An energy project risk prediction method and system based on a knowledge graph

Abstract

This invention discloses a knowledge graph-based method and system for predicting energy project risks. The knowledge graph-based method includes: constructing knowledge subgraphs for each approval branch and independently selecting the best migration source type for each approval branch based on graph kernel similarity; performing graph-domain adaptive adversarial migration training on each approval branch to generate branch-level migration enhancement embedding vectors; extracting general structural patterns of cross-branch cross-influences and inferring potential cross-dependencies to construct a multi-branch parallel knowledge graph; performing branch-aware graph attention encoding on the multi-branch parallel knowledge graph to generate migration collaborative fusion embedding vectors; and deeply integrating the fusion embedding vectors with progress features into a Bayesian inference network to output a comprehensive risk level and key bottleneck branch identifiers.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph and risk prediction technology, and more specifically, to a knowledge graph-based method and system for predicting energy project risks. Background Technology

[0002] In the field of energy project approval risk prediction, energy projects typically involve multiple parallel approval branches, such as safety reviews, environmental impact assessments, and hazardous chemical permits. Delays in each branch have a significant impact on the overall project schedule. Existing technologies usually model the entities and workflow relationships of approval nodes in the project approval process based on knowledge graphs. They collect historical approval data from similar energy projects to construct an approval knowledge graph, generate entity embedding representations through graph coding, and then predict the delay risk of each approval branch. When there are insufficient historical approval cases for the target energy project type (such as hydrogen production, storage, and transportation projects, compressed air energy storage projects, and other emerging energy projects), existing technologies compensate for data sparsity by transferring knowledge graph features from mature energy project types.

[0003] However, existing methods have the following technical problems: First, existing methods uniformly select a single migration source and perform global knowledge migration according to the overall project type. However, the technical characteristics of each approval branch of emerging energy projects are significantly different. Uniform migration leads to a mismatch between the knowledge graph of some branches and the characteristics of the target branches, which reduces the accuracy of risk prediction. Second, the cross-influence relationships between the approval branches of emerging energy projects lack historical data support in the knowledge graph. Existing methods cannot provide a reliable graph representation of the cross-dependencies between branches of emerging types, resulting in inaccurate identification of key bottleneck branches and a lack of uncertainty assessment in risk prediction results. Summary of the Invention

[0004] This invention provides a knowledge graph-based method and system for predicting energy project risks, addressing the technical problems in related technologies, such as the difficulty in accurately predicting approval risks of emerging energy projects due to a lack of historical approval data, the inability to effectively model cross-approval branch influence relationships, and the difficulty in quantifying the uncertainty of prediction results.

[0005] This invention discloses a knowledge graph-based method for predicting energy project risks, comprising: acquiring approval data of a target emerging energy project and historical approval data of multiple mature energy project types; constructing a branch knowledge subgraph for each approval branch according to the approval branch category; calculating the graph kernel similarity between the branch knowledge subgraph of the target emerging type and the corresponding branch knowledge subgraph of each mature type for each approval branch category; and independently selecting the best migration source type branch knowledge subgraph for each approval branch.

[0006] For each approval branch, graph domain adaptive adversarial training is performed based on the selected best migration source type branch knowledge subgraph and the branch knowledge subgraph corresponding to the target emerging type to generate branch-level migration enhancement embedding vectors of entities in each branch knowledge subgraph of the target emerging type.

[0007] Extract the general structural pattern of cross-branch cross-influence from the knowledge graph of each mature energy project type, infer the potential cross-dependency entities between branches based on the semantic features of the branch-level transfer enhanced embedding vector, and establish cross-influence relationship edges. Connect the knowledge subgraphs of each branch through the cross-influence relationship edges to generate a multi-branch parallel knowledge graph.

[0008] On a multi-branch parallel knowledge graph, graph attention convolution is performed inside each branch knowledge subgraph to update entity embeddings, and cross-branch attention propagation is performed along the cross-influence relationship edges. After multiple rounds of iteration, the migration collaborative fusion embedding vectors of each entity are generated.

[0009] Obtain the current progress data of each approval branch of the target emerging energy project, and input the current progress data and the migration collaborative fusion embedding vector into a deep integrated Bayesian inference network after feature concatenation. Output the overall approval comprehensive risk level and key bottleneck branch identifier.

[0010] Furthermore, the graph kernel similarity is calculated using the Weisfeiler-Lehman subtree kernel algorithm. The Weisfeiler-Lehman subtree kernel algorithm generates refined labels for each node by iteratively aggregating the label information of neighboring nodes in the graph, and counts the co-occurrence frequency of the same refined labels in the two graphs in each iteration round. The inner product of the refined label frequency vectors of the two graphs is used as the graph kernel similarity output value.

[0011] Furthermore, the graph domain adaptive adversarial training includes:

[0012] A branch graph attention encoder is pre-trained on the knowledge subgraph of the best transfer source type branch, and the source domain entity embedding vector is extracted using the approval risk prediction task loss as a supervision signal.

[0013] The domain discriminator and the branch graph attention encoder are subjected to adversarial training. The training objective of the domain discriminator is to correctly distinguish the source domain of the source domain embedding vector and the target domain embedding vector. The training objective of the branch graph attention encoder is to generate embedding vectors that prevent the domain discriminator from distinguishing the source domain. The total loss function of the adversarial training is the loss of the approval risk prediction task minus the product of the adversarial training weight coefficient and the domain classification loss. The value of the adversarial training weight coefficient is an open interval from zero to one.

[0014] After the adversarial training converges, the branch knowledge subgraph corresponding to the emerging target type is input into the trained branch graph attention encoder, and the branch-level transfer enhancement embedding vector is output.

[0015] Furthermore, in adversarial training, a gradient penalty term is added to the loss function of the domain discriminator. The gradient penalty term constrains the gradient norm of the domain discriminator at the random linear interpolation point between the source domain embedding vector and the target domain embedding vector. The gradient penalty term is defined as the expected value of the square of the difference between the L2 norm of the gradient of the domain discriminator with respect to the input at the random linear interpolation point and one.

[0016] Furthermore, the extraction of the general structural pattern of cross-branch cross-influence and the establishment of cross-influence relationship edges include:

[0017] For the knowledge graph of each mature energy project type, the cross-branch relationship edges connecting different approval branches are statistically analyzed, and the general structural pattern is extracted. The general structural pattern includes the directionality and influence intensity range of the cross-influence between approval branches. The influence intensity range is the interval formed by the mean of the weights of all cross-branch relationship edges of the same direction combination in each mature type plus or minus one standard deviation.

[0018] Based on branch-level transfer enhancement embedding vectors, cosine similarity is used to calculate the semantic similarity of entity pairs between different branches. When the semantic similarity exceeds a preset similarity threshold and the branch combination of the entity pair conforms to the cross-branch influence directionality in the general structural pattern, a cross-influence relationship edge is established from the source entity to the affected entity.

[0019] For each established cross-influence relationship edge, the weight is initialized to the median value of the influence intensity range in the corresponding direction in the general structural pattern.

[0020] Furthermore, when establishing the cross-influence relationship edge, a branch role consistency check is applied to the candidate entity pair: for the influencing source entity, it is checked whether the out-degree ranking of the influencing source entity in its respective branch knowledge subgraph is within a pre-preset ratio range; for the affected entity, it is checked whether the in-degree ranking of the affected entity in its respective branch knowledge subgraph is within a pre-preset ratio range; only when the out-degree ranking of the influencing source entity and the in-degree ranking of the affected entity are both within the pre-preset ratio range of their respective branch knowledge subgraphs is the cross-influence relationship edge confirmed to be established.

[0021] Furthermore, when updating entity embeddings by graph attention convolution within each branch knowledge subgraph, signal attenuation gating is applied to branch knowledge subgraphs with data sparsity exceeding a preset threshold. The data sparsity index is defined as the reciprocal of the ratio of the actual number of relation edges in the branch knowledge subgraph to the maximum number of relation edges formed by all entity pairs. When the data sparsity index exceeds the preset threshold, the branch embedding vector of each entity within the branch knowledge subgraph after graph attention convolution is multiplied by a signal attenuation gating coefficient, which is defined as the ratio of the preset threshold to the data sparsity index. When the data sparsity index does not exceed the preset threshold, the signal attenuation gating coefficient is set to one.

[0022] Furthermore, when performing cross-branch attention propagation along the edges of the cross-influence relationship, a directional mask constraint is also applied to the cross-branch attention coefficients: based on the directionality of cross-influence between branches recorded in the general structural pattern, the attention coefficients of reverse propagation that do not conform to the influence direction are forced to be set to zero, so that information is only propagated along the influence direction recorded in the general structural pattern.

[0023] Furthermore, the deeply integrated Bayesian inference network comprises multiple member networks corresponding to each approval branch. Each member network is a Bayesian neural network, and the network weight parameters follow a preset prior distribution. The posterior distribution of the weight parameters is learned through variational inference. During the inference phase, forward propagation is performed by multiple Monte Carlo samplings of the posterior distribution of the weights. The mean of the delay probabilities output from multiple samplings is taken as the independent delay probability, and the variance of the delay probabilities output from multiple samplings is taken as the branch-level cognitive uncertainty component. Based on evidence theory, the independent delay probabilities of each branch are dynamically weighted and fused according to the branch-level cognitive uncertainty component. The evidence credibility weight of each branch is defined as the ratio of one minus the branch-level cognitive uncertainty component of that branch to the sum of one minus the branch-level cognitive uncertainty components of all branches. The overall approval comprehensive risk index is the weighted sum of the evidence credibility weight of each branch and the independent delay probability. The overall approval comprehensive risk level is output according to the correspondence between the overall approval comprehensive risk index and the preset risk level threshold range. The branch with the highest independent delay probability and evidence credibility weight exceeding the preset credibility threshold is taken as the key bottleneck branch and the key bottleneck branch identifier is output.

[0024] This invention provides a knowledge graph-based energy project risk prediction system, comprising:

[0025] The branch knowledge subgraph construction and migration source screening module is used to obtain the approval data of the target emerging energy project and the historical approval data of multiple mature energy project types. It constructs branch knowledge subgraphs for each approval branch according to the approval branch category, calculates the graph kernel similarity for each approval branch category, and independently screens the best migration source type branch knowledge subgraph for each approval branch.

[0026] The branch-level adversarial transfer training module is used to perform graph domain adaptive adversarial training on each approval branch based on the selected best transfer source type branch knowledge subgraph and the branch knowledge subgraph corresponding to the target emerging type, and generate branch-level transfer enhanced embedding vectors.

[0027] The multi-branch parallel knowledge graph construction module is used to extract the general structural pattern of cross-branch cross-influence from the knowledge graphs of various mature energy project types. Based on the semantic features of the branch-level transfer enhanced embedding vector, it infers the potential cross-dependency entities between branches and establishes cross-influence relationship edges. The knowledge subgraphs of each branch are connected through cross-influence relationship edges to generate a multi-branch parallel knowledge graph.

[0028] The branch-aware graph attention encoding module is used to perform graph attention convolution to update entity embeddings within the knowledge subgraphs of each branch on a multi-branch parallel knowledge graph, perform cross-branch attention propagation along the edges of cross-influence relationships, and generate the migration and collaborative fusion embedding vectors of each entity through multiple rounds of iteration.

[0029] The approval risk prediction output module is used to obtain the current progress data of each approval branch of the target emerging energy project. After the current progress data and the migration collaborative fusion embedding vector are concatenated for feature splicing, they are input into the deep integrated Bayesian inference network to output the overall approval comprehensive risk level and key bottleneck branch identifiers.

[0030] This invention addresses the technical problem of mismatched graph knowledge and reduced accuracy of risk prediction caused by unified migration by independently selecting the best migration source type for each approval branch and performing graph-domain adaptive adversarial migration training separately. It achieves the technical effect that the entity embedding distribution of each approval branch can be aligned with its best-matching migration source domain. Furthermore, by extracting general structural patterns of cross-branch cross-influence from mature knowledge graphs and inferring potential cross-dependencies based on semantic features, it solves the technical problem of lacking historical data support for cross-influence relationships between branches in emerging energy projects. This achieves the technical effect of constructing a multi-branch parallel knowledge graph with reasonable prior constraints even under sparse data conditions. Finally, by employing a deep ensemble Bayesian inference network to output branch-level cognitive uncertainty components and performing dynamic weighted fusion based on evidence theory, it achieves the technical effect of providing a reliable quantifiable assessment of the migration knowledge of each branch along with the risk prediction results. Attached Figure Description

[0031] Figure 1 This is a flowchart of the knowledge graph-based energy project risk prediction method provided in an embodiment of the present invention;

[0032] Figure 2 This is a schematic diagram comparing the similarity between each approval branch and the three types of mature project diagrams provided in this embodiment of the invention;

[0033] Figure 3 This is a schematic diagram comparing the adversarial training parameters of the three approval branches provided in an embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of the cross-branch cross-influence edge weight distribution (including confidence interval) provided in an embodiment of the present invention;

[0035] Figure 5 This is a schematic diagram of the data sparsity and gating coefficient of the three approval branches provided in the embodiment of the present invention;

[0036] Figure 6 This is a schematic diagram comparing the current progress and planned time limit of each approval branch of PROJ-H001 provided in an embodiment of the present invention;

[0037] Figure 7 This is a schematic diagram comparing key indicators for risk prediction of various approval branches provided in this embodiment of the invention;

[0038] Figure 8 This is a schematic diagram illustrating the decomposition of the credibility weight of evidence for each approval branch and the overall risk contribution provided in the embodiments of the present invention. Detailed Implementation

[0039] In the field of energy project approval risk prediction, existing technologies typically model the entities and workflow relationships of approval nodes in the project approval process based on knowledge graphs, thereby predicting the delay risk of each approval branch. For energy projects involving multiple parallel approval branches (such as safety reviews, environmental impact assessments, and hazardous materials permits), existing methods first collect historical approval data from similar energy projects to construct an approval knowledge graph at the project type level. Then, they perform graph encoding based on the entities and relationships in the approval knowledge graph to generate entity embedding representations. Finally, they use the graph embedding results to predict the delay risk of each approval branch. When there are insufficient historical approval cases for the target energy project type (such as emerging energy projects like hydrogen production, storage, and transportation projects, and compressed air energy storage projects), existing technologies compensate for data sparsity by transferring knowledge graph features from mature energy project types.

[0040] However, existing knowledge graph-based energy project risk prediction methods have the following technical problems in emerging energy project scenarios.

[0041] First, existing methods uniformly select a single migration source and perform global knowledge transfer based on the overall project type. However, the technical characteristics of different approval branches of emerging energy projects vary significantly. For example, the safety review of hydrogen energy projects involves unique risk nodes specific to high-pressure hydrogen storage, while environmental impact assessments involve unique impact paths of hydrogen leakage and diffusion. The optimal migration source for different approval branches may correspond to different types of mature energy projects. Uniform migration leads to a mismatch between the knowledge transfer graph of some branches and the characteristics of the target branches, reducing the accuracy of risk prediction.

[0042] Second, the cross-influence relationships between various approval branches of emerging energy projects lack historical data support in the knowledge graph. Existing methods for cross-branch relationship modeling rely on sufficient data on the association between branches of the same type for training, which cannot provide a reliable graph representation of the cross-dependencies between branches of emerging types. This results in inaccurate identification of key bottleneck branches and a lack of uncertainty assessment in risk prediction results.

[0043] According to an embodiment of this invention, a knowledge graph-based method for predicting energy project risks is provided. This method addresses the aforementioned technical problems through branch-level differentiated knowledge transfer and cross-branch cross-influence relationship inference. It should be understood that the execution entity of this embodiment is a computing device or server with data processing capabilities. This computing device or server is equipped with graph storage and computing components, capable of performing data processing operations such as knowledge graph construction, graph encoding, model training, and inference.

[0044] At least one embodiment of the present invention discloses a knowledge graph-based method for predicting energy project risks, such as... Figure 1 As shown, it includes the following steps:

[0045] Step 1: Construct a knowledge subgraph for each approval branch and independently select the best migration source type for each approval branch;

[0046] Limited approval data for target emerging energy projects and historical approval data for multiple mature energy project types are obtained, and the data is labeled according to approval branch types. For both the target emerging type and each mature type, approval nodes of each approval branch are treated as entities, and approval flow relationships are treated as relation edges, constructing a branch knowledge subgraph for each approval branch. For each approval branch category, the graph kernel similarity between the branch knowledge subgraph of the target emerging type and the corresponding branch knowledge subgraph of each mature type is calculated. Based on the graph kernel similarity, the best migration source type branch knowledge subgraph with the most similar graph structure is independently selected for each approval branch.

[0047] Specifically, the aforementioned branch labeling refers to categorizing each piece of approval data according to its respective approval branch category (such as safety review branch, environmental impact assessment branch, hazardous chemical permit branch, etc.). For each approval branch category of each project type, all approval node data under that branch category are treated as an entity set, and the flow dependencies and preconditions between approval nodes are treated as a set of relation edges, thus constructing a branch knowledge subgraph for that approval branch.

[0048] For each approval branch category Calculate the branch knowledge subgraph of emerging target types With the Each maturity type corresponds to a branch knowledge subgraph. Kernel similarity between ,in Assign index numbers to the maturity types, sort the maturity types in descending order of graph kernel similarity, and select the knowledge subgraph of the branch corresponding to the maturity type ranked first as the approval branch. Optimal migration source type branch knowledge subgraph .

[0049] It should be noted that the graph kernel similarity calculation above uses the Weisfeiler-Lehman subtree kernel algorithm. The input of the Weisfeiler-Lehman subtree kernel algorithm is the branch knowledge subgraph of the emerging type and the corresponding branch knowledge subgraph of the mature type. The output is a scalar value representing the structural similarity between the two subgraphs. The Weisfeiler-Lehman subtree kernel algorithm generates refined labels for each node by iteratively aggregating the label information of the neighboring nodes. It counts the co-occurrence frequency of the same refined labels in the two graphs in each iteration round. The inner product of the refined label frequency vectors of the two graphs is used as the graph kernel similarity output value, thus efficiently measuring the structural similarity between the two graphs without explicitly enumerating the subgraph isomorphisms.

[0050] Step 2: Perform branch-level graph domain adaptive adversarial transfer training on each approval branch to generate branch-level transfer enhancement embedding vectors;

[0051] For each approval branch Knowledge subgraph based on the selected best migration source type Branch knowledge subgraphs corresponding to emerging target types , perform graph domain adaptive adversarial training to perform branch-level knowledge transfer.

[0052] The above graph domain adaptive adversarial training process includes the following sub-steps:

[0053] Step 201: In the knowledge subgraph of the best migration source type branch Pre-trained branch graph attention encoder Extract source domain entity embedding vectors. Branch graph attention encoder. A Graph Attention Network (GAT) is employed, using the initial feature vectors of each entity in the branch knowledge subgraph as input. The attention weights between each entity and its neighboring entities are calculated through graph attention convolutional layers. The weighted aggregation of neighboring entity features outputs the embedding vectors of each entity. The initial feature vectors of each entity are generated by mapping its attribute information (including approval node name code, approval time limit, required material quantity, etc.) through the embedding layer. Non-numerical attributes such as the approval node name are converted to numerical representations through coded classification data processing, while numerical attributes such as the approval time limit and required material quantity are standardized using Z-scores before participating in the embedding layer mapping. Branch Graph Attention Encoder Knowledge subgraph of the best migration source type branch The above is based on the risk prediction task loss of the approval. To pre-train the network using the supervisory signals, the Adam optimization algorithm is employed to update the network parameters. Specifically, the loss function for the approval risk prediction task is used. For branches The binary cross-entropy loss between the predicted risk output and the actual delay label is processed by a branch graph attention encoder. The fully connected output layer at the end of the connection outputs the delay risk probability of each entity, and calculates it using the historical delay label (1 for delay and 0 for no delay) of each approval node as the monitoring target.

[0054] Step 202: Domain discriminator With branch graph attention encoder Conduct adversarial training. Domain discriminator. This is a binary classification network consisting of two fully connected layers and a sigmoid activation function. It takes an entity embedding vector as input and outputs the probability that the entity embedding vector belongs to the source or target domain. Binary cross-entropy is used as the domain classification loss. Training is conducted. During adversarial training, the domain discriminator... The training objective is to correctly distinguish the source domain embedding vector from the source domain embedding vector, and the target domain embedding vector from the source domain. (Branch graph attention encoder) The training objective is to generate a domain discriminator. It cannot distinguish the embedding vectors from the source domain. The total loss function for the above adversarial training is defined as:

[0055]

[0056] in, For branches Risk prediction of approval tasks and losses Domain discriminator Domain classification loss, To counteract the training weights, which are used to balance the performance of the risk prediction task with the degree of domain alignment. The range of values ​​is A higher value indicates a greater emphasis on domain alignment. This is achieved by alternately optimizing the domain discriminator. and branch graph attention encoder The parameters are adjusted to gradually align the entity embedding distributions of the source and target domains. During adversarial training, the domain discriminator... Gradient ascent is used to optimize the classification loss in the domain. Maximize the domain discriminant The ability to discriminate between source and target domain embedding vectors, branch graph attention encoder Optimize the total loss using gradient descent To minimize the total loss, i.e., to maintain the performance of the risk prediction task while improving the domain discriminator The ability to distinguish between them decreases, thereby driving the embedding distribution of the source domain and the target domain to tend to be consistent.

[0057] Furthermore, to improve the stability of domain alignment during adversarial training, the domain discriminator is also adjusted in step 202 of the adversarial training. The loss function adds a gradient penalty term. gradient penalty term Domain discriminator The gradient norm at the linear interpolation point between the source and target domain embedding vectors is constrained to approach 1. Gradient penalty term. The definition of is:

[0058]

[0059] in, For each random linear interpolation point between the source domain embedding vector and the target domain embedding vector, Domain discriminator At the interpolation point The gradient with respect to the input. for Norm, For random linear interpolation points Calculate the expected value. Gradient penalty term. By constraining the gradient norm of the domain discriminator to around 1, the training instability caused by gradient explosion or vanishing during adversarial training is mitigated.

[0060] Step 203: After the adversarial training converges, construct the branch knowledge subgraph corresponding to the emerging target type. Input to the trained branch graph attention encoder In the output, the branch-level transfer enhancement embedding vectors of entities in each branch of the knowledge subgraph of the emerging target type are generated. .

[0061] Step 3: Construct a multi-branch parallel knowledge graph for the target emerging energy project;

[0062] Based on the existing data on cross-branch influence relationships in the knowledge graphs of various mature energy project types, a general structural pattern of cross-branch influence is extracted. Based on the semantic features of the branch-level migration enhancement embedding vectors of entities in each branch of the target emerging type, potential cross-dependent entities between branches are inferred and cross-influence relationship edges are established. The knowledge subgraphs of each branch are connected through these cross-influence relationship edges to generate a multi-branch parallel knowledge graph of the target emerging energy project.

[0063] Specifically, the above steps include the following sub-steps:

[0064] Step 301: For the knowledge graphs of each mature energy project type, statistically analyze the cross-branch relationship edges connecting different approval branches, and extract the general structural pattern of cross-branch cross-influence. The general structural pattern includes the directionality of cross-influence between approval branches (i.e., which branch's node is the source of influence and which branch's node is the affected party) and the range of influence intensity (i.e., the statistical interval of the weights of similar cross-branch relationship edges in each mature type). The statistical method for the range of influence intensity is as follows: for all cross-branch relationship edge weights of the same directional combination (such as the direction from the safety review branch to the environmental impact assessment branch) in the knowledge graph of each mature type, calculate their mean and standard deviation, and take the interval formed by adding and subtracting one standard deviation from the mean as the range of influence intensity of that directional combination.

[0065] Step 302: For entities in the branch knowledge subgraph of each approval branch of the target emerging type, enhance the embedding vector based on branch-level migration. Calculate the semantic similarity of entity pairs between different branches. For each branch... Entities in and branches Entities in ,in and For two different approval branch categories, cosine similarity is used to calculate semantic similarity. :

[0066]

[0067] in, For branches medium entity Branch-level transfer enhancement embedding vectors, For branches medium entity Branch-level transfer enhancement embedding vectors, for Norm. When semantic similarity Exceeding the preset similarity threshold ( When the branch combination of entity pairs conforms to the cross-branch influence directionality recorded in the general structural pattern, a cross-influence relationship edge is established from the source entity to the affected entity.

[0068] Furthermore, to mitigate the risk of introducing false cross-influence relationships, a branch role consistency check is applied to candidate entity pairs when establishing cross-influence relationships: for the source entity in a candidate entity pair, it is checked whether the source entity is located at a high out-degree key node position in its respective branch's knowledge subgraph; for the affected entity, it is checked whether the affected entity is located at a high in-degree dependent node position in its respective branch's knowledge subgraph. The cross-influence relationship is only confirmed when both the source entity's out-degree ranking and the affected entity's in-degree ranking are within the pre-preset range of their respective branch's knowledge subgraph. This check ensures that the cross-influence relationship connects key nodes with actual approval flow influence between branches.

[0069] Step 303: Using the influence intensity range in the general structural pattern as a priori, initialize the weights of each cross-influence relationship edge established in Step 302. Specifically, for branches... To branch For the cross-influence relationship edges of the directions, initialize the weight of the cross-influence relationship edges to the median value of the influence intensity range of the corresponding direction in the general structural pattern.

[0070] Furthermore, the median value of the aforementioned influence intensity range is taken from the midpoint of the interval formed by adding or subtracting one standard deviation from the mean of the corresponding directional combinations statistically analyzed in step 301. That is, the mean of this interval is used as the weight initialization value, so that the initial weight of the cross-influence relationship edge is taken within a reasonable range supported by the historical data of mature types.

[0071] Step 304: Connect the branch knowledge subgraphs of each approval branch of the target emerging type through cross-influence relationship edges to generate a multi-branch parallel knowledge graph of the target emerging energy project. .

[0072] Step 4: Perform branch-aware graph attention encoding processing on the multi-branch parallel knowledge graph to generate transfer co-fusion embedding vectors;

[0073] Multi-branch parallel knowledge graph The branch-aware graph attention encoding process is performed. Within each branch knowledge subgraph, graph attention convolution is performed based on branch-level transfer augmentation embedding vectors to update entity embeddings; cross-branch attention propagation is performed along the cross-influence relationship edges between branches, and the direction and intensity of information flow are guided by the pre-initialized weights of the cross-influence relationship edges during the propagation; after multiple rounds of iterative propagation, a multi-branch parallel knowledge graph is generated. The migration, collaborative fusion, and embedding vectors of various entities.

[0074] Specifically, the above steps include the following sub-steps:

[0075] Step 401: Perform branch-internal graph attention convolution within each branch knowledge subgraph. For each branch... Entities in In physical form Branch-level transfer enhancement embedding vector As input, compute entity With entity The attention coefficients among neighboring entities within the branch knowledge subgraph are used to weight and aggregate the embedding vectors of these neighboring entities, outputting the updated branch-internal embedding vector. :

[0076]

[0077] in, For entities In the branch The set of neighborhood entities within a knowledge subgraph For neighborhood entity set Neighborhood entities in The intra-branch attention coefficient, The characteristic transformation matrix within the branch. For neighboring entities In the branch Branch-level transfer enhancement embedding vectors in the middle The activation function. Intra-branch attention coefficients. By the entity with neighboring entities The embedded vectors are concatenated and then calculated using linear transformation and Softmax normalization. Specifically, let... ,in This is a learnable attention parameter vector. Indicates transpose. If we denote the vector concatenation operation, then:

[0078]

[0079] in, Represents an exponential function. For neighborhood entity set Neighborhood entities in For entities with neighboring entities The original attention score after linear transformation.

[0080] Furthermore, to reduce the interference of noise signals in sparse branches on the overall graph coding quality, signal attenuation gating is applied to the knowledge subgraphs of branches with data sparsity exceeding a preset threshold. For branches... Data sparsity index Defined as the reciprocal of the ratio of the actual number of relation edges in the knowledge subgraph of this branch to the maximum number of relation edges that can be formed by all entity pairs in the knowledge subgraph of this branch. The range of values ​​is The larger the value of the sparsity index, the sparser the data in the knowledge subgraph of that branch. Exceeding the preset threshold At that time, a decay gate is applied to the in-branch embedding vector of each entity in the knowledge subgraph of that branch after graph attention convolution:

[0081]

[0082] in, For branches The signal attenuation gating factor is defined as:

[0083]

[0084] when hour, A value of 1 means no attenuation is applied. This gating ensures that branches with higher data sparsity contribute less to cross-branch information propagation, thus reducing noise propagation.

[0085] Furthermore, the aforementioned preset threshold The value is determined based on the sparsity index of the branch knowledge subgraph. Exceeding the preset threshold When this occurs, it means that the actual relation edge density in the knowledge subgraph of this branch has decreased to a level that may introduce significant noise, at which point the attenuation gating coefficient is used. This reduces the magnitude of the embedding vector of the knowledge subgraph of that branch proportionally to its sparsity, thereby reducing the information contribution weight of that branch in subsequent cross-branch attention propagation.

[0086] Step 402: Perform cross-branch attention propagation along the edges that cross influence relationships between branches. For entities Collect relationships between entities through cross-influence edges. Embedded information of entities in other connected branches, based on prior initialization weights of cross-influence relationship edges. Guide the intensity of information flow and calculate cross-branch aggregated embedding vectors. :

[0087]

[0088] in, To influence relationships between edges and entities through cross-influence A connected set of cross-branch entities. For cross-branch entity collection Cross-branch entities in To be from the entity To the entity The prior initial weights of the cross-influence relationship edges, For cross-branch attention coefficients, This is the cross-branch feature transformation matrix. For entities In entity Belonging to branch The branch embedding vector updated after step 401 of the inner branch sequence. Cross-branch attention coefficients. By the entity With cross-branch entities The embedded vectors are concatenated and then calculated using linear transformation and Softmax normalization, let... ,in A learnable cross-branch attention parameter vector. To indicate transpose, then:

[0089]

[0090] in, Represents an exponential function. For cross-branch entity collection Cross-branch entities in For entities With cross-branch entities The original cross-branch attention score after linear transformation.

[0091] Furthermore, to ensure that the direction of information flow during cross-branch attention propagation more accurately reflects the approval dependencies between branches, the cross-branch attention coefficient was also adjusted. Apply directional masking constraints: Based on the directionality of inter-branch cross-influence recorded in the general structural pattern, attention coefficients that do not conform to the influence direction (i.e., back propagation from a non-influence source branch to an influence source branch) are forced to zero during back propagation. This constraint ensures that information propagates only along the influence direction recorded in the general structural pattern, avoiding unreasonable feature confusion introduced by back propagation.

[0092] Step 403: Embed the vector within the branch With cross-branch aggregation embedding vector Merge to generate entities The current round's fused embedding vector is used as the input embedding for the next round's iteration. After a preset number of rounds... After iterative propagation, the output is the migration collaborative fusion embedding vector of each entity. The fusion computing method is as follows:

[0093]

[0094] in, For layer normalization operation, This is the preset number of iteration rounds.

[0095] Furthermore, the aforementioned preset number of rounds The choice between intra-branch embedding vectors and cross-branch aggregated embedding vectors in multi-branch parallel knowledge graphs determines the relationship between intra-branch embedding vectors and cross-branch aggregated embedding vectors. The depth of information propagation within the process. After each iteration, the fused embedding vector for the current iteration... The input embedding of the corresponding entity is replaced and used as the starting point for the computation of the next round of steps 401 and 402. This allows the embedded representations of each entity to gradually integrate structural and semantic information from more distant cross-branch neighborhoods in each iteration, after a preset number of rounds. The final output of the transfer co-fusion embedding vector after rounds of iterations This refers to the migration, collaboration, fusion, and embedding vectors of each entity.

[0096] Step 5: Perform approval risk prediction and output comprehensive risk assessment results;

[0097] The current progress data of each approval branch of the target emerging energy project is obtained. The current progress data and the migration collaborative fusion embedding vector are then concatenated and input into a deep integrated Bayesian inference network to output the overall approval comprehensive risk level and key bottleneck branch identifiers.

[0098] Specifically, the above steps include the following sub-steps:

[0099] Step 501: Obtain the current progress data for each approval branch of the target emerging energy project. The current progress data includes the current approval node identifier, elapsed time, and the number of remaining approval nodes for each branch. Perform feature encoding on the current progress data to generate a progress feature vector for each branch. In this process, non-numerical attributes such as approval node identifiers are converted into numerical representations through coded classification data processing, while numerical attributes such as time consumed and the number of remaining approval nodes are standardized by Z-score before participating in feature encoding to eliminate the influence of different units on feature concatenation operations.

[0100] Furthermore, the aforementioned elapsed time refers to the actual time elapsed from the initiation of each approval branch to the current moment. The elapsed time, together with the remaining number of approval nodes, constitutes a dynamic characteristic reflecting the current approval progress status of each branch, making the progress feature vector... It can capture the actual progress pressure of each branch at the current time point, thereby enabling the subsequent synthesis of feature vectors. The construction process provides time-sensitive progress information for risk prediction.

[0101] Step 502: For each approval branch , progress feature vector The migration co-fusion embedding vectors of each entity in this branch The branch-level fusion representation vectors obtained after the graph readout operation (i.e., mean pooling of the migration collaborative fusion embedding vectors of all entities within the branch) are concatenated to generate the comprehensive feature vector of that branch. .

[0102] Step 503: Combine the feature vectors of each branch The input is fed into a deep ensemble Bayesian inference network. This deep ensemble Bayesian inference network contains multiple member networks, each corresponding to a different approval branch. ,in The total number of branches approved. For the first Each approval branch corresponds to a member network. Each member network... For a Bayesian neural network, the comprehensive feature vector of the corresponding branch is used. As input, each member network The network weights follow a predefined prior distribution. During training, the posterior distribution of the weights is learned through variational inference. The evidence lower bound (ELBO) loss is used as the training objective, and the Adam optimization algorithm is employed to update the network parameters. Each member network... The output layer is a single-node fully connected layer connected to a Sigmoid activation function, outputting the predicted delay probability of that branch. During training, the historical delay labels (1 for delay, 0 for no delay) of each approval branch are used as the supervision target. The evidence lower bound (ELBO) loss is composed of the difference between the reconstruction term (binary cross-entropy between the predicted delay probability and the true label) and the regularization term (KL divergence between the weight posterior distribution and the prior distribution). During the inference phase, forward propagation is performed by multiple Monte Carlo samplings of the weight posterior distribution, and the mean of the delay probabilities output by the multiple sampling forward propagation is used as the independent delay probability. The variance of the delay probability output from multiple sampling forward propagation is used as the branch-level cognitive uncertainty component. Among them, the branch-level cognitive uncertainty component This reflects the degree of uncertainty in model predictions caused by insufficient training data for this branch.

[0103] Furthermore, to enable member networks to share cross-branch information and improve the consistency of predictions across branches, additional measures are taken in each member network. A cross-member feature interaction layer is added between member networks: After the output of the intermediate hidden layers of each member network, the hidden layer feature vectors of each member network are concatenated and processed by a shared attention aggregation layer to generate a global context vector. This global context vector is then concatenated with the hidden layer feature vectors of each member network and input into the subsequent layers of each member network. This cross-member feature interaction layer allows the prediction of each branch to reference the state information of other branches, thereby improving the accuracy of identifying key bottleneck branches.

[0104] Step 504: Based on evidence theory, dynamically weight and fuse the independent delay probabilities of each branch according to the cognitive uncertainty components at the branch level. This involves adjusting the independent delay probabilities of each branch. Based on the branch-level cognitive uncertainty components of each branch Calculate the credibility weight of the evidence for this branch. :

[0105]

[0106] in, The branch index is used to sum all approval branches, representing a branch-level cognitive uncertainty component. The larger the value, the greater the credibility weight of the evidence in that branch. The lower the value. Based on the evidence credibility weights of each branch. Independent delay probability for each branch Weighted fusion is performed to calculate the overall approval comprehensive risk index. :

[0107]

[0108] Based on the overall approval risk index Based on the correspondence with the preset risk level threshold range, the overall approval comprehensive risk level is output.

[0109] Furthermore, the aforementioned preset risk level threshold range will be used to determine the overall approval comprehensive risk index. The range of values Divided into several consecutive sub-intervals, each sub-interval corresponds to a risk level, and the overall approval comprehensive risk index is calculated. A higher value corresponds to a higher risk level, thus increasing the overall risk index for continuous approvals. The risk level is mapped to a discrete output for direct use by decision-makers.

[0110] Step 505: Calculate the independent delay probability for each branch. Sort in descending order and select the evidence with the highest probability of independent delay and the highest credibility weight. Exceeding the preset confidence threshold ( The branch at each branch level is designated as a critical bottleneck branch, and a critical bottleneck branch identifier is output. Simultaneously, based on the cognitive uncertainty components at each branch level... Calculate the migration reliability score for each branch. And output:

[0111]

[0112] Migration reliability score The range of values ​​is A higher value indicates that the transferred knowledge of that branch is more reliable, where the branch-level cognitive uncertainty component... This is the variance, and its range is [value range missing]. This is guaranteed by the Monte Carlo sampling output of the Bayesian neural network. Decision-makers can score the transfer reliability. Identify the branch with the lowest prediction confidence and prioritize data supplementation or manual review for that branch.

[0113] In this implementation, because the kernel similarity between the target emerging type branch knowledge subgraph and the corresponding mature type branch knowledge subgraph is independently calculated for each approval branch in step 1, and the best migration source type is selected for each approval branch according to the kernel similarity, different approval branches can each be matched with the mature type branch most similar to their graph structure as the migration source. In step 2, because independent graph domain adaptive adversarial training is performed on each branch, the entity embedding distribution alignment process of each branch only involves the migration source domain that best matches the technical features of that branch. This avoids the situation where unified migration leads to a mismatch in the graph knowledge of some branches, and enables the entity embedding of each branch knowledge subgraph to obtain targeted migration enhancement. Therefore, it overcomes the problem of reduced risk prediction accuracy caused by unified migration.

[0114] In this implementation, a general structural pattern of cross-branch cross-influence is extracted from multiple mature type knowledge graphs in step 3. The potential cross-dependencies are inferred by combining the semantic features of each branch entity of the target emerging type. The influence intensity range in the general structural pattern is used as a priori to initialize the weights of the cross-influence relationship edges. Therefore, even when emerging energy projects lack historical data on cross-influence between branches, it can still provide a cross-branch association representation with reasonable prior constraints for multi-branch parallel knowledge graphs. Thus, it overcomes the problem of no historical data to support the cross-influence relationship between branches of emerging types.

[0115] Furthermore, in this implementation, because signal attenuation gating is applied to the knowledge subgraphs of branches with data sparsity exceeding a preset threshold in step 4, noise signals in sparse branches are suppressed during cross-branch propagation, reducing interference with the overall graph coding quality. Because a deep integrated Bayesian inference network is used in step 5 to output the branch-level cognitive uncertainty components of each branch, and dynamic weighted fusion is performed based on evidence theory, the risk prediction results not only include the overall approval risk level and key bottleneck branch identification, but also a quantitative assessment of the reliability of knowledge transfer in each branch. Therefore, this provides a basis for decision-makers to identify the branches with the lowest prediction confidence for priority data supplementation.

[0116] The following is an example of an application of the present invention, such as... Figure 2-8 As shown, the implementation process is as follows:

[0117] In 20XX, a provincial energy regulatory platform received a risk prediction task for the approval of a hydrogen production, storage, and transportation project (Project No.: PROJ-H001). This project involved three parallel approval branches: Safety Assessment (SAF), Environmental Impact Assessment (ENV), and Hazardous Chemicals Permit (HAZ). Because hydrogen production, storage, and transportation projects are emerging energy projects, the platform only had four valid historical approval case records, resulting in severely sparse data. The platform currently has historical approval data available for migration and reference for three mature energy project types: liquefied natural gas (LNG) storage and transportation projects, hydrogen refueling station construction projects (HRS), and compressed natural gas (CNG) projects, each with over 120 historical approval records. The platform needs to identify the optimal migration sources for each of the three approval branches of PROJ-H001, thereby predicting the overall comprehensive approval risk level and identifying key bottleneck branches.

[0118] The platform first labeled the historical data of the three approval branches of PROJ-H001 and the corresponding branches of the three mature projects of LNG, HRS and CNG, and then constructed a knowledge subgraph for each branch. Taking the safety review branch as an example, its subgraph entity set includes approval nodes such as "high pressure hydrogen storage tank inspection", "leakage and diffusion risk assessment" and "explosion-proof area delineation", and the relationship edges include flow relationships such as "preconditions" and "parallel dependencies".

[0119] For each approval branch, the platform calculates the Weisfeiler-Lehman subtree kernel similarity between the PROJ-H001 target subgraph and the corresponding branch subgraphs of the three maturity types, and determines the best migration source for each branch in descending order of similarity.

[0120] Table 1. Similarity calculation results of knowledge graph-based energy project risk prediction for each approval branch and selection of the best migration source:

[0121]

[0122] The three approval branches were matched with different optimal migration sources: the SAF branch was most similar to the HRS type (hydrogen refueling station construction projects) in terms of approval graph structure, with a kernel similarity of 0.847; the ENV branch was most similar to the LNG type (liquefied natural gas storage and transportation projects), with a kernel similarity of 0.783; and the HAZ branch was most similar to the CNG type (compressed natural gas projects), with a kernel similarity of 0.809. This branch differentiation result verifies the necessity of independently screening migration sources for each branch.

[0123] Taking the SAF branch as an example, the platform pre-trains the branch graph attention encoder based on the SAF branch knowledge subgraph of the HRS class. The SAF branch subgraph contains 28 approval node entities. The initial feature vector of each entity consists of three attributes: approval node name encoding, approval time limit (unit: working days), and required material quantity. Among them, the approval time limit and the required material quantity are used in the embedding layer mapping after Z-score standardization.

[0124] After pre-training convergence, the platform introduces a domain discriminator to perform adversarial training. The adversarial training weights are based on knowledge graph-based energy project risk prediction. The total loss function is:

[0125]

[0126] During adversarial training, a gradient penalty term is applied simultaneously, ensuring that the gradient norm of the discriminator in the constraint domain approaches 1 at the interpolation point, thus guaranteeing training stability. After the adversarial training converges, the SAF branch knowledge subgraph of PROJ-H001 is input into the encoder, outputting the branch-level transfer enhancement embedding vectors for each entity in the SAF branch. The ENV and HAZ branches perform the same process based on LNG and CNG class transfer sources, respectively.

[0127] Table 2 Key parameters for adversarial training of three approval branches in knowledge graph-based energy project risk prediction:

[0128]

[0129] After each of the three branches completed adversarial training, all entities in the three branches of PROJ-H001 obtained branch-level transfer-enhanced embedding vectors for their best transfer sources, providing a semantic foundation for the subsequent construction of a multi-branch parallel knowledge graph.

[0130] The platform first statistically analyzes the cross-branch relationships in the knowledge graphs of three mature project types: LNG, HRS, and CNG, extracting common structural patterns. Taking the "SAF branch → ENV branch" direction as an example, the weights of all cross-branch relationships in this direction across the three mature project types are calculated, yielding a mean of 0.54 and a standard deviation of 0.09. The influence intensity range is within the scope of knowledge graph-based energy project risk prediction. The initial weight value (i.e., the mean) is 0.54.

[0131] Table 3. Statistical results of the general structural pattern of cross-branch cross-impact in energy project risk prediction based on knowledge graphs:

[0132]

[0133] Subsequently, the platform calculated the cosine similarity of each pair of entities in PROJ-H001. Taking the entity "High-Pressure Hydrogen Storage Tank Inspection" (denoted as entity P) in the SAF branch and the entity "Hydrogen Leakage and Diffusion Impact Assessment" (denoted as entity Q) in the ENV branch as examples, their semantic similarity was calculated as follows:

[0134]

[0135] This value exceeds the preset similarity threshold for knowledge graph-based energy project risk prediction. Furthermore, the "SAF→ENV" direction conforms to the influence directionality recorded in the general structural pattern. Simultaneously, the "High-Pressure Hydrogen Storage Tank Inspection" node ranks in the top 20% of the out-degree in the SAF subgraph, and the "Hydrogen Leakage and Diffusion Impact Assessment" node ranks in the top 20% of the in-degree in the ENV subgraph. Through branch role consistency verification, the cross-influence relationship edge is confirmed, with an initial weight of 0.54. Finally, the three branch knowledge subgraphs are connected through all confirmed cross-influence relationship edges to generate a multi-branch parallel knowledge graph containing a total of 28 cross-branch relationship edges.

[0136] The platform first calculates the data sparsity index for the three branches. Taking the HAZ branch as an example, its subgraph contains 8 entities, and the maximum possible number of relational edges is [missing information - likely related to knowledge graph-based energy project risk prediction]. If the actual number of relation edges is 6, then the sparsity index is:

[0137]

[0138] Preset sparsity threshold for knowledge graph-based energy project risk prediction The sparsity index of the HAZ branch is used for energy project risk prediction based on knowledge graphs. The trigger signal attenuation is gated, and the gate coefficient is:

[0139]

[0140] Table 4. Data sparsity and gating coefficients for three approval branches in energy project risk prediction based on knowledge graphs:

[0141]

[0142] All three branches triggered signal attenuation gating, with the HAZ branch having the highest sparsity and the smallest gating coefficient (0.536), thus its contribution weight in cross-branch information propagation was compressed to the greatest extent.

[0143] After completing the intra-branch attention convolution and applying gating, cross-branch attention propagation is performed along the 28 cross-branch relationship edges. The prior weights initialized in Table 3 are used to guide the information flow intensity, while attention coefficients that do not conform to the influence direction of the general structural pattern are forcibly set to zero. Subsequently, the intra-branch embedding vector and the cross-branch aggregated embedding vector are fused through layer normalization. After three rounds of iterative propagation, the migration collaborative fusion embedding vector of all entities in PROJ-H001 is output.

[0144] The platform obtains the current progress data of PROJ-H001 on September 15, 20XX, and constructs progress feature vectors for the three branches respectively. The time consumed and the number of remaining approval nodes are standardized by Z-score.

[0145] Table 5. Current progress data of each approval branch of the knowledge graph-based energy project risk prediction PROJ-H001:

[0146]

[0147] The progress feature vector is concatenated with the mean pooling results of the collaborative fusion embedding vectors of each branch's migration to generate a comprehensive feature vector, which is then input into a deep ensemble Bayesian inference network. Each member network undergoes 50 Monte Carlo sampling forward propagation cycles, outputting the independent delay probability of each branch and the branch-level cognitive uncertainty component.

[0148] The credibility weight of evidence is calculated based on the cognitive uncertainty component of each branch, taking the SAF branch as an example:

[0149]

[0150] Table 6. Risk prediction outputs and integrated risk data for each approval branch of the knowledge graph-based energy project risk prediction:

[0151]

[0152] The overall approval risk index is calculated as follows:

[0153]

[0154] The preset risk level threshold range is: Knowledge graph-based energy project risk prediction corresponds to low risk. Knowledge graph-based risk prediction for energy projects corresponds to medium risk. Knowledge graph-based risk prediction for energy projects corresponds to high risk. The risk prediction of the energy project based on the knowledge graph fell into the high-risk range, and the platform output the overall approval risk level as "high risk".

[0155] The SAF branch had the highest probability of independent delay (0.82), and its evidence credibility weight of 0.373 exceeded the preset credibility threshold. (This is related to knowledge graph-based energy project risk prediction.) The platform outputs the SAF branch (Security Review) as a critical bottleneck branch. Meanwhile, the HAZ branch has the lowest migration reliability score (0.69), prompting decision-makers to prioritize supplementing the HAZ branch with historical data or conducting manual review.

[0156] Throughout the implementation process, the data starts from the limited historical approval records of PROJ-H001. Step 1 involves constructing a branch knowledge subgraph and calculating graph kernel similarity to determine differentiated migration sources (HRS, LNG, and CNG) for the three branches. Step 2 involves performing adversarial transfer training based on the best migration sources for each branch, generating branch-level embedding vectors carrying transfer knowledge. Step 3 utilizes the common structural patterns of mature projects to infer and initialize cross-branch relationship edges, connecting the three branch subgraphs into a multi-branch parallel knowledge graph. Step 4 uses branch-aware graph attention encoding with sparsity gating to fuse intra-branch structural information and cross-branch cross-influence information into a transfer collaborative fusion embedding vector. Step 5 concatenates the embedding vector with the current progress data and outputs the delay probability and uncertainty of each branch through a Bayesian inference network. Finally, through dynamic weighted fusion based on evidence theory, the comprehensive risk level "high risk", the key bottleneck branch "safety review", and the transfer reliability score of each branch are output, realizing a complete data flow from raw sparse data to quantifiable uncertainty risk prediction results.

[0157] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A knowledge graph-based method for predicting energy project risks, characterized in that, Includes the following steps: Obtain approval data for target emerging energy projects and historical approval data for multiple mature energy project types. Construct branch knowledge subgraphs for each approval branch according to the approval branch category. Calculate the graph kernel similarity between the branch knowledge subgraph of the target emerging type and the corresponding branch knowledge subgraph of each mature type for each approval branch category. Independently select the best migration source type branch knowledge subgraph for each approval branch. For each approval branch, graph domain adaptive adversarial training is performed based on the selected best migration source type branch knowledge subgraph and the branch knowledge subgraph corresponding to the target emerging type to generate branch-level migration enhancement embedding vectors of entities in each branch knowledge subgraph of the target emerging type. Extract the general structural pattern of cross-branch cross-influence from the knowledge graph of each mature energy project type, infer the potential cross-dependency entities between branches based on the semantic features of the branch-level transfer enhanced embedding vector, and establish cross-influence relationship edges. Connect the knowledge subgraphs of each branch through the cross-influence relationship edges to generate a multi-branch parallel knowledge graph. On a multi-branch parallel knowledge graph, graph attention convolution is performed inside each branch knowledge subgraph to update entity embeddings, and cross-branch attention propagation is performed along the cross-influence relationship edges. After multiple rounds of iteration, the migration collaborative fusion embedding vectors of each entity are generated. Obtain the current progress data of each approval branch of the target emerging energy project, and input the current progress data and the migration collaborative fusion embedding vector into a deep integrated Bayesian inference network after feature concatenation. Output the overall approval comprehensive risk level and key bottleneck branch identifier.

2. The knowledge graph-based energy project risk prediction method according to claim 1, characterized in that, The graph kernel similarity is calculated using the Weisfeiler-Lehman subtree kernel algorithm. The Weisfeiler-Lehman subtree kernel algorithm generates refined labels for each node by iteratively aggregating the label information of neighboring nodes in the graph, and counts the co-occurrence frequency of the same refined labels in the two graphs in each iteration round. The inner product of the refined label frequency vectors of the two graphs is used as the graph kernel similarity output value.

3. The knowledge graph-based energy project risk prediction method according to claim 1, characterized in that, The graph domain adaptive adversarial training includes: A branch graph attention encoder is pre-trained on the knowledge subgraph of the best transfer source type branch, and the source domain entity embedding vector is extracted using the approval risk prediction task loss as a supervision signal. The domain discriminator and the branch graph attention encoder are subjected to adversarial training. The training objective of the domain discriminator is to correctly distinguish the source domain of the source domain embedding vector and the target domain embedding vector. The training objective of the branch graph attention encoder is to generate embedding vectors that prevent the domain discriminator from distinguishing the source domain. The total loss function of the adversarial training is the loss of the approval risk prediction task minus the product of the adversarial training weight coefficient and the domain classification loss. The value of the adversarial training weight coefficient is an open interval from zero to one. After the adversarial training converges, the branch knowledge subgraph corresponding to the emerging target type is input into the trained branch graph attention encoder, and the branch-level transfer enhancement embedding vector is output.

4. The knowledge graph-based energy project risk prediction method according to claim 3, characterized in that, In adversarial training, a gradient penalty term is added to the loss function of the domain discriminator. The gradient penalty term constrains the gradient norm of the domain discriminator at the random linear interpolation point between the source domain embedding vector and the target domain embedding vector. The gradient penalty term is defined as the expected value of the square of the difference between the L2 norm of the gradient of the domain discriminator with respect to the input at the random linear interpolation point and one.

5. The knowledge graph-based energy project risk prediction method according to claim 1, characterized in that, The extraction of the general structural pattern of cross-branch cross-influence and the establishment of cross-influence relationship edges include: For the knowledge graph of each mature energy project type, the cross-branch relationship edges connecting different approval branches are statistically analyzed, and the general structural pattern is extracted. The general structural pattern includes the directionality and influence intensity range of the cross-influence between approval branches. The influence intensity range is the interval formed by the mean of the weights of all cross-branch relationship edges of the same direction combination in each mature type plus or minus one standard deviation. Based on branch-level transfer enhancement embedding vectors, cosine similarity is used to calculate the semantic similarity of entity pairs between different branches. When the semantic similarity exceeds a preset similarity threshold and the branch combination of the entity pair conforms to the cross-branch influence directionality in the general structural pattern, a cross-influence relationship edge is established from the source entity to the affected entity. For each established cross-influence relationship edge, the weight is initialized to the median value of the influence intensity range in the corresponding direction in the general structural pattern.

6. The knowledge graph-based energy project risk prediction method according to claim 5, characterized in that, When establishing the cross-influence relationship edge, a branch role consistency check is also applied to the candidate entity pair: for the influencing source entity, it is checked whether the out-degree ranking of the influencing source entity in its respective branch knowledge subgraph is within a preset proportion range; for the affected entity, it is checked whether the in-degree ranking of the affected entity in its respective branch knowledge subgraph is within a preset proportion range; only when the out-degree ranking of the influencing source entity and the in-degree ranking of the affected entity are both within the preset proportion range of their respective branch knowledge subgraphs is the cross-influence relationship edge established.

7. The knowledge graph-based energy project risk prediction method according to claim 1, characterized in that, When updating entity embeddings by performing graph attention convolution within each branch knowledge subgraph, signal attenuation gating is also applied to branch knowledge subgraphs where the data sparsity exceeds a preset threshold. The data sparsity index is defined as the reciprocal of the ratio of the actual number of relation edges in the branch knowledge subgraph to the maximum number of relation edges formed by all entity pairs. When the data sparsity index exceeds the preset threshold, the branch embedding vector of each entity within the branch knowledge subgraph after graph attention convolution is multiplied by a signal attenuation gating coefficient, which is defined as the ratio of the preset threshold to the data sparsity index. When the data sparsity index does not exceed the preset threshold, the signal attenuation gating coefficient is set to one.

8. The knowledge graph-based energy project risk prediction method according to claim 1, characterized in that, When performing cross-branch attention propagation along the edges of the cross-influence relationship, a directional mask constraint is also applied to the cross-branch attention coefficients: based on the directionality of cross-influence between branches recorded in the general structural pattern, the attention coefficients of reverse propagation that do not conform to the influence direction are forced to be set to zero, so that information is only propagated along the influence direction recorded in the general structural pattern.

9. The knowledge graph-based energy project risk prediction method according to claim 1, characterized in that, The deeply integrated Bayesian inference network comprises multiple member networks corresponding to each approval branch. Each member network is a Bayesian neural network, and the network weight parameters follow a preset prior distribution. The posterior distribution of the weight parameters is learned through variational inference. During the inference phase, forward propagation is performed by multiple Monte Carlo samplings of the posterior distribution of the weights. The mean of the delay probabilities output from multiple samplings is taken as the independent delay probability, and the variance of the delay probabilities output from multiple samplings is taken as the branch-level cognitive uncertainty component. Based on evidence theory, the independent delay probabilities of each branch are dynamically weighted and fused according to the branch-level cognitive uncertainty component. The evidence credibility weight of each branch is defined as the ratio of one minus the branch-level cognitive uncertainty component of that branch to the sum of one minus the branch-level cognitive uncertainty components of all branches. The overall approval comprehensive risk index is the weighted sum of the evidence credibility weights and independent delay probabilities of each branch. The overall approval comprehensive risk level is output according to the correspondence between the overall approval comprehensive risk index and the preset risk level threshold range. The branch with the highest independent delay probability and evidence credibility weight exceeding the preset credibility threshold is identified as the key bottleneck branch, and the key bottleneck branch identifier is output.

10. A knowledge graph-based energy project risk prediction system, used to execute the knowledge graph-based energy project risk prediction method according to any one of claims 1 to 9, characterized in that, include: The branch knowledge subgraph construction and migration source screening module is used to obtain the approval data of the target emerging energy project and the historical approval data of multiple mature energy project types. It constructs branch knowledge subgraphs for each approval branch according to the approval branch category, calculates the graph kernel similarity for each approval branch category, and independently screens the best migration source type branch knowledge subgraph for each approval branch. The branch-level adversarial transfer training module is used to perform graph domain adaptive adversarial training on each approval branch based on the selected best transfer source type branch knowledge subgraph and the branch knowledge subgraph corresponding to the target emerging type, and generate branch-level transfer enhanced embedding vectors. The multi-branch parallel knowledge graph construction module is used to extract the general structural pattern of cross-branch cross-influence from the knowledge graphs of various mature energy project types. Based on the semantic features of the branch-level transfer enhanced embedding vector, it infers the potential cross-dependency entities between branches and establishes cross-influence relationship edges. The knowledge subgraphs of each branch are connected through cross-influence relationship edges to generate a multi-branch parallel knowledge graph. The branch-aware graph attention encoding module is used to perform graph attention convolution to update entity embeddings within the knowledge subgraphs of each branch on a multi-branch parallel knowledge graph, perform cross-branch attention propagation along the edges of cross-influence relationships, and generate the migration and collaborative fusion embedding vectors of each entity through multiple rounds of iteration. The approval risk prediction output module is used to obtain the current progress data of each approval branch of the target emerging energy project. After the current progress data and the migration collaborative fusion embedding vector are concatenated for feature splicing, they are input into the deep integrated Bayesian inference network to output the overall approval comprehensive risk level and key bottleneck branch identifiers.

Images