Risk level assessment method for power resource infrastructure projects based on attention mechanism optimization

Abstract

The present invention discloses a method for assessing the risk level of power resource infrastructure projects under the optimization of attention mechanism, and relates to the technical field of engineering project risk assessment. The method comprises: establishing a risk index evaluation system for power resource infrastructure projects based on the system structure data of power resource infrastructure projects; taking the risk index evaluation system for power resource infrastructure projects as input, constructing a graph neural network model under the optimization of attention mechanism, the graph neural network model comprises a sequentially connected attention mechanism neural layer, a graph neural network layer of a multi-head attention mechanism, and an entity classification layer; training and testing the graph neural network model under the optimization of attention mechanism, and obtaining a risk model for power resource infrastructure projects, which is used to output the risk level of power resource infrastructure projects and the model prediction results. The method can assign attention weights to each neighbor node, thereby identifying important neighbor nodes, and overcomes the limitation that the graph convolutional neural network algorithm treats all neighbor nodes of a node equally.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering project risk assessment, and in particular to a method for risk level assessment of power resource infrastructure projects under attention mechanism optimization. Background Art

[0002] The risk level assessment of power resource infrastructure projects can help project decision makers fully understand and evaluate the various risks faced by the project, and take corresponding measures to reduce risks and improve the success rate and benefits of the project. However, the construction of power resource infrastructure projects is a complex and high-risk system engineering. The risk assessment methods currently used are relatively limited, mainly focusing on qualitative assessment and quantitative assessment. For complex project risk assessment, more flexible and comprehensive methods and tools are needed. Most of the existing risk assessment methods are based on expert experience. Different assessors may have different subjective perceptions and understandings of risks, which will lead to inconsistency and reduced credibility of the assessment results. The machine learning method based on non-human judgment not only avoids the errors of human experience, but also can analyze and evaluate various risk factors, understand their probability of occurrence, possible impact and loss, and thus formulate corresponding risk response measures in a targeted manner. Among them, the more advanced method is to use neural network related models, but traditional algorithms such as graph convolutional neural network algorithms are often less flexible, and there are also problems such as memory consumption that requires calling the complete graph structure for training and ignoring the different importance of different neighbor nodes when aggregating neighbor nodes. It is difficult to extend the application to large-scale network structures. Summary of the invention

[0003] In view of the above problems and technical requirements, the inventors have proposed a risk level assessment method for power resource infrastructure projects under the optimization of the attention mechanism. The technical solution of the present invention is as follows:

[0004] The risk level assessment method of power resource infrastructure projects under the optimization of attention mechanism includes the following steps:

[0005] Establish a risk indicator evaluation system for power resource infrastructure projects based on the system structure data of power resource infrastructure projects;

[0006] Taking the risk indicator evaluation system of power resource infrastructure projects as input, the attention mechanism is introduced to construct a graph neural network model optimized by the attention mechanism. The graph neural network model includes a sequentially connected attention mechanism neural layer, a multi-head attention mechanism graph neural network layer, and an entity classification layer.

[0007] The graph neural network model optimized by the attention mechanism is trained and tested to obtain the risk model of power resource infrastructure projects, which is used to output the risk level of power resource infrastructure projects and the model prediction results.

[0008] Its further technical solution is to use the risk indicator evaluation system of power resource infrastructure projects as input, including:

[0009] In the risk indicator evaluation system for power resource infrastructure projects, the engineering information of power resource infrastructure projects is used as the nodes of the graph neural network model, the secondary indicators are discretized as the features of the corresponding nodes, the engineering categories of power resource infrastructure projects and the information of different units / teams to which the projects belong are used as the edges between the corresponding nodes, and the risk levels of power resource infrastructure projects from level one to level five are used as the label data of the graph neural network model input data;

[0010] Among them, the engineering information of power resource infrastructure projects includes five first-level indicators: personnel factors, machinery and equipment factors, material factors, method factors and environmental factors; the first-level indicators are divided into 12 second-level indicators: supervision unit, construction unit, construction unit, construction machinery and equipment, safety system and tools, construction materials, document materials, operation / technical methods, management methods, operation environment, natural environment and social environment.

[0011] The further technical solution is that in the attention mechanism neural layer:

[0012] Based on the model's own nodes v i Features and neighbor nodes v j The attention coefficient of each node is calculated by the feature a ij ;

[0013] According to the calculated attention coefficient of each node a ij , and the weighted sum of the features is obtained , as the input data of the graph neural network layer of the multi-head attention mechanism;

[0014] The expression is:

[0015] ;

[0016] In the formula, T It is represented as the weight matrix of the node transforming from the input feature vector to the output feature vector, h j Neighbor node v j The features after linear transformation are V () represents the node set of the graph neural network model, and ReLU is the activation function.

[0017] Its further technical solution is to use the model's own nodes iFeatures and neighbor nodes j The attention coefficient of each node is calculated by the feature a ij ,include:

[0018] After linearly transforming the features of each node, perform self-attention operation on each node to obtain neighbor nodes v j For own node v i Importance e ij , expressed as:

[0019] ;

[0020] In the formula, R () represents the LeakyRelu activation function, β () is a function for calculating the correlation between two nodes. For own node vi and neighbor nodes vj The feature vector in high-dimensional space after linear transformation;

[0021] The calculated importance of nodes e ij Perform softmax normalization to obtain the attention coefficient a ij , expressed as:

[0022] ;

[0023] In the formula, the subscript k Represents the number of independent executions.

[0024] The further technical solution is that in the graph neural network layer of the multi-head attention mechanism:

[0025] In order to further improve the robustness of the graph neural network model, a multi-head attention mechanism is introduced into the model to accumulate the new features obtained by each attention head, which can be expressed as:

[0026] ;

[0027] In the formula, N i For the i A head of attention, Indicates n The attention coefficient calculated by the layer, Indicates n The weight matrix of the layer, For each node output by the graph neural network layer of the multi-head attention mechanismv i The cumulative sum of new features that incorporate information from its neighborhood;

[0028] Then the reconstructed new features are obtained by averaging As the input data of the entity classification layer, the calculation formula is:

[0029] ;

[0030] In the formula, N is the number of attention heads, The output of the graph neural network layer of the multi-head attention mechanism for each node v i New features that incorporate information from its neighborhood.

[0031] A further technical solution is that, in the entity classification layer:

[0032] Use the softmax function as the activation function to convert the output data of the graph neural network layer of the multi-head attention mechanism into The input data of this layer is processed to output the label category of the target node and the probability of belonging to each risk level. The calculation formula is:

[0033] ;

[0034] In the formula, f r The feature vector representing the target node Middle r The eigenvalues ​​of the rows, It is the probability value of the target node belonging to a certain risk level obtained after calculation by the entity classification layer. The index of the maximum value in the probability is the risk level assessment result of the target node.

[0035] The further technical solution is to train and test the graph neural network model optimized by the attention mechanism, including:

[0036] Set the training parameters, including the number of iterations, the number of attention heads, the number of samples per training, the learning rate, and the weight decay rate;

[0037] The parameters in the graph neural network model optimized by the attention mechanism are randomly initialized, and the momentum method is used to perform gradient optimization on the parameters. The formula is expressed as:

[0038] ;

[0039] In the formula, For the t The model parameters updated in the iterations are For thet- Model parameters updated in 1 iteration, is the learning rate, For the t The gradient of the iteration, is the momentum parameter.

[0040] Its further technical solution is that the method further comprises:

[0041] The data of the system structure of the power resource infrastructure project is cleaned. The cleaned data includes:

[0042] The level of safety inspection, the level of quality inspection, the intensity of review and execution, the timeliness of delivery of drawings / funds, changes in requirements during construction, the basic quality of operating personnel, the age of operating personnel, the training time of operating personnel, the basic quality of team leaders, the qualifications of team leaders, equipment maintenance, equipment load operation, equipment safety protection device inspection, equipment service life, the site and time involved in equipment transportation / assembly, system stability, system accuracy, random inspection of tools and protective equipment, quality of material random inspection, material replenishment and transportation time, material storage status and applicability, material hazard level, relevant documents / procedures / qualification approval processing time, document organization and backup, construction technology level, number of operating personnel, nature of operation, overtime construction, frequency / quality of accident hazard rectification and investigation, emergency plan and drill preparation, improvement of rules and regulations and safety inspection, salary reward and punishment system, water and electricity supply, construction site layout, operation area, operation period, geological and hydrological conditions, climate, epidemic situation and natural disasters, temperature, legal / policy requirements, residents' opinions, and opinions of relevant departments.

[0043] Its further technical solution is that the method further comprises:

[0044] The torch and dgl methods are used to convert the data in the risk indicator evaluation system of power resource infrastructure projects into graph structure data.

[0045] The beneficial technical effects of the present invention are:

[0046] The risk level assessment method for power resource infrastructure projects under the optimization of the attention mechanism proposed in this application can assign attention weights to each neighbor node, thereby identifying important neighbor nodes, overcoming the limitation of the graph convolutional neural network algorithm that all neighbor nodes of a node are treated equally. The results obtained by the method described in the present invention are greatly improved compared with the traditional graph convolutional neural network method, that is, the accuracy of the neural network model after the attention mechanism optimization is significantly higher than that of the traditional graph convolutional neural network model, and the convergence speed is faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1It is the overall flow chart of the risk level assessment method of power resource infrastructure projects under the optimization of the attention mechanism provided in this application;

[0048] Figure 2 This is a diagram showing the effect of the optimized graph neural network model using the attention mechanism provided in this application;

[0049] Figure 3 This is the effect diagram of the traditional graph convolutional neural network model. DETAILED DESCRIPTION

[0050] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0051] Please refer to Figure 1 As shown, this embodiment provides a method for risk level assessment of power resource infrastructure projects under attention mechanism optimization, which specifically includes the following steps:

[0052] Step 1: Establish a risk indicator evaluation system for power resource infrastructure projects based on the power resource infrastructure project system architecture data.

[0053] Specifically, the system structure data of the power resource infrastructure project includes relevant power platform data and actual on-site production data, etc. After desensitization, the data is cleaned. The cleaned data includes:

[0054] Safety inspection level, quality inspection level, review execution strength, timeliness of drawing / fund delivery, changes in construction requirements, basic quality of operators, age of operators, training time of operators, basic quality of team leaders, qualifications of team leaders, equipment maintenance, equipment load operation (equipment performance), equipment safety protection device inspection, equipment service life, equipment transportation / assembly site and time, system stability, system accuracy, random inspection of tools and protective equipment, quality of material random inspection, material replenishment and transportation time, material storage status and applicability, material hazard level, relevant documents The time for processing documents / procedures / qualification approval, document organization and backup, construction technology level, number of operators (single / multiple, etc.), nature of operations (general / temporary / emergency repair, etc.), overtime construction, frequency / quality of accident hazard rectification and investigation, emergency plans and drill preparations, improvement of rules and regulations and safety inspections, salary reward and punishment system, water and electricity supply, construction site layout (building location, material storage location, etc.), operation area (high altitude, crossing railways, etc.), operation time (day, night), geological and hydrological conditions, climate, epidemic situation and natural disasters, temperature, legal / policy requirements, residents' opinions, and opinions of relevant departments.

[0055] Furthermore, the five first-level indicators of "personnel factors", "mechanical equipment factors", "material factors", "method factors" and "environmental factors" of power resource infrastructure project engineering information are divided, and the 12 second-level indicators of "supervision unit", "construction unit", "construction unit", "construction machinery and equipment", "safety system and tools", "construction materials", "document materials", "operation / technical methods", "management methods", "operating environment", "natural environment" and "social environment" are used as entity attributes, and the characteristic data are obtained after discretization processing.

[0056] Furthermore, the architecture also includes seven power resource infrastructure project engineering categories and information on 25 different units or teams to which the projects belong.

[0057] The final established risk indicator evaluation system structure of power resource infrastructure projects is shown in Table 1:

[0058] Table 1. System structure of risk indicator evaluation for power resource infrastructure projects

[0059]

[0060] It should be noted that this application is based on the comprehensive expert experience and relevant standard documents of the power resource infrastructure project company, refers to the theoretical basis of risk assessment of the whole process of power engineering project construction, and takes into account factors such as the pre-construction preparation of the power resource infrastructure project, the construction site data, the social and economic benefits of the construction, etc., and divides the risk indicators of the power resource infrastructure project into one to five "RISK" levels.

[0061] Step 2: Taking the risk indicator evaluation system of power resource infrastructure projects as input, introduce the attention mechanism and construct a graph neural network model optimized by the attention mechanism.

[0062] Firstly, the torch and dgl methods are used to convert the data in the risk indicator evaluation system of power resource infrastructure projects into graph structure data.

[0063] Then, the engineering information of the power resource infrastructure projects in the system architecture is used as the nodes of the graph neural network model, the 12 discretized secondary indicators are used as the features of the corresponding nodes, the engineering categories of the power resource infrastructure projects and the information of the different units / teams to which the projects belong are used as the edges between the corresponding nodes, and the risk level values ​​of the power resource infrastructure projects from level one to level five are used as label data to associate with the nodes to complete the node information.

[0064] The graph neural network model optimized by the attention mechanism constructed in this application is used to perform feature aggregation and prediction on graph structure data. The model includes a sequentially connected attention mechanism neural layer, a multi-head attention mechanism graph neural network layer, and an entity classification layer.

[0065] <1> In the attention mechanism neural layer, linear transformation of data features and parameter initialization are performed based on the model’s own nodes. v i Features and neighbor nodes v j The attention coefficient of each node is calculated by the feature a ij , and then according to the calculated attention coefficient of each node a ij , perform weighted summation of the features as input data for the graph neural network layer of the multi-head attention mechanism.

[0066] Specifically, we first need to perform a linear transformation on the entity features, map each type of entity features into the same feature space, and obtain the feature vector mapped in the high-dimensional space, that is, , h 0i For the target entity v i (i.e., the original feature vector of the target node), is the feature transformation weight matrix of the learnable parameters, h i For the target entity v i After the matrix linear transformation, it is mapped to the feature vector of the high-dimensional space. Then, a self-attention operation is performed on each node to obtain the neighbor nodes v j For own node v i Importance e ij , expressed as:

[0067] (1)

[0068] In the formula, R () represents the LeakyRelu activation function, β () is a function for calculating the correlation between two nodes. For own node vi and neighbor nodes vj The eigenvector in high-dimensional space after linear transformation, T It is represented as a weight matrix that transforms a node from an input feature vector to an output feature vector.

[0069] In order to distribute weights more reasonably, the calculated importance of nodes e ij Perform softmax normalization to obtain the attention coefficient a ij , expressed as:

[0070] (2)

[0071] In the formula, the subscript k Represents the number of independent executions. Then, the node v i The output feature vector of is:

[0072] (3)

[0073] In the formula, h j Neighbor node v j The features after linear transformation are For Node v i The output feature vector of V () represents the node set of the graph neural network model, and ReLU is the activation function.

[0074] <2> In the graph neural network layer of the multi-head attention mechanism, in order to further improve the robustness of the graph neural network model, a multi-head attention mechanism is introduced into the model to accumulate the new features obtained by each attention head, which is expressed as:

[0075] (4)

[0076] In the formula, N i For the i A head of attention, Indicates n The attention coefficient calculated by the layer, Indicates n The weight matrix of the layer, For each node output by the graph neural network layer of the multi-head attention mechanism v i The cumulative sum of new features that incorporate information from its neighborhood.

[0077] In addition, this layer uses a multi-head attention mechanism to obtain more comprehensive information, greatly enhancing the stability of the model and obtaining new features through averaging. As the input data of the entity classification layer, the calculation formula is:

[0078] (5)

[0079] In the formula, N is the number of attention heads, The output of the graph neural network layer of the multi-head attention mechanism for each node v i New features that incorporate information from its neighborhood.

[0080] <3> In the entity classification layer, the softmax function is used as the activation function to convert the output data of the graph neural network layer of the multi-head attention mechanism into The input data of this layer is processed to output the label category of the target node and the probability of belonging to each risk level. The calculation formula is:

[0081] (6)

[0082] In the formula, f r The feature vector representing the target node Middle r The eigenvalues ​​of the rows, It is the probability value of the target node belonging to a certain risk level obtained after calculation by the entity classification layer. The index of the maximum value in the probability is the risk level assessment result of the target node.

[0083] Step 3: Train and test the graph neural network model optimized by the attention mechanism to obtain the risk model of the power resource infrastructure project, which is used to output the risk level of the power resource infrastructure project and the model prediction results.

[0084] Before training and testing, the training parameters were set, including the number of iterations set to 400, the number of attention heads set to 3, the batch size of each training sample set to 1, the learning rate initialized to 0.01, and the weight decay rate initialized to 0.001.

[0085] The parameters in the graph neural network model optimized by the attention mechanism are randomly initialized, and the parameters are gradient optimized using the Momentum method to solve the problem that the local optimal point gradient is 0, it cannot be continuously updated, and the oscillation amplitude is too large. The formula for using the Momentum gradient optimization algorithm is expressed as:

[0086] (7)

[0087] In the formula, For the t The model parameters updated in the iterations are For the t-Model parameters updated in 1 iteration, is the learning rate, For the t The gradient of the iteration, is the momentum parameter.

[0088] The risk model for power resource infrastructure projects aggregates and predicts features of graph structure data, including:

[0089] First, the attention mechanism neural layer of the model is used to calculate the importance of nodes by linearly transforming the features of the graph structure data itself and the features of neighboring nodes. e ij ; Secondly, the softmax function is used to normalize the importance of nodes e ij Perform normalization operation to obtain the normalized attention coefficient a ij ; Again, according to the calculated attention coefficient of each node a ij , perform weighted summation on the features, accumulate and average them through the graph neural network layer of the multi-head attention mechanism to obtain a reconstructed new feature vector, and then input the reconstructed new feature vector into the entity classification layer; finally, after being processed by the entity classification layer, the probability is output and sorted to determine the label category of the target node, realize the risk level classification of the target node, and output the risk level prediction result.

[0090] The beneficial effects of the above methods are explained and compared through simulation experiments and tabular data. Figure 2 As shown in the figure, the accuracy of the graph neural network model after the attention mechanism optimization increases with the number of iterations during the training of data. When the iteration reaches 150, the accuracy of the test set is close to the highest value. Although the accuracy fluctuates in subsequent training, it finally converges to 89.48%. At the same time, the loss value begins to decrease gradually after 150 iterations, and converges to a relatively stable value after 400 iterations.

[0091] Depend on Figure 3 As shown in the figure, during the training process of the traditional graph convolutional neural network model, the accuracy rate can only approach the maximum value of 85.37% after about 200 iterations after 400 iterations. In addition, the fluctuation range and frequency of the accuracy rate are significantly higher than those of the model after the attention mechanism is optimized. This shows that the traditional graph convolutional neural network model has poor flexibility in calculating the accuracy rate, and errors will be caused by the information of individual nodes themselves, and it is impossible to efficiently learn effective features from the data.

[0092] contrast Figure 2 , 3It can be seen that the risk level assessment method of power resource infrastructure projects under the optimization of attention mechanism proposed in the present invention adds a self-attention mechanism in the propagation process, and the method of the present invention increases the importance of nodes. e ij , an attention coefficient can be assigned to each neighbor node a ij , thereby identifying important neighbor nodes, overcoming the limitation of traditional graph convolutional neural networks that treat all neighbor nodes of a node equally. The results obtained by the method described in the present invention are significantly improved compared with the traditional graph convolutional neural network method, that is, the accuracy of the neural network model after the attention mechanism is optimized is significantly higher than that of the traditional graph convolutional neural network model, and a better convergence effect can be achieved with fewer iterations, overcoming the poor flexibility of the graph convolutional neural network algorithm, the memory consumption problem that the model needs to call the complete graph structure for training, and the problem of ignoring the different importance of different neighbor nodes when aggregating neighbor nodes.

[0093] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0094] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0095] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0096] The above is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiments. It is understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the protection scope of the present invention.

Claims

1. The risk level assessment method of power resource infrastructure projects under the optimization of attention mechanism is characterized by: The method comprises: A risk index evaluation system for power resource infrastructure projects is established based on the system architecture data of power resource infrastructure projects; in the evaluation system, the engineering information of power resource infrastructure projects includes five primary indicators, namely, personnel factors, machinery and equipment factors, material factors, method factors and environmental factors; the primary indicators are divided into 12 secondary indicators, namely, the supervision unit, construction unit, construction unit, construction machinery and equipment, safety system and tools, construction materials, document materials, operation / technical methods, management methods, operation environment, natural environment and social environment; Taking the risk indicator evaluation system of power resource infrastructure projects as input, introducing the attention mechanism, and constructing a graph neural network model optimized by the attention mechanism, the graph neural network model includes a sequentially connected attention mechanism neural layer, a multi-head attention mechanism graph neural network layer, and an entity classification layer; The graph neural network model optimized by the attention mechanism is trained and tested to obtain a risk model for power resource infrastructure projects, which is used to output the risk level of the power resource infrastructure projects and the model prediction results; In the electric power resource infrastructure project risk indicator evaluation system, the engineering information of the electric power resource infrastructure project is used as the node of the graph neural network model, the secondary indicators are discretized as the features of the corresponding nodes, the engineering categories of the electric power resource infrastructure project and the information of different units / teams to which the projects belong are used as the edges between the corresponding nodes, and the risk levels of the electric power resource infrastructure project from level one to level five are used as the label data of the graph neural network model input data; In the attention mechanism neural layer: Based on the model's own nodes v i Features and neighbor nodes v j The attention coefficient of each node is calculated by the feature a ij ; According to the calculated attention coefficient of each node a ij , and the weighted sum of the features is obtained , as the input data of the graph neural network layer of the multi-head attention mechanism; The attention coefficient of each node is calculated a ij ,include: After linearly transforming the features of each node, a self-attention operation is performed on each node to obtain the neighboring nodes v j For the own node v i Importance e ij , expressed as: ； In the formula, R () represents the LeakyRelu activation function, β () is a function for calculating the correlation between two nodes. For own node vi and neighbor nodes vj The eigenvector in high-dimensional space after linear transformation, T It is represented as the weight matrix of the node transforming from the input feature vector to the output feature vector; The calculated importance of nodes e ij Perform softmax normalization to obtain the attention coefficient a ij ; In the entity classification layer: Use the softmax function as the activation function to convert the output data of the graph neural network layer of the multi-head attention mechanism into The input data of this layer is processed to output the label category of the target node and the probability of belonging to each risk level. The calculation formula is: ； In the formula, f r The feature vector representing the target node Middle r The eigenvalues ​​of the rows, is the probability value of the target node belonging to a certain risk level obtained after calculation by the entity classification layer, and the index of the maximum value in the probability is the risk level assessment result of the target node.

2. The method for risk level assessment of power resource infrastructure projects under the optimization of the attention mechanism according to claim 1 is characterized in that: Said The expression is: ； In the formula, h j Neighbor node v j The features after linear transformation are V () represents the node set of the graph neural network model, and ReLU is the activation function.

3. The method for risk level assessment of power resource infrastructure projects under the optimization of the attention mechanism according to claim 2 is characterized in that: The attention coefficient a ij It is expressed as: ； In the formula, the subscript k Represents the number of independent executions.

4. The method for risk level assessment of power resource infrastructure projects under the optimization of the attention mechanism according to claim 2 is characterized in that: In the graph neural network layer of the multi-head attention mechanism: In order to further improve the robustness of the graph neural network model, a multi-head attention mechanism is introduced into the model to accumulate the new features obtained by each attention head, which can be expressed as: ； In the formula, N i For the i Attention head, Indicates n The attention coefficient calculated by the layer, Indicates n The weight matrix of the layer, For each node output by the graph neural network layer of the multi-head attention mechanism v i The cumulative sum of new features that incorporate information from its neighborhood; Then the reconstructed new features are obtained by averaging As the input data of the entity classification layer, the calculation formula is: ； In the formula, N is the number of attention heads, The output of the graph neural network layer of the multi-head attention mechanism for each node v i New features that incorporate information from its neighborhood.

5. The method for risk level assessment of power resource infrastructure projects under the optimization of attention mechanism according to claim 1 is characterized in that: The graph neural network model optimized by the attention mechanism is trained and tested, including: Set the training parameters, including the number of iterations, the number of attention heads, the number of samples per training, the learning rate, and the weight decay rate; The parameters in the graph neural network model optimized by the attention mechanism are randomly initialized, and the parameters are gradient optimized using the Momentum method. The formula is expressed as: ； In the formula, For the t The model parameters updated in the iterations are For the t- Model parameters updated in 1 iteration, is the learning rate, For the t The gradient of the iteration, is the momentum parameter.

6. The method for risk level assessment of power resource infrastructure projects under the optimization of attention mechanism according to claim 1 is characterized in that: The method further comprises: The data of the system structure of the power resource infrastructure project is cleaned, and the cleaned data includes: The level of safety inspection, the level of quality inspection, the intensity of review and execution, the timeliness of delivery of drawings / funds, changes in requirements during construction, the basic quality of operating personnel, the age of operating personnel, the training time of operating personnel, the basic quality of team leaders, the qualifications of team leaders, equipment maintenance, equipment load operation, equipment safety protection device inspection, equipment service life, the site and time involved in equipment transportation / assembly, system stability, system accuracy, random inspection of tools and protective equipment, quality of material random inspection, material replenishment and transportation time, material storage status and applicability, material hazard level, relevant documents / procedures / qualification approval processing time, document organization and backup, construction technology level, number of operating personnel, nature of operation, overtime construction, frequency / quality of accident hazard rectification and investigation, emergency plan and drill preparation, improvement of rules and regulations and safety inspection, salary reward and punishment system, water and electricity supply, construction site layout, operation area, operation period, geological and hydrological conditions, climate, epidemic situation and natural disasters, temperature, legal / policy requirements, residents' opinions, and opinions of relevant departments.

7. The method for risk level assessment of power resource infrastructure projects under the optimization of attention mechanism according to claim 1 is characterized in that: The method further comprises: The torch and dgl methods are used to convert the data in the risk indicator evaluation system of the power resource infrastructure project to graph structure data.

Images