A regional power grid collaboration method based on a knowledge graph

Abstract

The application discloses a kind of regional power grid cooperation methods based on knowledge graph driving, comprising the following steps: collecting regional power grid data, generating multi-source heterogeneous power grid data set;Construct regional power grid knowledge graph, initialize federation heterogeneous graph neural network and graph neural controlled differential equation joint framework;Perform federation heterogeneous graph neural network coding, generate local embedding, mode weight decoupling parameter and local coding parameter;Run graph neural controlled differential equation modeling, generate continuous time state trajectory and dynamics parameter set;Perform federation aggregation on client set results, generate global model parameters;Issue global model parameters and update local model;Based on knowledge graph data, control path input and updated model parameters, generate regional power grid cooperation result set.The application significantly improves the cross-regional power grid cooperation prediction accuracy and calculation efficiency, effectively guarantees the operation safety and dispatching reliability.

Description

Technical Field

[0001] This invention relates to the field of regional power grid dispatching and optimization, and in particular to a knowledge graph-driven regional power grid coordination method. Background Technology

[0002] With the continuous expansion of regional power grids, cross-regional power resource dispatch and optimized operation have become core tasks for ensuring the safe, stable, and economically efficient operation of the power system. Existing regional power grid coordination methods largely rely on centralized dispatch models, characterized by aggregating operational monitoring information and dispatch plans through a unified data center before performing power flow calculations and control decisions. However, this centralized approach suffers from problems such as excessive communication load, data privacy risks, and difficulty in real-time integration of cross-regional dynamic information, making it difficult to meet the dual requirements of real-time performance and security for modern power systems.

[0003] At the information processing level, current methods typically employ single data mining or statistical analysis techniques to model power grid operation characteristics, lacking the ability to uniformly represent multi-source heterogeneous data. Especially when dealing with multi-dimensional information such as equipment ledgers, metering curves, and the external environment, traditional modeling methods struggle to achieve structured integration, leading to data silos in cross-regional collaborative computing and impacting the overall accuracy of the model.

[0004] In the field of intelligent computing, while existing research has introduced partial graph computing models or federated learning methods, these are often limited to static graphs or independent training frameworks, lacking a detailed characterization of the temporal evolution of power grid operation. Furthermore, inadequate parameter alignment and dynamic constraint handling across regions often lead to significant deviations in collaborative results, making it difficult to support power grid security margin assessment and power exchange optimization in complex scenarios.

[0005] Therefore, how to provide a knowledge graph-driven regional power grid coordination method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a knowledge graph-driven regional power grid collaboration method. This invention combines federated heterogeneous graph neural networks and graph neural controlled differential equations to model and align multi-source heterogeneous power grid data within a joint framework, generating continuous-time state trajectories, global model parameters, and cross-regional collaborative inference results. It describes in detail the entire process of regional power grid collaborative computing and has the advantages of strong data privacy protection, high dynamic modeling accuracy, and excellent cross-regional collaborative efficiency.

[0007] A knowledge graph-driven regional power grid coordination method according to an embodiment of the present invention includes the following steps:

[0008] Collect regional power grid data, perform data preprocessing operations, and generate a multi-source heterogeneous power grid data set;

[0009] Based on a multi-source heterogeneous power grid data set, a regional power grid knowledge graph data is constructed, a client set and a client heterogeneous graph pattern set are established, a joint framework of federated heterogeneous graph neural network and graph neural controlled differential equation is initialized, and initial entity embedding, initial relation embedding and continuous time control path input are generated.

[0010] Within the joint framework, based on the client set, the federated heterogeneous graph neural network encoding process is executed. Based on the regional power grid knowledge graph data and the client heterogeneous graph pattern set, the initial entity embedding and initial relation embedding are locally updated to generate local entity embedding, local relation embedding, pattern weight decoupling parameters and local encoding parameters.

[0011] Within the joint framework, based on local entity embedding, local relation embedding, and continuous-time control path input, the graph neural controlled differential equation modeling process is executed to generate a continuous-time state trajectory and dynamic parameter set.

[0012] Federated alignment and aggregation are performed on the local encoding parameters, mode weight decoupling parameters, continuous-time state trajectories and dynamic parameter sets of the client set to generate global model parameters;

[0013] The global model parameters are distributed to the client set to update the parameters in the federated heterogeneous graph neural network encoding process and the graph neural controlled differential equation modeling process, and to generate refined local model parameters.

[0014] Based on regional power grid knowledge graph data, continuous-time control path inputs, and refined local model parameters, a cross-regional collaborative reasoning input is constructed, cross-regional parameter alignment and constraint calculations are performed, and a set of regional power grid collaborative results is generated.

[0015] Optionally, the generation of the multi-source heterogeneous power grid dataset includes:

[0016] Collect operational monitoring data, scheduling plan data, equipment ledger data, metering curve data, and external environment data, and establish a unified timestamp sequence and unit standard to form an initial data collection set;

[0017] Perform data cleaning operations on the initial collected dataset, including outlier removal, missing value imputation, and duplicate sample removal, to generate a cleaned dataset;

[0018] Standardization processing operations are performed on the cleaned dataset, including unification of units and normalization of numerical ranges, to generate a standardized dataset;

[0019] Perform pattern mapping operations on a standardized dataset to map fields to node and edge attributes of regional power grid knowledge graph data, forming a mapping structure consistent with the topology of the regional power grid.

[0020] The pattern mapping results are aggregated into a multi-source heterogeneous power grid data set.

[0021] Optionally, the generation of the initial entity embedding, initial relation embedding, and continuous-time control path input includes:

[0022] Based on a multi-source heterogeneous power grid data set, a regional power grid knowledge graph data was constructed. This involved defining entity and relation types, aligning entities and relations, unifying unique identifier codes and timestamps, and mapping node attributes and edge attributes to generate regional power grid knowledge graph data.

[0023] The regional power grid knowledge graph data is divided into multiple client subgraphs according to the publicly available regional boundaries, forming a client set, and a set of client heterogeneous graph patterns is extracted within each client subgraph;

[0024] Initialize the joint framework of federated heterogeneous graph neural network and graph neural controlled differential equation, load the client set and client heterogeneous graph pattern set, and set the random seed and parameter range;

[0025] On the regional power grid knowledge graph data, under the constraints of random seeds and parameter range, the zero-mean unit variance random normal distribution initialization method is adopted to generate the initial entity embedding and the initial relation embedding.

[0026] Based on operation monitoring data and scheduling plan data, a continuous time control path input with a fixed time step is constructed under a unified timestamp sequence and aligned with the client subgraph.

[0027] Optionally, the generation of the local entity embedding, local relation embedding, pattern weight decoupling parameters, and local encoding parameters includes:

[0028] Within the joint framework, based on the client set, regional power grid knowledge graph data and client heterogeneous graph pattern set are loaded, and local graph structure encoding and local relation structure encoding are performed on the initial entity embedding and initial relation embedding.

[0029] In the process of encoding the map structure, the initial entity embedding is updated and the local entity embedding is generated through adjacency aggregation and multi-layer convolutional propagation mechanism.

[0030] During the local relation structure encoding process, the initial relation embedding is updated based on the weight calculation results of different types of relations in the client heterogeneous graph pattern set, and the local relation embedding is generated.

[0031] During the joint update of local entity embedding and local relation embedding, the pattern weight decoupling parameter is calculated to separate the contribution of different relation types in the client heterogeneous graph pattern set to the local relation embedding.

[0032] After the update is completed, local encoding parameters are generated based on local entity embedding, local relation embedding, and pattern weight decoupling parameters.

[0033] Optionally, the generation of the continuous-time state trajectory and dynamic parameter set includes:

[0034] Within the joint framework, local entity embeddings, local relation embeddings, and continuous time control path inputs are loaded, and time indices are aligned according to a unified timestamp sequence to form a graph neural controlled differential equation modeling input sequence.

[0035] Based on the graph neural controlled differential equation modeling input sequence, a fixed time step is configured, and a unified timestamp sequence is used as the iterative input to prepare the running conditions for the graph neural controlled differential equation modeling process;

[0036] The state is updated according to a fixed time step. The local entity embedding, local relation embedding and the continuous time control path input at the current moment are used as inputs to the graph neural controlled differential equation modeling process, and the continuous time state trajectory is generated iteratively step by step.

[0037] Based on the input sequence modeled by the continuous-time state trajectory and the graph neural controlled differential equation, the set of dynamic parameters is calculated, and the consistency with the continuous-time state trajectory is verified.

[0038] Optionally, the generation of the global model parameters includes:

[0039] Within the joint framework, the local encoding parameters, mode weight decoupling parameters, continuous time state trajectories and dynamic parameter sets of the client set are loaded, and time alignment is completed according to a unified timestamp sequence;

[0040] Federated alignment and aggregation are performed on local encoding parameters and pattern weight decoupling parameters to generate global relation parameters and global entity transformation parameters.

[0041] Federated alignment and aggregation are performed on the continuous-time state trajectory and dynamic parameter set to generate global dynamic parameters and complete the consistency verification with the continuous-time state trajectory.

[0042] Global relational parameters, global entity transformation parameters, and global dynamic parameters are combined to generate global model parameters.

[0043] Optionally, the generation of the refined local model parameters includes:

[0044] Within the joint framework, global model parameters are received and then distributed to the client set.

[0045] Load global model parameters into the client collection, configure global relation parameters and global entity transformation parameters into the federated heterogeneous graph neural network encoding process, and configure global dynamic parameters into the graph neural controlled differential equation modeling process;

[0046] In the client set, based on local entity embedding, local relation embedding and continuous-time control path input, the parameters of the federated heterogeneous graph neural network encoding process are updated to generate updated local encoding parameters;

[0047] In the client set, based on the continuous-time control path input, the parameter update of the graph neural controlled differential equation modeling process is performed to generate the updated set of dynamic parameters;

[0048] The updated set of local encoded parameters and the updated set of dynamic parameters are combined to generate a refined set of local model parameters.

[0049] Optionally, the generation of the regional power grid coordination result set includes:

[0050] Within the joint framework, regional power grid knowledge graph data, continuous-time control path inputs, and refined local model parameters are loaded to construct cross-regional collaborative reasoning inputs and complete the vectorization encoding of the inputs;

[0051] Based on the cross-regional collaborative reasoning input, cross-regional parameter alignment calculation is performed to align the local encoded parameters in the client set with the updated dynamic parameter set in a unified space, generating a cross-regional unified representation;

[0052] Based on the unified cross-regional representation, constraint calculation is performed, and global relational parameters and global entity transformation parameters are introduced into the inference process. At the same time, global dynamic parameters are introduced into the inference process to obtain cross-regional inference output.

[0053] From the cross-regional reasoning output, a set of regional power grid collaborative results is generated based on regional power grid knowledge graph data and continuous time state trajectories.

[0054] The beneficial effects of this invention are:

[0055] First, this invention enables efficient fusion of multi-source heterogeneous data during regional power grid collaboration. By unifying the processing of operation monitoring, scheduling plans, equipment ledgers, metering curves, and external environment data, it effectively solves the problems of data silos and information inconsistencies in existing technologies, thereby improving the completeness and accuracy of power grid collaborative modeling.

[0056] Secondly, this invention introduces a combination mechanism of federated heterogeneous graph neural networks and graph neural controlled differential equations within a joint framework. This not only ensures the privacy and security of cross-regional data, but also fully characterizes the dynamic evolution of power grid operation. It makes up for the shortcomings of traditional centralized modeling in handling time continuity and dynamic constraints, and significantly enhances the reliability and real-time performance of cross-regional calculation results.

[0057] Furthermore, this invention generates a set of regional power grid collaborative results, including power flow distribution prediction, voltage stability index, power exchange plan and safety margin assessment, through cross-regional parameter alignment and constraint calculation. This enables accurate analysis and optimization support for cross-regional operation status, improves the overall dispatch efficiency and safety protection capability of the power grid, and provides an innovative solution for the collaborative operation of new power systems. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is an overall flowchart of a knowledge graph-driven regional power grid coordination method proposed in this invention;

[0060] Figure 2 This is a schematic diagram of the encoding process and mode weight decoupling parameter calculation structure of the federated heterogeneous graph neural network proposed in this invention;

[0061] Figure 3 This is a schematic diagram of the graph neural controlled differential equation modeling process proposed in this invention, used to illustrate the process of generating a continuous-time state trajectory and dynamic parameter set based on local embedding and continuous-time control path input. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1-3 A knowledge graph-driven regional power grid coordination method includes the following steps:

[0064] Collect regional power grid data, including operation monitoring data, dispatch plan data, equipment ledger data, metering curve data and external environment data, perform data preprocessing operations, including cleaning, standardization and pattern mapping, and generate a multi-source heterogeneous power grid data set;

[0065] Based on a multi-source heterogeneous power grid data set, a regional power grid knowledge graph data is constructed, a client set and a client heterogeneous graph pattern set are established, a joint framework of federated heterogeneous graph neural network and graph neural controlled differential equation is initialized, and initial entity embedding, initial relation embedding and continuous time control path input are generated.

[0066] Within the joint framework, based on the client set, the federated heterogeneous graph neural network encoding process is executed. Based on the regional power grid knowledge graph data and the client heterogeneous graph pattern set, the initial entity embedding and initial relation embedding are locally updated to generate local entity embedding, local relation embedding, pattern weight decoupling parameters and local encoding parameters. The output is used for graph neural controlled differential equation modeling and federated aggregation.

[0067] Within the joint framework, based on local entity embedding, local relation embedding and continuous-time control path input, the graph neural controlled differential equation modeling process is executed to generate a continuous-time state trajectory and dynamic parameter set, and output the continuous-time state trajectory and dynamic parameter set, together with the local encoding parameters and pattern weight decoupling parameters, for use in federated aggregation.

[0068] Within the joint framework, federated alignment and aggregation are performed on the local encoding parameters, pattern weight decoupling parameters, continuous-time state trajectories and dynamic parameter sets of the client set to generate global model parameters, including global relation parameters, global entity transformation parameters and global dynamic parameters.

[0069] Within the joint framework, global model parameters are distributed to the client set to update the parameters in the federated heterogeneous graph neural network encoding process and the graph neural controlled differential equation modeling process, generating refined local model parameters;

[0070] Within the joint framework, based on regional power grid knowledge graph data, continuous-time control path inputs, and refined local model parameters, cross-regional collaborative reasoning inputs are constructed, cross-regional parameter alignment and constraint calculations are performed, a set of regional power grid collaborative results is generated, and the results are output and archived.

[0071] In this embodiment, the generation of the multi-source heterogeneous power grid data set includes:

[0072] Collect operational monitoring data, scheduling plan data, equipment ledger data, metering curve data, and external environment data, and establish a unified timestamp sequence and unit standard to form an initial data collection set;

[0073] Data cleaning operations are performed on the initial collected data set, including outlier removal, missing value imputation, and duplicate sample removal, to generate a cleaned data set. Outlier removal targets transient abrupt changes and measurement points that do not conform to the physical constraints of the power grid in the operation monitoring data, and invalid data is identified by comparison with the power system balance equation. Missing value imputation is achieved by using time series interpolation within the same client set and correlation modeling with neighboring devices to ensure that continuous time control path inputs can be accurately constructed. Duplicate sample removal is achieved through hash verification and timestamp consistency detection to prevent redundant entries generated by multi-source acquisition devices during parallel recording from entering the multi-source heterogeneous power grid data set.

[0074] Standardization operations are performed on the cleaned dataset, including unit unification and numerical range normalization, to generate a standardized dataset. Unit unification uses a unified International System of Units (SI) to convert voltage, power, load, and meteorological data from different data sources, so as to facilitate feature alignment in the federated heterogeneous graph neural network. Numerical range normalization uses a linear scaling method to compress the input vector to the [0,1] interval, which helps the graph neural network controlled differential equation maintain numerical stability during continuous-time state trajectory modeling, while avoiding gradient distortion in federated aggregation.

[0075] A pattern mapping operation is performed on a standardized dataset to map fields to node and edge attributes of the regional power grid knowledge graph data, forming a mapping structure consistent with the regional power grid topology. The pattern mapping defines the ontology structure of the power grid knowledge graph data, mapping the collected equipment ledger fields to node attributes, such as the voltage level of bus nodes and the rated capacity of line nodes, and mapping the operation monitoring data and scheduling plan data to edge attributes, such as the power flow and scheduling relationship of lines. This mapping enables multi-source heterogeneous power grid datasets to be directly embedded into the regional power grid knowledge graph data, providing structured input for the federated heterogeneous graph neural network to perform local encoding and pattern weight decoupling, and also providing topological semantic support based on physical constraints for the graph neural network's controlled differential equations.

[0076] The pattern mapping results are aggregated into a multi-source heterogeneous power grid data set, which is then output for subsequent use in constructing regional power grid knowledge graph data.

[0077] In this embodiment, the generation of the initial entity embedding, initial relation embedding, and continuous-time control path input includes:

[0078] Based on a multi-source heterogeneous power grid dataset, a regional power grid knowledge graph was constructed. This involved defining entity and relation types, aligning entities and relations, unifying unique identifiers and timestamps, and mapping node and edge attributes. Specifically, the construction of the regional power grid knowledge graph ensured the integrity of the power grid topology by defining entities such as buses, lines, transformers, switches, and loads, and relation types such as connections, power supply, control, and adjacency. Entity alignment used unique identifiers to bind the same device from different data sources, and relation alignment ensured edge accuracy through cross-validation of scheduling plan data and operation monitoring data. Unified timestamps aligned all data with the same sampling interval within the same time series. Node attributes were generated from equipment ledgers and operation monitoring data, while edge attributes were generated from power flow, status, and scheduling relationships, thus ensuring that the regional power grid knowledge graph data possessed both topological constraints and physical meaning.

[0079] The regional power grid knowledge graph data is divided into multiple client subgraphs according to the publicly available regional boundaries, forming a client set. Within each client subgraph, a client heterogeneous graph pattern set is extracted. The extraction of the client heterogeneous graph pattern set is generated by identifying different entity types and their relationship patterns in the client subgraphs. For example, if the same client subgraph contains both electrical equipment entities and meteorological entities, it forms a heterogeneous pattern. This set is used to guide the federated heterogeneous graph neural network to distinguish different types of nodes and edges during local modeling, ensuring that pattern weight decoupling and cross-client pattern alignment can be performed correctly.

[0080] The joint framework of federated heterogeneous graph neural network and graph neural controlled differential equation is initialized, the client set and client heterogeneous graph pattern set are loaded, and a random seed and parameter range are set. The initialization of the joint framework loads the client set and client heterogeneous graph pattern set globally to ensure that the data structure of different clients is consistent within the model. The random seed is used to control the reproducibility of the initial parameter generation process and avoid excessive differences in model training results under different experimental conditions. The parameter range is used to limit the value boundaries of the initial parameters to prevent gradient explosion caused by excessively large embedding values ​​or gradient vanishing caused by excessively small embedding values, thereby ensuring the numerical stability of subsequent embedding generation and continuous-time modeling.

[0081] For the unified identification of regional power grid knowledge graph data, under the constraints of random seeds and parameter ranges, a zero-mean unit variance random normal distribution initialization method is adopted to generate initial entity embeddings and initial relation embeddings. A one-to-one correspondence is established between the numbers of the initial entity embeddings and initial relation embeddings and the nodes and edges of the client subgraph. The zero-mean unit variance random normal distribution ensures that the initial entity embeddings and initial relation embeddings have a balanced numerical distribution, making the parameter updates of the subsequent federated heterogeneous graph neural network more stable. The one-to-one correspondence between the numbers and the nodes and edges ensures that each entity and relation has a unique vector representation and that there are no conflicts between client sets, which provides a foundation for subsequent federated alignment and aggregation.

[0082] The zero-mean, unit-variance random normal distribution initialization method refers to independently sampling each dimension of the initial embedding vector according to a normal distribution with a mean of zero and a variance of one under the constraint of a random seed. This ensures that the numerical distribution center of the overall embedding vector is close to zero and the numerical fluctuation range is controlled. By setting the parameter range, a truncation operation is performed on the sampling results to restrict the values ​​to a fixed boundary interval, avoiding excessively large or small initial values ​​that could lead to instability in subsequent training. The initial entity embeddings and initial relation embeddings generated by this method are numerically balanced and reproducible, and can ensure consistency and alignment of the training process between different client subgraphs in the federated environment.

[0083] Based on the operation monitoring data and scheduling plan data under a unified timestamp sequence, a continuous time control path input with a fixed time step is constructed and aligned with the client subgraph. The continuous time control path input with a fixed time step is formed by slicing the operation monitoring data and scheduling plan data according to the same timestamp to form a progressively evolving control signal sequence. This input serves as the driving term of the graph neural controlled differential equation, ensuring that the continuous time state trajectory can truly reflect the power grid operation process. At the same time, alignment with the client subgraph ensures that the input data of different regional power grids are consistent in the federated environment.

[0084] Output the initial entity embedding, initial relation embedding, and continuous time control path input for use in subsequent steps.

[0085] In this embodiment, the generation of local entity embedding, local relation embedding, pattern weight decoupling parameters, and local encoding parameters includes:

[0086] Within the joint framework, based on the client set, regional power grid knowledge graph data and client heterogeneous graph pattern set are loaded, and local graph structure encoding and local relation structure encoding are performed on the initial entity embedding and initial relation embedding.

[0087] In the process of encoding the map structure, the initial entity embedding is updated and local entity embedding is generated through adjacency aggregation and multi-layer convolutional propagation mechanism. Specifically, the local graph structure encoding process uses adjacency aggregation to perform a weighted average of the feature vectors of adjacent nodes in the client subgraph, and then uses a multi-layer convolutional propagation mechanism to gradually diffuse the information of neighbors at different levels, layer by layer to form a higher-order representation of the node. This process can capture the complex electrical correlation characteristics between nodes while maintaining the topological structure constraints of the regional power grid knowledge graph data, thereby generating local entity embeddings that reflect the local structure of the regional power grid.

[0088] During the local relation structure encoding process, the initial relation embedding is updated based on the weight calculation results of different types of relations in the client heterogeneous graph pattern set, and local relation embeddings are generated. In the local relation structure encoding process, for different relation types (such as power supply relations, control relations, and adjacent relations) in the client heterogeneous graph pattern set, their weights in the connection mode are calculated respectively, and the weights are used to correct the initial relation embedding, so that the influence of each type of relation in the vector space is modeled separately. In this way, the updated local relation embedding can distinguish the differentiated contributions of different relations to the power grid operation and improve the accuracy of subsequent federated alignment.

[0089] In the joint update process of local entity embeddings and local relation embeddings, a pattern weight decoupling parameter is calculated to separate the contributions of different relation types in the client's heterogeneous graph pattern set to the local relation embedding. The joint update process refers to simultaneously adjusting the local entity embedding and local relation embedding within the same training iteration to maintain a consistent optimization direction. The pattern weight decoupling parameter is used to identify and separate the independent contributions of different relation types in the client's heterogeneous graph pattern set to the local relation embedding, thereby preventing any type of relation from being overemphasized or weakened in the embedding. Through the pattern weight decoupling parameter, the model can perform fine-grained control over heterogeneous patterns, achieving effective alignment of cross-client relation patterns.

[0090] After the update is completed, local encoding parameters are generated based on local entity embedding, local relation embedding, and pattern weight decoupling parameters. These parameters are then used as inputs to graph neural controlled differential equation modeling and federated aggregation. The generation of local encoding parameters involves using local entity embedding and local relation embedding as feature inputs, while simultaneously incorporating pattern weight decoupling parameters for constraint, resulting in a comprehensive set of encoding parameters that includes node representation, relation representation, and pattern weight information. These local encoding parameters are subsequently input into the graph neural controlled differential equation modeling module and the federated aggregation process, enabling both dynamic modeling and global parameter aggregation to utilize the multi-dimensional features learned locally, thereby improving the effectiveness and robustness of cross-regional power grid collaboration.

[0091] In this embodiment, the generation of the continuous-time state trajectory and dynamic parameter set includes:

[0092] Within the joint framework, local entity embeddings, local relation embeddings, and continuous time control path inputs are loaded, and time indices are aligned according to a unified timestamp sequence to form the input sequence for graph neural controlled differential equation modeling. The unified timestamp sequence alignment with the time index means that all timestamps in the operation monitoring data and scheduling plan data are projected into a unified sampling sequence, so that different data sources are synchronized on the same time axis, avoiding time misalignment when cross-client sets, and ensuring that the input sequence for graph neural controlled differential equation modeling can maintain continuity and comparability.

[0093] Based on the graph neural network controlled differential equation modeling input sequence, a fixed time step is configured, and a unified timestamp sequence is used as the iterative input to prepare the running conditions for the graph neural network controlled differential equation modeling process. Configuring a fixed time step refers to setting a stable evolution interval in continuous-time modeling to ensure the stability of the numerical solution. Preparing the running conditions for the graph neural network controlled differential equation modeling process means loading a fixed time step and a unified timestamp sequence within the joint framework, so that subsequent state updates can be executed under strict time scheduling, ensuring that the evolutionary pattern of the regional power grid knowledge graph data in the time dimension can be completely modeled.

[0094] State updates are performed at fixed time steps, using local entity embeddings, local relation embeddings, and the current continuous-time control path input as inputs to the graph neural controlled differential equation modeling process. This iteratively generates a continuous-time state trajectory. Performing state updates at fixed time steps means that at each preset time interval, the state from the previous moment and the current continuous-time control path input are input into the model, allowing the state variables to evolve continuously over time. Iteratively generating the continuous-time state trajectory means continuously superimposing the state update results from each time step to form a continuous-time state evolution curve covering the entire time range, thereby accurately depicting the dynamic operation of the regional power grid.

[0095] Based on the input sequence of the continuous-time state trajectory and the controlled differential equation modeling of the graph neural network, the set of dynamic parameters is calculated and the consistency with the continuous-time state trajectory is verified.

[0096] The calculation of the dynamic parameter set includes three parts:

[0097] First, the transmission rate is calculated by comparing the changes in the states of the nodes at both ends of the edge in adjacent time slices and normalizing them by a fixed time step to obtain the average transmission rate of the edge in the time dimension.

[0098] Secondly, the calculation of the dynamic coupling coefficient takes the previous state of each node and the previous state of its neighboring nodes as inputs, predicts the next state of the node, and thus obtains the influence strength of different neighbors on the state change of the node. The coupling relationship of all nodes is combined to form a dynamic coupling coefficient matrix.

[0099] Finally, the stability index is calculated based on the overall effect of the dynamic coupling coefficient matrix to evaluate the convergence of the system under continuous iteration, and obtain a stability index that can characterize the long-term stable operation capability of the system.

[0100] The set of dynamic parameters consists of a transmission rate vector, a dynamic coupling coefficient matrix, and a stability index, and is used to describe the dynamic characteristics and stability of the regional power grid in the time dimension.

[0101] The consistency verification with the continuous-time state trajectory refers to comparing the values ​​of the dynamic parameter set with the evolution law of the state trajectory to ensure that the dynamic parameter set can accurately reflect the changing trend of the state trajectory, thereby ensuring the physical rationality and robustness of the model in the regional power grid collaborative task.

[0102] Output the continuous-time state trajectory and set of dynamic parameters.

[0103] In this embodiment, the generation of global model parameters includes:

[0104] Within the joint framework, the local coding parameters, mode weight decoupling parameters, continuous time state trajectories, and dynamic parameter sets of the client sets are loaded and time-aligned according to a unified timestamp sequence. This means mapping all the local coding parameters, mode weight decoupling parameters, continuous time state trajectories, and dynamic parameter sets generated by different client sets during independent operation to the same time baseline. This ensures that the data of each client remains consistent in the time dimension and avoids data asynchrony during cross-regional power grid collaboration.

[0105] Federated alignment and aggregation are performed on local encoding parameters and pattern weight decoupling parameters to generate global relation parameters and global entity transformation parameters. Specifically, the federated alignment and aggregation of local encoding parameters and pattern weight decoupling parameters refers to normalizing and mapping the feature parameters and pattern weight parameters obtained by each client in local encoding through a federated learning mechanism without exchanging the original data. Then, the parameters are aggregated using weighted averaging or optimization to obtain unified global parameters. The generated global relation parameters are used to uniformly characterize the connection strength of relation edges in the cross-regional power grid knowledge graph, and the global entity transformation parameters are used to uniformly adjust the entity embedding representation of different clients, thereby ensuring the consistency of the global model representation across the cross-regional scope.

[0106] Federated alignment and aggregation are performed on the continuous-time state trajectory and dynamic parameter set to generate global dynamic parameters, and consistency verification with the continuous-time state trajectory is completed. Federated alignment and aggregation of the continuous-time state trajectory and dynamic parameter set refers to aligning and matching parameters and distributions among the dynamic modeling results of each client, correcting differences in the temporal evolution patterns of different clients before global aggregation. The generated global dynamic parameters can reflect the overall dynamic evolution of the regional power grid in the time dimension. Consistency verification with the continuous-time state trajectory involves comparing the evolution trends of the continuous-time state trajectories of each client using the aggregated global dynamic parameters to ensure that the parameter aggregation results are consistent with the actual trajectories, thereby avoiding deviations when the global model is applied across regions.

[0107] Global relational parameters, global entity transformation parameters, and global dynamic parameters are combined to generate global model parameters. The combination to generate global model parameters refers to integrating the global relational parameters, global entity transformation parameters, and global dynamic parameters learned across clients according to a predetermined structure to form a complete parameter set. These global model parameters contain both structured information from the cross-regional power grid knowledge graph and dynamic features from continuous-time modeling, providing unified input conditions for subsequent updates and cross-regional collaborative reasoning.

[0108] Output global model parameters for use in subsequent steps.

[0109] In this embodiment, the generation of the refined local model parameters includes:

[0110] Within the joint framework, global model parameters are received and then distributed to the client set.

[0111] The global model parameters are loaded into the client set, and the global relation parameters and global entity transformation parameters are configured into the federated heterogeneous graph neural network encoding process. The global dynamic parameters are configured into the graph neural controlled differential equation modeling process. Specifically, configuring the global relation parameters and global entity transformation parameters into the federated heterogeneous graph neural network encoding process means embedding the aggregated global unified relation representation and entity mapping rules into the client's local encoding network, so that it can maintain the consistency of cross-regional power grid knowledge graph relations and entity representations in subsequent training and inference. Configuring the global dynamic parameters into the graph neural controlled differential equation modeling process means injecting the aggregated global dynamic evolution law as a priori conditions into the local differential equation modeling module to constrain the local continuous-time state update process, thereby improving the physical consistency of dynamic modeling.

[0112] Within the client set, based on local entity embeddings, local relation embeddings, and continuous-time control path inputs, parameter updates are performed in the federated heterogeneous graph neural network encoding process to generate updated local encoding parameters. Parameter updates in the federated heterogeneous graph neural network encoding process refer to modifying and retraining existing local embeddings on the client side using global relation parameters and global entity transformation parameters, ensuring that the local graph structure representation remains aligned with global knowledge. Generating updated local encoding parameters means obtaining new local network weights and feature mapping coefficients through this retraining process, providing a consistent feature foundation for subsequent cross-regional aggregation.

[0113] Within the client set, based on the continuous-time control path input, parameter updates are performed during the graph neural controlled differential equation modeling process, generating an updated set of dynamic parameters. Parameter updates during the graph neural controlled differential equation modeling process refer to retraining or calibrating the differential equation module using global dynamic parameters as initialization conditions, driven by the local continuous-time control path input, so that the model can better fit the dynamic characteristics of the local power grid. The generated updated set of dynamic parameters refers to the new set of dynamic coefficients obtained locally after correction, which combines global dynamic laws with local differences.

[0114] The updated local coding parameters and the updated dynamic parameters are combined to generate refined local model parameters. The combination of updated local coding parameters and updated dynamic parameters means integrating the two updated results within a unified client to form a parameter set that includes both structured features and dynamic evolution features. The generation of refined local model parameters means that the integrated result can be used in subsequent cross-regional collaborative reasoning to highlight local details while maintaining global consistency, thereby improving the accuracy and adaptability of cross-regional power grid collaborative reasoning.

[0115] In this embodiment, the generation of the regional power grid coordination result set includes:

[0116] Within the joint framework, regional power grid knowledge graph data, continuous-time control path inputs, and refined local model parameters are loaded to construct cross-regional collaborative reasoning inputs and complete input vectorization encoding. Constructing cross-regional collaborative reasoning inputs means combining entity, relationship, and topology information from the regional power grid knowledge graph data with continuous-time control path inputs and refined local model parameters to form a unified input structure. Completing input vectorization encoding means transforming this input structure into a high-dimensional vector representation that can be processed by the joint framework, thereby ensuring that cross-regional data can be used for unified reasoning within the same feature space.

[0117] Based on the input of cross-regional collaborative reasoning, cross-regional parameter alignment calculation is performed to align the local encoded parameters in the client set with the updated set of dynamic parameters in a unified space, generating a unified cross-regional representation. Cross-regional parameter alignment calculation refers to comparing and correcting the differences between the local encoded parameters of different clients and the global dynamic parameters while ensuring the privacy of each client is not compromised. Alignment in a unified space means projecting data from different sources onto the same representation space through parameter mapping and normalization mechanisms. Generating a unified cross-regional representation means that the aligned result reflects the uniformity of all clients in terms of structure and dynamic features, thus providing a consistent input basis for cross-regional collaborative reasoning.

[0118] Based on a unified cross-regional representation, constraint computation is performed, introducing global relation parameters and global entity transformation parameters into the inference process to ensure consistency between the relation structure and entity representation. Simultaneously, global dynamic parameters are introduced into the inference process to ensure the matching of temporal evolution patterns with continuous-time state trajectories, resulting in cross-regional inference output. Constraint computation refers to simultaneously introducing global relation parameters, global entity transformation parameters, and global dynamic parameters into the inference process of the unified cross-regional representation to impose structural and dynamic constraints on the unified representation. This ensures that the inference process satisfies both the relational consistency of the power grid knowledge graph and the dynamic evolution patterns of continuous-time state trajectories. The resulting cross-regional inference output means that the output generated under these constraints accurately reflects the operational characteristics of the cross-regional power grid under unified conditions.

[0119] From the cross-regional inference output, based on regional power grid knowledge graph data and continuous-time state trajectories, a set of regional power grid collaborative results is generated. This set includes cross-regional power flow distribution prediction, cross-regional voltage stability indicators, cross-regional power exchange plans, and cross-regional safety margin assessments. Among these:

[0120] Cross-regional power flow distribution prediction: Based on the node status information in the cross-regional unified representation, power flow is calculated using the power transmission relationship between nodes to obtain the distribution of active and reactive power on different nodes and lines within the cross-regional range. This is used to predict the power transmission mode of the regional power grid under given operating conditions, providing data support for cross-regional power dispatch and load balancing.

[0121] Cross-regional voltage stability index: Under dynamic constraints, the voltage fluctuation of each node in the continuous time state trajectory is analyzed over time. By statistically analyzing the voltage deviation amplitude and convergence rate, a set of indexes reflecting the voltage stability level of the system is generated. This is used to evaluate the ability of cross-regional power grids to maintain voltage stability under disturbances or load changes, and to provide a reference for voltage control strategies and stability maintenance.

[0122] Cross-regional power exchange plan: Based on the cross-client relationship parameter mapping results, the power flow direction and exchange quota between regions are determined. The power exchange matrix between each region is calculated through energy flow allocation to form a cross-regional power exchange plan, which is used to guide the power allocation and trading between regions, ensure the rational distribution and transmission of electricity between different regions, and improve the overall operating efficiency.

[0123] Cross-regional safety margin assessment: The continuous-time state trajectory and dynamic parameter set are jointly input into the reasoning process, and the operating state is compared with the critical stability condition to calculate the safety margin value. This is used to quantify the risk space of the system under extreme operating conditions, to judge the safe operation level of the regional power grid under peak load or sudden disturbance, and to provide a basis for cross-regional emergency dispatch and risk warning.

[0124] The results of regional power grid coordination will be collected, output, and archived for subsequent cross-regional scheduling and optimization.

[0125] Example 1:

[0126] To verify the feasibility of this invention in practice, it was applied to the scenario of cross-regional collaborative operation of a regional power grid. A regional power grid contains operational monitoring data, dispatching plan data, equipment ledger data, metering curve data, and external environmental data. These data come from diverse sources and have inconsistent formats, exhibiting multi-source heterogeneity and complexity. In existing centralized modeling methods, data from different regions is difficult to process uniformly, and parameters across regions are difficult to align, resulting in significant deviations in power exchange calculations and inaccurate voltage stability assessments, failing to meet the needs of large-scale power grid collaborative operation. This invention, by constructing a regional power grid knowledge graph and combining a federated heterogeneous graph neural network with graph neural controlled differential equations, achieves unified modeling and cross-regional reasoning of multi-source heterogeneous data while ensuring data privacy. This results in significant improvements in tasks such as power flow prediction, power exchange planning, and safety margin assessment.

[0127] In practical applications, the operation monitoring data and dispatch plan data of the regional power grid are first cleaned and standardized to ensure consistency in timestamp sequences and numerical ranges. Then, a regional power grid knowledge graph is constructed, mapping equipment attributes from the equipment ledger to node features, and embedding metering curves and external environmental data as edge attributes into the graph structure, forming a knowledge representation consistent with the regional topology. Based on this, a federated heterogeneous graph neural network is used to locally encode each regional subgraph, generating local entity embeddings and relation embeddings. Pattern weight decoupling parameters are used to distinguish the contribution of different relations to the results. Finally, combined with a graph neural controlled differential equation model, a continuous-time state trajectory and dynamic parameter set are generated based on continuous-time control path inputs to characterize the dynamic evolution characteristics of power grid operation.

[0128] In the experimental data, three regional power grids were selected as test objects, each containing 1200 to 1800 nodes, totaling approximately 4800 nodes and 6200 edges, with a total data volume of approximately 5TB. Using the method of this invention, the performance differences between the experiment and traditional centralized power flow calculation methods and ordinary graph neural network methods were compared, focusing on the accuracy of power flow distribution prediction, voltage stability assessment error, power exchange plan deviation, safety margin calculation deviation, and overall computational efficiency.

[0129] The table below shows the comparison results of the experimental data:

[0130] Table 1 Performance Comparison Results of Regional Power Grid Coordination Methods

[0131]

[0132] The comparative results show that the present invention significantly outperforms traditional methods in all indicators. The power flow distribution prediction accuracy reaches 91.1%, which is about 13.8 percentage points higher than the traditional centralized method and about 6.6 percentage points higher than the ordinary graph neural network method. This improvement is mainly due to the role of knowledge graphs in aligning multi-source heterogeneous data and the advantages of cross-regional information sharing under the federated learning mechanism. The average error of voltage stability assessment was reduced from 10.9% in the traditional method to 5.4%, indicating that under dynamic constraints, controlled differential equations can better characterize the evolution of node voltage over time. In the power exchange plan results, this invention controls the cross-regional power deviation to 25.5MW, which is nearly 51% lower than the traditional centralized method, reflecting the role of mode weight decoupling parameters in relation mapping and making the energy allocation in different regions more reasonable. In terms of safety margin calculation, the deviation of this invention is 5.1%, which is close to half of the traditional method, indicating that cross-regional parameter alignment and constraint calculation play a key role in risk assessment. In terms of computational efficiency, this invention completes full-network inference in only 13.9 minutes, while the traditional method requires 28.4 minutes and the ordinary graph neural network method requires 21.6 minutes. The efficiency improvement is due to the fact that this invention achieves parallel inference within a federated framework and reduces redundant computation through time series modeling. In addition, in terms of data privacy protection, this invention adopts a federated learning mechanism, in which each client only uploads parameters rather than raw data, achieving comprehensive protection, while traditional centralized methods have no privacy protection mechanism at all.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A regional power grid coordination method based on a knowledge graph, characterized in that, Includes the following steps: Collect regional power grid data, perform data preprocessing operations, and generate a multi-source heterogeneous power grid data set; Based on a multi-source heterogeneous power grid data set, a regional power grid knowledge graph data is constructed, a client set and a client heterogeneous graph pattern set are established, a joint framework of federated heterogeneous graph neural network and graph neural controlled differential equation is initialized, and initial entity embedding, initial relation embedding and continuous time control path input are generated. Within the joint framework, based on the client set, the federated heterogeneous graph neural network encoding process is executed. Based on the regional power grid knowledge graph data and the client heterogeneous graph pattern set, the initial entity embedding and initial relation embedding are locally updated to generate local entity embedding, local relation embedding, pattern weight decoupling parameters and local encoding parameters. Within the joint framework, based on local entity embedding, local relation embedding, and continuous-time control path input, the graph neural controlled differential equation modeling process is executed to generate a continuous-time state trajectory and dynamic parameter set. Federated alignment and aggregation are performed on the local encoding parameters, mode weight decoupling parameters, continuous-time state trajectories and dynamic parameter sets of the client set to generate global model parameters; The global model parameters are distributed to the client set to update the parameters in the federated heterogeneous graph neural network encoding process and the graph neural controlled differential equation modeling process, and to generate refined local model parameters. Based on regional power grid knowledge graph data, continuous-time control path inputs, and refined local model parameters, a cross-regional collaborative reasoning input is constructed, cross-regional parameter alignment and constraint calculations are performed, and a set of regional power grid collaborative results is generated. 2.The regional power grid coordination method based on knowledge graph driving according to claim 1, characterized in that, The generation of the multi-source heterogeneous power grid data set includes: Collect operational monitoring data, scheduling plan data, equipment ledger data, metering curve data, and external environment data, and establish a unified timestamp sequence and unit standard to form an initial data collection set; Perform data cleaning operations on the initial collected dataset, including outlier removal, missing value imputation, and duplicate sample removal, to generate a cleaned dataset; Standardization processing operations are performed on the cleaned dataset, including unification of units and normalization of numerical ranges, to generate a standardized dataset; Perform pattern mapping operations on a standardized dataset to map fields to node and edge attributes of regional power grid knowledge graph data, forming a mapping structure consistent with the topology of the regional power grid. The pattern mapping results are aggregated into a multi-source heterogeneous power grid data set.

3. The knowledge graph-driven regional power grid coordination method according to claim 1, characterized in that, The generation of the initial entity embedding, initial relation embedding, and continuous-time control path input includes: Based on a multi-source heterogeneous power grid data set, a regional power grid knowledge graph data was constructed. This involved defining entity and relation types, aligning entities and relations, unifying unique identifier codes and timestamps, and mapping node attributes and edge attributes to generate regional power grid knowledge graph data. The regional power grid knowledge graph data is divided into multiple client subgraphs according to the publicly available regional boundaries, forming a client set, and a set of client heterogeneous graph patterns is extracted within each client subgraph; Initialize the joint framework of federated heterogeneous graph neural network and graph neural controlled differential equation, load the client set and client heterogeneous graph pattern set, and set the random seed and parameter range; On the regional power grid knowledge graph data, under the constraints of random seeds and parameter range, the zero-mean unit variance random normal distribution initialization method is adopted to generate the initial entity embedding and the initial relation embedding. Based on operation monitoring data and scheduling plan data, a continuous time control path input with a fixed time step is constructed under a unified timestamp sequence and aligned with the client subgraph.

4. The knowledge graph-driven regional power grid coordination method according to claim 1, characterized in that, The generation of local entity embedding, local relation embedding, pattern weight decoupling parameters, and local encoding parameters includes: Within the joint framework, based on the client set, regional power grid knowledge graph data and client heterogeneous graph pattern set are loaded, and local graph structure encoding and local relation structure encoding are performed on the initial entity embedding and initial relation embedding. In the process of encoding the map structure, the initial entity embedding is updated and the local entity embedding is generated through adjacency aggregation and multi-layer convolutional propagation mechanism. During the local relation structure encoding process, the initial relation embedding is updated based on the weight calculation results of different types of relations in the client heterogeneous graph pattern set, and the local relation embedding is generated. During the joint update of local entity embedding and local relation embedding, the pattern weight decoupling parameter is calculated to separate the contribution of different relation types in the client heterogeneous graph pattern set to the local relation embedding. After the update is completed, local encoding parameters are generated based on local entity embedding, local relation embedding, and pattern weight decoupling parameters.

5. The knowledge graph-driven regional power grid coordination method according to claim 1, characterized in that, The generation of the continuous-time state trajectory and dynamic parameter set includes: Within the joint framework, local entity embeddings, local relation embeddings, and continuous time control path inputs are loaded, and time indices are aligned according to a unified timestamp sequence to form a graph neural controlled differential equation modeling input sequence. Based on the graph neural controlled differential equation modeling input sequence, a fixed time step is configured, and a unified timestamp sequence is used as the iterative input to prepare the running conditions for the graph neural controlled differential equation modeling process; The state is updated according to a fixed time step. The local entity embedding, local relation embedding and the continuous time control path input at the current moment are used as inputs to the graph neural controlled differential equation modeling process, and the continuous time state trajectory is generated iteratively step by step. Based on the input sequence modeled by the continuous-time state trajectory and the graph neural controlled differential equation, the set of dynamic parameters is calculated, and the consistency with the continuous-time state trajectory is verified.

6. The knowledge graph-driven regional power grid coordination method according to claim 1, characterized in that, The generation of the global model parameters includes: Within the joint framework, the local encoding parameters, mode weight decoupling parameters, continuous time state trajectories and dynamic parameter sets of the client set are loaded, and time alignment is completed according to a unified timestamp sequence; Federated alignment and aggregation are performed on local encoding parameters and pattern weight decoupling parameters to generate global relation parameters and global entity transformation parameters. Federated alignment and aggregation are performed on the continuous-time state trajectory and dynamic parameter set to generate global dynamic parameters and complete the consistency verification with the continuous-time state trajectory. Global relational parameters, global entity transformation parameters, and global dynamic parameters are combined to generate global model parameters.

7. The knowledge graph-driven regional power grid coordination method according to claim 1, characterized in that, The generation of the refined local model parameters includes: Within the joint framework, global model parameters are received and then distributed to the client set. Load global model parameters into the client collection, configure global relation parameters and global entity transformation parameters into the federated heterogeneous graph neural network encoding process, and configure global dynamic parameters into the graph neural controlled differential equation modeling process; In the client set, based on local entity embedding, local relation embedding and continuous-time control path input, the parameters of the federated heterogeneous graph neural network encoding process are updated to generate updated local encoding parameters; In the client set, based on the continuous-time control path input, the parameter update of the graph neural controlled differential equation modeling process is performed to generate the updated set of dynamic parameters; The updated set of local encoded parameters and the updated set of dynamic parameters are combined to generate a refined set of local model parameters.

8. The knowledge graph-driven regional power grid coordination method according to claim 1, characterized in that, The generation of the regional power grid coordination result set includes: Within the joint framework, regional power grid knowledge graph data, continuous-time control path inputs, and refined local model parameters are loaded to construct cross-regional collaborative reasoning inputs and complete the vectorization encoding of the inputs; Based on the cross-regional collaborative reasoning input, cross-regional parameter alignment calculation is performed to align the local encoded parameters in the client set with the updated dynamic parameter set in a unified space, generating a cross-regional unified representation; Based on the unified cross-regional representation, constraint calculation is performed, and global relational parameters and global entity transformation parameters are introduced into the inference process. At the same time, global dynamic parameters are introduced into the inference process to obtain cross-regional inference output. From the cross-regional reasoning output, a set of regional power grid collaborative results is generated based on regional power grid knowledge graph data and continuous time state trajectories.

Images