A power grid stability evaluation method, system, device and storage medium

Abstract

The application discloses a power grid stability evaluation method, system, device and storage medium, the method comprises the following steps: modeling an attribute graph based on power grid topology and operation data to obtain a set of expected scenarios; a graph neural network preventive safety evaluation model is constructed to evaluate all expected scenarios in the scenario set, determine high-risk scenarios and submit them to a physical simulation system for complete time domain simulation to obtain physical simulation calculation results; compare the evaluation results with the physical simulation calculation results scenario by scenario, and obtain high-value calibration samples according to the comparison results; based on the high-value calibration samples, incremental learning is performed on the model during off-peak periods of the power grid, and the original model is replaced with the updated model the next day. The application closes loop and fuses the rapid evaluation capability of the graph neural network and the high-precision calculation capability of the physical simulation system, realizes the power grid stability evaluation through the day-level closed-loop iterative mechanism of incremental calibration.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power system security and stability analysis and artificial intelligence. Specifically, it relates to a power grid stability assessment method, system, equipment and storage medium for online preventive security assessment and risk warning in power grid dispatch scenarios. Background Technology

[0002] With the continued advancement of the "dual carbon" target and the large-scale grid connection of high-proportion renewable energy sources, the power grid operation exhibits complex characteristics of strong randomness, high volatility, and deep coupling across multiple scenarios. Traditional offline safety and stability analysis models based on typical operating modes face the following challenges:

[0003] (1) The number of anticipated scenarios is enormous, and the full-scale time-domain simulation calculation time is unacceptable;

[0004] (2) The operation mode changes rapidly, and the offline analysis conclusions lag behind the actual operation status;

[0005] (3) The continuous access of new devices leads to frequent changes in topology and dynamic characteristics, making it difficult to fully cover historical experience.

[0006] In existing technologies, pure AI data-driven methods have fast inference speed but lack physical constraints and have insufficient interpretability; pure physical simulation methods have high accuracy but are computationally time-consuming and difficult to meet real-time operation requirements; and hybrid computation methods that combine AI simulation and physical simulation allow the AI ​​model and physical simulation to run independently without forming a continuous calibration relationship with the physical simulation system (PSDB).

[0007] In summary, the existing technology has not yet disclosed a comprehensive method that can establish a daily closed-loop iterative mechanism for rapid evaluation of graph neural networks and accurate verification of physical simulation systems. Summary of the Invention

[0008] The purpose of this invention is to provide a power grid stability assessment method, system, device, and storage medium. Without replacing the deployed physical simulation system, it deeply integrates the rapid assessment capability of graph neural networks with the high-precision computing capability of physical simulation systems in a closed loop. The graph neural network serves as a rapid preventive security assessment engine, and the physical simulation system serves as a high-precision benchmark truth source. Through a daily closed-loop iterative mechanism with incremental calibration, it achieves power grid stability assessment.

[0009] To achieve the above objectives, the solution of the present invention is:

[0010] A power grid stability assessment method, comprising,

[0011] Based on SCADA (Supervisory Control and Data Acquisition) data, the power grid topology and operational data are modeled as attribute graphs, and a set of anticipated scenarios is constructed based on these attribute graphs.

[0012] A graph neural network preventive security assessment model integrating edge features is constructed to assess all anticipated scenarios in the anticipated scenario set, and high-risk scenarios are determined based on the assessment results.

[0013] The high-risk scenario is submitted to the physical simulation system to perform a complete time-domain simulation, and the physical simulation calculation results are obtained.

[0014] The evaluation results are compared with the physical simulation calculation results for each scenario, and high-risk scenarios are selected based on the comparison results to obtain high-value calibration samples.

[0015] During off-peak hours of the power grid, incremental learning is performed on the graph neural network preventive safety assessment model based on the high-value calibration samples, and the updated model is used as the new graph neural network preventive safety assessment model the next day.

[0016] Among these methods, based on SCADA data, the power grid topology and operational data are modeled as attribute graphs, including:

[0017] SCADA data is acquired in real time from a unified data platform, and the power grid topology and operating status are modeled as attribute graphs. ,in, For a set of nodes, Let be the set of edges. The node feature matrix, The edge feature matrix; For a set of nodes The total number of bus nodes in the system; For the set of edges The total number of transmission branches in the system; The node feature dimension is the length of the feature vector of each bus node. is the edge feature dimension, that is, the feature vector length of each transmission branch.

[0018] The process of constructing a set of envisioned scenarios based on the attribute graph includes,

[0019] Define the set of envisioned scenarios Each anticipated scenario By triplet Description, in which Location of the fault. For fault clearance time, Fault type; among which The index represents the total number of anticipated scenarios. Number the scene. ;

[0020] Based on the power grid topology connections and operational data contained in the attribute graph, several anticipated scenarios are generated using N-1 basic scanning, operational risk enhancement, and / or historical risk backtracking methods, forming a set of anticipated scenarios. Among them, the N-1 basic scanning method refers to applying three-phase short-circuit faults to all operating lines and transformers one by one to obtain several scenarios; the operational risk enhancement method refers to superimposing N-2 combined faults and extreme operating conditions based on maintenance plans, new energy prediction fluctuation ranges, and load prediction results to obtain several scenarios; and the historical risk backtracking method refers to extracting typical fault scenarios that have triggered alarms from historical operating records.

[0021] This includes constructing a graph neural network-based preventative security assessment model that integrates edge features, and evaluating all anticipated scenarios in the anticipated scenario set, including...

[0022] The graph neural network preventive security assessment model that integrates edge features includes the following structure in sequence.

[0023] The input layer receives the power grid topology and operation data contained in the attribute graph, as well as the condition vectors of the expected scenarios in the expected scenario set;

[0024] The message passing layer contains multiple graph attention convolutional layers that fuse edge features. Each layer aggregates neighbor node information through an attention-weighted message passing mechanism and updates node feature representations layer by layer.

[0025] The graph-level readout layer aggregates the final feature vectors of all nodes into a fixed-dimensional global graph representation vector through mean readout operations.

[0026] The prediction layer concatenates the global graph representation vector with the condition vector of the anticipated scene, and then obtains the evaluation result through two independent multilayer perceptron prediction heads.

[0027] Among these, high-risk scenarios identified based on the assessment results include:

[0028] The evaluation result is a normalized stability margin estimate. Scenarios with a normalized stability margin estimate lower than the high-risk screening threshold are marked as high-risk scenarios.

[0029] Specifically, the evaluation results are compared with the physical simulation calculation results scene by scene. Based on the comparison results, high-risk scenarios are selected to obtain high-value calibration samples, including:

[0030] The evaluation results include the stability probability and normalized stability margin estimate for each high-risk scenario, and the physical simulation calculation results include the stability determination result and the baseline true value of the stability margin for each high-risk scenario.

[0031] If the binary classification result obtained based on the stability probability of a high-risk scenario is inconsistent with the stability judgment result of the high-risk scenario, or if the margin deviation between the normalized stability margin estimate and the true value of the stability margin benchmark exceeds the calibration sample screening threshold, the high-risk scenario will be regarded as a high-value calibration sample.

[0032] Among them, the binary classification result obtained based on stable probability includes,

[0033] stability probability pass The function rounds the result to obtain the binary classification result. .

[0034] Specifically, during off-peak hours of the power grid, incremental learning is performed on the graph neural network preventive safety assessment model based on the high-value calibration samples. The updated model is then used as the new graph neural network preventive safety assessment model the following day.

[0035] During off-peak hours of the power grid, all high-value calibration samples obtained on that day are merged with the cumulative samples within a recently set time range to obtain an incremental training set;

[0036] Let the incremental learning loss function be,

[0037]

[0038] in, For classifying losses, For margin regression loss, Strengthen the regularization term for flexible weights; For loss weights;

[0039] The graph neural network preventive safety assessment model is trained based on the incremental training set to obtain the updated model;

[0040] The updated model was used the following day as the new graph neural network preventive security assessment model.

[0041] A power grid stability assessment system, comprising,

[0042] The pre-conception scenario set construction module is configured to model the power grid topology and operation data into an attribute graph based on SCADA data, and construct a pre-conception scenario set based on the attribute graph;

[0043] The high-risk scenario acquisition module is configured to construct a graph neural network preventive security assessment model that integrates edge features, assess all the expected scenarios in the expected scenario set, and determine the high-risk scenarios based on the assessment results;

[0044] The physical simulation calculation result acquisition module is configured to submit the high-risk scenario to the physical simulation system to perform a complete time-domain simulation and obtain the physical simulation calculation results;

[0045] The high-value calibration sample acquisition module is configured to compare the evaluation results with the physical simulation calculation results scene by scene, and select high-risk scenarios to obtain high-value calibration samples based on the comparison results; and...

[0046] The incremental training and update module is configured to perform incremental learning on the graph neural network preventive safety assessment model based on the high-value calibration samples during off-peak hours of the power grid, and use the updated model as the new graph neural network preventive safety assessment model the next day.

[0047] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, it implements the steps of the power grid stability assessment method as described above.

[0048] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the steps of the power grid stability assessment method as described above.

[0049] By adopting the above scheme, this invention achieves continuous and autonomous improvement in the accuracy of online preventive safety assessment of the power grid through a daily closed-loop mechanism that uses graph neural networks to quickly predict and screen high-risk scenarios for all anticipated scenarios, performs physical simulation for accurate verification only for high-risk scenarios, and drives incremental calibration through virtual-real comparison. At the same time, it ensures the interpretability and traceability of the assessment results.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] (1) Closed-loop fusion rather than one-way substitution: A daily-level two-way closed-loop calibration mechanism between graph neural network and physical simulation system (PSDB) was established, so that the model evaluation accuracy can continuously and autonomously evolve with the changes in power grid operating conditions.

[0052] (2) Reuse of existing systems with zero modification: The physical simulation system (PSDB) can be directly reused as a high-precision reference truth source without modifying the existing production system.

[0053] (3) Graph neural network coarse screening greatly improves efficiency: after millisecond-level evaluation, only about 5% to 10% of high-risk scenarios are submitted to physical simulation for fine calculation, reducing the amount of physical simulation calculation by more than an order of magnitude.

[0054] (4) The evaluation results are interpretable and traceable: The final evaluation results of high-risk scenarios come from the deterministic calculation of physical simulation, which fundamentally solves the problem of the unreliability of data-driven methods.

[0055] (5) Incremental learning mechanism based on elastic weight consolidation to deal with distribution shift: while learning new knowledge, it maintains the ability to evaluate historical working conditions. Attached Figure Description

[0056] Figure 1 This is a flowchart of the present invention;

[0057] Figure 2 This is a flowchart of virtual-to-real comparison and error feedback. Detailed Implementation

[0058] The technical solution and beneficial effects of the present invention will be described in detail below with reference to the accompanying drawings.

[0059] This invention provides a power grid stability assessment method, comprising the following steps:

[0060] Step S1: Generate real-time anticipatory scenarios based on multi-source information: Integrate multi-source information such as power grid maintenance plans, power generation plans, new energy and load forecasts, and model the power grid topology and operating status as an attribute graph. Through three methods, namely N-1 basic scanning, operation risk enhancement and historical risk backtracking, the early warning scenario generator automatically generates a set of anticipatory scenarios.

[0061] Step S2: Rapid Prediction and High-Risk Scenario Labeling using Graph Neural Networks: A preventative safety assessment model based on graph neural networks that integrates edge features is constructed. This model integrates edge features when calculating the attention coefficient, enabling it to perceive the impact of line parameters and power flow status on topological adjacency relationships. The model takes the pre-fault section attribute map and scenario condition vector as input, performs millisecond-level rapid assessments of all anticipated scenarios, and outputs the stability probability and normalized stability margin estimate for each scenario. Scenarios with a normalized stability margin estimate lower than the high-risk screening threshold are labeled as high-risk scenarios, while the remaining low-risk scenarios directly output the graph neural network assessment results.

[0062] Step S3: Accurate verification by the physical simulation system: Submit the list of high-risk scenarios marked in step S2 to the physical simulation system to perform a complete time-domain simulation, obtain the stability judgment results and the stability margin benchmark true value, and output them as high-precision benchmark true values.

[0063] Step S4: Virtual-Real Comparison and Model Calibration: The graph neural network evaluation results in Step S2 are compared with the physical simulation results in Step S3 for each scenario. High-value calibration samples are selected and added to the incremental training sample library. At the same time, the evaluation results of the two are integrated to generate a graded risk assessment report with traceability.

[0064] Step S5: Closed-loop iteration: During off-peak hours of the power grid, incremental learning is performed on the graph neural network preventive safety assessment model based on the incremental training sample library accumulated in Step S4. The incremental learning loss function is...

[0065]

[0066] In the formula, For classifying losses, For margin regression loss, To reinforce the regularization term for flexible weights, , , These are the weighting coefficients; the updated model, after verification, will be deployed the following day to replace the current model.

[0067] Steps S1 to S5 constitute a daily iterative closed loop: Steps S1 to S4 are executed during the daily operation cycle to achieve online evaluation and virtual-real comparison. During off-peak hours, Step S5 is executed to complete model calibration. The calibrated model is launched online the next day, forming a continuous evolution mechanism of "scene generation - graph neural network prediction - physical simulation verification - virtual-real comparison - closed loop iteration".

[0068] In step S4, physical simulation results are used as the final evaluation basis for high-risk scenarios, while graph neural network evaluation results are used as the standard for low-risk scenarios, and "AI evaluation" label and confidence level are marked. All results are accompanied by full-link traceability information such as model version, training data coverage, and simulation parameters.

[0069] The following study uses a power grid in northern China as an example. The resulting attribute graph has approximately 1200 bus nodes and 1800 transmission branches. In this grid, renewable energy capacity accounts for over 44% of the total installed capacity. During the winter heating season, the system's regulation capacity is limited under the "heat-driven power generation" operation mode, leading to significant pressure on renewable energy consumption and power balance. This creates an urgent need for real-time and accurate online preventative safety assessments. The grid's dispatch center has deployed the PSDB physical simulation system for offline safety and stability analysis.

[0070] like Figure 1 As shown, the embodiments of the present invention include the following steps:

[0071] Step 1: Generate real-time anticipated scenarios based on multi-source information such as power grid maintenance plans, power generation plans, and new energy and load forecasts. Specific steps include:

[0072] Step 101: Obtain SCADA data in real time from the unified data platform, and model the power grid topology and operating status as an attribute graph, specifically as follows:

[0073] Modeling the power grid topology and operating state as a property graph ,in For the set of nodes (the mother line). For the edge set (transmission lines and transformers). The node feature matrix, The edge feature matrix; For a set of nodes The total number of bus nodes in the system; For the set of edges The total number of transmission branches in the system; The node feature dimension is the length of the feature vector of each bus node. is the edge feature dimension, that is, the feature vector length of each transmission branch.

[0074] Node feature matrix The The line is the first Feature vectors of each bus node subscript Number the busbar nodes. Node feature vectors Includes: voltage amplitude Voltage phase angle Active injection Reactive power injection Node type encoding Steady-state quantities of the pre-fault section; among which For load nodes (given active power) and reactive power ), For generator nodes (given active power) and voltage amplitude ), A balancing node (given voltage magnitude and voltage phase angle, used to balance the power deficit in the system).

[0075] Edge feature matrix Each row corresponds to a feature vector of a transmission branch. Suppose a branch connects to a bus node. and bus node (in , The bus node number is determined by the index in the node characteristics. (The meanings are the same, but different letters are used here to distinguish them), then the edge feature vector of this branch. Includes: line resistance Line reactance Line to ground susceptance Rated capacity of the line Current trend of merit and load rate ;in branch road Apparent power on Its amplitude; This is the rated transmission capacity of the branch.

[0076] Topological adjacency matrix It is updated in real time according to the power grid operation mode, reflecting the current topological connection relationship.

[0077] In this embodiment, the above-mentioned node feature dimensions Edge feature dimension The data update cycle is 5 minutes.

[0078] Step 102: Attribute graph constructed based on step 101 The included power grid topology connections and equipment parameter information, combined with maintenance plans, renewable energy forecasts, and other scheduling information, define a set of anticipated scenarios, specifically:

[0079] Define the set of envisioned scenarios Each anticipated scenario By triplet Description, in which The location of the fault (line or bus number). For fault clearance time, The fault type (three-phase short circuit, single-phase grounding, etc.); among which The index represents the total number of anticipated scenarios. Number the scene. .

[0080] The set of envisioned scenarios is automatically generated in the following ways:

[0081] (a) Basic N-1 scan: Apply three-phase short-circuit faults to all operating lines and transformers one by one, approximately 1800 scenarios;

[0082] (b) Enhanced operational risks: Based on maintenance plans, new energy source forecast fluctuation ranges, and load forecast results, plus N-2 combined faults and extreme operating conditions, approximately 2400 scenarios are considered;

[0083] (c) Historical risk retrospective: Extract typical fault scenarios that have triggered alarms from historical operation records, approximately 800 scenarios.

[0084] Total generated An anticipated scenario.

[0085] Step 2: Rapid prediction and high-risk scene labeling using graph neural networks. Specific steps include:

[0086] Step 201: Construct a graph neural network preventive safety assessment model that integrates edge features. In this embodiment, a graph attention network (GAT) is used as the specific implementation, with pre-fault cross-sectional data and scene condition vectors as input:

[0087] Constructing a preventative security assessment model based on graph attention networks (GAT) For each anticipated scenario Conduct a rapid safety assessment under the current pre-failure section.

[0088] Model The overall structure includes the following parts:

[0089] Input layer: Receives the property graph constructed in step 101 Node feature matrix Edge feature matrix Topological adjacency relationships, and condition vectors for the current anticipated scenario. (Includes scenario parameter information such as fault location code, fault type code, and fault clearing time);

[0090] Message Passing Layer: Stack The graph attention convolutional layer that fuses edge features (see step 202) aggregates neighbor node information through an attention-weighted message passing mechanism in each layer and updates the node feature representation layer by layer.

[0091] Level readout layer: via After layer message passing, the final feature vectors of all nodes are aggregated into a fixed-dimensional global graph representation vector through the mean read operation. (See step 203 for details);

[0092] Prediction layer: With condition vector After being stitched together, the data is fed into two independent multilayer perceptron (MLP) prediction heads—the classification prediction head outputs a stable probability. Margin prediction head output normalized stable margin estimation The core innovation of this architecture lies in the fact that the calculation of the attention coefficient in step 202 incorporates edge features, enabling the model to perceive the differentiated impact of line parameters and power flow state on topological adjacency relationships.

[0093] Step 202: The model's... The layer message passing mechanism is as follows:

[0094] Attention coefficient calculation:

[0095]

[0096] Node feature update:

[0097]

[0098] in, For the first After layer message passing, the node eigenvectors, For the first After layer message passing, the node (node The feature vectors of the neighboring nodes, (That is, the features of the nodes in layer 0 are the input node feature vectors). For the first Layer nodes For nodes Attention coefficient, representing the node Information for nodes Update contribution weight; For the linear transformation matrix of node features, For the edge characteristic linear transformation matrix, The attention weight vector consists of three learnable parameters. For nodes The neighborhood group, This is a vector concatenation operation. The activation function (e.g., LeakyReLU). The summation variable in the denominator of the attention coefficient formula. For traversing nodes The summation index of all neighbors, corresponding to the scene number in the previous text. The meanings are different. (The first...) Layer node features By the Characteristics of all neighbors of the floor It is obtained through attention-weighted aggregation, that is, the output of the previous layer is used as input for each layer, and the node representation is updated layer by layer.

[0099] Step 203: After passing through a multi-layer graph attention network, obtain the global representation through graph-level readout operations:

[0100]

[0101] in The global graph representation vector is an aggregated representation of the final features of all nodes in the graph. For the After layer message passing, the node The final feature vector; The total number of layers in the graph attention network; For graph-level readout functions, this embodiment uses mean pooling, which means taking the arithmetic mean of the final feature vectors of all nodes in the graph.

[0102] The global representation and the scene conditional vector are concatenated and then fed into a dual-head prediction network.

[0103]

[0104]

[0105] in The stable probability is obtained by processing with a probability mapping activation function such as Sigmoid. This is a normalized stability margin estimate. For the first A conditional vector for a pre-defined scenario, containing scenario parameter information such as fault location code, fault type code, and fault clearing time; For stable probabilistic prediction heads, multilayer perceptron networks A multilayer perceptron network with a margin prediction head.

[0106] Step 204: In this embodiment, the model performs parallel inference on all 5000 scenarios on a domestically produced computing power cluster, taking approximately 500ms. A high-risk screening threshold is set. Normalized stability margin estimation Below The scenarios were marked as high-risk scenarios, and about 400 high-risk scenarios (about 8% of the total) were selected. The remaining 4,600 low-risk scenarios were directly output as graph neural network evaluation results.

[0107] It should be noted that the model also outputs the stability probability. and normalized stability margin estimation The two indicators play different roles in subsequent processes. Normalized stability margin estimation. The high-risk scenario screening used in this step and the margin mean square error calculation in step 402 reflect the quantitative distance of the scenario from the stability boundary; stability probability Used for calculating the consistency rate of stability classification in step 402, through After rounding to the nearest integer, the classification result is determined as stable / unstable, etc., and then compared with the stability determination result of the physical simulation system. The two indicators describe the safety status of the scene from different dimensions and jointly serve the subsequent virtual-real comparison and incremental calibration.

[0108] Step 3: Accurate Verification of the Physical Simulation System

[0109] Step 301: Submit the 400 high-risk scenarios marked in Step 2 to the PSDB physics simulation system. PSDB uses 50 computing nodes deployed in parallel to perform a complete time-domain simulation calculation for each high-risk scenario (simulation duration 10 seconds), with a single scenario calculation time of approximately 50 seconds. With 50 nodes computing in parallel, the 400 scenarios are completed in approximately 7 minutes, obtaining the stability assessment results and stability margin baseline true values ​​for each high-risk scenario.

[0110] Step 4: Virtual-to-real comparison and model calibration. This step is only for the high-risk scene set marked in Step 2. (Approximately 400 scenarios) are compared scenario by scenario. Low-risk scenarios, which have not been calculated by the physical simulation system, are not included in this comparison step.

[0111] The specific steps include:

[0112] Step 401: As Figure 2 The process shown compares the graph neural network prediction results from step 2 with the physical simulation calculation results from step 3 scene by scene to construct an error sample set:

[0113]

[0114] in The stable probability output in step 2 ( (interval continuous values) The stability determination result output in step 3 is a binary result, where 1 indicates stability and 0 indicates instability. The normalized stability margin estimate output from step 2 ( (interval continuous values) This is the baseline true value (continuous value) of the stability margin output from step 3. Stability probability. With stability determination results The comparison method is as follows: pass Function rounding to convert to binary classification result , and then with Perform a consistency comparison.

[0115] Step 402: Calculate the following error index:

[0116] Stability classification error rate:

[0117]

[0118] Mean squared error of margin estimation:

[0119]

[0120] In this embodiment, the classification consistency rate of 400 high-risk scenarios is 97.8% (approximately 9 scenarios are inconsistent in classification), and the margin mean square error is 0.018.

[0121] Step 403: Screening for high-value calibration samples: Nine scenarios with inconsistent classifications and margin deviations exceeding the calibration sample screening threshold are excluded. Approximately 35 scenarios (after deduplication) from 26 scenarios (based on the 90th percentile of the historical error distribution) were added to the incremental training sample library. It should be noted that the calibration sample selection threshold here... Compared with the high-risk screening threshold in step 204 These are two different thresholds: Application to normalized stability margin estimation This is used to classify high-risk / low-risk scenarios; The absolute deviation between the GNN margin estimate and the PSDB margin. They are used to screen high-value calibration samples. They differ in physical meaning, numerical value, and target application.

[0122] Step 404: Integrate the graph neural network evaluation results with the physical simulation results to output the final risk assessment report. For 400 high-risk scenarios, the PSDB results are used as the standard; for 4600 low-risk scenarios, the graph neural network results are used as the standard, labeled with "AI Assessment" and confidence level. Risk levels are categorized from high to low based on the stability margin benchmark true value and stability determination results: red (Level 1), orange (Level 2), yellow (Level 3), and blue (Level 4). Specific thresholds are dynamically configured according to the power grid operation regulations. All assessment results include full-link traceability information such as model version, training data coverage, and simulation parameters.

[0123] Step 5: Closed-loop iteration, the specific steps include:

[0124] Step 501: Incremental training starts at 2:30 AM. The 35 newly added high-value calibration samples from the current day are merged with the accumulated high-value calibration samples from the past 7 days (approximately 250 samples) to form the incremental training set. The incremental learning loss function is:

[0125]

[0126] in:

[0127] Classification loss (cross-entropy loss):

[0128]

[0129] Marginal regression loss (mean squared error loss):

[0130]

[0131] Elastic Weight Consolidation (EWC) regularization term:

[0132]

[0133] in These are the current model parameters. These are the parameters for the previous version of the model; The diagonal approximation of the Fisher information matrix measures the parameters. The importance of historical tasks This is the incremental training set.

[0134] In this embodiment, the loss weight is set as It is 1.0. 0.5 With a learning rate of 0.1 and a learning rate of 1e-4, training for 25 epochs takes approximately 12 minutes.

[0135] Step 502: After training, the accuracy of the new and old models was compared on the reserved validation set. The classification accuracy of the new model improved from 97.8% to 98.3%, and the mean squared error margin decreased from 0.018 to 0.016. After offline validation and online canary release, it was launched the next day.

[0136] After 30 days of closed-loop iterative operation, the graph neural network model constructed in step 201... The accuracy rate of the online preventive safety assessment classification of the power grid remained stable at over 98.5%, and the mean square error of the margin estimation converged to below 0.015, verifying the effectiveness of the method of the present invention.

[0137] This invention also provides a power grid stability assessment system, comprising:

[0138] The pre-conception scenario set construction module is configured to model the power grid topology and operation data into an attribute graph based on SCADA data, and construct a pre-conception scenario set based on the attribute graph;

[0139] The high-risk scenario acquisition module is configured to construct a graph neural network preventive security assessment model that integrates edge features, assess all the expected scenarios in the expected scenario set, and determine the high-risk scenarios based on the assessment results;

[0140] The physical simulation calculation result acquisition module is configured to submit the high-risk scenario to the physical simulation system to perform a complete time-domain simulation and obtain the physical simulation calculation results;

[0141] The high-value calibration sample acquisition module is configured to compare the evaluation results with the physical simulation calculation results scene by scene, and select high-risk scenarios to obtain high-value calibration samples based on the comparison results; and...

[0142] The incremental training and update module is configured to perform incremental learning on the graph neural network preventive safety assessment model based on the high-value calibration samples during off-peak hours of the power grid, and use the updated model as the new graph neural network preventive safety assessment model the next day.

[0143] Each module implements the content of steps 1-5 mentioned above, which will not be repeated here.

[0144] This invention also provides another computer device, including a processor and a memory configured to store a computer program capable of running on the processor; wherein, when the processor is configured to run the computer program, it performs the method steps described in the foregoing embodiments.

[0145] In an exemplary embodiment, the present invention also provides a computer-readable storage medium for storing a computer program.

[0146] Optionally, the computer-readable storage medium can be applied to any of the methods in the embodiments of the present invention, and the computer program causes the computer to execute the corresponding processes implemented by the processor in the various methods of the embodiments of the present invention. For the sake of brevity, these will not be described in detail here.

[0147] In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

Claims

1. A method for assessing power grid stability, characterized in that: include, Based on SCADA data, the power grid topology and operation data are modeled as attribute graphs, and a set of anticipated scenarios is constructed based on the attribute graphs. A graph neural network preventive security assessment model integrating edge features is constructed to assess all anticipated scenarios in the anticipated scenario set, and high-risk scenarios are determined based on the assessment results. The high-risk scenario is submitted to the physical simulation system to perform a complete time-domain simulation, and the physical simulation calculation results are obtained. The evaluation results are compared with the physical simulation calculation results for each scenario, and high-risk scenarios are selected based on the comparison results to obtain high-value calibration samples. During off-peak hours of the power grid, incremental learning is performed on the graph neural network preventive safety assessment model based on the high-value calibration samples, and the updated model is used as the new graph neural network preventive safety assessment model the next day.

2. The method as described in claim 1, characterized in that: Based on SCADA data, the power grid topology and operational data are modeled as a property graph, including, SCADA data is acquired in real time from a unified data platform, and the power grid topology and operating status are modeled as attribute graphs. ,in, For a set of nodes, Let be the set of edges. The node feature matrix, The edge feature matrix; For a set of nodes The total number of bus nodes in the system; For the set of edges The total number of transmission branches in the system; The node feature dimension is the length of the feature vector of each bus node. is the edge feature dimension, that is, the feature vector length of each transmission branch.

3. The method as described in claim 1, characterized in that: Based on the attribute graph, a set of envisioned scenarios is constructed, including: Define the set of envisioned scenarios Each anticipated scenario By triplet Description, in which Location of the fault. For fault clearing time, The fault type; among which The index represents the total number of anticipated scenarios. Number the scene. ; Based on the power grid topology connections and operational data contained in the attribute graph, several anticipated scenarios are generated using N-1 basic scanning, operational risk enhancement, and / or historical risk backtracking methods, forming a set of anticipated scenarios. Among them, the N-1 basic scanning method refers to applying three-phase short-circuit faults to all operating lines and transformers one by one to obtain several scenarios; the operational risk enhancement method refers to superimposing N-2 combined faults and extreme operating conditions based on maintenance plans, new energy prediction fluctuation ranges, and load prediction results to obtain several scenarios; and the historical risk backtracking method refers to extracting typical fault scenarios that have triggered alarms from historical operating records.

4. The method as described in claim 1, characterized in that: A graph neural network-based preventative security assessment model incorporating edge features is constructed to evaluate all anticipated scenarios in the aforementioned scenario set, including: The graph neural network preventive security assessment model that integrates edge features includes the following structure in sequence. The input layer receives the power grid topology and operation data contained in the attribute graph, as well as the condition vectors of the expected scenarios in the expected scenario set; The message passing layer contains multiple graph attention convolutional layers that fuse edge features. Each layer aggregates neighbor node information through an attention-weighted message passing mechanism and updates node feature representations layer by layer. The graph-level readout layer aggregates the final feature vectors of all nodes into a fixed-dimensional global graph representation vector through mean readout operations. The prediction layer concatenates the global graph representation vector with the condition vector of the anticipated scene, and then obtains the evaluation result through two independent multilayer perceptron prediction heads.

5. The method as described in claim 1, characterized in that: High-risk scenarios were identified based on the assessment results, including: The evaluation result is a normalized stability margin estimate. Scenarios with a normalized stability margin estimate lower than the high-risk screening threshold are marked as high-risk scenarios.

6. The method as described in claim 1, characterized in that: The evaluation results are compared with the physical simulation calculation results for each scenario. Based on the comparison results, high-risk scenarios are selected to obtain high-value calibration samples, including: The evaluation results include the stability probability and normalized stability margin estimate for each high-risk scenario, and the physical simulation calculation results include the stability determination result and the baseline true value of the stability margin for each high-risk scenario. If the binary classification result obtained based on the stability probability of a high-risk scenario is inconsistent with the stability judgment result of the high-risk scenario, or if the margin deviation between the normalized stability margin estimate and the true value of the stability margin benchmark exceeds the calibration sample screening threshold, the high-risk scenario will be regarded as a high-value calibration sample.

7. The method as described in claim 6, characterized in that: The binary classification result is obtained based on the stable probability, including: stability probability pass The function rounds the result to obtain the binary classification result. .

8. The method as described in claim 1, characterized in that: During off-peak hours of the power grid, incremental learning is performed on the graph neural network preventive safety assessment model based on the high-value calibration samples. The updated model is then used as the new graph neural network preventive safety assessment model the following day, including... During off-peak hours of the power grid, all high-value calibration samples obtained on that day are merged with the cumulative samples within a recently set time range to obtain an incremental training set; Let the incremental learning loss function be, ， in, For classifying losses, For margin regression loss, Strengthen the regularization term for flexible weights; For loss weights; The graph neural network preventive safety assessment model is trained based on the incremental training set to obtain the updated model; The updated model was used the following day as the new graph neural network preventive security assessment model.

9. A power grid stability assessment system, characterized in that: include, The pre-conception scenario set construction module is configured to model the power grid topology and operation data into an attribute graph based on SCADA data, and construct a pre-conception scenario set based on the attribute graph; The high-risk scenario acquisition module is configured to construct a graph neural network preventive security assessment model that integrates edge features, assess all the expected scenarios in the expected scenario set, and determine the high-risk scenarios based on the assessment results; The physical simulation calculation result acquisition module is configured to submit the high-risk scenario to the physical simulation system to perform a complete time-domain simulation and obtain the physical simulation calculation results; The high-value calibration sample acquisition module is configured to compare the evaluation results with the physical simulation calculation results scene by scene, and select high-risk scenarios to obtain high-value calibration samples based on the comparison results; and... The incremental training and update module is configured to perform incremental learning on the graph neural network preventive safety assessment model based on the high-value calibration samples during off-peak hours of the power grid, and use the updated model as the new graph neural network preventive safety assessment model the next day.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that: When the processor executes the computer program, it implements the steps of the power grid stability assessment method as described in any one of claims 1 to 8.

11. A computer-readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, it implements the steps of the power grid stability assessment method as described in any one of claims 1 to 8.

Images