-source status data in the future period, and then judges its faults and performance.

[0071] In one embodiment, the state data of the plurality of power grid devices are collected and processed by edge computing technology to obtain multi-source state data, including:

[0072] The deployed first regression model and the first deep neural network model are trained on the terminal mobile device, and the deployed second regression model and the second deep neural network model are trained on the edge server to obtain the trained first regression model, the trained first deep neural network model, the trained second regression model and the trained second deep neural network model; the first regression model and the second regression model have different numbers of deep neural network layers; the first deep neural network model and the second deep neural network model both have multiple exit points;

[0073] Based on the measured bandwidth and delay requirements between the terminal mobile device and the edge server, searching for a first optimal branch point of the first deep neural network model that has been trained through a first regression model that has been trained, searching for a second optimal branch point of the second deep neural network model that has been trained through a second regression model that has been trained, and deploying corresponding branch deep neural network models on the terminal mobile device and the edge server according to the first optimal branch point and the second optimal branch point, so that the terminal mobile device that has been deployed collects status data of a plurality of the power grid devices and transmits the data to the edge server that has been deployed;

[0074] The state data is processed by the edge server that has been deployed to obtain multi-source state data, so that the edge server that has been deployed can store the multi-source state data to each edge node through a content distribution network based on user traffic requirements.

[0075] Specifically, the present invention collects and processes the multi-source heterogeneous state data of power grid equipment in real time through the Edgent framework, and the framework is divided into three stages: Offline Training Stage, Online Optimization Stage and Co-Inference Stage. In Offline Training, some time-consuming tasks are deployed in advance: in order to balance the running speed and real-time accuracy, the Exit Point 2 balancing network is introduced to realize the multi-branch network. Edgent inputs the deployed deep network (DNN) into the static / dynamic configurator component and obtains the corresponding online tuning configuration. Specifically, in order to train the first regression model and the first deep neural network model on the terminal mobile device, the second regression model and the second deep neural network model are also trained on the edge server to obtain four trained models; wherein, the first regression model and the second regression model have different numbers of deep neural network layers; the first deep neural network model and the second deep neural network model both have multiple exit points; and then these trained models are statically configured where the bandwidth remains stable during DNN reasoning.

[0076] Considering the fixed nature of key power grid equipment, it is necessary to design a static configurator for Edgent. The key idea of ​​the static configurator is to train a regression model to predict hierarchical inference latency and train a branch model to enable an early exit mechanism. In the offline configuration phase, in order to generate a static configuration, the static configurator starts two tasks: (1) configure the hierarchical inference latency on the mobile device and edge server respectively, and train the regression models of different types of DNN layers (such as Convolution, Fully-Connected, etc.) accordingly; (2) train a DNN model with multiple exit points through the BranchyNet framework to obtain a branch DNN. The analysis process is to record the inference latency of each layer, rather than the inference latency of the entire model. Since the hierarchical inference latency depends on the infrastructure, and DNN training is application-related, Edgent only needs to initialize the above two tasks once for a specific DNN inference.

[0077] In the Online Optimization Stage, the static configuration is used to search for the best exit and partition points. That is, the trained regression models corresponding to the terminal mobile device and the edge server are used to search for the best branching points of the trained deep neural network model to maximize the accuracy while ensuring the execution deadline through three inputs: (1) static configuration, (2) measured bandwidth between the edge server and the terminal device, and (3) latency requirements.

[0078] In one embodiment, searching for a first optimal branch point of a first deep neural network model that has been trained by using a first regression model that has been trained based on the measured bandwidth and latency requirements between the terminal mobile device and the edge server includes:

[0079] Select any branch layer of the first deep neural network model that has completed training, and predict the running delay of the branch layer on the terminal mobile device and the edge server respectively through the first regression model that has completed training;

[0080] Based on the measured bandwidth between the terminal mobile device and the edge server, respectively calculate the communication delay of the terminal mobile device and the edge server, and combine the running delay to obtain a total delay;

[0081] The branch layer is iteratively updated to calculate the total delay through the updated branch layer, and the branch layer with the smallest total delay is used as the first optimal branch point of the first deep neural network model that has completed training.

[0082] Specifically, for terminal mobile devices with The DNN model with exit points, the branch of selecting the i-th exit point is layer, is the output of the pth layer, and the running delay of the jth layer on the terminal mobile device is predicted by the first regression model that has completed training and the running delay of layer j on the server Then, based on the measured bandwidth between the terminal mobile device and the edge server, the terminal mobile device and the edge server are enabled to input and interact with each other, and the communication delay between the terminal mobile device and the edge server can be calculated respectively. The total delay can be calculated by combining the communication delay and the operation delay. , update the branch layer and repeat the calculation of the total delay until all branch layers are traversed, and the branch layer with the smallest total delay is used as the first optimal branch point of the first deep neural network model, and the pth layer is represented as the split point of the branch with the i-th exit point. When p=1, Indicates that the total reasoning process is only executed on the mobile terminal device, and when p= hour, Indicates that the total calculation is completed only on the edge server; by exhaustively searching the points, the best partition point with the minimum delay of the i-th exit point can be obtained. Since model partitioning does not affect the inference accuracy, DNN inference with different exit layers (i.e., with different precisions) can be tested continuously, and the model with the maximum precision while meeting the delay requirements can be found. Since the regression model for hierarchical delay prediction has been pre-trained and linear search operations, data can be processed quickly. Similarly, the search process for the second best branch point on the edge server can refer to the search process for the first best branch point, but the model used is different.

[0083] In addition, in the Online Optimization Stage, data from different sources can also be cleaned, such as intelligent filling, outlier marking, etc. Edgent measures the current bandwidth status and jointly optimizes the DNN partition and the appropriate size of the DNN according to the given latency requirements and the configuration obtained offline, with the goal of maximizing the inference accuracy under the given latency requirements.

[0084] In one embodiment, the edge server that has completed deployment stores the multi-source state data to each edge node through a content distribution network based on user traffic requirements, including:

[0085] The user traffic is monitored in real time by the deployed edge server, and the user traffic is analyzed by a machine learning model to predict the user traffic demand;

[0086] Building a content distribution network, and configuring each edge node in the content distribution network according to the user traffic demand;

[0087] The multi-source status data are classified according to access frequencies, so as to store the multi-source status data in corresponding edge nodes according to the classification results.

[0088] Specifically, in the Co-Inference Stage, based on the collaborative reasoning plan generated in the online tuning stage (i.e., the selected exit point and partition point), the layers before the partition point will be executed on the edge server, and the remaining layers will remain on the device, so that the deployed terminal mobile devices can collect the status data of various power grid equipment and transmit it to the deployed edge server. The edge server transmits the processed status data to the edge node and caches the commonly used content on the edge node in advance, so that it can respond quickly when the user requests it, without having to obtain the content from the remote server every time. The status data of the power grid equipment includes at least historical data, behavior patterns, location information, etc. from multiple sources such as user equipment, sensors, application logs, etc., and should cover the equipment's operating parameters, alarm information, fault records, maintenance records, etc.

[0089] In this embodiment, the edge server first monitors and collects user traffic data in real time, including user visits, access patterns, data request types, etc., and uses machine learning algorithms or statistical analysis methods to predict user traffic demand to determine traffic trends and hotspot status data in the future. These models can use various algorithms, such as regression, classification, clustering, etc., to continuously train and optimize based on historical traffic, status data and real-time traffic, status data, thereby improving prediction accuracy; then build a CDN network consisting of multiple distributed edge nodes, and the edge nodes can be other servers. These nodes can be distributed in different geographical locations around the world to cover a wider range of user groups, and minimize latency and minimize response speed. The goal is to select edge nodes with sufficient computing power according to the predicted user traffic demand, select edge nodes with sufficient storage capacity according to the access frequency and size of the data, and ensure that the edge nodes have sufficient network bandwidth for physical and network node configuration; and classify multi-source status data according to the access frequency, and cache the status data with the highest access frequency on the edge nodes that are closer to the user according to the priority order corresponding to the classification results, so as to quickly respond to user requests; among them, the distance of the edge node from the user is hierarchically set according to the priority of the data. For example, the fault data of the transformer belongs to the highest priority, so it is deployed on the edge node within the first preset range (such as within ten meters) from the user for the user to quickly retrieve.

[0090] In addition, the present invention can not only reduce data transmission delay and bandwidth consumption in the network by placing computing and data storage resources on the network edge node closest to the user; it can also continuously monitor and evaluate user traffic requirements, as well as indicators such as cache hit rate, and dynamically adjust cache strategies and machine learning models according to real-time conditions to adapt to changes in traffic demand; it can also improve user experience through network optimization, thereby delivering content to users faster, effectively optimizing resource allocation, reducing the delay in downloading information from remote data centers, and providing high-performance data collection and transmission for edge computing.

[0091] The present invention can ensure the immediacy of data by monitoring user traffic in real time, while the prediction of the machine learning model improves the accuracy of user traffic demand and contributes to the efficient allocation of resources; dynamically adjusting the configuration of edge nodes in CDN according to the predicted user traffic demand can ensure the efficiency and stability of content distribution, reduce latency, and improve user experience; storing multi-source status data in edge nodes according to access frequency can optimize the data storage structure, increase data access speed, and reduce storage costs; through accurate prediction and dynamic configuration, CDN resources can be used more effectively to avoid resource waste, while improving the overall performance and response speed of the system.

[0092] In one embodiment, step S1 includes:

[0093] Integrating the multi-source status data to obtain multi-source status integrated data with a unified format and structure;

[0094] Setting the data mode and entity relationship of the multi-source state data to construct the multi-source state integration data into a state data graph;

[0095] Based on the Laplace matrix of the state data map, characterization data is extracted from the multi-source state data.

[0096] Specifically, after obtaining the multi-source status data of the power grid equipment, it is first cleaned to remove duplicate data, invalid data and abnormal data, fill or interpolate the missing data, and standardize and normalize to ensure the consistency and comparability of the data; then the processed data is integrated to form a unified data format and structure. The integrated data should contain the unique identification of the equipment, status monitoring parameters, timestamp and other information; then the data model and entity relationship of the multi-source status data are set; wherein the data model includes entity type, attribute type and relationship type, entity type can include equipment, alarm, fault, maintenance task, etc., attribute type can include equipment model, manufacturer, installation location, operating parameters, etc., relationship type can include the association relationship between equipment and alarm, the causal relationship between equipment and fault, the execution relationship between equipment and maintenance task, etc.; and entities and relationships are extracted from the integrated data according to the data model to construct an entity relationship network, which should be able to clearly represent the association relationship between equipment, equipment status changes and equipment maintenance history, etc.; the integrated data and the defined entity relationship are imported into the selected graph construction tool to obtain the state data graph of the power grid equipment.

[0097] The present invention is similar to graph neural network, targeting multi-modal multi-state data ,in, Represents the dimension of a certain mode, and optimizes it together with the Laplace matrix of the state data map to obtain the representation data matrix U v , so that it has the characteristic information of the original data and retains the unique attributes of the original information, so as to extract the beneficial features of the target to the greatest extent.

[0098] The present invention integrates multi-source status data, eliminates differences in data formats and structures, forms a unified data view, facilitates subsequent data analysis and processing, and improves data quality and availability; by constructing a status data graph, complex data relationships can be presented in an intuitive manner, which is convenient for understanding and analyzing the status of power grid equipment, and the nodes and edges in the graph can represent equipment, attributes and the relationship between them, which helps to reveal the intrinsic connection between data; the Laplace matrix based on the graph is used to extract and characterize data, which can capture key features and patterns in the data; by integrating and analyzing multi-source status data, power grid operators can have a more comprehensive understanding of the operating status of the equipment, so as to make more accurate and timely decisions, which helps to improve the stability and security of the power grid and reduce operation and maintenance costs.

[0099] S2, constructing an objective function based on the representation data, and fusing the multi-source state data through the objective function to obtain a multi-source state data fusion matrix;

[0100] Specifically, the present invention uses multimodal data to enhance the ability of risk prediction and fault diagnosis, and obtains multifaceted information by combining data collected by various sensors (such as sound, vibration, and temperature sensors), so as to more comprehensively understand the system operation status. For example, sound sensors can capture abnormal noise, vibration sensors can detect the vibration of mechanical parts, and temperature sensors can provide heat distribution during equipment operation. Then, through the deep learning model, data of different modes are comprehensively analyzed to dig out the patterns and laws hidden behind the data. This comprehensive analysis capability enables the present invention to more accurately identify potential signs of faults and provide early warnings and diagnoses at an early stage.

[0101] The present invention is analogous to a graph neural network, and models are built based on the representation data to obtain the initial objective function:

[0102]

[0103] In the formula, is the multi-source state data fusion matrix; m and n are integers; is the norm operation; x v is multi-source state data; u v To characterize the data; w v is the weight; is the regularization parameter; I is the identity matrix; , are the data of the i-th and j-th columns of the data matrix respectively; s ij is the data in the i-th row and j-th column of the multi-source state data fusion matrix; F is the label matrix; L is the Laplace matrix of the state data spectrum.

[0104] in, Indicates that multi-view fusion is performed directly by characterizing data, and the multi-source state fusion matrix is ​​constrained , ensuring that the fusion matrix extracts rich structural information; It represents the constraints on the fusion matrix, so that it has a compact structure and is convenient for multi-view clustering.

[0105] According to Fan Ji's theorem, the initial objective function is transformed into the objective function. The schematic diagram of its construction process is as follows: Figure 2 As shown, it can be expressed by the following formula:

[0106]

[0107] In the formula, is a hyperparameter; Tr is the number of traces; T is the transposition operation; the label matrix contains the number of targets and categories, and KMeans clustering can be performed to achieve cluster analysis of the targets; the adaptive weight w v That is, to fuse the single data matrix and realize the sufficiency of the multi-source state data fusion matrix characteristics. The present invention draws on the idea of ​​multi-view neural network and uses an approximate matrix, namely the multi-source state data fusion matrix, to approximate the original data, thereby effectively fusing the multi-source state data.

[0108] In one embodiment, fusing the multi-source state data through the objective function to obtain a multi-source state data fusion matrix includes:

[0109] Under the constraint conditions, the parameters in the objective function are updated by an alternating iteration method until the objective function reaches convergence, thereby obtaining an optimized weight and an optimized matrix representing data; the parameters include weights and a matrix representing data, and the expression of the weights is represented by a Lagrangian function of the objective function;

[0110] Based on the optimization weights, the individual representation data optimization matrices are fused to obtain the multi-source state data fusion matrix, so as to achieve fusion of the multi-source state data.

[0111] Specifically, the multi-source state data fusion process is as follows: Figure 3 As shown, the present invention is subject to the constraints Under this condition, the alternating iteration method is directly used to update the parameters in the objective function until the objective function converges. The parameters of the objective function include weights, characterization data matrix, label matrix and multi-source state data fusion matrix. The following is the specific process of optimizing each parameter variable:

[0112] The representation data matrix U v To optimize and update , other parameters need to be considered as constants. The optimization process is as follows:

[0113]

[0114] Where: is a multi-source state data matrix;

[0115] Where I is the unit matrix of corresponding size. According to the above formula, we can know that the size of the unit matrix is ​​n*n, and the size of L is n*n. Since the size of the original multi-source state data is d*n, the obtained representation data is also d*n, which is consistent with the size of the original data and directly retains a large number of distinct features, thereby alleviating the sensitivity problem of the original data to a certain extent.

[0116] For weight w v To optimize and update , other parameters need to be considered as constants. In order to maintain consistency in multiple views and ensure that the fusion matrix is ​​adaptively generated through data representation, the Lagrangian function of the objective function can be directly used to obtain:

[0117]

[0118] It can be seen that the adaptive weight w v For information about the characterization data matrix U and the multi-source state data fusion matrix S, see Figure 2 and Figure 3 , through the adaptive weight w v The optimized weights obtained after optimization fuse the optimized matrix of representation data obtained after the optimization of the representation data matrix, so as to realize the sufficiency of the characteristics of the multi-source state fusion data matrix S, obtain the multi-source state fusion data matrix, and realize the fusion of multi-source state data. v The range is between [0, 1], which is conducive to the rapid convergence of the model.

[0119] To optimize and update the label matrix F, other parameters need to be considered as constants, then:

[0120]

[0121] In the formula, k is the label category, which is used to distinguish different categories;

[0122] At this time, the objective function is equivalent to the eigenvalue decomposition of the Laplace matrix. Since the label matrix is ​​n*k, it is only necessary to take the eigenvectors corresponding to the first k smallest eigenvalues ​​of L. The label matrix F contains the number and category of targets. The clustering achieved by the label matrix does not require post-processing such as KMeans, thereby avoiding the randomness and contingency of the experimental results, and further enhancing the credibility and scientificity of the proposed scheme.

[0123] To optimize and update the multi-source state data fusion matrix S, other parameters need to be considered as constants, then:

[0124]

[0125]

[0126] Where g is the multi-source vector; f is the data in the label matrix.

[0127] The parts containing S are all represented by the elements of the matrix, and then the summation sign is changed, so that multiple source vectors can be obtained. The objective function is constrained: The solution can be directly obtained by combining the Lagrangian function with the KKT condition and the multi-source vector. At this point, all parameter variables have been optimized.

[0128] Through the above optimization process of the objective function parameters, it can be obtained that the data matrix and weight are convex functions, and their local solution is the optimal solution; due to the label matrix The second-order partial derivative of is positive, so the label matrix is ​​also a convex function. Therefore, the key to whether the objective function converges is whether there is an optimal value for the multi-source state data fusion matrix: Let Expressed as the latest result of each iteration, then there must be:

[0129]

[0130] Combined with Fan Ji's theorem , then there is

[0131]

[0132] Adding the above two inequalities together, we get:

[0133]

[0134] Adding the following equation to both sides of the inequality , we can get:

[0135]

[0136] So far, the above inequality shows that the multi-source state data fusion matrix S is constantly decreasing, so the entire objective function value is monotonically decreasing with the increase in the number of runs, so the function value is convergent. , updating the weights requires , updating the label matrix requires , updating the multi-source state data fusion matrix requires , where d is the feature dimension of each view, m is the number of majority graphs, and c is the number of clusters. Therefore, the time complexity of the objective function is , indicating that it can be applied to practical engineering.

[0137] The present invention uses representation data to replace original data for modeling, thereby effectively avoiding the initial value sensitivity problem caused by the original state data. The representation data that retains the key feature information and attribute information of the original state data has a high information density and representativeness, which helps to reduce the impact of redundant data on the prediction results and provides a solid foundation for subsequent data fusion and state prediction. By constructing an objective function to fuse multi-source state data, it helps to discover the correlation and complementarity between different data sources, thereby more accurately reflecting the actual operating status of power grid equipment.

[0138] S3, inputting the multi-source state data fusion matrix into a lightweight neural network model for processing, and outputting a state prediction result of the power grid equipment;

[0139] In one embodiment, the structure of the lightweight neural network model is as follows: Figure 4 As shown, it includes deep separable convolutional networks and recurrent neural networks;

[0140] Wherein, step S3 comprises:

[0141] Extracting features from the multi-source state data fusion matrix through the deep separable convolutional network to obtain fusion state features;

[0142] The fusion state feature is modeled in time series by the recursive neural network to obtain a state prediction result of the power grid equipment.

[0143] Specifically, the deep separable convolutional network used in the present invention has fewer network parameters than conventional convolution based on the same effect, realizes the lightweight of the model, and thus improves the operation efficiency of the model. The rich features obtained by the deep separable convolutional network are sent to the recurrent neural network for time series modeling to predict the state of the power grid equipment. The present invention uses a deep separable convolutional network to extract features from sensor data or images connected to the power grid equipment. For example, for power production equipment, CNNs can be used to extract key features from sensor data such as temperature, pressure, and vibration, or feature extraction can be performed on visual images of power equipment, such as wear and cracks; the extracted features are modeled in time series using a recurrent neural network (RNN) to predict equipment performance degradation and failure risks. By inputting past sensor data sequences into the RNN, the model can learn the dynamic changes of equipment status and predict possible future performance degradation or failure conditions. Compared with traditional prediction models based on rules or statistical methods, the present invention uses CNNs for feature extraction to better capture complex patterns and associated information in the data; using RNN to model time series data can take into account the dynamic changes of data and the dependencies between sequences.

[0144] In one embodiment, the recursive neural network includes a first update gate, a reset gate, and a second update gate;

[0145] The step of performing time series modeling on the fusion state feature through the recursive neural network to obtain a state prediction result of the power grid device includes:

[0146] The fusion state feature and its weight at the current time step and the hidden feature and its weight corresponding to the fusion state feature at the previous time step are processed respectively through the first update gate to obtain a first processing result;

[0147] The fusion state feature and its weight at the current time step and the hidden feature and its weight corresponding to the fusion state feature at the previous time step are processed respectively through the reset gate to obtain a second processing result;

[0148] The first processing result and the second processing result are processed by the second update gate to obtain a state prediction result of the power grid device.

[0149] Specifically, the recursive neural network in the present invention is also a gated neural network, which includes a first update gate , the update gate for time step t is calculated by the following formula: , when the data s in the multi-source state data fusion matrix t The weights input to the network unit and itself Multiplied, for s t The corresponding hidden state , that is, the state data corresponding to the data in the multi-source state data fusion matrix is ​​also processed in the same way, keeping the previous The information of each unit is divided by its own weight The two results are added together, and then the sigmoid activation function is applied to compress the result between 0 and 1 to obtain the first processed result. In this way, the first update gate helps the model determine how much past information (from the previous time step) needs to be passed into the future, so that the model can decide to copy all the information from the past and eliminate the risk of the gradient vanishing problem.

[0150] The gated neural network also includes a reset gate r t ,pass Used to decide how much past information to forget. The formula is the same as the update gate formula. The difference lies in the use of weights and gates. A new memory content is introduced into the current memory content. The calculation formula is: , using the reset gate to store relevant information from the past, the input data s t , After multiplying by its own weight, the gate r is reset t Multiply them element-wise with their weights, add the results together and apply a non-linear activation function , and obtain the second processing result.

[0151] The gated neural network also includes a second update gate: , for the update gate and Perform element-by-element multiplication and then and Perform element-by-element multiplication and finally add them together, and the status and performance of the power grid equipment can be obtained through the recurrent neural network.

[0152] In order to solve the long-term dependency problem, the present invention uses a gated recurrent unit (GRU) to replace the traditional long short-term memory network (LSTM). GRU has a similar gating mechanism, but has fewer parameters and lower computational cost. At the same time, it can still effectively capture the long-term dependencies in time series data, improve the model's modeling ability for time series data, and reduce computational cost. Through this method of comprehensive utilization of deep learning technology, the present invention can more accurately predict the performance degradation and failure risk of key equipment, thereby realizing a more effective predictive maintenance strategy.

[0153] The present invention uses edge computing to realize real-time collection and processing of multi-source status data. With the help of the Edge framework, data can be collected and processed simultaneously, providing a basis for the fusion of multi-source status data. Drawing on the idea of ​​multi-view neural network, the original data is approximated by using an approximate matrix, thereby effectively and fully fusion of multi-source status data. The approximate multi-source status data fusion matrix is ​​transmitted to the lightweight network for feature extraction and device status prediction. In the feature extraction stage, a deep separable convolutional network is used to replace the traditional convolutional neural network, which not only reduces the network parameters of the model, but also improves the training speed. It is combined with a gated recurrent unit to effectively solve the long-term dependency problem, so as to achieve accurate prediction of the device status.

[0154] It should be noted that although the steps in the above flowchart are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders.

[0155] In summary, the present invention discloses a method for predicting the state of power grid equipment based on edge intelligence and multi-source data fusion, the method comprising collecting and processing the state data of a plurality of the power grid equipment through edge computing technology to obtain multi-source state data, and extracting characterization data from the multi-source state data; the characterization data has characteristic information and attribute information of the multi-source state data; constructing an objective function based on the characterization data, and fusing the multi-source state data through the objective function to obtain a multi-source state data fusion matrix; inputting the multi-source state data fusion matrix into a lightweight neural network model for processing, and outputting the state prediction result of the power grid equipment. The present invention combines edge computing and deep learning technology to realize the rapid processing of the state data of power grid equipment and improve the accuracy of state prediction.

[0156] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The above-mentioned embodiments only express several preferred implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principle of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the protection scope of the claims.

Claims

1. A power grid equipment status prediction method based on edge intelligence and multi-source data fusion, characterized in that: include: The state data of the plurality of power grid devices are collected and processed by edge computing technology to obtain multi-source state data, and characterization data is extracted from the multi-source state data; the characterization data has characteristic information and attribute information of the multi-source state data; Constructing an objective function based on the representation data, and fusing the multi-source state data through the objective function to obtain a multi-source state data fusion matrix; Inputting the multi-source state data fusion matrix into a lightweight neural network model for processing, and outputting a state prediction result of the power grid equipment; The extracting the representation data from the multi-source state data comprises: Integrating the multi-source status data to obtain multi-source status integrated data with a unified format and structure; Setting the data mode and entity relationship of the multi-source state data to construct the multi-source state integration data into a state data graph; Extracting representation data from the multi-source state data based on the Laplace matrix of the state data map; The objective function is expressed by the following formula: In the formula, is the multi-source state data fusion matrix; U v is the data matrix; F is the label matrix; m and n are integers; is the norm operation; x v is multi-source state data; u v To characterize the data; w v is the weight; is the regularization parameter; is the Laplace matrix of the multi-source state data fusion matrix; is a hyperparameter; Tr is the trace number; T is the transpose operation; I is the unit matrix; , are the data of the i-th and j-th columns of the data matrix respectively; s ij is the data of the i-th row and j-th column of the multi-source state data fusion matrix; The step of fusing the multi-source state data through the objective function to obtain a multi-source state data fusion matrix includes: Under the constraint conditions, the parameters in the objective function are updated by an alternating iteration method until the objective function reaches convergence, thereby obtaining an optimized weight and an optimized matrix representing data; the parameters include weights and a matrix representing data, and the expression of the weights is represented by a Lagrangian function of the objective function; Based on the optimization weights, the individual characterization data optimization matrices are fused to obtain the multi-source state data fusion matrix, so as to achieve fusion of the multi-source state data.

2. A method for predicting power grid equipment status based on edge intelligence and multi-source data fusion according to claim 1, characterized in that: The state data of the plurality of power grid devices are collected and processed by edge computing technology to obtain multi-source state data, including: The deployed first regression model and the first deep neural network model are trained on the terminal mobile device, and the deployed second regression model and the second deep neural network model are trained on the edge server to obtain the trained first regression model, the trained first deep neural network model, the trained second regression model and the trained second deep neural network model; the first regression model and the second regression model have different numbers of deep neural network layers; the first deep neural network model and the second deep neural network model both have multiple exit points; Based on the measured bandwidth and delay requirements between the terminal mobile device and the edge server, searching for a first optimal branch point of the first deep neural network model that has been trained through a first regression model that has been trained, searching for a second optimal branch point of the second deep neural network model that has been trained through a second regression model that has been trained, and deploying corresponding branch deep neural network models on the terminal mobile device and the edge server according to the first optimal branch point and the second optimal branch point, so that the terminal mobile device that has been deployed collects status data of a plurality of the power grid devices and transmits the data to the edge server that has been deployed; The state data is processed by the edge server that has been deployed to obtain multi-source state data, so that the edge server that has been deployed can store the multi-source state data to each edge node through a content distribution network based on user traffic requirements.

3. A method for predicting power grid equipment status based on edge intelligence and multi-source data fusion according to claim 2, characterized in that: The step of searching for a first optimal branch point of a first deep neural network model that has been trained based on the measured bandwidth and delay requirements between the terminal mobile device and the edge server through a first regression model that has been trained includes: Select any branch layer of the first deep neural network model that has completed training, and predict the running delay of the branch layer on the terminal mobile device and the edge server respectively through the first regression model that has completed training; Based on the measured bandwidth between the terminal mobile device and the edge server, respectively calculate the communication delay of the terminal mobile device and the edge server, and combine the running delay to obtain a total delay; The branch layer is iteratively updated to calculate the total delay through the updated branch layer, and the branch layer with the smallest total delay is used as the first optimal branch point of the first deep neural network model that has completed training.

4. A method for predicting power grid equipment status based on edge intelligence and multi-source data fusion according to claim 2, characterized in that: The edge server that has completed the deployment stores the multi-source state data to each edge node through the content distribution network based on user traffic requirements, including: The user traffic is monitored in real time by the deployed edge server, and the user traffic is analyzed by a machine learning model to predict the user traffic demand; Building a content distribution network, and configuring each edge node in the content distribution network according to the user traffic demand; The multi-source status data are classified according to access frequencies, so as to store the multi-source status data in corresponding edge nodes according to the classification results.

5. The method for predicting power grid equipment status based on edge intelligence and multi-source data fusion according to claim 1 is characterized in that: The parameters also include a label matrix and a multi-source state data fusion matrix; wherein, The updating of the parameters in the objective function by the alternating iteration method comprises: The characterization data matrix is ​​updated, and the updating process is expressed by the following formula: Where: is a multi-source state data matrix; The weight is updated, and the updating process is expressed by the following formula: The label matrix is ​​updated, and the updating process is expressed by the following formula: In the formula, k is the label category; The multi-source state data fusion matrix is ​​updated, and the updating process is expressed by the following formula: Where g is the multi-source vector; f is the data in the label matrix.

6. A method for predicting power grid equipment status based on edge intelligence and multi-source data fusion according to claim 1, characterized in that: The lightweight neural network model includes a deep separable convolutional network and a recurrent neural network; The step of inputting the multi-source state data fusion matrix into a lightweight neural network model for processing and outputting a state prediction result of the power grid equipment includes: Extracting features from the multi-source state data fusion matrix through the deep separable convolutional network to obtain fusion state features; The fusion state feature is modeled in time series by the recursive neural network to obtain a state prediction result of the power grid equipment.

7. A method for predicting power grid equipment status based on edge intelligence and multi-source data fusion according to claim 6, characterized in that: The recursive neural network includes a first update gate, a reset gate and a second update gate; The step of performing time series modeling on the fusion state feature through the recursive neural network to obtain a state prediction result of the power grid device includes: The fusion state feature and its weight at the current time step and the hidden feature and its weight corresponding to the fusion state feature at the previous time step are processed respectively through the first update gate to obtain a first processing result; The fusion state feature and its weight at the current time step and the hidden feature and its weight corresponding to the fusion state feature at the previous time step are processed respectively through the reset gate to obtain a second processing result; The first processing result and the second processing result are processed by the second update gate to obtain a state prediction result of the power grid device.

Images