Power grid node fault diagnosis method and device based on deep hybrid neural network

Abstract

The present invention is a method and device for diagnosing power grid node faults based on a deep hybrid neural network, and is in the field of artificial intelligence prediction technology. The device includes a data acquisition module, a data storage module, a data preprocessing module, and a prediction module; the method includes the following steps: S1: collecting power grid node fault data in the area; S2: preprocessing the power grid node fault data; S3: using a coding method to obtain a knowledge graph; S4: using GNN to extract node information in the knowledge graph, and aggregate and output the final node representation; S5: constructing a multi-layer MLP loss function and training; S6: processing and analyzing in real time to predict or judge the fault condition of the power grid node. The method of the present invention uses GNN to achieve accurate expression of the semantics of power grid connection fault data; uses multi-layer MLP combined with cross-validation learning, which greatly improves the model learning ability and accuracy; and can accurately predict or judge the fault condition of the power grid node.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence prediction technology, and is particularly applicable to a power grid node fault diagnosis method and device based on deep hybrid neural network. Background Art

[0002] Fault prediction between power grid nodes is an important part of power management. The node fault prediction data it provides is extremely important for the control, operation and planning of power systems. Using data to accurately predict faults between power grid nodes can not only serve as an important reference to take countermeasures in advance to actively prevent faults in the power system, but also ensure the stable operation of the power system, optimize resource allocation and improve energy utilization efficiency. This measure has important economic and social value. In addition, the accuracy of fault prediction between power grid nodes has a direct impact on many aspects of the power system, including but not limited to the safety, reliability, economy and power quality of system operation. At the same time, in the past decade, the intensification of market competition, the increase in user electricity demand and the optimization of staff allocation strategies have made fault prediction not only more important but also more difficult. For a long time, domestic scholars have conducted extensive and in-depth research on the theory of fault prediction between power system nodes and proposed many effective methods, such as regression analysis, time series method, neural network method, wavelet analysis method, etc. Among them, for a certain prediction problem, multiple prediction methods can be established, and different prediction methods can be used to provide different prediction information and different prediction accuracy to achieve predictions in multiple dimensions and levels.

[0003] However, the shallow neural network widely used in the above prediction methods is limited by the over-smoothing effect, which limits the learning ability of the model, resulting in relatively poor prediction accuracy and low robustness of power grid load data. High-precision grid node fault prediction is a key technical problem that needs to be solved in power management. How to use the monitoring data of existing grid nodes and the characteristic information of grid nodes to solve this key technical problem with deep neural networks is the focus of current research. Summary of the invention

[0004] In view of this, in order to solve the above technical problems, the present invention provides a power grid node fault diagnosis method and device based on deep hybrid neural network, which aims to model the knowledge graph composed of power grid nodes, and at the same time adopts deep neural network to improve the model learning ability, so as to ultimately achieve high-precision power grid node fault prediction.

[0005] In order to achieve the above-mentioned object, the present invention provides a power grid node fault diagnosis method based on deep hybrid neural network, which is implemented by a power grid node fault diagnosis device based on deep hybrid neural network; Figure 1The power grid node fault diagnosis device based on deep hybrid neural network is composed of a data acquisition module, a data storage module, a data preprocessing module and a prediction module; the data acquisition module reads the data of the power grid node sensor connected to it through communication to realize the acquisition of power grid node fault data; the data storage module is composed of a power grid node data storage unit and a prediction data storage unit; the power grid node data storage unit is a memory, which is connected to the data acquisition module and the data preprocessing module respectively, and is used to store and transmit the power grid node fault data collected by the data acquisition module, and the prediction data storage unit is a memory, which is connected to the prediction module, and is used to store and transmit the data fed back by the prediction module; the data preprocessing module is connected to the prediction module, and is used for preprocessing the power grid node fault data; the prediction module is a processor with a deep hybrid neural network in memory, and the deep hybrid neural network is composed of a multi-layer perceptron (MLP) with a graph neural network (GNN) connected in series with N layers of cascade structure, which is used to predict or judge the fault condition of the power grid node.

[0006] Combination Figure 2 , characterized in that it comprises the following steps:

[0007] S1: The data acquisition module collects the fault data of the power grid nodes in the area and stores it in the power grid node data storage unit;

[0008] S2: The data preprocessing module cleans the grid node fault data in the grid node data storage unit and analyzes the attributes of the grid node fault and the relationship between the node and the attribute;

[0009] S3: The data preprocessing module uses encoding methods to convert the attributes of power grid nodes, power grid node faults, and the relationship between nodes and attributes into a knowledge graph;

[0010] S4: The prediction module uses GNN to extract node information in the knowledge graph for aggregation and outputs the final node representation;

[0011] S5: The prediction module constructs the loss function of the MLP with an N-layer cascade structure, uses the final node representation of the historical power grid node fault data for training, stores the obtained parameters of the deep hybrid neural network in the prediction data storage unit, and updates the parameters in a recursive manner;

[0012] S6: The prediction module uses the trained deep hybrid neural network to process and analyze the grid node data collected by the data acquisition module in real time, and predict or judge the fault conditions of the grid nodes.

[0013] Each piece of power grid node fault data at least includes: node ID, node current data, node voltage data, number of tributaries of the node, service life of the node circuit, and node fault event.

[0014] Further, the step S2 is specifically as follows: the data preprocessing module counts the grid node fault data in the grid node data storage unit, counts the probability of occurrence of fault events corresponding to all nodes, and after cleaning the grid node fault data corresponding to the extremely low probability fault events, all fault events corresponding to the remaining grid node fault data constitute the attributes of the grid node fault; the clustering model is used to perform hierarchical clustering of the node fault events on the remaining grid node fault data, that is, the level number of the node fault event is set, and the node current data corresponding to the node ID, the node voltage data, the number of tributaries of the node, and the service life of the node circuit are used as inputs of the clustering model.

[0015] Preferably, the clustering model is a K-means algorithm.

[0016] Furthermore, the knowledge graph described in step S3 is in the form of triples<n,a,r> , where n represents the grid node, corresponding to the node ID; a represents the attribute of the grid node fault, corresponding to the node current data, node voltage data, number of tributaries of the node, service life of the node circuit, and node fault event; r represents the relationship between the node and the attribute, corresponding to the level number of the node fault event.

[0017] The step S3 is specifically: encoding the attributes of the grid node, the grid node fault, and the relationship between the node and the attribute to obtain the corresponding encoding vector Concatenate all the encoding vectors to get the initial matrix of the knowledge graph in, d is an artificially set encoding dimension; N1, N2, and N3 represent the number of grid nodes, the number of attributes of grid node failures, and the number of relationships between nodes and attributes, respectively.

[0018] Preferably, the encoding method described in step S3 is a random encoding method.

[0019] Further, the step S4 is specifically as follows:

[0020] S401: Set the number of attributes of candidate power grid nodes NS, and randomly select any power grid node n i (where 1≤i≤N1) attributes are sampled to obtain attributes of a fixed number of power grid nodes as candidate attributes of the power grid nodes.

[0021] The sampling described in step S401 is specifically as follows: (1) When the number of attributes TS of the power grid node is greater than or equal to NS, NS attributes are randomly selected from the corresponding attributes as candidate attributes of the power grid node, and the relationship between the power grid node and the candidate attributes of the power grid node is established; (2) When the number of attributes TS of the power grid node is less than NS, (NS-TS) attributes are randomly and repeatedly selected from the existing TS attributes, associated with the power grid node, as candidate attributes of the power grid node, and the relationship between the power grid node and the candidate attributes of the power grid node is established.

[0022] S402: Establish an L-layer GNN, with the input being the initial matrix E of the knowledge graph g .

[0023] S403: Repeat step S401 to update the node information of each layer of GNN to achieve node information aggregation.

[0024] Specifically, the node information of any (l)th layer (1≤l≤L) in step S403 is aggregated as follows: Where: σ(·) represents the ReLU activation function; Represents a splicing operation; is a trainable weight matrix; is a bias vector; is the grid node n i The node information in the (l-1)th layer of GNN is initially Aggregate functions is the grid node n in the (l-1)th layer of GNN i Candidate attribute a i Embedding vector of Ns(n i ) is the grid node n i The candidate attributes of the grid node n constitute the sampling set to which they belong; i The relationship between the candidate attributes i Corresponding semantic feature scores sum[·] is the sum of vector elements; is the grid node n in the (l-1)th layer of GNN i The relationship between the candidate attributes i The embedding vector of is a trainable weight matrix; is a bias vector; p is the number of fully connected layers in the (l)th layer of GNN, is element-wise multiplication.

[0025] S404: Concatenate the final representations of all nodes to obtain the output final node representation as Among them, for the grid node n i The final node is represented as

[0026] Furthermore, the prediction output of any i-th layer of the N-layer cascaded MLP described in step S5 is And the input is satisfy Among them, w i , b i is the trainable weight matrix and bias vector of this layer; the loss function can be: Lc(·) is the cross entropy validation function.

[0027] Preferably, the loss function in step S5 may be: in, p i is the predicted value of the MLP at the i-th layer The probability of being consistent with the true label y; β and γ are set parameters; pci is the predicted value of MLP at layer i The number of elements that are the same as the true label y vector; Ts is an artificially set positive integer threshold.

[0028] Usually γ=1, β=2 are selected.

[0029] Preferably, the power grid node fault data will be divided in a ratio of 8:1:1, 80% of the data set is used as a training set for training the model; 10% of the data set is used as a validation set to determine the final number of layers of the model; and 10% of the data set is used as a test set to test the performance of the model.

[0030] Furthermore, the number of layers N of the cascade structured MLP depends on the prediction accuracy of the power grid node fault obtained through the validation set, and the initial value is a positive integer greater than 2. When the prediction accuracy in the validation set increases, the number of layers of the hybrid neural network will be deepened until the prediction accuracy no longer increases or decreases; that is, Acc N Represents the prediction accuracy of power grid node failure obtained through the validation set when the number of layers is N.

[0031] The beneficial effects of the present invention are as follows: the present invention provides a method and device for diagnosing power grid node faults based on a deep hybrid neural network, and uses a deep hybrid neural network to model the knowledge graph composed of power grid nodes, so as to accurately predict or judge the fault conditions of power grid nodes; wherein, based on a multi-level deep hybrid neural network, GNN is first used to accurately express the semantics of power grid connection fault data, and then a multi-layer MLP is used in combination with cross-validation learning, which greatly improves the model learning ability and improves the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to make the purpose and technical solution of the present invention, the present invention provides the following drawings for explanation:

[0033] Figure 1 It is a framework diagram of a power grid node fault diagnosis device based on a deep hybrid neural network of the present invention;

[0034] Figure 2 It is a flow chart of the power grid node fault diagnosis method based on deep hybrid neural network of the present invention;

[0035] Figure 3 This is a schematic diagram of the power supply system in the substation area in Example 1 of the present invention;

[0036] Figure 4 This is a diagram of the working principle of the deep hybrid neural network in Example 2 of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose and technical solution of the present invention more clearly understood, the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0038] Example 1: Combination Figure 3 , the power supply system in a certain area is monitored as a whole by the master station system, which has k=8 substations under its jurisdiction, and each substation supplies power to several users. An electric energy meter is installed at the user terminal, and the data is uploaded to the collector through RS485 communication, and then the data is uploaded to the concentrator by power line carrier communication, and finally the data is sent to the master station system by wireless communication. The master station system, substations, and users are regarded as grid nodes, and the service life of the circuits between substations and within substations in the region is the installation life of the circuit. When a circuit fails, the electric energy meter will feedback the corresponding node ID, the node current data, the node voltage data, the number of tributaries of the node, and the node fault event. In order to assist and effectively realize the troubleshooting and timely fault identification of power grid node faults in the region, the present invention provides a "power grid node fault diagnosis device based on deep hybrid neural network".

[0039] Combination Figure 1, specifically comprising: a data acquisition module (1), a data storage module (2), a data preprocessing module (3) and a prediction module (4); the data acquisition module (1) is a corresponding concentrator, which reads data from sensors of power grid nodes connected thereto through RS485 communication and power line carrier communication to realize the acquisition of power grid node fault data, wherein the sensor is an electric energy meter and the power grid node corresponds to a user end. The data storage module (2) is composed of a power grid node data storage unit (21) and a prediction data storage unit (22); the power grid node data storage unit (21) is part of the memory of the main station system, and is respectively connected to the data acquisition module (1) and the data preprocessing module (3) through wireless communication, and is used to store and transmit the power grid node fault data collected by the data acquisition module (1); the prediction data storage unit (22) is part of the memory of the main station system, and is connected to the prediction module (4), and is used to store and transmit the data fed back by the prediction module (4); the data preprocessing module (3) is connected to the prediction module (4), and is used for preprocessing the power grid node fault data; the prediction module (4) is a processor with a deep hybrid neural network in the main station system, and the deep hybrid neural network is composed of an MLP with a GNN series N-layer cascade structure, and is used to predict or judge the fault condition of the power grid node.

[0040] Example 2: Based on the device of Example 1, the present invention also provides a "grid node fault diagnosis method based on deep hybrid neural network" to predict or judge the fault condition of the grid node.

[0041] Combination Figure 2 , which specifically includes the following steps:

[0042] Step 1: The data collection module (1) collects fault data of power grid nodes in the area and stores the data in the power grid node data storage unit (21).

[0043] The service life of the node circuit has been manually extracted in advance and stored in the power grid node data storage unit (21). During implementation, it is only necessary to splice the other data corresponding to the power grid node ID.

[0044] Each piece of power grid node fault data at least includes: node ID, node current data, node voltage data, number of tributaries of the node, service life of the node circuit, and node fault event.

[0045] Step 2: The data preprocessing module (3) cleans the grid node fault data in the grid node data storage unit (21), and analyzes the attributes of the grid node fault and the relationship between the node and the attribute.

[0046] Further, specifically: the data preprocessing module (3) performs statistics on the grid node fault data in the grid node data storage unit (21), and the probability of occurrence of fault events corresponding to all nodes is counted. After the grid node fault data corresponding to the extremely low probability fault events are cleaned, all the fault events corresponding to the remaining grid node fault data constitute the attributes of the grid node fault; the K-means algorithm model is used to perform hierarchical clustering of the node fault events on the remaining grid node fault data, that is, the level number of the node fault events is set, and the node current data corresponding to the node ID, the node voltage data, the number of tributaries of the node, and the service life of the node circuit are used as inputs of the K-means algorithm model.

[0047] Step 3: Combination Figure 4 ,The data preprocessing module (3) uses a random coding method to convert the attributes of ,grid node faults and the relationship between nodes and attributes into a ,knowledge graph.

[0048] The knowledge graph is in the form of triples<n,a,r> , where n represents the grid node, corresponding to the node ID; a represents the attribute of the grid node fault, corresponding to the node current data, node voltage data, number of tributaries of the node, service life of the node circuit, and node fault event; r represents the relationship between the node and the attribute, corresponding to the level number of the node fault event.

[0049] Furthermore, step three is specifically as follows: randomly encode the attributes of the grid nodes, grid node faults, and the relationship between the nodes and the attributes to obtain the corresponding encoding vector Concatenate all the encoding vectors to get the initial matrix of the knowledge graph in, d is an artificially set encoding dimension; N1, N2, and N3 represent the number of grid nodes, the number of attributes of grid node failures, and the number of relationships between nodes and attributes, respectively.

[0050] Step 4: Combination Figure 4 , the prediction module (4) uses GNN to extract node information in the knowledge graph for aggregation and outputs the final node representation. Specifically:

[0051] S401: Set the number of attributes of candidate power grid nodes NS, and randomly select any power grid node n i (where 1≤i≤N1) attributes are sampled to obtain attributes of a fixed number of power grid nodes as candidate attributes of the power grid nodes.

[0052] The sampling described in step S401 is specifically as follows: (1) When the number of attributes TS of the power grid node is greater than or equal to NS, NS attributes are randomly selected from the corresponding attributes as candidate attributes of the power grid node, and the relationship between the power grid node and the candidate attributes of the power grid node is established; (2) When the number of attributes TS of the power grid node is less than NS, (NS-TS) attributes are randomly and repeatedly selected from the existing TS attributes, associated with the power grid node, as candidate attributes of the power grid node, and the relationship between the power grid node and the candidate attributes of the power grid node is established.

[0053] S402: Establish an L-layer GNN, with the input being the initial matrix E of the knowledge graph g .

[0054] S403: Repeat step S401 to update the node information of each layer of GNN to achieve node information aggregation.

[0055] Specifically, the node information of any (l)th layer (1≤l≤L) in step S403 is aggregated as follows: Where: σ(·) represents the ReLU activation function; Represents a splicing operation; is a trainable weight matrix; is a bias vector; is the grid node n i The node information in the (l-1)th layer of GNN is initially Aggregate functions is the grid node n in the (l-1)th layer of GNN i The embedding vector of the candidate attribute ai; Ns(n i ) is the grid node n i The candidate attributes of the grid node n constitute the sampling set to which they belong; i The semantic feature score corresponding to the relationship ri between the candidate attributes sum[·] is the sum of vector elements; is the grid node n in the (l-1)th layer of GNN i The embedding vector of the relation r between the candidate attribute; is a trainable weight matrix; is a bias vector; p is the number of fully connected layers in the (l)th layer of GNN, is element-wise multiplication.

[0056] S404: Concatenate the final representations of all nodes to obtain the output final node representation as Among them, for the grid node n i The final node is represented as

[0057] Step 5: Combine Figure 4 The prediction module (4) constructs a loss function of an N-layer cascaded MLP, uses the final node representation of historical power grid node fault data for training, stores the obtained parameters of the deep hybrid neural network in the prediction data storage unit (22), and recursively updates the parameters.

[0058] The predicted output of any i-th layer of the N-layer cascaded MLP is And the input is satisfy Among them, w i , b i is the trainable weight matrix and bias vector for this layer.

[0059] The loss function can be: in, p i is the predicted value of the MLP at the i-th layer The probability of being consistent with the true label y; γ = 1, β = 2; pci is the predicted value of MLP at layer i The number of elements that are the same as the true label y vector; Ts is an artificially set positive integer threshold.

[0060] The grid node fault data will be divided in a ratio of 8:1:1, 80% of the data set will be used as a training set for training the model; 10% of the data set will be used as a validation set to determine the final number of layers of the model; and 10% of the data set will be used as a test set to test the performance of the model.

[0061] The number of layers N of the cascade structured MLP depends on the prediction accuracy of the power grid node fault obtained through the validation set, and the initial value is a positive integer greater than 2. When the prediction accuracy in the validation set increases, the number of layers of the hybrid neural network will be deepened until the prediction accuracy no longer increases or decreases; that is, Acc N Represents the prediction accuracy of power grid node failure obtained through the validation set when the number of layers is N.

[0062] Step 6: The prediction module (4) uses the trained deep hybrid neural network to process and analyze the grid node data collected in real time by the data acquisition module (1) to predict or determine the fault condition of the grid node.

[0063] In the experiment of historical data set of regional power grid, some parameters are set as follows: the number of training iterations is 100, the learning rate is 0.01, NS=10, the random coding dimension is d=128; in order to avoid over-smoothing, the number of hidden layers of GNN network is 1, the fully connected layer is p=2, MLP is a single-layer structure, and a total of 7 MLP cascades. After experimental verification, the present invention can effectively improve the accuracy of prediction and judgment of fault conditions of power grid nodes.

[0064] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A power grid node fault diagnosis method based on deep hybrid neural network, characterized in that: The method comprises the following steps: S1: The data acquisition module (1) acquires the fault data of the power grid nodes in the area and stores it in the power grid node data storage unit (21); S2: The data preprocessing module (3) cleans the grid node fault data in the grid node data storage unit (21), and analyzes the attributes of the grid node fault and the relationship between the node and the attribute; S3: The data preprocessing module (3) uses an encoding method to convert the attributes of the power grid nodes, the power grid node faults, and the relationship between the nodes and the attributes into a knowledge graph; S4: The prediction module (4) uses GNN to extract node information in the knowledge graph, aggregates it, and outputs the final node representation; S5: The prediction module (4) constructs a loss function of an N-layer cascade structured MLP, uses the final node representation of the historical power grid node fault data for training, stores the obtained parameters of the deep hybrid neural network in the prediction data storage unit (22), and recursively updates the parameters; S6: The prediction module (4) uses the trained deep hybrid neural network to process and analyze the grid node data collected in real time by the data acquisition module (1), and predict or determine the fault condition of the grid node; Wherein, each piece of grid node fault data at least includes: node ID, node current data, node voltage data, number of tributaries of the node, service life of node circuit, and node fault event; The step S2 is specifically as follows: the data preprocessing module (3) performs statistics on the grid node fault data in the grid node data storage unit (21), and the probability of occurrence of fault events corresponding to all nodes is counted. After the grid node fault data corresponding to the extremely low probability fault events are cleaned, all the fault events corresponding to the remaining grid node fault data constitute the attributes of the grid node fault; the remaining grid node fault data is hierarchically clustered into node fault events using a clustering model, that is, the level number of node fault events is set, and the node current data corresponding to the node ID, the node voltage data, the number of tributaries of the node, and the service life of the node circuit are used as inputs of the clustering model; The knowledge graph described in step S3 is in the form of triples<n,a,r> , where n represents the grid node, corresponding to the node ID; a represents the attribute of the grid node fault, corresponding to the node current data, node voltage data, number of tributaries of the node, node circuit service life, node fault event; r represents the relationship between the node and the attribute, corresponding to the level number of the node fault event; The step S3 is specifically: encoding the attributes of the grid node, the grid node fault, and the relationship between the node and the attribute to obtain the corresponding encoding vector Concatenate all the encoding vectors to get the initial matrix of the knowledge graph in, d is an artificially set encoding dimension; N1, N2, and N3 represent the number of grid nodes, the number of attributes of grid node failures, and the number of relationships between nodes and attributes, respectively; The step S4 is specifically as follows: S401: Set the number of attributes of candidate power grid nodes NS, and randomly select any power grid node n i The attributes of the grid nodes are sampled to obtain the attributes of a fixed number of grid nodes as candidate attributes of grid nodes; where 1≤i≤N1; S402: Establish an L-layer GNN, with the input being the initial matrix E of the knowledge graph g ; S403: Repeat step S401 to update the node information of each layer of GNN to achieve node information aggregation; S404: Concatenate the final representations of all nodes to obtain the output final node representation as Among them, for the grid node n i The final node is represented as The sampling described in step S401 is specifically as follows: (1) when the number of attributes TS of the power grid node is greater than or equal to NS, randomly select NS attributes from the corresponding attributes as candidate attributes of the power grid node, and establish the relationship between the power grid node and the candidate attributes of the power grid node; (2) when the number of attributes TS of the power grid node is less than NS, randomly select NS-TS attributes from the existing TS attributes repeatedly, associate them with the power grid node, and establish the relationship between the power grid node and the candidate attributes of the power grid node; In step S403, the node information of any layer l is aggregated into Where: 1≤l≤L; σ(·) represents the ReLU activation function; Represents a splicing operation; is a trainable weight matrix; is a bias vector; is the grid node n i The node information in the l-1th layer of GNN is initially Aggregate functions is the grid node n in the (l-1)th layer of GNN i Candidate attribute a i Embedding vector of Ns(n i ) is the grid node n i The candidate attributes of the grid node n constitute the sampling set to which they belong; i The relationship between the candidate attributes i Corresponding semantic feature scores sum[·] is the sum of vector elements; is the grid node n in the l-1th layer of GNN i The relationship between the candidate attributes i The embedding vector of is a trainable weight matrix; is a bias vector; p is the number of fully connected layers in the (l)th layer of GNN, and ⊙ is the element-by-element multiplication; The prediction output of any i-th layer of the N-layer cascaded MLP described in step S5 is And the input is satisfy Among them, w i 、b i is the trainable weight matrix and bias vector of this layer; the loss function can be: Lc(·) is the cross entropy validation function.

2. The power grid node fault diagnosis method based on deep hybrid neural network according to claim 1 is characterized in that: The clustering model described in step S2 is the K-means algorithm.

3. The power grid node fault diagnosis method based on deep hybrid neural network according to claim 1 is characterized in that: The encoding method described in step S3 is a random encoding method.

4. The power grid node fault diagnosis method based on deep hybrid neural network according to claim 1 is characterized in that: The loss function described in step S5 may be: in, p i is the predicted value of the MLP at the i-th layer The probability of being consistent with the true label y; β and γ are set parameters; pc i is the predicted value of the MLP at the i-th layer The number of elements that are the same as the true label y vector; Ts is an artificially set positive integer threshold.

5. The power grid node fault diagnosis method based on deep hybrid neural network according to claim 1 is characterized in that: The grid node fault data will be divided in a ratio of 8:1:1, and 80% of the data set will be used as a training set for training the model; 10% of the dataset was used as a validation set to determine the final number of layers of the model, and 10% of the dataset was used as a test set to test the performance of the model.

6. The power grid node fault diagnosis method based on deep hybrid neural network according to claim 5 is characterized in that: The number of layers N of the cascade structured MLP depends on the prediction accuracy of the power grid node failure obtained through the validation set, and the initial value is a positive integer greater than 2. When the prediction accuracy in the validation set increases, the number of layers of the hybrid neural network will be deepened until the prediction accuracy no longer increases or decreases; that is, Acc N Represents the prediction accuracy of power grid node failure obtained through the validation set when the number of layers is N.

7. A power grid node fault diagnosis device based on a deep hybrid neural network based on the method according to any one of claims 1 to 6, characterized in that: The device comprises a data acquisition module (1), a data storage module (2), a data preprocessing module (3) and a prediction module (4); the data acquisition module (1) reads data from a grid node sensor connected thereto through communication to realize the acquisition of grid node fault data; the data storage module (2) is composed of a grid node data storage unit (21) and a prediction data storage unit (22); the grid node data storage unit (21) is a memory, connected to the data acquisition module (1) and the data preprocessing module (3) respectively, and used for storing and transmitting grid node fault data collected by the data acquisition module (1); the prediction data storage unit (22) is a memory, connected to the prediction module (4), and used for storing and transmitting data fed back by the prediction module (4); the data preprocessing module (3) is connected to the prediction module (4) and used for preprocessing grid node fault data; the prediction module (4) is a processor with a deep hybrid neural network in the memory, and the deep hybrid neural network is composed of an MLP with a GNN series N-layer cascade structure, and is used for predicting or judging the fault condition of the grid node.

Images