A power system risk assessment method based on a hierarchical regional graph neural network

Abstract

This invention relates to the field of electronic digital data processing technology and discloses a power system risk assessment method based on a hierarchical regional graph neural network. The method includes: dividing the power system into multiple regions and constructing a local graph structure for each region; establishing connections between regions to form a regional-level graph; acquiring multi-source heterogeneous operational data of the power system as input features for corresponding nodes; using a graph neural network to learn and aggregate the node features in each local graph to obtain a regional feature representation for the corresponding operational region; aggregating the regional feature representations in the inter-regional connection graph based on a graph attention mechanism to fuse risk information; and generating node-level, regional-level, and system-level risk assessment results based on the feature representations of nodes and regions, as well as the risk information. This method solves the problems of effective regional layering, unclear risk propagation mechanisms, and lack of multi-level output, achieving the goals of multi-source data fusion, multi-level assessment, and dynamic adaptation to topology changes.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a power system risk assessment method based on a hierarchical regional graph neural network. Background Technology

[0002] In recent years, global climate change and energy security issues have become increasingly severe, prompting countries worldwide to accelerate energy structure transformation and actively promote green and low-carbon development strategies. Guided by the "dual carbon" goal, renewable energy sources such as wind and solar power, with their clean, renewable, and zero-emission characteristics, have become crucial supports for new power systems. Building a new power system primarily based on new energy sources has become the core path to achieving green and low-carbon development in the power industry. Power systems are subject to the risk of uncontrollable random failures during operation, such as equipment failures, load fluctuations, external damage, and system instability. Therefore, risk assessment is an essential step for power systems. Power system risk assessment is the process of quantitatively evaluating the probability of power system failures and the severity of their consequences. Power system risk assessment ensures the safe and stable operation of the system, reduces the probability of random failures, avoids the risk of localized or large-scale load shedding, thereby stabilizing residential and industrial loads and reducing socio-economic losses. Therefore, risk assessment has become a challenging fundamental task for power systems. Against the backdrop of global energy transition, the rapid growth in the proportion of new energy sources has led to a decline in the peak-shaving capacity of power systems, insufficient reactive power and voltage support, and increased system stability risks. These issues were not considered in the context of traditional risk assessment. Therefore, traditional power system risk assessment theories cannot fully adapt to new power systems, especially given the new characteristics of the introduction of new energy power electronic equipment and the volatility and randomness brought about by a high proportion of new energy integration.

[0003] For example, Chinese patent CN118428716A discloses a novel power system risk assessment method based on multi-head graph attention networks, providing the following technical solution: First, a comprehensive scenario considering equipment failures under the influence of new energy output and weather factors is constructed. Then, a deep learning framework is built, employing a multi-head attention mechanism to improve the accuracy and efficiency of the novel power system risk assessment. In this framework, numerous components of the power system are mapped to the graph topology, and risk points are accurately captured and assessed through iterative training. Finally, this invention performs novel power system risk assessment based on this deep learning framework, providing the power industry with an efficient and accurate risk management and decision-making tool. However, the aforementioned novel power system risk assessment method based on multi-head graph attention networks uses a single-layer graph structure, making it difficult to effectively model the hierarchical relationships between power grid regions and the systemic risk propagation mechanism. Its multi-source data fusion capability is limited, and it lacks multi-level risk assessment output. Furthermore, it fails to fully consider the hierarchical relationships between power grid regions and the systemic risk propagation mechanism, making it difficult to adapt to the needs of scenarios such as multi-regional collaborative operation. Summary of the Invention

[0004] This invention addresses the problems of ineffective regional layering, unclear risk propagation mechanisms, weak multi-source data fusion, and lack of multi-level output in existing technologies. It proposes a power system risk assessment method based on a hierarchical regional graph neural network, achieving the goals of multi-source data fusion, multi-level assessment, and dynamic adaptation to topology changes.

[0005] Furthermore, this invention aims to construct a hierarchical regional graph neural network to achieve risk assessment of multi-level and multi-source data of the power system, and provide comprehensive risk indicators at the node, regional, and system levels to support more accurate power grid risk early warning and dispatch decisions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A power system risk assessment method based on hierarchical regional graph neural networks includes: The power system is divided into multiple regions, and a local graph structure is constructed for each region. Nodes represent power equipment, and edges represent the connections between equipment. The connections between regions are established to form a region-level graph. Multi-source heterogeneous operation data of the power system are obtained as input features for the corresponding nodes; graph neural networks are used to learn and aggregate the node features in each local graph to obtain the regional feature representation of the corresponding operation area. Based on the graph attention mechanism, the feature representations of each region are aggregated in the inter-region connectivity graph to integrate risk information; Based on the characteristic representations of nodes and regions and risk information, risk assessment results are generated at the node level, region level, and system level, respectively.

[0007] By dividing the system into hierarchical regions and constructing local graphs, the topology and inter-regional dependencies of the power system are effectively captured, improving modeling accuracy. The integration of multi-source heterogeneous data enhances the model's adaptability to complex real-world scenarios. Graph neural networks and graph attention mechanisms ensure the efficient dissemination and fusion of risk information. Multi-level risk assessment outputs provide a comprehensive risk perspective, supporting refined monitoring and decision-making of the power grid.

[0008] Preferably, the method of learning and aggregating node features in each local graph using a graph neural network includes: learning node features in the local graph of each region using a graph attention network with residual connections, wherein the residual connections fuse the input and output features of the graph attention layer, and after multi-layer feature propagation and aggregation, the final embedding vector of each node is obtained; performing a region pooling operation on the final embedding vectors of all nodes in the region to obtain a region feature vector representing the overall operational risk status of the region.

[0009] Introducing graph attention networks with residual connections can alleviate the problems of gradient vanishing and feature degradation in deep network training, improve the stability and effectiveness of feature learning, and effectively aggregate node-level information through region pooling operations to generate region-level feature representations, providing a foundation for subsequent cross-regional risk fusion and enhancing the model's hierarchical perception capability.

[0010] Preferably, the aggregation of the feature representations of each region includes: constructing a region-level graph with each region as a node and the electrical or operational coupling relationship between regions as edges; applying a graph attention mechanism on the region-level graph to dynamically calculate the attention weight between any two connected regions, thereby achieving adaptive fusion of cross-regional risk information; the attention weight serves as a risk impact intensity factor, and is dynamically calculated and adjusted based on historical linkage fault data, real-time power flow direction and switch status, and scheduling authority.

[0011] A regional-level graph was constructed and a graph attention mechanism was used to realize dynamic modeling of risk propagation between regions. The dynamic adjustment of attention weights is based on real-time data, which enhances the model's adaptability to changes in system state, improves the accuracy and interpretability of risk fusion, and effectively captures the systemic risk propagation mechanism.

[0012] Preferably, the regional-level graph further includes: triggering dynamic updates of the regional-level graph when at least one of the following triggering conditions is detected: a change in system topology, a change in regional operating status, or a significant change in regional load and power supply structure; the dynamic updates of the regional-level graph include: re-identifying the electrical connection relationships between regions to update the edge set, deleting invalid connection edges, and recalculating the weight factors of the connection edges between regions based on the updated power flow direction, electrical coupling strength, and scheduling association.

[0013] The dynamic update mechanism enables the model to respond in real time to changes in power system topology and operating status, maintain the timeliness and robustness of risk assessment, ensure that it can still provide accurate risk assessment during system adjustments or failures, and improve the practicality of the model in real-world applications.

[0014] Preferably, in generating node-level, region-level, and system-level risk assessment results, generating node-level risk assessment results includes: inputting the final feature embedding vector obtained by the node through graph neural network learning into a risk mapping function to calculate the quantitative value of the probability or severity of the node experiencing a preset risk type; the risk mapping function adopts a linear mapping function, a nonlinear activation function, or a lightweight feedforward neural network; the preset risk type includes voltage over-limit risk or power flow over-limit risk.

[0015] Node-level risk assessment transforms feature vectors into specific risk indicators through a risk mapping function, providing device-level risk quantification and supporting local fault prevention and optimization.

[0016] Preferably, in generating node-level, regional-level, and system-level risk assessment results, generating regional-level risk assessment results includes: performing aggregation calculations on the node-level risk index values ​​of all nodes within the region to obtain a comprehensive risk index characterizing the overall risk level of the region; the aggregation calculation adopts weighted summation, average aggregation, or attention-weighted aggregation, and the weights are determined according to the importance, capacity, and operational sensitivity of the equipment represented by the nodes.

[0017] Regional risk assessment aggregates node-level indicators and comprehensively considers equipment importance to generate an overall regional risk level, facilitating regional monitoring and scheduling decisions.

[0018] Preferably, the generation of node-level, regional-level, and system-level risk assessment results includes generating regional-level risk assessment results by weighting and integrating regional-level risk indicators for all regions, with the weight coefficients for each region set according to its scale, load ratio, and criticality level.

[0019] By weighting and integrating regional indicators, a system-level risk assessment is generated. The weighting coefficients are based on regional importance, ensuring a comprehensive evaluation of system-level risks, providing a basis for global scheduling, taking into account the impact of different regions on system security, and enhancing the comprehensiveness of risk assessment and decision support capabilities.

[0020] Preferably, after acquiring the multi-source heterogeneous operation data of the power system, the method further includes: modeling the uncertainties of wind speed and light intensity using probability distribution models, generating random scenarios of new energy output through Monte Carlo simulation, and extracting multiple representative typical scenarios from the random scenarios using the K-clustering algorithm as a set of inputs to the graph neural network model.

[0021] By using probabilistic modeling and Monte Carlo simulation to address uncertainties in new energy sources and generating typical scenarios, the robustness of the model to fluctuations in new energy sources is enhanced, and the comprehensiveness and reliability of risk assessment are improved.

[0022] Preferably, dividing the power system into multiple regions includes: dividing the power system into at least two types of regions, namely, a main grid region, a distribution network region, and a microgrid region containing distributed power sources, based on the geographical distribution, voltage level differences, or operational functional characteristics of the power system.

[0023] Preferably, in acquiring multi-source heterogeneous operation data of the power system as input features for corresponding nodes, the multi-source heterogeneous operation data includes: electrical state parameters including voltage amplitude, phase angle, frequency, and current; new energy characteristics, including at least wind speed and irradiance, new energy output, and output prediction error; energy storage and V2G characteristics including state of charge, battery capacity, and charging and discharging power; load characteristics including active load and reactive load, and load prediction values; external disturbances including meteorological factors such as temperature and wind speed; for nodes in different regions, the dimension and type of their feature vectors are allowed to be heterogeneous, and the model can automatically adapt the dimension.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0025] 1. This invention divides the system into multiple regions, such as the main network and distribution network, and constructs local graphs and inter-regional connection graphs for each region, thereby achieving multi-scale risk transmission modeling from equipment nodes to operating areas and then to the overall system. It aligns with the physical topology and scheduling hierarchy of the power system, enabling the model to simultaneously capture local anomalies and global impacts.

[0026] 2. This invention integrates multiple features such as electrical status, new energy output, energy storage, and meteorological environment, and designs a fusion method based on attention mechanism, which has the ability to fuse multi-source heterogeneous data and dynamically adapt. The connection graph between regions can dynamically adjust the edges and weights according to scenarios such as topology changes and island operation, thereby significantly improving the robustness of the model in dealing with new energy fluctuations and network structure changes, and ensuring the real-time and accuracy of risk assessment results.

[0027] 3. This invention can simultaneously generate multi-granularity risk indicators at the node, regional, and system levels, rather than a single comprehensive score. This hierarchical output not only reveals the specific location and scope of impact of risks, but also directly supports differentiated scheduling decisions, such as system-level early warning, regional autonomous control, or local protection of key nodes, realizing an effective closed loop of risk assessment from perception to decision-making. Attached Figure Description

[0028] Figure 1This is an overall flowchart of a power system risk assessment method based on a hierarchical regional graph neural network according to the present invention.

[0029] Figure 2 This is a schematic diagram illustrating the risk information fusion of a power system risk assessment method based on a hierarchical regional graph neural network according to the present invention. Detailed Implementation

[0030] See Figures 1-2 As shown, a power system risk assessment method based on a hierarchical regional graph neural network includes: The power system is divided into multiple regions, and a local graph structure is constructed for each region. Nodes represent power equipment, and edges represent the connections between equipment. The connections between regions are established to form a region-level graph. Multi-source heterogeneous operation data of the power system are obtained as input features for the corresponding nodes; graph neural networks are used to learn and aggregate the node features in each local graph to obtain the regional feature representation of the corresponding operation area. Based on the graph attention mechanism, the feature representations of each region are aggregated in the inter-region connectivity graph to integrate risk information; Based on the characteristic representations of nodes and regions and risk information, risk assessment results are generated at the node level, region level, and system level, respectively.

[0031] Existing risk assessment methods, when facing new power systems, face challenges due to the rapid increase in the large-scale integration of new energy sources, insufficient local grid absorption, and significant gaps in system peak-shaving capacity. This leads to a decline in the power system's frequency regulation capability, posing risks of frequency exceeding limits or even stability disruption. Furthermore, the reactive power and voltage regulation capabilities of new energy sources such as wind and solar power are far lower than those of conventional thermal power units, exposing the system to voltage instability risks. The high proportion of new energy integration, coupled with inter-regional AC / DC interconnection and distributed microgrid integration, makes the local transient energy impact characteristics under system disturbance events more complex, easily triggering global stability risks.

[0032] Most current power system risk assessment methods based on graph neural networks still rely on single-layer graph structures, focusing on node-level or device-level risk modeling. They fail to fully consider the hierarchical relationships between power grid regions and the systemic risk propagation mechanism, making them ill-suited for scenarios such as widespread distributed resource integration, increased spatiotemporal fluctuations in new energy sources, and multi-regional collaborative operation. Furthermore, existing methods suffer from weak integration capabilities and poor structural flexibility when handling multi-source heterogeneous data, making it difficult to achieve high-precision, interpretable, and deployable risk assessments under conditions of high-proportion new energy integration. Therefore, this invention proposes a risk assessment method with hierarchical structure perception capabilities, support for multi-regional information fusion, and adaptability to multi-source data input, to adapt to the complex operating characteristics of new power systems.

[0033] like Figure 1 and Figure 2 In one embodiment shown, Figure 1 This is an overall flowchart of a power system risk assessment method based on a hierarchical region graph neural network according to the present invention. Figure 2 This diagram illustrates the risk information fusion of a power system risk assessment method based on a hierarchical regional graph neural network, as described in this invention. This invention designs a power system risk assessment method based on a hierarchical regional graph neural network, which achieves multi-level risk assessment from local equipment to the overall system by fusing multi-source heterogeneous data and multi-layer graph structure modeling. Specifically, it includes: First, based on the geographical distribution, voltage level differences, or operational characteristics of the power system, it is divided into multiple physical or functional regions, such as the main grid region, distribution network region, and microgrid region containing distributed generation. Within each region, a local graph structure representing equipment and their connections is constructed. Nodes can represent power equipment such as buses, substations, renewable energy access points, energy storage devices, or electric vehicle access points, while edges represent connections such as branches, transformers, or cables. Simultaneously, electrical or operational coupling relationships between regions are established, forming a higher-level inter-regional connectivity graph. This allows for the simultaneous capture of both the topological characteristics within a region and the interactive dependencies between regions during the modeling process.

[0034] In the data preparation phase, multi-source heterogeneous operational data corresponding to each node in the power system are acquired as input features. These features may include electrical state parameters, renewable energy characteristics, energy storage and vehicle-to-grid interaction characteristics, load characteristics, and external environmental disturbances. To handle the uncertainty of renewable energy output, probability distribution models can be used to model wind speed and solar intensity separately. A large number of random scenarios are generated through Monte Carlo simulation, and representative typical scenarios are extracted from them using clustering algorithms such as K-medoids, serving as a set of inputs for subsequent graph neural network models. Node features in different regions are allowed to be heterogeneous in dimension and type, and the model has the ability to automatically adapt.

[0035] In the feature learning stage, a graph neural network is used to perform node-level feature learning and aggregation on the local graph of each region. A graph attention network structure with residual connections is preferably introduced, which is used to fuse the input and output features of the graph attention layer, helping to alleviate the gradient vanishing and feature degradation problems in deep network training. After multi-layer feature propagation and aggregation, the final feature embedding vector of each node is obtained. Furthermore, by performing region pooling on the embedding vectors of all nodes within the region, a region feature vector representing the overall operational risk state of the region is generated.

[0036] To model the transmission and interaction of risks between regions, a region-level graph is constructed, with each region as a node and the coupling relationships between regions as edges. A graph attention mechanism is applied to this graph to dynamically calculate the attention weights between connected regions, which serve as risk impact strength factors, enabling adaptive fusion of cross-regional risk information. The attention weights can be dynamically calculated and adjusted based on historical linked fault data, real-time power flow direction, switch status, and scheduling authority relationships. Furthermore, this method supports dynamic updates to the region-level graph: when the system topology changes, the regional operating state changes, or the regional load and power supply structure changes significantly, the graph is automatically updated. This re-identifies inter-regional connections, removes invalid edges, and recalculates connection weights based on the updated power flow direction, electrical coupling strength, and scheduling relationships, thereby improving the model's adaptability and robustness to changes in system operating state.

[0037] In the risk assessment output phase, risk indicators at the node, region, and system levels are generated based on the learned multi-level feature representations. Node-level risk is calculated by embedding the node's final features into a vector and inputting it into a risk mapping function, quantifying the likelihood or severity of specific risks such as voltage exceedances or power flow exceedances at that node. Region-level risk is obtained by aggregating the node-level risk indicators of all nodes within the region; the aggregation weights can be determined based on the importance, capacity, and operational sensitivity of the equipment represented by the nodes. System-level risk is obtained by weighted fusion of the region-level risk indicators of all regions; the weight coefficients for each region can be set according to its size, load share, and criticality level. Ultimately, the system can output comprehensive risk assessment results covering the node, region, and system levels, providing multi-granular information support for power grid operation monitoring, risk early warning, and dispatching decisions.

[0038] In another embodiment, the present invention relates to a method for multi-source risk assessment of power systems, comprising the following steps: S110: Divide the power system into multiple physical or functional regions and construct a local graph structure for each region to represent the equipment and connection relationships.

[0039] In this embodiment, the main grid, distribution network, and microgrid have different structural and control characteristics. To improve modeling accuracy, the power system is first divided into regions based on geography, voltage level, or operational function.

[0040] Specifically, within each region, a local graph structure is constructed, where nodes represent busbars, substations, new energy access points, energy storage devices, or electric vehicle access points, and edges represent branches, transformers, or cables.

[0041] Furthermore, this step constructs inter-regional connections, forming a region-level graph structure to capture the coupling dependencies between regions. This step ensures that the subsequent graph neural network can learn both the topological characteristics within regions and reflect the interactions between regions.

[0042] After dividing the region, a local graph is constructed for each region, and the connection relationships between different regions in the original system are preserved for subsequent multi-layer modeling.

[0043] S120: Extract multi-source features such as the operating status of power equipment, the output of new energy sources, and the environment, and use them as node input information for the graph neural network.

[0044] In one implementation, a graph neural network model based on a message-passing mechanism is used to propagate and aggregate node features. The graph neural network can be one or more combinations of Graph Convolutional Network (GCN), Graph Attention Network (GAT), or GraphSAGE based sampling aggregation. The graph neural network includes at least one graph feature update layer, where each layer iteratively updates node representations through weighted aggregation of features from neighboring nodes. The weight coefficients can be adaptively learned based on topological relationships or feature correlations between nodes. The number of layers, hidden feature dimensions, activation function type, and parameter initialization method of the graph neural network can be set according to the region size, number of nodes, and computational resource conditions, and do not constitute a limitation on the scope of protection of this invention.

[0045] The input feature vector for each node includes: Electrical status parameters: voltage amplitude, phase angle, frequency, current; New energy characteristics: wind speed, irradiance, wind / solar power output, prediction error; Energy storage / V2G characteristics: State of charge (SOC), battery capacity, charge / discharge power; Load characteristics: active / reactive load, load forecast; External disturbances: meteorological factors such as temperature, wind speed, and cloud cover.

[0046] Features can be heterogeneous in different regions, and the model can automatically adapt to the dimensions.

[0047] To improve timeliness and scenario robustness, a typical new energy scenario construction mechanism is introduced. For wind speed, a two-parameter Weibull distribution is used for modeling; for photovoltaics, a Beta distribution is used to simulate light intensity, and the photovoltaic output scenario is converted by combining the PV module efficiency curve.

[0048] A large number of new energy output scenarios were generated through Monte Carlo simulation, and K-medoids clustering was used to extract representative typical scenarios as input for subsequent models.

[0049] S130: In the local maps of each region, based on the graph neural network framework described in S120, a graph attention network (GAT) structure with residual connections is preferably used to learn and aggregate node features, thereby obtaining node-level and region-level risk feature representations.

[0050] In one implementation, a lightweight graph neural network is used for feature encoding in each region, and local risk state representation learning is achieved by utilizing node features and topological relationships.

[0051] In one implementation, a graph attention network (GAT) structure with residual connections is used for node feature learning in the local graphs of each region. The residual connections are used to fuse the input features of the graph attention layer with its output features to alleviate the gradient vanishing and feature degradation problems during deep feature propagation.

[0052] The residual connection can be set as an intra-layer residual connection between adjacent graph attention layers, or as a cross-layer residual connection spanning multiple graph attention layers, or both of the above forms can be used simultaneously. The specific connection method can be configured according to the model depth and region size.

[0053] After several layers of propagation, the final embedding vector of the node is obtained, and the region center vector is obtained by using a region pooling operation (such as average pooling). This vector represents the overall operational risk status of the region.

[0054] S140: Construct an inter-regional connectivity graph and use the graph attention mechanism (GAT) to model the fusion and propagation of risk information between different regions.

[0055] In one implementation, inter-regional risk transmission modeling is used to capture potential impact relationships between the main grid and distribution network, and among multiple sub-regions. This is based on the regional vector z obtained in the previous stage. r Constructing the regional map G region Each node represents a region, and edges represent their connections. A graph attention mechanism is used for cross-region feature aggregation.

[0056] Furthermore, to enhance the interpretability and stability of interactions between regions, this invention introduces a weighting factor w into the region graph. rk This weight is used to measure the intensity of risk impact between region r and its neighboring region k. This weight can be dynamically generated based on historical linkage fault data, power flow direction, switch status, or dispatch authority relationship.

[0057] The region map structure supports dynamic adjustment based on changes in system operating status.

[0058] In one implementation, the dynamic update of the regional map structure is triggered when at least one of the following triggering conditions is detected: 1. The system topology changes, including line switching, switch status changes, and transformer operation status adjustments; 2. The regional operation status changes, including regional islanding, regional merging, or disconnection; 3. The regional load or power supply structure changes significantly, including large-scale load reconfiguration, centralized grid connection of new energy sources, or disconnection.

[0059] After triggering a dynamic update of the regional graph structure, the edge set and its connection weights in the regional graph are adjusted according to the updated system operation information. Specifically, this includes: 1. Re-identifying the electrical connection relationships between regions and updating the connection edges between regional nodes; 2. Deleting the corresponding connection edges when there are no effective electrical or operational coupling relationships between regions; 3. Recalculating the weight factors of the connection edges between regions based on the updated power flow direction, electrical coupling strength, or scheduling association.

[0060] The above methods allow for automatic updates of edge sets and connection strengths in scenarios such as system topology changes, regional silos, and load reconfiguration, thereby improving the model's practicality and robustness.

[0061] Finally, the system merges the updated feature vectors {z} from all regions. r '} serves as a high-level representation input for the overall system risk assessment, driving subsequent indicator outputs.

[0062] S150: Based on the multi-layer feature representations obtained in steps S130 and S140, generate node-level, regional-level, and system-level risk assessment results for power grid operation monitoring and decision support.

[0063] In one implementation, the node-level risk index value is calculated based on the final feature embedding vector corresponding to the node. The node feature representation is converted into a specific risk index value through a risk mapping function. The risk mapping function may include a linear mapping, a nonlinear activation function, or a lightweight feedforward neural network, which is used to characterize the probability or severity of risks such as voltage overrun or power flow overrun at the node.

[0064] Regional-level risk index values ​​are obtained by aggregating node-level risk indicators within the region. The aggregation method can be weighted summation, average aggregation, or attention-weighted aggregation, where the weights can be determined based on node importance, equipment capacity, or operational sensitivity, thereby forming a comprehensive risk index that characterizes the overall operational risk level of the region.

[0065] System-level risk index values ​​are calculated based on regional feature vectors obtained through risk fusion between regions. By fusing multiple regional risk indicators, a comprehensive risk assessment result at the system level is obtained. The fusion method can adopt a weighted fusion approach, where the weight coefficients can be set according to regional size, load share, or criticality level to reflect the degree of impact of different regions on system operational safety.

[0066] In one implementation, the system-level comprehensive risk level is obtained by weighted fusion of node-level, regional-level, and system-level risk indicators. The weighting coefficients reflect the relative importance of different risk indicators to system operational safety, and their determination can include at least one of the following: 1. Offline assessment of the importance of different risk indicators and generation of corresponding weights based on historical operational data and accident statistics; 2. Dynamic adjustment of the weighting coefficients based on the current system operating status, according to load levels, renewable energy penetration rates, or the degree of operational constraint tension; 3. Pre-setting a weighting configuration scheme based on operational strategies or scheduling objectives, and correcting it during operation based on monitoring results. The weighting coefficients can be adaptively updated during operation to adjust the focus on different risk types.

[0067] Each region can output a separate overall risk level R. r This data is combined with the local risk scores of key nodes or equipment in the region to form a regional risk map, which is used to assist in the initiation of regional autonomous regulation or local protection strategies.

[0068] This invention achieves hierarchical risk perception through multi-scale risk transmission modeling at the system, region, and node levels; it possesses strong multi-source fusion capabilities through unified modeling of heterogeneous features such as new energy sources, energy storage, V2G, load, and meteorology; it provides comprehensive risk indicators, outputting multi-granularity indicators at the node, region, and system levels to support early warning and scheduling; and it enhances robustness in responding to new energy fluctuations by adapting to topological changes and new energy fluctuations through residual aggregation and dynamic regional maps.

[0069] In another embodiment, a power system comprising a main grid area, a distribution network area, and a microgrid area containing new energy sources is taken as the research object. First, the system is divided into multiple operating areas according to geographical location and voltage level, and a corresponding local graph structure is constructed in each area. At the same time, the connection relationship between areas is constructed to form a regional-level graph structure.

[0070] The system collects electrical operation data, renewable energy output data, energy storage and electric vehicle access status data, load data, and corresponding meteorological and environmental data from each node during operation, using these as node input features. Based on this, the system's operating status is modeled under different renewable energy fluctuation conditions by constructing various typical renewable energy output scenarios.

[0071] Within the region, a graph attention network with residual connections is used to learn node features and obtain node-level risk feature representations. Then, a region-level risk feature vector is generated through region pooling. Subsequently, risk information between regions is fused in the region-level graph, and the region graph structure and connection weights are dynamically adjusted according to changes in system topology and operating status.

[0072] Finally, based on the risk assessment results at the node, regional, and system levels, a comprehensive assessment of system operation risks is conducted. The results show that the method of this invention can maintain the stability and consistency of risk assessment results under conditions of fluctuations in renewable energy output or adjustments to the system topology, and can simultaneously output risk information at the node, regional, and system levels, providing effective support for operation monitoring, risk early warning, and scheduling decisions.

[0073] All data collection and extraction in this invention are carried out under compliant and legal conditions.

Claims

1. A power system risk assessment method based on hierarchical regional graph neural networks, characterized in that, include: The power system is divided into multiple regions, and a local graph structure is constructed for each region, where nodes represent power equipment and edges represent the connection relationships between equipment. Establish connections between regions to create a regional-level map; Obtain multi-source heterogeneous operation data of the power system as input features for the corresponding nodes; A graph neural network is used to learn and aggregate the node features of each local graph to obtain the regional feature representation of the corresponding operating region; Based on the graph attention mechanism, the feature representations of each region are aggregated in the inter-region connectivity graph to integrate risk information; Based on the characteristic representations of nodes and regions and risk information, risk assessment results are generated at the node level, region level, and system level, respectively.

2. The power system risk assessment method based on hierarchical regional graph neural network according to claim 1, characterized in that, The method of learning and aggregating node features in each local graph using a graph neural network includes: learning node features in each local graph of a region using a graph attention network with residual connections, wherein the residual connections fuse the input and output features of the graph attention layer, and after multi-layer feature propagation and aggregation, the final embedding vector of each node is obtained; performing a region pooling operation on the final embedding vectors of all nodes in the region to obtain a region feature vector representing the overall operational risk status of the region.

3. A power system risk assessment method based on a hierarchical regional graph neural network according to claim 1 or 2, characterized in that, The aggregation of regional feature representations includes: constructing a regional-level graph with each region as a node and the electrical or operational coupling relationship between regions as edges; applying a graph attention mechanism to the regional-level graph to dynamically calculate the attention weight between any two connected regions, thereby achieving adaptive fusion of cross-regional risk information; the attention weight serves as a risk impact intensity factor, which is dynamically calculated and adjusted based on historical linkage fault data, real-time power flow direction and switch status, and scheduling authority.

4. The power system risk assessment method based on a hierarchical regional graph neural network according to claim 3, characterized in that, The regional-level map also includes: triggering dynamic updates of the regional-level map when at least one of the following triggering conditions is detected: a change in system topology, a change in regional operating status, or a significant change in regional load and power structure. The dynamic update of the regional graph includes re-identifying the electrical connection relationships between regions, updating the edge set, deleting invalid connection edges, and recalculating the weight factors of the connection edges between regions based on the updated power flow direction, electrical coupling strength, and scheduling association.

5. The power system risk assessment method based on a hierarchical regional graph neural network according to claim 4, characterized in that, The generation of node-level, region-level, and system-level risk assessment results includes generating node-level risk assessment results by: inputting the final feature embedding vector obtained by the node through graph neural network learning into the risk mapping function, and calculating the quantitative value of the probability or severity of the node experiencing a preset risk type; the risk mapping function adopts a linear mapping function, a nonlinear activation function, or a lightweight feedforward neural network; the preset risk type includes voltage over-limit risk or power flow over-limit risk.

6. A power system risk assessment method based on a hierarchical regional graph neural network according to claim 4 or 5, characterized in that, The generation of node-level, regional-level, and system-level risk assessment results includes generating regional-level risk assessment results by: aggregating the node-level risk index values ​​of all nodes in the region to obtain a comprehensive risk index that characterizes the overall risk level of the region; the aggregation operation adopts weighted summation, average aggregation, or attention-weighted aggregation, and the weights are determined according to the importance, capacity, and operational sensitivity of the equipment represented by the nodes.

7. The power system risk assessment method based on a hierarchical regional graph neural network according to claim 6, characterized in that, The generation of node-level, regional-level, and system-level risk assessment results includes the generation of regional-level risk assessment results by weighted fusion of regional-level risk indicators for all regions, with the weight coefficients for each region set according to its scale, load ratio, and criticality level.

8. The power system risk assessment method based on a hierarchical regional graph neural network according to claim 7, characterized in that, After acquiring the multi-source heterogeneous operation data of the power system, the method further includes: modeling the uncertainties of wind speed and light intensity using probability distribution models, generating random scenarios of new energy output through Monte Carlo simulation, and extracting multiple representative typical scenarios from the random scenarios using the K-clustering algorithm as a set of inputs to the graph neural network model.

9. A power system risk assessment method based on a hierarchical regional graph neural network according to claim 8, characterized in that, The division of the power system into multiple regions includes: dividing the power system into at least two types of regions, namely, the main grid region, the distribution network region, and the microgrid region containing distributed power sources, based on the geographical distribution, voltage level differences, or operational functional characteristics of the power system.

10. A power system risk assessment method based on a hierarchical regional graph neural network according to claim 8, characterized in that, The acquisition of multi-source heterogeneous operation data of the power system as input features for corresponding nodes includes: electrical state parameters including voltage amplitude, phase angle, frequency, and current; new energy characteristics, including at least wind speed and irradiance, new energy output, and output prediction error; energy storage and V2G characteristics including state of charge, battery capacity, and charging and discharging power; load characteristics including active load and reactive load, and load prediction values; external disturbances including meteorological factors such as temperature and wind speed; for nodes in different regions, the dimension and type of their feature vectors are allowed to be heterogeneous, and the model can automatically adapt the dimension.

Images