Operation risk early warning method and system based on power grid information system

Abstract

The invention provides an operation risk early warning method and system based on a power grid information system, and relates to the technical field of power grid operation risk assessment. According to the method, equipment risk characterization and future risk score prediction are realized by constructing a risk perception graph fusing node and edge features; and identifying a high-risk region and a diffusion path by combining a clustering algorithm, and triggering a hierarchical response strategy based on a rule base. According to the invention, real-time early warning and intelligent management and control of the operation risk of the power grid are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid operation risk assessment, and in particular to an operation risk early warning method and system based on a power grid information system. Background Art

[0002] Currently, the risk assessment of power grid operation mainly relies on manual experience and dispatching rules. Although some regions have introduced operation monitoring, state estimation and event log analysis systems, most of them are limited to static safety verification and post-fault handling. It is difficult to form a dynamic early warning mechanism covering real-time data streams. In addition, data between systems are isolated, responses are delayed, and predictive capabilities are weak.

[0003] With the advancement of automated grid dispatching, integrated information systems, and intelligent device sensing, operational risk management is evolving from traditional post-event response to proactive sensing, trend identification, and early warning based on multi-source information across the entire network. Especially in complex operational scenarios such as large-scale renewable energy grid integration, electric vehicle load integration, and coupled meteorological disturbances, real-time, refined, and adaptive risk identification is crucial.

[0004] However, existing methods lack temporal coupling and semantic fusion between the various types of operational data in power grid information systems (such as telesignaling, telemetering, alarms, operation records, and planned outages). Furthermore, early warning logic often relies on fixed rule matching, making it difficult to reflect new and uncertain risk events. Risk quantification often relies on static indicators and lacks dynamic distribution modeling and uncertainty expression mechanisms. Therefore, an operational risk early warning method that integrates data-driven and expert rules, and takes into account both structural characteristics and distribution evolution, is urgently needed. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide an operation risk warning method and system based on the power grid information system, which realizes real-time risk prediction and active warning based on multi-source data of the entire network, makes up for the defect of delayed response of traditional static rule systems, and provides an expandable intelligent risk control framework for complex operation scenarios such as high penetration of new energy and centralized charging of electric vehicles.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] An operation risk early warning method based on a power grid information system, comprising:

[0008] Collect scheduling data, equipment status data, operation monitoring data, historical fault record data, and power grid topology data to form a multi-source heterogeneous raw data set, and perform data cleaning, standardization, and time synchronization on the raw data set to obtain a risk analysis data set;

[0009] Defining nodes and their edge relationships using the power grid topology data, mapping equipment operation indicators in the risk analysis dataset into node features, and mapping topological connectivity and power flow dependency into edge features to construct a risk perception feature graph;

[0010] Using a graph neural network to train the risk perception feature graph, and outputting a risk representation vector corresponding to each device node;

[0011] The risk representation vector is input into a time series analysis model integrated with a multi-scale attention mechanism to predict the operation risk score of each device node within a preset time window in the future;

[0012] Based on the risk scoring results and the grid topology data, a density-peak clustering combined with a hierarchical clustering algorithm is used to dynamically identify high-risk areas and potential risk diffusion paths;

[0013] Establish a risk warning rule base, match the operation risk score of each device node with the rule base, and trigger the four warning levels of normal, attention, warning or severe according to the matching results, and generate corresponding response strategies.

[0014] Preferably, it also includes:

[0015] The warning level and the corresponding response strategy are pushed to the power grid dispatching platform and operation and maintenance terminal to achieve auxiliary decision-making and automatic handling of operation risks.

[0016] Preferably, the grid topology data is used to define nodes and their edge relationships, the equipment operation indicators in the risk analysis data set are mapped into node features, and the topological connectivity and power flow dependency relationships are mapped into edge features to construct a risk perception feature graph, including:

[0017] Based on the power grid topology data, extract the unique identifier and electrical connection relationship of each device, define the nodes and edges in the graph structure, and construct an initial power grid device topology graph;

[0018] Aligning the equipment operating indicators corresponding to each node in the risk analysis data set, extracting real-time current, voltage, active power, operating temperature, and the number of historical faults, and encoding the real-time current, voltage, active power, operating temperature, and the number of historical faults into a node feature vector;

[0019] Extracting edge attribute data based on the flow direction, electrical connection type, and historical linkage fault information between adjacent nodes in the initial power grid device topology diagram, and encoding the attribute data into an edge feature vector;

[0020] The node feature vector and the edge feature vector are embedded into the initial power grid equipment topology map to construct the risk perception feature map containing structural information and operation attribute information.

[0021] Preferably, based on the power flow direction, electrical connection type and historical linkage fault information between adjacent nodes in the initial power grid equipment topology diagram, edge attribute data is extracted and encoded into an edge feature vector, including:

[0022] Read the real-time power flow direction data corresponding to the adjacent nodes. If the real-time power flow direction data is flowing from the first node to the second node, set the power flow direction flag to +1; otherwise, set it to -1;

[0023] Generate an electrical connection type code value according to a preset mapping table based on the line voltage level, conductor cross-section and switch topology state between the adjacent nodes;

[0024] Retrieve the fault events containing the adjacent nodes from the historical fault records, count the number of occurrences of the linked faults, and perform proportional normalization using the number of faults in the total observation period as the denominator to obtain a linked fault probability value;

[0025] Feature splicing is performed in the order of the power flow direction identifier, the electrical connection type code value, and the linkage failure probability value to form the edge feature vector.

[0026] Preferably, the risk perception feature graph is trained using a graph neural network to output a risk representation vector corresponding to each device node, including:

[0027] Assume that the initial node feature matrix is ​​F (0) , the adjacency matrix of the edge connection relationship is A, the identity matrix is ​​I, and in each graph convolution layer l, the node feature update is calculated as follows:

[0028]

[0029] in is the adjacency matrix after adding self-loops, for The degree matrix, W (l) is the trainable weight matrix of the lth layer, σ(·) is the nonlinear activation function;

[0030] In the training sample, each node in the graph is assigned a risk level label Y within the historical observation time window i ∈{low,medium,high}, for supervised learning;

[0031] When the network propagates forward to the last layer L, the node output matrix Z = F is obtained (L) W (out)The output vector Z of each node i is calculated as i,* The predicted probability of belonging to risk level c is calculated using the following formula:

[0032]

[0033] The cross-entropy loss is constructed as where N is the total number of nodes participating in training, 1(·) is an indicator function; c' represents traversing all possible risk level categories in the normalization denominator;

[0034] The gradient of the loss function is calculated, and all weight matrices W are updated using an iterative optimization algorithm (l) and W (out) until convergence is reached or a preset number of iterations is reached;

[0035] After training is complete, the row vector corresponding to each node in the final node output matrix F (L) is fixed as the risk representation vector of the node.

[0036] Preferably, the risk representation vector is input into a time series analysis model that fuses a multi-scale attention mechanism to predict the operating risk score of each device node within a future preset time window, including:

[0037] Based on the risk representation vector, an input sequence is organized according to device node number and consecutive time steps where each represents the risk representation vector at time t, T is the length of the history, and d is the feature dimension;

[0038] The sequence {x t} is input in parallel to multiple sub-channels, each sub-channel corresponding to a different time window scale k∈{k1,k2,…,k m}, each scale corresponding to a local context modeling capability with different step lengths; where m represents the number of scales in the multi-scale attention mechanism

[0039] At each scale k, the context of each time position t in the input sequence is weighted using an attention mechanism, and the calculation formula is:

[0040]

[0041] where are the query and key vectors of time step t and its context position τ, respectively, represents the context window index set centered on t at scale k, represents the attention weight of position τ in scale k at time step t, and τ' represents the index variable used for normalizing the weight in the same window.

[0042] The weighted results at all scales are concatenated and integrated into a unified temporal code z through linear mapping t , the formula is as follows:

[0043]

[0044] in is a value vector, Linear(·) is a linear fusion operation;

[0045] The fused sequence Input the classification layer and output the operation risk score of each device node within the preset future time window Where Δ is the prediction step size;

[0046] Using the historically known risk score r T+Δ To supervise the labels, construct the mean square error loss function And optimize the model parameters through back propagation until the loss function converges; Represents real risk, Represents predicted risk.

[0047] Preferably, based on the risk scoring results and the grid topology data, a density-peak clustering combined with hierarchical clustering algorithm is used to dynamically identify high-risk areas and potential risk diffusion paths, including:

[0048] At each device node defined by the power grid topology data, the risk scores of the node and its directly adjacent nodes are aggregated to obtain the local risk density value of the node;

[0049] Comparing the local risk density value of each node with the distance between it and the highest density node closest to it in topology, when a node has both a high local risk density and a distance to a node with a higher density reaches a minimum distance threshold, the node is marked as a risk peak node;

[0050] Taking all the risk peak nodes as cluster centers, the surrounding nodes are absorbed in order from near to far along the topological shortest path to generate several initial high-risk clusters;

[0051] For any two initial high-risk clusters, if the topological distance between their boundary nodes is not greater than the preset merging threshold, and the difference in the average risk scores between the two clusters is less than the preset difference threshold, they will be merged in a bottom-up manner until there is no cluster pair that meets the conditions.

[0052] Each cluster obtained after hierarchical merging is defined as a high-risk area, and its central node, boundary node and cluster device list are recorded;

[0053] According to the equipment connection direction and power flow direction in the power grid topology data, a recursive search is performed outward from the central node of the high-risk area, and a sequence of nodes adjacent to the identified high-risk area with continuously increasing risk scores is output to form a potential risk diffusion path.

[0054] Preferably, the judgment method of having a higher local risk density and a distance to a higher density node reaching a minimum distance threshold is:

[0055] Sort the local risk density values ​​of all nodes from large to small;

[0056] Nodes whose local risk density values ​​are in the top 10% of the entire network distribution are selected as nodes with higher local risk density;

[0057] For each node with a higher local risk density, calculate its shortest topological distance to all nodes with higher density;

[0058] When the shortest topological distance is greater than a preset minimum distance threshold, the node is confirmed to be a risk peak node.

[0059] Preferably, a risk warning rule base is established, and the operational risk score of each device node is matched with the rule base. According to the matching results, four warning levels (normal, concern, warning, or severe) are triggered, and corresponding response strategies are generated, including:

[0060] Based on expert experience, grid operation procedures, and historical failure cases, several rules are formulated. Each rule includes: risk score threshold range, applicable equipment type, associated operating environment conditions, corresponding warning level, and response strategy template.

[0061] The operational risk scores of device nodes are divided into four levels, corresponding to the four warning levels of normal, concern, warning, and severe. Each warning level has a unique and non-overlapping risk score range.

[0062] For each device node, first identify the level interval to which the current risk score belongs, and then retrieve the rule items that match it based on the node type and the area it is located in;

[0063] If multiple rule items meet the matching conditions, the rule with the lower upper limit of the risk score range is prioritized to ensure the warning sensitivity and ultimately determine the warning level of the node;

[0064] According to the determined warning level, the corresponding response strategy template is extracted from the rule item to generate the response strategy including alarm information push, patrol task dispatch, equipment load adjustment, protection device linkage or planned maintenance.

[0065] An operation risk early warning system based on a power grid information system, comprising:

[0066] A data processing unit is used to collect scheduling plan data, equipment status data, operation monitoring data, historical fault record data, and power grid topology data to form a multi-source heterogeneous raw data set, and perform data cleaning, standardization, and time synchronization on the raw data set to obtain a risk analysis data set;

[0067] a feature graph construction unit, configured to define nodes and their edge relationships using the power grid topology data, map equipment operation indicators in the risk analysis dataset into node features, map topological connectivity and power flow dependency into edge features, and construct a risk perception feature graph;

[0068] A graph neural network training unit is used to train the risk perception feature graph using a graph neural network and output a risk representation vector corresponding to each device node;

[0069] A time series risk prediction unit, configured to input the risk representation vector into a time series analysis model integrated with a multi-scale attention mechanism to predict the operation risk score of each device node within a preset future time window;

[0070] A high-risk area identification unit is used to dynamically identify high-risk areas and potential risk diffusion paths based on the risk scoring results and the power grid topology data by using a density-peak clustering combined with a hierarchical clustering algorithm;

[0071] The warning strategy generation unit is used to establish a risk warning rule base, match the operation risk score of each device node with the rule base, and trigger the four warning levels of normal, concern, warning or severe according to the matching results, and generate corresponding response strategies.

[0072] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0073] By introducing graph neural networks and multi-scale attention time series models, this method realizes the three-dimensional dynamic risk measurement of "structure-state-time" in the power grid scenario for the first time: on the one hand, it uniformly cleans and integrates multi-source data such as scheduling plans, telemetry, alarms, fault records, etc., breaking the data silos between existing systems; on the other hand, it explicitly encodes the power grid topology relationship and flow direction as edge features, so that the risk score not only reflects the node's own status, but also captures the risk propagation trend in the network; combined with density-peak and hierarchical clustering, high-risk areas and their diffusion paths can be locked in real time. Finally, through the warning level determination driven by expert rules and machine learning, the response strategy is leapfrogged from "post-fault disposal" to "pre-fault prevention", significantly improving the warning accuracy and disposal timeliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below only illustrate some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0075] Figure 1 The method flowchart provided for the embodiments of the present application is as shown in

[0076] Figure 2 The system structure schematic diagram provided for the embodiments of the present application is as shown in DETAILED DESCRIPTION

[0077] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0078] The purpose of the present application is to provide a power grid information system-based operation risk early warning method and system, which realizes real-time risk prediction and active early warning based on full-network multi-source data, makes up for the defects of response lag of traditional static rule systems, and provides an extensible intelligent risk control framework for complex operation scenarios such as new energy high penetration and electric vehicle concentrated charging.

[0079] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0080] Figure 1 The method flowchart provided for the embodiments of the present application is as shown in Figure 1 The present application provides a power grid information system-based operation risk early warning method, which comprises the following steps:

[0081] Step 100: Collecting scheduling plan data, equipment state data, operation monitoring data, historical fault record data and power grid topology structure data, forming a multi-source heterogeneous original data set, and performing data cleaning, standardization and time synchronization processing on the original data set to obtain a risk analysis data set;

[0082] Step 200: Defining nodes and their edge relationship with the power grid topology structure data, mapping the equipment operation indexes in the risk analysis data set into node features, mapping the topological connectivity and power flow dependency relationship into edge features, and constructing a risk perception feature map;

[0083] Step 300: Use a graph neural network to train the risk perception feature graph and output a risk representation vector corresponding to each device node;

[0084] Step 400: Input the risk representation vector into a time series analysis model integrated with a multi-scale attention mechanism to predict the operational risk score of each device node within a future preset time window;

[0085] Step 500: Based on the risk scoring results and the grid topology data, a density-peak clustering combined with a hierarchical clustering algorithm is used to dynamically identify high-risk areas and potential risk diffusion paths;

[0086] Step 600: Establish a risk warning rule base, match the operation risk score of each device node with the rule base, and trigger the four warning levels of normal, attention, warning or severe according to the matching results, and generate corresponding response strategies.

[0087] Preferably, it also includes:

[0088] Step 700: Push the warning level and the corresponding response strategy to the power grid dispatching platform and operation and maintenance terminal to achieve auxiliary decision-making and automatic handling of operation risks.

[0089] In this embodiment, an event-driven streaming access method is first used to collect scheduling plans, telemetry / telesignaling, equipment status, alarm logs, and historical fault work orders in real time. This embodiment achieves consistent positioning of data across business systems by constructing a triple mapping relationship of "scheduling unit ID-topology node code-equipment instance code"; and embeds a topology consistency check operator in the data stream to perform a dual-channel check on the transaction timestamp and GPS time of the access record, and combines the "parent-child" inheritance relationship of the topology hierarchy to perform forward filling of missing fragments, filtering out false outliers caused by maintenance or planned power outages. For abnormal high-frequency bands, the embodiment triggers a dynamic window scaling algorithm based on fault labels to refine the sampling granularity, thereby obtaining a continuous and complete time series chain without losing key fluctuation characteristics.

[0090] Subsequently, this embodiment applies a self-built multi-source indicator standardization template library to perform unified dimensional conversion on the original indicators of different sampling periods and measurement units; for the operation monitoring data, a hierarchical Z-score iterative normalization method is used to offset the differences in equipment power levels; for historical fault work orders, a semantic parser is used to split the free text into triplets of fault type code, failure mode code and repair action code, and hash association is performed with the topological nodes.

[0091] In step 200 of this embodiment, the topology snapshot interface between the dispatching master station and the energy management system (EMS) is first called to extract the latest topology table for the day using the three-layer index of "equipment global unique identifier - busbar code - phase connection number". Through an incremental comparison algorithm, online updates are only performed on newly added, removed, or isolated edges for maintenance, avoiding the computational overhead caused by redrawing the entire network. Its innovation lies in: writing the flow direction into the edge attributes during graph initialization. The direction value is automatically determined by the real-time flow calculation framework based on the node voltage phase angle difference and is synchronously refreshed with the 15-second real-time settlement cycle to ensure that the edge characteristics are consistent with the operating status. At the same time, the fault event tracing link is used to perform a sliding update on the linkage fault counts of adjacent nodes, providing high-confidence edge attributes for subsequent risk diffusion modeling.

[0092] In terms of node feature mapping, this embodiment adopts a "one node, one cache" mechanism, reserving a fixed-length sparse vector slot for each node in the GPU memory: real-time current, phase voltage, active power, and operating temperature are truncated and normalized according to the 99.5 percentile threshold and then written directly into the floating-point slot; the number of historical faults is updated in real time through exponential decay counting and written into the discrete slot to ensure the immediate impact of sudden faults on feature distribution. Subsequently, the system uses zero-copy technology to directly embed node features and edge features into the initial topology graph stored in the form of a compressed sparse matrix, generating a risk-aware feature graph that contains both structural semantics and high-dimensional state information, ensuring that subsequent graph neural networks can complete batch forward reasoning in milliseconds.

[0093] In this embodiment, the system first subscribes to the real-time flow of the power flow from the dispatching master station, and automatically locates the power direction of the adjacent nodes with the help of the unique device identifier; when the power flows from the upstream substation to the downstream load node, the power flow direction identifier is set to +1, and in the opposite direction it is set to -1, and the identifier is automatically refreshed in each 15-second power flow settlement cycle. Subsequently, the system calls the solidified "voltage level-conductor cross-section-switch status" three-dimensional mapping table in the line archive library, and directly returns the corresponding electrical connection type code value, realizing the second-level mapping of complex line combinations without the need for additional rule reasoning. The innovation here is: by bidirectionally binding the power flow direction to the line archive, the edge attributes can be synchronized instantly as the operating mode switches, avoiding the delay and mismatch caused by the traditional offline topology overload.

[0094] Next, this embodiment retrieves linked fault records by node pair from the fault event table in the operation and maintenance data lake and counts the number of faults involving adjacent nodes using a sliding time window. The system divides this number by the total number of fault events within the window to obtain a linked fault probability value, which is used to quantify the contagiousness of the risk. Finally, the system concatenates three pieces of data in a fixed order: "flow direction identifier → electrical connection type code value → linked fault probability value" to generate an edge feature vector of length 3. This vector is written to the graph cache in sparse matrix form. This provides the subsequent graph neural network with high-confidence edge features that simultaneously reflect directionality, electrical properties, and fault coupling without adding additional preprocessing overhead, ensuring the real-time and accurate inference of the model.

[0095] Optionally, in step 300 of this embodiment, a multi-layer graph convolutional neural network structure is first initialized based on the node features and edge structure constructed in the risk-aware feature graph, and each layer performs node information aggregation and nonlinear activation operations. In order to improve the stability of training, the embodiment introduces a self-connection mechanism in the feature propagation process, so that each node can retain its original features at the same time, and update the features according to the connection structure and edge attributes of its adjacent nodes. Each layer of convolution operation is provided with a trainable weight parameter to control the impact of different adjacency relationships on the update of node features. The training samples are supervised and labeled by assigning a clear risk level label (such as low risk, medium risk, high risk) to each node in the historical time window to ensure that the network can fully learn the mapping relationship between structural information and risk semantics during the forward propagation process.

[0096] After obtaining the final output of the graph convolutional network, the embodiment maps each node to the predicted probability of three risk levels through normalization processing, and constructs a cross-entropy loss function to measure the deviation between the predicted result and the true label. In order to improve the efficiency of model training and generalization ability, an iterative optimization algorithm is used to perform reverse gradient updates on the network parameters of each layer until the loss value converges or reaches a preset number of iterations. After the training is completed, the result vector corresponding to each node in the output layer is fixed as the risk characterization vector of the node. This characterization vector integrates the local operating status of the node, the surrounding structural relationship and the historical risk performance, and can be used as a high-dimensional input for the downstream time series risk prediction model, effectively supporting the accuracy and real-time performance of subsequent early warning analysis tasks.

[0097] In step 400 of this embodiment, the risk characterization vector is first reorganized into a time-series input sequence according to the device node number and continuous time steps; this embodiment constructs multiple sub-channels in parallel, and each sub-channel corresponds to a time window scale of different lengths, so that the model can simultaneously model the risk evolution law at multiple time granularities. At each scale, this embodiment is based on the attention mechanism, takes the current time step as a reference, calculates its correlation with each context time point in the window, and generates a normalized weight distribution accordingly; then, based on the weight, a weighted summation is performed on the historical risk characterization vector to obtain the time aggregation representation at this scale. Through this mechanism, the risk change characteristics of the device under short-term fluctuations, medium-term adjustments and long-term trends can be captured respectively, thereby improving the response sensitivity and stability of the model to complex operating situations.

[0098] Next, this embodiment concatenates the weighted results of all scales in time scale order, inputs them into the linear mapping layer, and integrates them into a unified time series coding vector; this coding vector is input into the risk score prediction module to output the operating risk score of the device node within the preset time window in the future. During the training process, this embodiment uses the real risk score in the historical period as the supervision signal, constructs the mean square error loss function, and optimizes the parameters of all attention mechanisms and mapping layers through back propagation until the prediction error converges. This method can fully express the distribution characteristics of risks at different time scales, realizes the joint modeling of operating trends, fluctuation risks and implicit uncertainties, and provides risk warning capabilities with foresight and accuracy advantages for complex power grid operation scenarios.

[0099] In step 500 of this embodiment, the risk scores of each device node and its one-hop adjacent nodes are first weighted and summed using the power grid topology relationship to obtain the local risk density of the node; then the nodes are sorted from high to low according to density, and the shortest topological distance from the current node to the node with higher density is calculated. When the local risk density of a node ranks in the top 10% of the entire network, and its shortest topological distance to any node with higher density is not less than a preset distance threshold, this embodiment determines the node as a risk peak node. Subsequently, all risk peak nodes are used as cluster centers, and the surrounding nodes are absorbed in order from near to far according to the topological shortest path to generate an initial cluster; if the topological distance between the boundary nodes of the two initial clusters does not exceed the merging threshold, and the difference in the average risk scores of the two clusters is lower than the difference threshold, they are merged in a bottom-up manner until the merging condition is no longer met.

[0100] After the hierarchical merging is completed, the embodiment defines each convergent cluster as a high-risk area, records the center node, boundary node and cluster device list, and based on the power flow direction and device connection direction, recursively searches the node sequence with continuously rising risk score from the cluster center outward, and outputs the potential risk diffusion path. Through the "density priority-hierarchical adjustment" two-stage clustering framework, the embodiment can quickly locate the local high-risk aggregation area, and dynamically identify the cross-regional diffusion link, avoiding the misjudgment and omission of the traditional single threshold method for complex topology, ensuring the accuracy and real-time performance of risk identification and diffusion prediction

[0101] In step 600 of the embodiment, first, the local risk density values of all nodes are sorted in descending order, and the nodes ranked in the top 10% of the whole network are selected as high-density candidate nodes. Then, for each high-density candidate node, the shortest path distance from it to all higher-density nodes is calculated based on the power grid topology structure data. When the shortest path distance is greater than the preset minimum distance threshold, it is considered that the node has significant risk aggregation and structural independence in space, and it is confirmed as a risk peak node. Through the density and distance dual judgment mechanism, the embodiment effectively avoids the mutual shielding problem between high-density nodes, and improves the representativeness of the peak node and the stability of the initial cluster division of the high-risk area.

[0102] In the embodiment, first, a structured risk early warning rule library is constructed based on expert experience, power grid operation procedures and historical fault cases. Each rule item clearly contains five elements: risk score threshold interval, applicable device type, related operating environment condition, corresponding early warning level and response strategy template. Among them, the risk score threshold interval is divided into four levels from low to high, corresponding to normal, attention, warning and serious, respectively, and ensuring that each level interval is unique and does not overlap. The embodiment matches the operating risk score of each device node with the rule library, first judges the score interval to which the score belongs, and then further filters the adaptive rule items in combination with the device type and the area where it is located. If there are multiple rules that meet the conditions at the same time, the rule with a lower upper limit of the score interval is preferred to enhance the sensitive response ability to potential risks, so as to accurately determine the current early warning level of the node.

[0103] In step 700 of the embodiment, based on the foregoing risk level determination result and the generated response strategy, the system pushes the early warning level and the corresponding response strategy to the power grid dispatching platform and the operation and maintenance terminal. The dispatching platform can automatically update the operation situation atlas and the risk distribution view after receiving the early warning level, assisting the dispatching personnel to make plan adjustment or regional load optimization; the operation and maintenance terminal sends out a patrol work order, executes equipment load limiting, starts a protection strategy or sends out a maintenance instruction according to the pushed response strategy, and realizes the closed-loop response from alarm identification to disposal execution. Through the mechanism, the system barrier between the risk prediction result and the control execution is broken, the risk early warning is changed from static prompt to dynamic intervention, and the safety, response efficiency and intelligent level of the power grid operation are improved.

[0104] Corresponding to the above method, as shown in Figure 2 The embodiment also provides a power grid information system-based operation risk early warning system, which comprises:

[0105] A data processing unit is configured to collect dispatching plan data, equipment state data, operation monitoring data, historical fault record data and power grid topology structure data, form a multi-source heterogeneous original data set, and perform data cleaning, standardization and time synchronization processing on the original data set to obtain a risk analysis data set;

[0106] A feature map construction unit is configured to define nodes and their edge relationship based on the power grid topology structure data, map equipment operation indexes in the risk analysis data set to node features, map topological connectivity and power flow dependency relationship to edge features, and construct a risk perception feature map;

[0107] A graph neural network training unit is configured to train the risk perception feature map by using a graph neural network, and output a risk representation vector corresponding to each equipment node;

[0108] A time series risk prediction unit is configured to input the risk representation vector into a time series analysis model with a multi-scale attention mechanism, and predict an operation risk score of each equipment node within a preset time window in the future;

[0109] A high-risk area identification unit is configured to dynamically identify a high-risk area and a potential risk diffusion path based on the risk score result and the power grid topology structure data by using an algorithm combining density-peak clustering and hierarchical clustering;

[0110] An early warning strategy generation unit is configured to establish a risk early warning rule library, match the operation risk score of each equipment node with the rule library, trigger a normal, attention, early warning or serious four-level early warning level according to a matching result, and generate a corresponding response strategy.

[0111] The beneficial effects of the present application are as follows:

[0112] (1) This invention constructs for the first time a power grid risk perception model with graph neural network as the core. It can achieve deep modeling and accurate characterization of risk correlation relationships in complex networks based on the integration of power grid topology and equipment operating status, and significantly improves the ability to identify potential risk patterns in high-dimensional, multi-source operating data.

[0113] (2) The present invention introduces a time series prediction model that integrates a multi-scale attention mechanism, supports parallel modeling of risk evolution trends at different time granularities, effectively captures the dynamic characteristics of the coexistence of sudden disturbances and slow-changing trends, has good predictive foresight and adaptability, and meets the requirements for refined risk trend identification in high-frequency fluctuation scenarios.

[0114] (3) The present invention adopts a regional identification algorithm that combines density-peak clustering with hierarchical clustering, which not only ensures the precise positioning of high-risk areas, but also realizes the dynamic identification of potential risk diffusion paths, breaking through the limitation of traditional static regional division that is poorly adaptable to the risk propagation laws in complex topological scenarios.

[0115] (4) The present invention constructs a rule-driven risk warning level classification mechanism, combines the scoring range, equipment type and environmental conditions to realize the automatic matching and push of warning response strategies, and fully opens up the closed-loop chain of "risk perception-level determination-strategy execution", which significantly improves the intelligence level and response time of power grid operation risk management.

[0116] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0117] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. An operation risk early warning method based on a power grid information system, characterized in that: include: Collect scheduling data, equipment status data, operation monitoring data, historical fault record data, and power grid topology data to form a multi-source heterogeneous raw data set, and perform data cleaning, standardization, and time synchronization on the raw data set to obtain a risk analysis data set; Defining nodes and their edge relationships using the power grid topology data, mapping equipment operation indicators in the risk analysis dataset into node features, and mapping topological connectivity and power flow dependency into edge features to construct a risk perception feature graph; Using a graph neural network to train the risk perception feature graph, and outputting a risk representation vector corresponding to each device node; The risk representation vector is input into a time series analysis model integrated with a multi-scale attention mechanism to predict the operation risk score of each device node within a preset time window in the future; Based on the risk scoring results and the grid topology data, a density-peak clustering combined with a hierarchical clustering algorithm is used to dynamically identify high-risk areas and potential risk diffusion paths; Establish a risk warning rule base, match the operation risk score of each device node with the rule base, and trigger the four warning levels of normal, attention, warning or severe according to the matching results, and generate corresponding response strategies.

2. The operation risk early warning method based on the power grid information system according to claim 1 is characterized in that: Also includes: The warning level and the corresponding response strategy are pushed to the power grid dispatching platform and operation and maintenance terminal to achieve auxiliary decision-making and automatic handling of operation risks.

3. The operation risk early warning method based on the power grid information system according to claim 1 is characterized in that: The nodes and their edge relationships are defined using the power grid topology data, the equipment operation indicators in the risk analysis data set are mapped into node features, and the topological connectivity and power flow dependency relationships are mapped into edge features to construct a risk perception feature graph, including: Based on the power grid topology data, extract the unique identifier and electrical connection relationship of each device, define the nodes and edges in the graph structure, and construct an initial power grid device topology graph; Aligning the equipment operating indicators corresponding to each node in the risk analysis data set, extracting real-time current, voltage, active power, operating temperature, and the number of historical faults, and encoding the real-time current, voltage, active power, operating temperature, and the number of historical faults into a node feature vector; Extracting edge attribute data based on the flow direction, electrical connection type, and historical linkage fault information between adjacent nodes in the initial power grid device topology diagram, and encoding the attribute data into an edge feature vector; The node feature vector and the edge feature vector are embedded into the initial power grid equipment topology map to construct the risk perception feature map containing structural information and operation attribute information.

4. The operation risk early warning method based on the power grid information system according to claim 3 is characterized in that: Extracting edge attribute data based on the flow direction, electrical connection type, and historical linkage fault information between adjacent nodes in the initial power grid device topology diagram, and encoding the attribute data into an edge feature vector, including: Read the real-time power flow direction data corresponding to the adjacent nodes. If the real-time power flow direction data is flowing from the first node to the second node, set the power flow direction flag to +1; otherwise, set it to -1; Generate an electrical connection type code value according to a preset mapping table based on the line voltage level, conductor cross-section and switch topology state between the adjacent nodes; Retrieve the fault events containing the adjacent nodes from the historical fault records, count the number of occurrences of the linked faults, and perform proportional normalization using the number of faults in the total observation period as the denominator to obtain a linked fault probability value; Feature splicing is performed in the order of the power flow direction identifier, the electrical connection type code value, and the linkage failure probability value to form the edge feature vector.

5. The operation risk early warning method based on the power grid information system according to claim 1 is characterized in that: The risk perception feature graph is trained using a graph neural network to output a risk representation vector corresponding to each device node, including: Assume that the initial node feature matrix is ​​F (0) , the adjacency matrix of the edge connection relationship is A, the identity matrix is ​​I, and in each graph convolution layer l, the node feature update is calculated as follows: in is the adjacency matrix after adding self-loops, for The degree matrix, W (l) is the trainable weight matrix of the lth layer, σ(·) is the nonlinear activation function; In the training sample, each node in the graph is assigned a risk level label Y within the historical observation time window i ∈{low,medium,high}, for supervised learning; When the network propagates forward to the last layer L, the node output matrix Z = F is obtained (L) W (out) , for each node i the output vector Z i,* The predicted probability of belonging to risk level c is calculated using the following formula: Constructing cross entropy loss Where N is the total number of nodes participating in the training, 1(·) is the indicator function; c′ represents traversing all possible risk level categories in the normalized denominator; For the loss function Find the gradient and use iterative optimization algorithm to update all weight matrices W (l) and W (out) , until Convergence or reaching the preset number of iterations; After training is completed, the final node output matrix F (L) The row vector corresponding to each node in is fixed as the risk characterization vector of the node.

6. The operation risk early warning method based on the power grid information system according to claim 1 is characterized in that: The risk representation vector is input into a time series analysis model integrated with a multi-scale attention mechanism to predict the operational risk score of each device node within a preset future time window, including: Based on the risk characterization vector, the input sequence is organized according to the device node number and continuous time step. Each of these Represents the risk characterization vector at time t, T is the length of historical time, and d is the characteristic dimension; The sequence {x t } Input multiple sub-channels in parallel, each sub-channel corresponds to a different time window scale k∈{k1,k2,…,k m }, each scale corresponds to a local context modeling capability of different step lengths; where m represents the number of scales in the multi-scale attention mechanism At each scale k, the attention mechanism is used to weight the context of each time position t in the input sequence. The calculation formula is: in are the query and key vectors at time step t and its context position τ, respectively, represents the context window index set centered on t at scale k, represents the attention weight of position τ in scale k at time step t, τ ′ Represents the index variable used for normalized weights under the same window; The weighted results at all scales are concatenated and integrated into a unified temporal code z through linear mapping t , the formula is as follows: in is a value vector, Linear(·) is a linear fusion operation; The fused sequence Input the classification layer and output the operation risk score of each device node within the preset future time window Where Δ is the prediction step size; Using the historically known risk score r T+Δ To supervise the labels, construct the mean square error loss function And optimize the model parameters through back propagation until the loss function converges; Represents real risk, Represents predicted risk.

7. The operation risk early warning method based on the power grid information system according to claim 1 is characterized in that: Based on the risk scoring results and the grid topology data, a density-peak clustering combined with a hierarchical clustering algorithm is used to dynamically identify high-risk areas and potential risk diffusion paths, including: At each device node defined by the power grid topology data, the risk scores of the node and its directly adjacent nodes are aggregated to obtain the local risk density value of the node; Comparing the local risk density value of each node with the distance between it and the highest density node closest to it in topology, when a node has both a high local risk density and a distance to a node with a higher density reaches a minimum distance threshold, the node is marked as a risk peak node; Taking all the risk peak nodes as cluster centers, the surrounding nodes are absorbed in order from near to far along the topological shortest path to generate several initial high-risk clusters; For any two initial high-risk clusters, if the topological distance between their boundary nodes is not greater than the preset merging threshold, and the difference in the average risk scores between the two clusters is less than the preset difference threshold, they will be merged in a bottom-up manner until there is no cluster pair that meets the conditions. Each cluster obtained after hierarchical merging is defined as a high-risk area, and its central node, boundary node and cluster device list are recorded; According to the equipment connection direction and power flow direction in the power grid topology data, a recursive search is performed outward from the central node of the high-risk area, and a sequence of nodes adjacent to the identified high-risk area with continuously increasing risk scores is output to form a potential risk diffusion path.

8. The operation risk early warning method based on the power grid information system according to claim 7 is characterized in that: The judgment method for nodes with higher local risk density and whose distance to higher density nodes reaches the minimum distance threshold is as follows: Sort the local risk density values ​​of all nodes from large to small; Nodes whose local risk density values ​​are in the top 10% of the entire network distribution are selected as nodes with higher local risk density; For each node with a higher local risk density, calculate its shortest topological distance to all nodes with higher density; When the shortest topological distance is greater than a preset minimum distance threshold, the node is confirmed to be a risk peak node.

9. The operation risk early warning method based on the power grid information system according to claim 1 is characterized in that: Establish a risk warning rule base, match the operational risk score of each device node with the rule base, and trigger four warning levels of normal, concern, warning, or severe according to the matching results, and generate corresponding response strategies, including: Based on expert experience, grid operation procedures, and historical failure cases, several rules are formulated. Each rule includes: risk score threshold range, applicable equipment type, associated operating environment conditions, corresponding warning level, and response strategy template. The operational risk scores of device nodes are divided into four levels, corresponding to the four warning levels of normal, attention, warning, and severe. Each warning level has a unique and non-overlapping risk score range. For each device node, first identify the level interval to which the current risk score belongs, and then retrieve the matching rule items based on the node type and the area where it is located; If multiple rule items meet the matching conditions, the rule with the lower upper limit of the risk score range is prioritized to ensure the warning sensitivity and ultimately determine the warning level of the node; According to the determined warning level, the corresponding response strategy template is extracted from the rule item to generate the response strategy including alarm information push, patrol task dispatch, equipment load adjustment, protection device linkage or planned maintenance.

10. An operation risk early warning system based on a power grid information system, characterized in that: include: A data processing unit is used to collect scheduling plan data, equipment status data, operation monitoring data, historical fault record data, and power grid topology data to form a multi-source heterogeneous raw data set, and perform data cleaning, standardization, and time synchronization on the raw data set to obtain a risk analysis data set; a feature graph construction unit, configured to define nodes and their edge relationships using the power grid topology data, map equipment operation indicators in the risk analysis data set into node features, map topological connectivity and power flow dependency into edge features, and construct a risk perception feature graph; A graph neural network training unit is used to train the risk perception feature graph using a graph neural network and output a risk representation vector corresponding to each device node; A time series risk prediction unit, configured to input the risk representation vector into a time series analysis model integrated with a multi-scale attention mechanism to predict the operation risk score of each device node within a preset future time window; A high-risk area identification unit is used to dynamically identify high-risk areas and potential risk diffusion paths based on the risk scoring results and the power grid topology data by using a density-peak clustering combined with a hierarchical clustering algorithm; The warning strategy generation unit is used to establish a risk warning rule base, match the operation risk score of each device node with the rule base, and trigger the four warning levels of normal, concern, warning or severe according to the matching results, and generate corresponding response strategies.

Images