Graph neural network method and system for power generation multi-equipment linkage analysis

Abstract

The invention discloses a graph neural network method and system for power generation multi-device linkage analysis, and the method comprises the following steps: S1, constructing a graph data structure according to a power generation multi-device linkage relationship; s2, endowing each node with a feature vector and initializing the feature vector, wherein the vector comprises running state information of corresponding equipment; s3, establishing a graph neural network model which is used for processing the graph data structure and learning interaction between devices; s4, inputting the graph data structure and the node feature vector into the graph neural network model; and S5, outputting a linkage analysis result of each device through the graph neural network model. According to the method, a graph data structure comprehensively considering physical connection and operation logic between equipment is constructed, rich feature vectors are given to each equipment node, interaction between the equipment is deeply learned through the graph neural network, and the accuracy and efficiency of analysis and optimization of the power generation system are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of linkage of multiple power generation devices based on graph neural network model, and more specifically, it particularly relates to a graph neural network method for linkage analysis of multiple power generation devices. At the same time, the present invention also relates to a graph neural network system for linkage analysis of multiple power generation devices. Background Art

[0002] With the rapid development of the energy industry, power generation systems are becoming more and more complex, involving the close linkage of multiple devices and processes. In China, with the advancement of smart grids and smart energy systems, higher requirements are placed on the monitoring and optimization control of power generation equipment. Traditional power generation equipment management and optimization methods mainly rely on the independent monitoring and control of individual devices, lacking a deep understanding and comprehensive optimization of the interactions between devices in the entire system. This approach shows limitations in dealing with complex interactions between devices, resulting in inefficiency and delayed response.

[0003] In recent years, graph neural networks have been applied in many fields due to their excellent performance, especially their high efficiency in processing graph structured data. However, there is no precedent for applying graph neural networks to the analysis of multi-device linkage in power generation systems, especially the lack of a systematic method to build data models, initialize device features, and effectively learn the interactions between devices.

[0004] For example, Chinese patent document CN106568415A discloses a wind power tower foundation settlement monitoring system and monitoring method, which monitors the equipment in real time through distributed sensors and data acquisition systems. The patent covers a data collection method for collecting and analyzing equipment operation data. However, the method focuses on the monitoring of a single device, rather than understanding and optimizing the interaction between devices as a whole;

[0005] Another example is Chinese patent document CN110113379A, which proposes a microgrid monitoring system based on the OneNet platform and its implementation method for real-time monitoring and analysis of the operating status of power generation equipment. Although the system can collect and analyze data from multiple devices, it cannot use graph neural networks to learn and analyze the complex relationships between devices;

[0006] Therefore, we propose a graph neural network method and system for the linkage analysis of multiple devices in power generation, which can effectively identify and predict the key linkage effects in the power generation system, thereby guiding more accurate and efficient system optimization control strategies. Summary of the invention

[0007] The purpose of the present invention is to solve the shortcomings existing in the prior art, and the proposed graph neural network method and system for linkage analysis of multiple power generation devices not only constructs a graph data structure that comprehensively considers the physical connection and operation logic between devices, but also gives each device node rich feature vectors, and deeply learns the interaction between devices through the graph neural network. At the same time, an attention mechanism is introduced, which allows the model to assign different weights according to the importance of neighboring nodes, further enhancing the processing ability and expressiveness of the model, and can better process large-scale graph data. It is suitable for inductive learning scenarios, significantly improving the accuracy and efficiency of power generation system analysis and optimization.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The graph neural network method for linkage analysis of multiple power generation devices includes the following steps:

[0010] S1. Construct a graph data structure based on the linkage relationship between multiple power generation devices, where nodes represent devices in the power generation system and edges represent the connection relationship between devices;

[0011] S2. Assign a feature vector to each node and initialize it. The feature vector contains the operating status information of the corresponding device.

[0012] S3. Establishing a graph neural network model, which is used to process the graph data structure and learn the interactions between devices;

[0013] S4, inputting the graph data structure and node feature vector into the graph neural network model;

[0014] S5. Output the linkage analysis results of each device through the graph neural network model, and the results are used to guide the optimal control of the power generation system.

[0015] Preferably, step S1 specifically comprises:

[0016] S11, determine the vertex number, and number several devices as 1, 2, ..., n respectively;

[0017] S12. Construct an adjacency matrix, initialize an n×n matrix, where n is the number of devices, traverse all device pairs, and check whether there is a physical connection. For an undirected graph, the adjacency matrix is ​​symmetric; for a directed graph, it is not necessarily symmetric;

[0018] Among them, the element definition of the adjacency matrix A is expressed as:

[0019]

[0020] Where E is the set of edges in the graph. If vertex v i and vertex v jIf there is an edge between them, then the corresponding matrix element a ij is 1; if vertex v i and vertex v j If there is no edge between them, then the corresponding matrix element a ij is 0;

[0021] S13. Implement an adjacency matrix, and then update the values ​​in the matrix according to the connection status between devices.

[0022] Preferably, in step S2, in the graph data structure, each node represents an entity, and features need to be determined for each node, the features including device temperature, pressure, and operating time;

[0023] Combine the features of all nodes into a feature matrix X. If the graph contains N nodes and each node has D features, then the feature matrix X will be N×D dimensional;

[0024] The connection relationship of the graph is represented by the adjacency matrix A. For an undirected graph, the adjacency matrix is ​​symmetric, but for a directed graph it is not necessarily symmetric.

[0025] Preferably, the specific process of step S3 is:

[0026] S31. Determine the number of layers of the graph neural network and the number of neurons in each layer;

[0027] S32, initialize the weight matrix W(l) and the bias vector b(l), where l represents the lth layer;

[0028] S33. For each layer, perform the following operations:

[0029] S331, calculate the product of the input matrix X(l) of the current layer and the weight matrix W(l), and add the bias vector b(l) to obtain the output matrix Y(l) of the current layer;

[0030] S332, processing the output matrix Y(l) through the activation function σ to obtain the hidden state matrix H(l) of the current layer;

[0031] S333, using the hidden state matrix H(l) as the input matrix X(l+1) of the next layer, and repeating steps S331-S332 until all layers are processed;

[0032] S34. Use the last layer hidden state matrix H(L) as the output of the entire graph neural network model.

[0033] Preferably, during the establishment of the graph network neural model, the unstandardized matrix is ​​calculated using the following formula:

[0034] L = DA;

[0035] In the formula, A represents the adjacency matrix of the graph, and D is the degree matrix, which means that the elements on the diagonal of the matrix are the degree of each node, which is the number of edges connected to the node;

[0036] Symmetric normalized matrix, the expression is:

[0037]

[0038] Where L is the unstandardized matrix, D is the degree matrix, and in the symmetric standardized matrix, the factor It is a diagonal matrix consisting of the square root elements of the inverse matrix of the degree matrix D;

[0039] The graph convolution operation is performed on the feature matrix and the adjacency matrix to integrate the state information of each device and its neighbors, which is expressed as:

[0040]

[0041] In the formula, H (l) is the hidden state matrix of the lth layer, where each row represents the feature vector of a node; W (l) is the weight matrix from layer l to layer l+1, used for linear transformation; σ is the activation function, used to introduce nonlinearity; is the inverse square root of the degree matrix D; is a symmetric normalized matrix

[0042] Preferably, step S4 is specifically:

[0043] S41, check that the graph data structure has been correctly constructed, including the adjacency matrix A and the node feature matrix X. For each node, ensure that its feature vector has been initialized according to step S2 and contains the operating status information of the corresponding device;

[0044] S42, taking the adjacency matrix A, the degree matrix D, and the node feature matrix X as inputs and passing them into the graph neural network model;

[0045] S43, forward propagation is performed through the defined graph convolution operation, and nonlinearity is introduced through the activation function to update the feature vector H of the node, including multiple rounds of iterations, each round of iteration will update the feature vector of the current node according to the information of the neighboring nodes until a predetermined number of layers is reached;

[0046] S44. After the graph convolution operation, the feature vector of each node will contain rich local and global graph structure information, and the final node feature matrix H will be used as the output of the model.

[0047] Preferably, in step S4, each node updates its own state by exchanging information with its neighbor nodes, and the expression is:

[0048]

[0049] In the formula, is the state of node i at time t, f is the update function, AGG j It is an aggregation function of neighbor node status;

[0050] The graph neural network model inductive learning adopts an aggregation method, which allows the graph neural network model to process unseen nodes and edges. The expression is:

[0051]

[0052] In the formula, N(v i ) represents node v i The set of neighbor nodes, W agg is the learnable weight matrix, h j is node v j The eigenvector of agg represents the feature vector of the target node after aggregation, and σ represents the activation function;

[0053] The graph neural network model introduces the attention mechanism based on the aggregation method. The attention mechanism allows the model to assign different weights according to the importance of neighboring nodes. According to the above expression, it is updated as follows:

[0054]

[0055] In the formula, a ij is the attention weight;

[0056] The above method can efficiently process large-scale graphs and is suitable for inductive learning scenarios. The attention mechanism gives the model stronger expressive power and automatically learns the importance of neighbor nodes.

[0057] A graph neural network system for linkage analysis of multiple power generation devices, the system is used to implement the above method, including:

[0058] Data collection and preprocessing module, which is responsible for collecting real-time operating data of power generation equipment, including temperature, pressure, and operating time, and performing preliminary cleaning and standardization processing; data is obtained through sensors and acquisition equipment; the preprocessing includes removing outliers, filling missing values, and normalization processing to ensure data quality;

[0059] The graph construction and feature initialization module constructs a graph data structure based on the physical connection and operation logic of the power generation equipment. Each device is a node, and the connection relationship between devices constitutes an edge. At the same time, this module is used to assign a feature vector to each node, which contains the operating status information of the device. After these feature vectors are initialized, they will be used as the input of the graph neural network;

[0060] A graph neural network processing module, which uses a graph convolutional network to learn the interactions between devices. The graph neural network processing module accepts an adjacency matrix, a degree matrix, and a feature matrix as inputs, and gradually extracts high-order interaction features of devices through multi-level graph convolution operations;

[0061] The model training and optimization module is responsible for the training, verification and testing of the graph neural network model. The model training and optimization module uses back propagation and gradient descent methods to optimize model parameters.

[0062] Result output and application module: the model outputs the linkage analysis results of each device, including equipment status prediction, abnormal warning, and performance evaluation. The results will be directly applied to the optimization control of the power generation system;

[0063] The user interaction and visualization module provides a user-friendly interface, allowing operators to easily set analysis parameters, view and understand analysis results, and integrate data visualization tools to intuitively display equipment status and linkage effects.

[0064] Preferably, the data collection and preprocessing module respectively performs data collection, data cleaning, data standardization, data integration, and data output operations, specifically:

[0065] Data collection: sensors deployed on power generation equipment continuously monitor key indicators such as temperature, pressure, and operating time, and data collection equipment collects data regularly or in real time;

[0066] Data cleaning, identifying and removing outliers and filling missing values, using the mean or median of adjacent data points or estimating missing data through interpolation methods;

[0067] Data standardization: The data is normalized so that it falls within the standard range. The expression is:

[0068]

[0069] Among them, x is the original data point, x min and x max are the minimum and maximum values ​​of the data respectively;

[0070] Data integration: The processed data is integrated into an adjacency matrix and a feature matrix. The adjacency matrix reflects the connection relationship between devices, and the feature matrix includes the operating status information of the devices. The matrix construction expression is:

[0071]

[0072] In the formula, t i ,p i ,r irepresent the temperature, pressure and operating time of the i-th equipment respectively;

[0073] Data output: output the preprocessed data set for further analysis and learning by the graph neural network model.

[0074] The graph neural network method and system for linkage analysis of multiple power generation devices provided by the present invention have the following advantages compared with the conventional technology:

[0075] Improve the accuracy of analysis. By constructing a graph data structure, the devices in the power generation system are used as nodes, and the connection relationships between devices are used as edges. The present invention can more accurately simulate and analyze the interactions between devices. The graph neural network model can identify high-order interaction features between devices by learning these relationships, thereby providing more accurate analysis results than traditional methods.

[0076] Enhanced prediction capability: By assigning rich feature vectors to each node, such as temperature, pressure, and operating time, the present invention enables the graph neural network to comprehensively consider various operating status information; this method improves the model's ability to predict equipment status, abnormal warning, and performance evaluation;

[0077] Optimize resource allocation. By analyzing the linkage effect between devices, the present invention can guide more accurate optimization control strategies of power generation systems, such as adjusting equipment operation strategies or preventive maintenance arrangements, thereby achieving optimal resource allocation and reducing operation and maintenance costs;

[0078] Improve system efficiency. The graph neural network model of the present invention can process large-scale graph data and is suitable for inductive learning scenarios. Even in a constantly changing power generation system, the model can adapt to new devices and connection relationships and continuously improve the operating efficiency of the system.

[0079] Fault diagnosis and prevention: By introducing the attention mechanism, the present invention can identify the devices or connections that have a greater impact on system operation and give priority to monitoring these key nodes; it helps to detect potential faults in a timely manner and prevent them, reducing downtime and maintenance costs;

[0080] User-friendly interaction and visualization: The user interaction and visualization module provided by the present invention enables operators to easily set analysis parameters, view and understand analysis results; the integrated data visualization tool intuitively displays device status and linkage effects, improving user experience and operational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 This is a flow chart of the graph neural network method for linkage analysis of multiple power generation devices of the present invention;

[0082] Figure 2 A flowchart of constructing a graph data structure in step S1 of the embodiment;

[0083] Figure 3 This is a graph neural network system architecture diagram for the linkage analysis of multiple power generation devices in the present invention. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0085] The present invention provides a graph neural network method and system for linkage analysis of multiple power generation devices, which has the following effects:

[0086] Flexibility and scalability. The modular design of the present invention allows the system to be flexibly configured and expanded according to actual needs. It is easy to maintain and upgrade. The modular structure enables each part of the system to operate independently, simplifying the maintenance and upgrade process. Each module can be optimized or replaced individually without affecting the operation of the entire system. It reduces complexity. The complex power generation equipment linkage analysis task is decomposed into multiple modules. Each module is responsible for a specific function, which reduces the complexity of the system and helps to improve the stability and reliability of the system. It improves efficiency. The modular design makes parallel processing possible. Each module can work simultaneously, improving the overall processing speed and efficiency. It promotes knowledge sharing. Due to the low coupling between modules, the design and implementation knowledge of each module can be shared and reused independently of other modules, which promotes the accumulation of knowledge and the dissemination of technology.

[0087] Based on the above, if Figure 1 As shown in FIG. 1 , the graph neural network method for linkage analysis of multiple power generation devices includes the following steps:

[0088] S1. Construct a graph data structure based on the linkage relationship between multiple power generation devices, where nodes represent devices in the power generation system and edges represent the connection relationship between devices;

[0089] Further, such as Figure 2 As shown, step S1 is specifically as follows:

[0090] S11, determine the vertex number, and number several devices as 1, 2, ..., n respectively;

[0091] S12. Construct an adjacency matrix, initialize an n×n matrix, where n is the number of devices, traverse all device pairs, and check whether there is a physical connection. For an undirected graph, the adjacency matrix is ​​symmetric; for a directed graph, it is not necessarily symmetric;

[0092] Among them, the element definition of the adjacency matrix A is expressed as:

[0093]

[0094] Where E is the set of edges in the graph. If vertex v i and vertex v j If there is an edge between them, then the corresponding matrix element a ij is 1; if vertex v i and vertex v j If there is no edge between them, then the corresponding matrix element a ij is 0;

[0095] S13. Implement an adjacency matrix, and then update the values ​​in the matrix according to the connection status between devices.

[0096] For example, if the power generation equipment is set to 4 groups of equipment, the equipment is numbered 1, 2, 3, and 4 respectively;

[0097] Initialize a 4×4 matrix, because the number of devices n = 4. Initially, all elements are set to 0, indicating no edges;

[0098] Traverse all device pairs and check whether there is a physical connection between them. If there is a connection, set the corresponding position in the adjacency matrix to 1;

[0099] For example, if device 1 is connected to device 2 and device 4, device 2 is connected to device 1 and device 3, device 3 is connected to device 2, and device 4 is connected to device 1, the following adjacency matrix A is constructed:

[0100]

[0101] Update the values ​​in the matrix based on the connections between devices. In this example, the adjacency matrix is ​​filled based on the connections between devices.

[0102] If at some point in the future, the connection between devices changes, specifically a new connection is added between device 3 and device 4, the adjacency matrix A needs to be updated, and the corresponding matrix elements a34 and a43 are set to 1, expressed as:

[0103]

[0104] In the above example, the rows and columns of the adjacency matrix represent the vertices in the graph, and the element value of the matrix indicates whether there is an edge between the corresponding two vertices. In practical applications, this adjacency matrix will be provided as input to the graph neural network model to learn the interactions between devices and the dynamic characteristics of the entire system.

[0105] S2, assigning a feature vector to each node and initializing it, the vector contains the operating status information of the corresponding device; in step S2, in the graph data structure, each node represents an entity, and it is necessary to determine features for each node, the features include device temperature, pressure, and operating time;

[0106] Combine the features of all nodes into a feature matrix X. If the graph contains N nodes and each node has D features, then the feature matrix X will be N×D dimensional;

[0107] The connection relationship of the graph is represented by the adjacency matrix A. For an undirected graph, the adjacency matrix is ​​symmetric, but for a directed graph it is not necessarily symmetric.

[0108] S3. Establish a graph neural network model, which is used to process the graph data structure and learn the interaction between devices; the specific process of step S3 is:

[0109] S31. Determine the number of layers of the graph neural network and the number of neurons in each layer;

[0110] S32, initialize the weight matrix W(l) and the bias vector b(l), where l represents the lth layer;

[0111] S33. For each layer, perform the following operations:

[0112] S331, calculate the product of the input matrix X(l) of the current layer and the weight matrix W(l), and add the bias vector b(l) to obtain the output matrix Y(l) of the current layer;

[0113] S332, processing the output matrix Y(l) through the activation function σ to obtain the hidden state matrix H(l) of the current layer;

[0114] S333, using the hidden state matrix H(l) as the input matrix X(l+1) of the next layer, and repeating steps S331-S332 until all layers are processed;

[0115] During the establishment of the graph network neural model, the unstandardized matrix is ​​calculated using the following formula:

[0116] L = DA;

[0117] In the formula, A represents the adjacency matrix of the graph, and D is the degree matrix, which means that the elements on the diagonal of the matrix are the degree of each node, which is the number of edges connected to the node;

[0118] Symmetric normalized matrix, the expression is:

[0119]

[0120] Where L is the unstandardized matrix, D is the degree matrix, and in the symmetric standardized matrix, the factor It is a diagonal matrix consisting of the square root elements of the inverse matrix of the degree matrix D;

[0121] The graph convolution operation is performed on the feature matrix and the adjacency matrix to integrate the state information of each device and its neighbors, which is expressed as:

[0122]

[0123] In the formula, H (l) is the hidden state matrix of the lth layer, where each row represents the feature vector of a node; W (l) is the weight matrix from layer l to layer l+1, used for linear transformation; σ is the activation function, used to introduce nonlinearity; is the inverse square root of the degree matrix D; is a symmetric normalized matrix.

[0124] S34. Use the last layer hidden state matrix H(L) as the output of the entire graph neural network model.

[0125] S4, inputting the graph data structure and node feature vector into the graph neural network model; step S4 is specifically as follows:

[0126] S41, check that the graph data structure has been correctly constructed, including the adjacency matrix A and the node feature matrix X. For each node, ensure that its feature vector has been initialized according to step S2 and contains the operating status information of the corresponding device;

[0127] S42, taking the adjacency matrix A, the degree matrix D, and the node feature matrix X as inputs and passing them into the graph neural network model;

[0128] S43, forward propagation is performed through the defined graph convolution operation, and nonlinearity is introduced through the activation function to update the feature vector H of the node, including multiple rounds of iterations, each round of iteration will update the feature vector of the current node according to the information of the neighboring nodes until a predetermined number of layers is reached;

[0129] S44. After the graph convolution operation, the feature vector of each node will contain rich local and global graph structure information, and the final node feature matrix H will be used as the output of the model;

[0130] It should be noted that, in step S4, each node updates its own state by exchanging information with its neighbor nodes, and the expression is:

[0131]

[0132] In the formula, is the state of node i at time t, f is the update function, AGG j It is an aggregation function of neighbor node status;

[0133] The graph neural network model inductive learning adopts an aggregation method, which allows the graph neural network model to process unseen nodes and edges. The expression is:

[0134]

[0135] In the formula, N(v i ) represents node v i The set of neighbor nodes, W agg is the learnable weight matrix, h j is node v j The eigenvector of agg represents the feature vector of the target node after aggregation, and σ represents the activation function;

[0136] The graph neural network model introduces the attention mechanism based on the aggregation method. The attention mechanism allows the model to assign different weights according to the importance of neighboring nodes. According to the above expression, it is updated as follows:

[0137] h agg =σ(∑ υj∈N(υi) a ij ·h j );

[0138] In the formula, a ij is the attention weight;

[0139] The above method can efficiently process large-scale graphs and is suitable for inductive learning scenarios. The attention mechanism gives the model stronger expressive power and automatically learns the importance of neighbor nodes.

[0140] As mentioned above, in the analysis of power generation systems, this node status update and information exchange mechanism can effectively simulate the interaction between devices; for example, when the temperature of a generator rises, it may affect the cooling system directly connected to it through the system network; by aggregating the status information of adjacent devices and weighting this information using the attention mechanism, the graph neural network model can accurately predict and analyze the impact of such temperature changes on the performance of the entire system;

[0141] In the graph neural network system for linkage analysis of multiple power generation devices, how to efficiently process large-scale graph data through node status updates, information exchange, aggregation methods and attention mechanisms, and apply them to inductive learning scenarios; these methods work together to improve the model's ability to predict equipment status, abnormal warnings and performance evaluation, thereby providing strong support for the optimal control of power generation systems.

[0142] S5. Outputting linkage analysis results of each device through the graph neural network model, and the results are used to guide the optimization control of the power generation system;

[0143] Furthermore, the linkage analysis results include device status prediction and system performance evaluation indicators, the specific form of which depends on the application scenario and requirements; specifically, the model can output the future state of each device, or the overall performance indicators of the system;

[0144] The above-mentioned graph neural network model outputs the linkage analysis results of each device, predicting how to adjust the operation strategies of other devices to maintain stable system operation, or how to adjust the equipment configuration to optimize system efficiency and reduce energy consumption when some equipment has abnormalities; the analysis results provide decision support for operators, helping them to manage and control the power generation system more effectively.

[0145] This embodiment also provides a graph neural network system for power generation multi-device linkage analysis, and the system is used to implement the above method, such as Figure 3 As shown, including:

[0146] Data collection and preprocessing module, which is responsible for collecting real-time operating data of power generation equipment, including temperature, pressure, and operating time, and performing preliminary cleaning and standardization processing; data is obtained through sensors and acquisition equipment; the preprocessing includes removing outliers, filling missing values, and normalization processing to ensure data quality;

[0147] Furthermore, the data collection and preprocessing modules respectively perform data collection, data cleaning, data standardization, data integration, and data output operations, specifically:

[0148] Data collection: sensors deployed on power generation equipment continuously monitor key indicators such as temperature, pressure, and operating time, and data collection equipment collects data regularly or in real time;

[0149] Data cleaning, identifying and removing outliers and filling missing values, using the mean or median of adjacent data points or estimating missing data through interpolation methods;

[0150] Data standardization: The data is normalized so that it falls within the standard range. The expression is:

[0151]

[0152] Among them, x is the original data point, x min and x max are the minimum and maximum values ​​of the data respectively;

[0153] Data integration: The processed data is integrated into an adjacency matrix and a feature matrix. The adjacency matrix reflects the connection relationship between devices, and the feature matrix includes the operating status information of the devices. The matrix construction expression is:

[0154]

[0155] In the formula, t i ,p i ,r i represent the temperature, pressure and operating time of the i-th equipment respectively;

[0156] Data output: output the preprocessed data set for further analysis and learning by the graph neural network model.

[0157] The graph construction and feature initialization module constructs a graph data structure based on the physical connection and operation logic of the power generation equipment. Each device is a node, and the connection relationship between devices constitutes an edge. At the same time, this module is used to assign a feature vector to each node, which contains the operating status information of the device. After these feature vectors are initialized, they will be used as the input of the graph neural network;

[0158] Furthermore, in the graph neural network system for linkage analysis of multiple power generation devices, the graph construction and feature initialization module is responsible for constructing the graph data structure according to the physical connection and operation logic of the power generation equipment, and assigning a feature vector to each node, specifically:

[0159] Graph construction:

[0160] Node definition,In the power generation system, each device is defined as a node in the graph;,generators, transformers, cooling systems, and control systems can all be,separate nodes;

[0161] Edge construction, constructs edges based on the physical connections and interactions between devices; if there is a cable connection between the generator and the transformer, an edge is created between these two nodes;

[0162] Undirected graph and directed graph, according to the connection characteristics between devices, if the connection is bidirectional, the graph is undirected, and if the connection is unidirectional, the graph is directed. In this power generation system, the interaction between devices is bidirectional, so it is usually constructed as an undirected graph;

[0163] Edge weights, which are weighted according to the strength or frequency of the interaction between devices. If two devices often work at the same time, the edge between them can be given a higher weight.

[0164] Feature initialization:

[0165] Feature selection: select features that can describe the operating status of each device node, including temperature, pressure, current, voltage, and operating time;

[0166] Feature vector construction, combining these features into a feature vector. For a generator node, its feature vector can be in the form of temperature, current, and operating time;

[0167] Feature standardization: In order to improve the training efficiency and accuracy of the model, the feature vectors need to be standardized so that their values ​​fall within a similar range;

[0168] Dynamic update: The feature vector should be dynamically updated according to the real-time data of the device to ensure that the graph neural network can learn the latest device status information;

[0169] Outlier processing,For outliers in device data, proper processing is required, including replacing or deleting them with the average value,,to avoid interfering with model training.

[0170] As mentioned above, through the graph construction and feature initialization module, the equipment of the power generation system is effectively converted into a graph data structure, which contains the physical connection relationship and operating status information between the equipment; this information provides rich input data for the graph neural network, enabling the model to learn and analyze the complex linkage effects between equipment; the output of this module will be directly used as the input of the graph neural network model for subsequent status updates, information aggregation and linkage analysis, and ultimately guide the optimization control of the power generation system.

[0171] A graph neural network processing module, which uses a graph convolutional network to learn the interactions between devices. The graph neural network processing module accepts an adjacency matrix, a degree matrix, and a feature matrix as inputs, and gradually extracts high-order interaction features of devices through multi-level graph convolution operations;

[0172] Through the graph neural network processing module, the complex interactions between devices in the power generation system are effectively learned and analyzed. The high-order interaction features provide in-depth insights into the system, helping operators understand how devices affect each other and optimize the operating strategy of the power generation system accordingly. The output of this module will be directly used to guide the optimal control of the power generation system and improve the stability and efficiency of the system.

[0173] The model training and optimization module is responsible for the training, verification and testing of the graph neural network model. The model training and optimization module uses back propagation and gradient descent methods to optimize model parameters;

[0174] Specifically,

[0175] Considering that the operating data of power generation equipment will change over time, the model training and optimization module adopts an online learning strategy to continuously learn and adapt from new data, and can regularly retrain and update the model to adapt to changes in system configuration or new operating conditions;

[0176] Through the model training and optimization module, the graph neural network model for linkage analysis of multiple power generation devices can remain efficient and accurate in a constantly changing environment. The output of this module is a fully trained and optimized model that can provide accurate linkage analysis results and guide the optimal control of the power generation system.

[0177] Result output and application module: The model outputs the linkage analysis results of each device, including device status prediction, abnormal warning, and performance evaluation. The results will be directly applied to the optimization control of the power generation system. This module is responsible for applying the linkage analysis results of each device output by the model, including device status prediction, abnormal warning, and performance evaluation, directly to the optimization control of the power generation system. Specifically:

[0178] Equipment status prediction: The model can predict the status of each device in the future, such as the expected changes in key operating parameters such as temperature and pressure. This prediction can help operators prepare in advance and adjust equipment operation plans;

[0179] Abnormal warning: By analyzing the normal operating mode of the equipment, the model can detect and warn in real time any pattern that deviates from normal behavior, thereby promptly identifying potential equipment failures or performance degradation;

[0180] Performance evaluation: The model evaluates the overall performance of the power generation system, including efficiency and energy consumption indicators, to identify bottlenecks in the system and optimize resource allocation and workflow.

[0181] Specific optimization application strategies include:

[0182] Optimization control, based on the analysis results of the model, automatically adjusts the equipment's operating parameters or operation strategies to optimize system performance, including adjusting the output power of the generator or changing the operating intensity of the cooling system;

[0183] Maintenance scheduling, using the model's abnormal warning to develop preventive maintenance plans to reduce equipment failures and downtime;

[0184] Energy management, through performance evaluation, optimizes energy consumption, such as reducing the operating load of certain equipment during off-peak hours to reduce energy costs;

[0185] Real-time monitoring: Establish a real-time monitoring system to display the model's prediction and warning results to operators in real time to provide decision support;

[0186] Feedback loop, which feeds actual application results into the model training process, continuously optimizes and adjusts the model to improve its accuracy and reliability.

[0187] Through the result output and application module, the graph neural network model for linkage analysis of multi-devices in power generation is not only a theoretical analysis tool, but has become an indispensable part of the operation and maintenance of the power generation system. The output results of this module will directly affect the operating efficiency and reliability of the power generation system, and help achieve smarter and more efficient power generation management.

[0188] The user interaction and visualization module provides a user-friendly interface, allowing operators to easily set analysis parameters, view and understand analysis results, and integrates data visualization tools to intuitively display equipment status and linkage effects; This module provides a user-friendly interface, allowing operators to easily set analysis parameters, view and understand analysis results, and integrates data visualization tools to intuitively display equipment status and linkage effects, specifically:

[0189] Visual interface:

[0190] Interactive design, using the interactive design concept, ensures that operators can set analysis parameters through intuitive click and drag operations, improving user experience;

[0191] Real-time data display: the operating status and linkage effect of the equipment are displayed in real time on the interface, including temperature curve and pressure graph, so that the operator can quickly understand the system status;

[0192] Result summary: Provides a summary of the model analysis results on the interface. The summary includes the current values ​​and forecast trends of key indicators, helping operators quickly understand key information;

[0193] Historical data comparison allows operators to view historical data and compare it with current data to better understand the changing trends of equipment performance;

[0194] Responsive layout: The interface adopts responsive design to ensure good visibility and operability on different devices and screen sizes.

[0195] Instructions:

[0196] Step-by-step instructions provide detailed step-by-step instructions to guide operators on how to set parameters, start analysis, and view results;

[0197] Help document: built-in help document explains the system functions, operation procedures and the meaning of analysis results for operators to refer to at any time;

[0198] Video tutorials, which provide video tutorials to train operators on how to use the system through practical demonstrations;

[0199] FAQ: Set up a FAQ section to centrally answer common questions that operators may encounter.

[0200] Customized reports:

[0201] Report templates: multiple report templates are provided, and operators can choose the appropriate report format according to their needs;

[0202] Automatic generation: the system automatically generates reports based on the analysis results, saving operators time in writing reports;

[0203] Export function allows operators to export reports to PDF and Excel formats for easy sharing and archiving;

[0204] Customized content: Operators can add customized content, analysis notes, recommendations and conclusions to the report.

[0205] In summary, compared with the conventional technology, the present invention has the following substantial improvements:

[0206] Improve the accuracy of analysis. By constructing a graph data structure, the devices in the power generation system are used as nodes, and the connection relationships between devices are used as edges. The present invention can more accurately simulate and analyze the interactions between devices. The graph neural network model can identify high-order interaction features between devices by learning these relationships, thereby providing more accurate analysis results than traditional methods.

[0207] Enhanced prediction capability: By assigning rich feature vectors to each node, such as temperature, pressure, and operating time, the present invention enables the graph neural network to comprehensively consider various operating status information; this method improves the model's ability to predict equipment status, abnormal warning, and performance evaluation;

[0208] Optimize resource allocation. By analyzing the linkage effect between devices, the present invention can guide more accurate optimization control strategies of power generation systems, such as adjusting equipment operation strategies or preventive maintenance arrangements, thereby achieving optimal resource allocation and reducing operation and maintenance costs;

[0209] Improve system efficiency. The graph neural network model of the present invention can process large-scale graph data and is suitable for inductive learning scenarios. Even in a constantly changing power generation system, the model can adapt to new devices and connection relationships and continuously improve the operating efficiency of the system.

[0210] Fault diagnosis and prevention: By introducing the attention mechanism, the present invention can identify the devices or connections that have a greater impact on system operation and give priority to monitoring these key nodes; it helps to detect potential faults in a timely manner and prevent them, reducing downtime and maintenance costs;

[0211] User-friendly interaction and visualization: The user interaction and visualization module provided by the present invention enables operators to easily set analysis parameters, view and understand analysis results; the integrated data visualization tool intuitively displays device status and linkage effects, improving user experience and operational efficiency.

[0212] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A graph neural network method for linkage analysis of multiple power generation devices, characterized by: The steps include: S1. Construct a graph data structure based on the linkage relationship between multiple power generation devices, where nodes represent devices in the power generation system and edges represent the connection relationship between devices; S2. Assign a feature vector to each node and initialize it. The feature vector contains the operating status information of the corresponding device. S3. Establishing a graph neural network model, which is used to process the graph data structure and learn the interactions between devices; S4, inputting the graph data structure and node feature vector into the graph neural network model; S5. Output the linkage analysis results of each device through the graph neural network model, and the results are used to guide the optimal control of the power generation system.

2. The graph neural network method for linkage analysis of multiple power generation devices according to claim 1 is characterized in that: Step S1 is specifically as follows: S11, determine the vertex number, and number several devices as 1, 2, ..., n respectively; S12. Construct an adjacency matrix, initialize an n×n matrix, where n is the number of devices, traverse all device pairs, and check whether there is a physical connection. For an undirected graph, the adjacency matrix is ​​symmetric; for a directed graph, it is not necessarily symmetric; Among them, the element definition of the adjacency matrix A is expressed as: Where E is the set of edges in the graph. If vertex v i and vertex v j If there is an edge between them, then the corresponding matrix element a ij is 1; if vertex v i and vertex v j If there is no edge between them, then the corresponding matrix element a ij is 0; S13. Implement an adjacency matrix, and then update the values ​​in the matrix according to the connection status between devices.

3. The graph neural network method for power generation multi-device linkage analysis according to claim 1 is characterized in that: In step S2, in the graph data structure, each node represents an entity, and features need to be determined for each node, including device temperature, pressure, and operating time; Combine the features of all nodes into a feature matrix X. If the graph contains N nodes and each node has D features, then the feature matrix X will be N×D dimensional; The connection relationship of the graph is represented by the adjacency matrix A. For an undirected graph, the adjacency matrix is ​​symmetric, but for a directed graph it is not necessarily symmetric.

4. The graph neural network method for linkage analysis of multiple power generation devices according to claim 1 is characterized in that: The specific process of step S3 is as follows: S31. Determine the number of layers of the graph neural network and the number of neurons in each layer; S32, initialize the weight matrix W(l) and the bias vector b(l), where l represents the lth layer; S33. For each layer, perform the following operations: S331, calculate the product of the input matrix X(l) of the current layer and the weight matrix W(l), and add the bias vector b(l) to obtain the output matrix Y(l) of the current layer; S332, processing the output matrix Y(l) through the activation function σ to obtain the hidden state matrix H(l) of the current layer; S333, using the hidden state matrix H(l) as the input matrix X(l+1) of the next layer, and repeating steps S331-S332 until all layers are processed; S34. Use the last layer hidden state matrix H(L) as the output of the entire graph neural network model.

5. The graph neural network method for linkage analysis of multiple power generation devices according to claim 4 is characterized by: During the establishment of the graph network neural model, the unstandardized matrix is ​​calculated using the following formula: L = DA; In the formula, A represents the adjacency matrix of the graph, and D is the degree matrix, which means that the elements on the diagonal of the matrix are the degree of each node, which is the number of edges connected to the node; Symmetric normalized matrix, the expression is: Where L is the unstandardized matrix, D is the degree matrix, and in the symmetric standardized matrix, the factor It is a diagonal matrix consisting of the square root elements of the inverse matrix of the degree matrix D; The graph convolution operation is performed on the feature matrix and the adjacency matrix to integrate the state information of each device and its neighbors, which is expressed as: In the formula, H (l) is the hidden state matrix of the lth layer, where each row represents the feature vector of a node; W (l) is the weight matrix from layer l to layer l+1, used for linear transformation; σ is the activation function, used to introduce nonlinearity; is the inverse square root of the degree matrix D; is a symmetric normalized matrix.

6. The graph neural network method for power generation multi-device linkage analysis according to claim 1 is characterized in that: Step S4 is specifically as follows: S41, check that the graph data structure has been correctly constructed, including the adjacency matrix A and the node feature matrix X. For each node, ensure that its feature vector has been initialized according to step S2 and contains the operating status information of the corresponding device; S42, taking the adjacency matrix A, the degree matrix D, and the node feature matrix X as inputs and passing them into the graph neural network model; S43, forward propagation is performed through the defined graph convolution operation, and nonlinearity is introduced through the activation function to update the feature vector H of the node, including multiple rounds of iterations, each round of iteration will update the feature vector of the current node according to the information of the neighboring nodes until a predetermined number of layers is reached; S44. After the graph convolution operation, the feature vector of each node will contain rich local and global graph structure information, and the final node feature matrix H will be used as the output of the model.

7. The graph neural network method for power generation multi-device linkage analysis according to claim 6 is characterized in that: In step S4, each node updates its own state by exchanging information with its neighbor nodes, and the expression is: In the formula, is the state of node i at time t, f is the update function, AGG j It is an aggregation function of neighbor node status; The graph neural network model inductive learning adopts an aggregation method, which allows the graph neural network model to process unseen nodes and edges. The expression is: In the formula, N(v i ) represents node v i The set of neighbor nodes, W agg is the learnable weight matrix, h j is node v j The eigenvector of agg represents the feature vector of the target node after aggregation, and σ represents the activation function; The graph neural network model introduces the attention mechanism based on the aggregation method. The attention mechanism allows the model to assign different weights according to the importance of neighboring nodes. According to the above expression, it is updated as follows: In the formula, a ij is the attention weight; The above method can efficiently process large-scale graphs and is suitable for inductive learning scenarios. The attention mechanism gives the model stronger expressive power and automatically learns the importance of neighbor nodes.

8. A graph neural network system for linkage analysis of multiple power generation devices, the system being used to implement the method described in any one of claims 1 to 7, characterized in that: include: Data collection and preprocessing module, which is responsible for collecting real-time operating data of power generation equipment, including temperature, pressure, and operating time, and performing preliminary cleaning and standardization processing; The data is obtained through sensors and collection equipment; the preprocessing includes removing outliers, filling missing values ​​and normalizing to ensure data quality; The graph construction and feature initialization module builds a graph data structure based on the physical connection and operation logic of the power generation equipment. Each device is a node, and the connection relationship between devices constitutes an edge. At the same time, this module is used to assign a feature vector to each node, which contains the operating status information of the device. After these feature vectors are initialized, they will be used as the input of the graph neural network; A graph neural network processing module, which uses a graph convolutional network to learn the interactions between devices. The graph neural network processing module accepts an adjacency matrix, a degree matrix, and a feature matrix as inputs, and gradually extracts high-order interaction features of devices through multi-level graph convolution operations; The model training and optimization module is responsible for the training, verification and testing of the graph neural network model. The model training and optimization module uses back propagation and gradient descent methods to optimize model parameters. Result output and application module: the model outputs the linkage analysis results of each device, including equipment status prediction, abnormal warning, and performance evaluation. The results will be directly applied to the optimization control of the power generation system; The user interaction and visualization module provides a user-friendly interface, allowing operators to easily set analysis parameters, view and understand analysis results, and integrate data visualization tools to intuitively display equipment status and linkage effects.

9. The graph neural network system for linkage analysis of multiple power generation devices according to claim 8 is characterized in that: The data collection and preprocessing modules respectively perform data collection, data cleaning, data standardization, data integration, and data output operations, specifically: Data collection: sensors deployed on power generation equipment continuously monitor key indicators such as temperature, pressure, and operating time, and data collection equipment collects data regularly or in real time; Data cleaning, identifying and removing outliers and filling missing values, using the mean or median of adjacent data points or estimating missing data through interpolation methods; Data standardization: The data is normalized so that it falls within the standard range. The expression is: Where x is the original data point, x min and x max are the minimum and maximum values ​​of the data respectively; Data integration: The processed data is integrated into an adjacency matrix and a feature matrix. The adjacency matrix reflects the connection relationship between devices, and the feature matrix includes the operating status information of the devices. The matrix construction expression is: In the formula, t i ,p i ,r i represent the temperature, pressure and operating time of the i-th equipment respectively; Data output: output the preprocessed data set for further analysis and learning by the graph neural network model.

Images