Power grid probability equivalent model construction method and device based on multi-scene clustering

Abstract

The invention relates to the technical field of data processing, in particular to a power grid probability equivalent model construction method and device based on multi-scene clustering, and the method comprises the steps: obtaining a power grid multi-dimensional data source, executing a spatial-temporal feature decoupling operation, and outputting a spatial-temporal feature tensor with a weight identifier; executing graph-driven enhanced clustering operation based on the tensor, constructing a topological risk graph structure, and generating a safety margin offset thermodynamic diagram through a deep reinforcement learning decision feature compression strategy; executing a double-algorithm optimization traceability operation, adopting a genetic algorithm to encode a feature cutting operation, taking a margin deviation value as a fitness function, optimizing a time window parameter by a butterfly algorithm, and outputting a feature operation traceability matrix; and feeding back the matrix to the enhanced clustering operation to update the decision rule. According to the method, the problem of dynamic security risk quantification distortion caused by scene clustering dimension reduction is solved under the constraint of limited computing resources through risk perception feature processing, enhanced resource allocation and a closed-loop architecture of double-algorithm traceability optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a power grid probability equivalent model construction method and device based on multi-scene clustering. BACKGROUND

[0002] The power grid probability equivalent model is used in power system engineering to cope with system uncertainty, such as load random fluctuation or device failure risk. Its construction process relies on probability theory to simplify the complex power grid into an equivalent structure, and to realize probability distribution modeling through a calculation simulation tool. The power grid probability equivalent model is used in large-scale data processing scenarios, such as power grid stability analysis. The reasoning logic includes uncertainty quantification, equivalent model generation, and computer-aided decision support, thereby improving the safety and efficiency of system planning.

[0003] In the multi-scene clustering construction process of the power grid probability equivalent model, the clustering algorithm is limited by the computational resource constraint and cannot simultaneously quantify the key dynamic security risk mapping accuracy in the dimension reduction scene. For example, the compression of the typical scene library may cause the loss of data features describing extreme working conditions such as superposition of typhoon and DC blocking. At the same time, there is a lack of standard method to verify the deviation of the equivalent model topology structure in representing the original power grid transient stability margin. In engineering practice, there is a multi-objective conflict. When the boundary scene coverage rate is improved, the calculation node dimension needs to be greatly increased. The amount of calculation required for real-time analysis exceeds the data throughput threshold of the dispatching decision system. Whether the model accurately retains the core security indicators such as voltage collapse critical point after reducing the complexity lacks quantitative criteria, forming a decision risk closed loop. The essential pain point is attributed to the fact that the existing method has not established a dynamic correlation rule base of boundary scene coverage, dynamic performance fidelity, and computational resource consumption. The system response capability becomes an unverifiable implicit parameter, which is specifically reflected in that if the equivalent model database fails to include the time sequence characteristics of a certain chain failure evolution, the simulation cannot restore the dynamic process of system oscillation out of step under the condition of limited computing power. Moreover, the decision interface lacks graphical tools to show the influence trajectory of the simplification operation on the credibility of the defense strategy, thereby weakening the accuracy of the preplan generation. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a power grid probability equivalent model construction method and device based on multi-scene clustering, which solves the dynamic security risk quantization distortion problem caused by scene clustering dimension reduction of the power grid probability equivalent model under the constraint of limited computing resources.

[0005] To solve the above technical problems, the specific technical solutions of the present application are as follows:

[0006] In a first aspect, the present application provides a power grid probability equivalent model construction method based on multi-scene clustering, comprising:

[0007] Obtaining a multi-dimensional data source of a power grid, the multi-dimensional data source including static topology, dynamic time sequence, and environmental parameters;

[0008] Performing a space-time feature decoupling operation on the multi-dimensional data source, separating time correlation features, space correlation features, and composite risk features, assigning each feature a weight identifier, and outputting a space-time feature tensor with a weight identifier;

[0009] Performing a graph-driven reinforcement clustering operation based on the space-time feature tensor, including: constructing a topology risk graph structure, selecting a feature compression strategy through deep reinforcement learning decision, and outputting a reduced dimension scene feature set containing a high-risk scene label;

[0010] Inputting the reduced dimension scene feature set into a hierarchical dynamic verification operation, performing differential verification according to the high-risk scene label, and generating a safety margin offset heat map;

[0011] Performing a double-algorithm optimization tracing operation on the safety margin offset heat map, including: genetic algorithm encoding feature pruning operation, butterfly algorithm optimizing time window parameters, and outputting a feature operation tracing matrix;

[0012] Feeding the feature operation tracing matrix back to the graph-driven reinforcement clustering operation to update the deep reinforcement learning decision rule.

[0013] Further, the power grid probability equivalent model construction method based on multi-scene clustering provided by the present application includes:

[0014] Adding a voltage recovery curve monotonicity identifier weight to the separated time correlation features;

[0015] Adding a power transfer key section weight to the separated space correlation features;

[0016] Adding an environmental mutation and device response coupling weight to the separated composite risk features.

[0017] Further, the power grid probability equivalent model construction method based on multi-scene clustering provided by the present application includes:

[0018] Analyzing the topology risk graph structure based on the space-time feature tensor, wherein:

[0019] Defining the node attribute as the bus voltage transient sensitivity;

[0020] Defining the edge weight as the fault propagation strength calculation value;

[0021] Based on the topology risk graph structure, responding to the real-time calculation resource occupancy rate, and performing a feature compression strategy selection;

[0022] Output the dimensionality-reduced scene feature set containing the high-risk scene label to the hierarchical dynamic verification operation.

[0023] Further, the multi-scene clustering-based power grid probabilistic equivalent model construction method, the execution feature compression strategy selection includes:

[0024] If the feature type is low risk, execute wavelet threshold dimensionality reduction operation;

[0025] If the feature contains a composite risk label, execute anti-aliasing protection sampling operation;

[0026] When the dimension of the computing node exceeds the preset threshold, perform clustering compression operation on the medium-risk features.

[0027] Further, the multi-scene clustering-based power grid probabilistic equivalent model construction method, the differential verification according to the high-risk scene label is executed in the following order:

[0028] According to the existence state of the high-risk scene label:

[0029] When it exists, execute full dynamic time domain simulation to obtain unit regulation response time series data;

[0030] When it does not exist, execute quasi-steady-state reduced-order verification to obtain static parameter cumulative error;

[0031] Perform spatial mapping operation on the unit regulation response time series data or the static parameter cumulative error to generate a three-dimensional color temperature layer.

[0032] Further, the multi-scene clustering-based power grid probabilistic equivalent model construction method, the three-dimensional color temperature layer generation performs the following mapping operations:

[0033] Convert the power angle deviation exceeding position to physical coordinates, and associate with the geographical coordinates of the generator group;

[0034] Match the voltage recovery delay position to the equipment identifier, and associate with the bus number of the substation;

[0035] Render the color temperature gradient on the geographical wiring diagram to identify the margin deviation percentage interval;

[0036] The topological risk graph structure and the three-dimensional color temperature layer cooperatively perform the following analysis operations:

[0037] Perform reverse mapping calculation on the power angle deviation exceeding coordinates to locate the topological edges with fault propagation intensity > 0.8;

[0038] Perform attribute association analysis on the voltage recovery delay position to match the node transient sensitivity attribute;

[0039] Perform risk level rendering calculation on the topological nodes within 50km range of the typhoon path, and update the three-dimensional color temperature layer.

[0040] Further, the power grid probability equivalence model construction method based on multi-scene clustering, the execution of the double algorithm optimization traceability operation includes:

[0041] The genetic algorithm is adopted to perform feature pruning coding, and the margin deviation value is the fitness function input;

[0042] The butterfly algorithm is adopted to perform time window parameter optimization;

[0043] The generation feature operation and the quantitative association path of the security error are generated.

[0044] Further, the power grid probability equivalence model construction method based on multi-scene clustering, the generation of the quantitative association path executes the following storage operation:

[0045] The quantitative error value marking operation is performed in the direct association path;

[0046] The risk propagation conditional probability weight recording operation is performed in the indirect association path;

[0047] The optimization compensation recording identification operation is performed in the high-impact area coordinate.

[0048] Further, the power grid probability equivalence model construction method based on multi-scene clustering, the update of the deep reinforcement learning decision rule executes the following judgment operation:

[0049] When the risk propagation conditional probability weight recording operation shows that the disconnection probability is distorted, the operation of adding an ice coating rate feature protection item to the reward function is performed;

[0050] When the quantitative error value marking operation shows that the deviation decreases by more than 50%, the operation of expanding the action space to increase the anti-aliasing sampling is performed.

[0051] Secondly, the present application provides a power grid probability equivalence model construction device based on multi-scene clustering, which is applied to the power grid probability equivalence model construction method based on multi-scene clustering and includes:

[0052] A data acquisition unit acquires a power grid multi-dimensional data source, and the multi-dimensional data source includes static topology, dynamic time sequence and environmental parameters;

[0053] A data processing unit performs a space-time feature decoupling operation on the multi-dimensional data source, separates time association features, space association features and composite risk features, assigns each feature a weight identifier, and outputs a space-time feature tensor with a weight identifier;

[0054] The clustering unit performs a graph-driven reinforcement clustering operation based on the spatiotemporal feature tensor, including: constructing a topological risk graph structure, deciding a feature compression strategy through deep reinforcement learning, and outputting a reduced dimension scene feature set containing a high-risk scene label;

[0055] The heat map generation unit inputs the reduced dimension scene feature set into a hierarchical dynamic verification operation, performs differential verification according to the high-risk scene label, and generates a safety margin offset heat map;

[0056] The optimization unit performs a double-algorithm optimization tracing operation on the safety margin offset heat map, including: a genetic algorithm encoding feature pruning operation, a butterfly algorithm optimizing a time window parameter, and outputting a feature operation tracing matrix;

[0057] The updating unit feeds back the feature operation tracing matrix to the graph-driven reinforcement clustering operation, and updates the deep reinforcement learning decision rule.

[0058] Advantages of the present application

[0059] The present application has the advantages that: BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained from the drawings without creative labor.

[0061] Figure 1 The flowchart of the power grid probability equivalent model construction method based on multi-scene clustering provided by the embodiments of the present application. DETAILED DESCRIPTION

[0062] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below in combination with specific embodiments of the present application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application. The technical solutions provided by each embodiment of the present application will be described in detail below in combination with the drawings. In order to better understand the objects of the present application, the present application will be further described in detail below.

[0063] In a first aspect, referring to Figure 1 The present application provides a power grid probability equivalence model construction method based on multi-scene clustering, comprising:

[0064] Step 1, obtaining a power grid multi-dimensional data source, the multi-dimensional data source comprising static topology, dynamic time sequence and environmental parameters;

[0065] Step 2, performing a space-time feature decoupling operation on the multi-dimensional data source, separating time correlation features, spatial correlation features and composite risk features, assigning each feature a weight identifier, and outputting a space-time feature tensor with a weight identifier;

[0066] Step 3, performing a graph-driven reinforcement clustering operation based on the space-time feature tensor, comprising: constructing a topological risk graph structure, determining a feature compression strategy through deep reinforcement learning, and outputting a reduced dimension scene feature set containing a high-risk scene label;

[0067] Step 4, inputting the reduced dimension scene feature set into a hierarchical dynamic verification operation, performing differential verification according to the high-risk scene label, and generating a safety margin offset heat map;

[0068] Step 5, performing a double-algorithm optimization tracing operation on the safety margin offset heat map, comprising: genetic algorithm encoding feature pruning operation, butterfly algorithm optimization time window parameter, and outputting a feature operation tracing matrix;

[0069] Step 6, feeding back the feature operation tracing matrix to the graph-driven reinforcement clustering operation, and updating the deep reinforcement learning decision rule.

[0070] In the operation of obtaining a power grid multi-dimensional data source, the static topology data contains power grid node connection relationship and device parameters, the dynamic time sequence data records voltage and current transient waveform, and the environmental parameters cover meteorological and geographical information. Through data cleaning and format conversion operation, the multi-source data time stamp is aligned and the dimension is standardized, providing structured input for feature decoupling.

[0071] When performing spatio-temporal feature decoupling on multi-dimensional data sources, the time correlation feature extracts the monotonicity of the post-fault voltage recovery curve, the spatial correlation feature locates the power transfer key section caused by DC blocking, and the compound risk feature fuses the wind speed sudden change event and the unit regulation lag time window. In the feature weight identification operation, the voltage recovery monotonicity identification weight is calibrated according to the historical fault database, the power transfer key section weight is calculated according to the power flow sensitivity, and the environmental mutation and device response coupling weight is quantified through the risk probability model.

[0072] When performing graph-driven reinforcement clustering operation based on spatio-temporal feature tensor, the topological risk graph structure construction process defines the node attribute as the bus voltage transient sensitivity trained by support vector machine, and the edge weight is mapped as the product of the outage probability and the power transfer amount in the typhoon passing area. The deep reinforcement learning decision feature compression strategy responds to the real-time CPU utilization of the dispatching system, and when the resource occupancy rate exceeds the threshold, the distribution network load curve dimension is automatically reduced to the typical mode.

[0073] When inputting the reduced dimension scene feature set into the hierarchical dynamic verification operation, the high-risk scene label triggers full dynamic time domain simulation, and the implicit trapezoidal integration method is used to solve the unit rotor motion equation; the scene without high-risk label activates the quasi-steady-state reduced order model, and the inter-provincial tie line power deviation is calculated by Thevenin equivalent. The spatial mapping operation converts the simulation results to geographic coordinates, the power angle deviation exceeds the position associated with the generator unit latitude and longitude data, the voltage recovery delay locates the substation bus number, and the color temperature gradient is distributed according to the red, yellow and green spectrum according to the deviation percentage interval.

[0074] When performing double-algorithm optimization tracing operation on the safety margin deviation heat map, the genetic algorithm encodes the wind power sampling rate compression operation as a chromosome, and the margin deviation value is input as the fitness function; the butterfly algorithm optimizes the pre-fault time window parameters, and avoids local optimization through quantum tunneling effect. The output quantifies the associated path record direct error value and indirect risk propagation weight, and the probability distortion of the broken line caused by the loss of icing rate features is marked as a high-weight path.

[0075] When feeding the feature operation trace matrix back to the graph-driven reinforcement clustering operation, the reinforcement learning decision rule is dynamically updated: a new icing rate feature protection item is added to the reward function, and the anti-aliasing sampling operation is expanded to the action space. The edge weight of the topological risk graph is recalibrated according to the color temperature warning area of the heat map, the risk level of the nodes within 50 kilometers of the typhoon path is raised to the highest level, forming a closed-loop optimization of risk quantification and resource scheduling.

[0076] Specifically, the power grid probabilistic equivalent model construction method based on multi-scene clustering provided by the application comprises:

[0077] The separated time correlation feature is subjected to voltage recovery curve monotonicity identification weight addition;

[0078] Perform power transfer key section weight addition on the separated space correlation features;

[0079] Perform environment mutation and device response coupling weight addition on the separated composite risk features.

[0080] When performing voltage recovery curve monotonicity identification weight addition operation on the separated time correlation features, analyze the voltage waveform change trend after the fault, and identify whether there is non-monotonic fluctuation phenomenon in the voltage recovery process. The voltage recovery curve monotonicity identification is generated according to the frequency of change of the derivative symbol, and is marked as a high weight feature when the derivative continuously maintains positive and negative consistency. The weight value is calibrated through historical fault data set, reflecting the stability difference of voltage recovery ability of different bus positions.

[0081] When performing power transfer key section weight addition operation on the separated space correlation features, locate the power flow redistribution path caused by DC blocking event, and identify the power transmission section in the power grid topology whose power transfer exceeds the preset threshold. The power transfer key section weight calculation is normalized by combining the line thermal stability limit parameter according to the section power flow sensitivity coefficient. The weight assignment reflects the influence degree of system safety margin of the section under N-1 fault.

[0082] When performing environment mutation and device response coupling weight addition operation on the separated composite risk features, an association model of meteorological parameter change and power equipment dynamic response is established. The environmental mutation factors include wind speed gradient change rate and line icing growth rate, and the device response factors include unit regulation rate and protection action delay. The coupling weight is quantified by joint probability distribution model, and the weight is automatically increased to the highest level in high-risk period such as typhoon passage period.

[0083] In the voltage recovery curve monotonicity identification weight addition operation, the time correlation feature extraction uses sliding time window algorithm, and the window length is dynamically adjusted according to the system inertia time constant. When it is detected that the voltage recovery curve exists multiple oscillations, the monotonicity identification weight is inversely adjusted according to the number of oscillations, which provides device regulation stability index for subsequent clustering decision.

[0084] In the power transfer key section weight addition operation, the spatial correlation feature positioning uses graph theory shortest path algorithm, and the optimal power transfer path after fault is calculated combined with line impedance parameters. The key section weight value is dynamically updated with the change of system operation mode, and the weight promotion coefficient is automatically corrected according to the day-ahead dispatching plan in the load peak period.

[0085] In the operation of coupling the environmental mutation and the device response weight, the composite risk feature fusion adopts a space-time convolutional neural network, and the timestamps of the weather radar data and the PMU measurement data are forcibly aligned. The coupling weight generation module outputs a risk probability heat map, and the weight of the device response lag feature within 50 kilometers of the typhoon path is increased to 3 times the baseline value.

[0086] Specifically, the power grid probability equivalence model construction method based on multi-scene clustering includes:

[0087] Based on the spatio-temporal feature tensor, the topology risk graph structure is analyzed, wherein:

[0088] The node attribute is defined as the bus voltage transient sensitivity;

[0089] The edge weight is defined as the fault propagation strength calculation value;

[0090] Based on the topology risk graph structure, the feature compression strategy selection is executed in response to the real-time calculation resource occupancy rate;

[0091] The reduced scene feature set containing the high-risk scene label is output to the hierarchical dynamic verification operation.

[0092] When analyzing the topology risk graph structure based on the spatio-temporal feature tensor, the node attribute definition operation adopts a support vector machine model to analyze historical fault data. The bus voltage transient sensitivity is calculated based on the first swing voltage drop depth and recovery rate after fault removal. The sensitivity value is mapped to the 0-1 interval, and the sensitivity of the main grid hub substation node is increased to more than 0.8. The edge weight definition operation integrates meteorological warning information and real-time operation data. The fault propagation strength calculation value integrates the line outage probability and the power transfer amount product factor. The edge weight of the power transmission channel within 50 kilometers of the typhoon path is automatically increased by 30%.

[0093] When executing the feature compression strategy selection in response to the real-time calculation resource occupancy rate, the scheduling system API periodically collects CPU utilization and memory occupancy data. The resource monitoring period is set to 200 milliseconds. When the calculation node dimension exceeds the preset threshold, the hierarchical compression strategy is triggered: the low-risk feature adopts wavelet threshold dimension reduction operation, and the components with energy percentage of more than 95% in the main frequency band are retained; the feature containing the composite risk label activates the anti-aliasing protection sampling mechanism, and the sampling frequency of the key time window before and after the fault is maintained at 10kHz.

[0094] In the deep reinforcement learning decision-making process, the state space contains the topology risk graph node sensitivity distribution and real-time resource occupancy. The action space is designed as a three-dimensional discrete structure: the first dimension selects the operation type, the second dimension sets the dimension reduction threshold or sampling frequency parameter, and the third dimension specifies the feature risk level. The reward function receives the margin offset thermal map data of the feedback of the hierarchical dynamic verification operation, and the voltage trajectory deviation coefficient and power angle deviation coefficient are weighted to calculate the immediate reward value.

[0095] The high-risk scenario label generation operation is based on the node sensitivity threshold and edge weight association rule. When the bus voltage transient sensitivity is greater than 0.7 and the associated edge fault propagation strength is greater than 0.8, it is marked as a high-risk scenario. When the high-risk label is output, the original data resolution of the high-risk label is preserved, and the feature dimension of the non-high-risk area is compressed to 30%-60% of the reference value. The compression ratio is dynamically optimized through reinforcement learning decision-making.

[0096] The topology risk graph structure updating mechanism and reinforcement learning decision-making form a closed loop, and the feature compression strategy selection result is fed back to the graph structure database in real time. During the typhoon passage, the node sensitivity weight of the affected area is dynamically corrected according to the meteorological warning level. The node sensitivity reference value of the red warning area is increased by 20%, which drives the subsequent feature compression strategy to preferentially protect the data integrity of this area.

[0097] Specifically, the power grid probability equivalence model construction method based on multi-scene clustering, the execution feature compression strategy selection includes:

[0098] If the feature type is low risk, execute wavelet threshold dimension reduction operation;

[0099] If the feature contains a composite risk label, execute anti-aliasing protection sampling operation;

[0100] When the node dimension exceeds the preset threshold, perform clustering compression operation on the medium-risk feature.

[0101] When executing the feature compression strategy selection, the judgment basis for low-risk features is the device abnormal record frequency in the historical fault database and the real-time monitoring alarm level. When the feature is classified as low risk, the wavelet threshold dimension reduction operation uses discrete wavelet transform to decompose the signal, selects the highest energy concentration frequency band component to reconstruct the data. The main energy distribution area is preserved during dimension reduction, and high-frequency noise components are actively filtered out, and non-critical data dimensions such as load fluctuation curves are compressed to 30% of the original value.

[0102] The feature with a composite risk label triggers an anti-aliasing protection sampling operation, identifying the spatio-temporal coupling area of environmental mutations and equipment failures. The protection sampling operation sets a critical time window of 2 seconds before failure to 1 second after failure, maintaining a millisecond-level sampling frequency during this period. The composite risk label is derived from the risk probability heat map output by the spatio-temporal convolutional neural network, and the feature dimension compression operation is automatically frozen when the risk probability exceeds the 0.8 threshold.

[0103] When the computing node dimension exceeds the preset threshold, real-time monitoring of the CPU utilization and memory occupancy of the scheduling system is performed. The preset threshold is dynamically adjusted according to the scheduling decision period, and the threshold is lowered to 75% of the system computing power upper limit during peak load periods. The K-means algorithm is used for clustering and compression of medium-risk features, with 3-5 typical operating modes set as the number of cluster centers, and power distribution network load curves and other medium-risk data are classified into the nearest neighbor center point.

[0104] In the wavelet threshold dimension reduction operation, the low-risk feature selection uses a sliding time window analysis algorithm, and the window length is adaptively adjusted according to the system fundamental period. The dimension reduction threshold is dynamically set according to the signal energy spectrum entropy value, and the retained components need to cover more than 95% of the energy distribution of the original signal. The reconstructed data retains the key change trend of voltage amplitude, and high-frequency harmonic components are actively discarded.

[0105] When the anti-aliasing protection sampling operation is executed, the length of the critical time window is dynamically configured according to the fault type. The sampling window is set to 2 seconds before and 1 second after for DC blocking faults, and to 1 second before and 2 seconds after for line short circuit faults. The sampling frequency is maintained at 10 kHz to meet the Nyquist theorem requirement, preventing spectral aliasing distortion in the transient process of unit regulation.

[0106] When the clustering compression operation is activated, the medium-risk feature selection is based on the device failure probability interval division, and features with a probability value in the 0.3-0.7 interval are classified as medium-risk. The K-means clustering initialization center point selects the extreme points of the historical typical daily load curve, and the Euclidean distance calculation gives higher weight to the peak and valley periods, and the compressed feature set retains the key dynamic characteristics of the typical operating modes.

[0107] Specifically, the power grid probabilistic equivalence model construction method based on multi-scene clustering, and the differential verification according to the high-risk scene label is performed in the following order:

[0108] According to the existence state of the high-risk scene label:

[0109] When it exists, perform full-dynamic time-domain simulation to obtain unit regulation response time series data;

[0110] When it does not exist, perform quasi-steady-state reduced-order verification to obtain static parameter cumulative error;

[0111] The space mapping operation is performed on the unit regulation response time sequence data or static parameter cumulative error amount to generate a three-dimensional color temperature layer.

[0112] The determination basis of the existence state of the high-risk scenario label is the correlation rule of the node sensitivity and the edge weight in the topology risk graph. When the bus voltage transient sensitivity is greater than 0.7 and the associated edge fault propagation strength is greater than 0.8, the high-risk scenario label state is marked as existing. At this time, the full dynamic time domain simulation operation is triggered, and the parallel differential algebraic equation solver is used to process the unit rotor motion equation and the network algebraic equation. The simulation process intercepts the key time window from 1.5 seconds before the fault to 8 seconds after the fault, accurately restores the action track of the excitation system limiting link and the response characteristic of the speed regulator, and outputs the millisecond-level unit regulation response time sequence data.

[0113] When the high-risk scenario label does not exist, the quasi-steady-state reduction verification operation is activated, and the power transmission network is simplified into a Thevenin equivalent voltage source-impedance model. The transient process of the generator stator winding is ignored, and the difference in key section power flow distribution is calculated. The static parameter cumulative error amount is obtained by comparing the short-circuit current amplitude of the equivalent model and the original system, and the sum of the absolute values of the inter-provincial tie-line power deviation is calculated to generate the error accumulation amount.

[0114] In the space mapping operation of the unit regulation response time sequence data or the static parameter cumulative error amount, the power angle deviation exceeding position is converted into the geographical coordinates of the generator group. The coordinate conversion is performed according to the longitude and latitude records of the units in the power system geographical information database, and the coordinate points with a deviation value exceeding 3 degrees are marked as red warning areas. The voltage recovery delay position is matched with the bus number of the substation, and the bus number with a delay exceeding 200 milliseconds is automatically associated with the node identifier of the dispatching system topology graph.

[0115] The three-dimensional color temperature layer generation operation superimposes a dynamic rendering layer on the power grid geographical connection diagram. The color temperature gradient is classified according to the percentage interval of the margin deviation: the area with a deviation greater than 10% is rendered in red spectrum, the area with a deviation of 5%-10% is mapped to yellow spectrum, and the area with a deviation less than 5% is displayed in green spectrum. The high-risk areas within 50 kilometers of the typhoon path are identified by pulse flicker, enhancing the visual warning effect.

[0116] In the calculation resource dynamic allocation mechanism of the full dynamic time domain simulation, if the simulation time exceeds 20% of the dispatching period, the quasi-steady-state reduction verification mode is automatically switched. The resource switching operation retains the key unit regulation response data, and the non-core area network parameters are simplified by the Thevenin equivalent method, balancing the calculation accuracy and real-time constraints.

[0117] Specifically, the power grid probabilistic equivalent model construction method based on multi-scenario clustering generates a three-dimensional color temperature layer by performing the following mapping operations:

[0118] Convert the power angle deviation exceeding position to physical coordinates, and associate the geographical coordinates of the generator group;

[0119] Match the voltage recovery delay position to the equipment identification, and associate the bus number of the substation;

[0120] Render the color temperature gradient on the geographical wiring diagram, and identify the margin deviation percentage interval;

[0121] The topology risk map structure cooperates with the three-dimensional color temperature layer to perform the following analysis operations:

[0122] Perform reverse mapping calculation on the power angle deviation exceeding coordinates to locate the topology edges with fault propagation strength > 0.8;

[0123] Perform attribute association analysis on the voltage recovery delay position to match the node transient sensitivity attribute;

[0124] Perform risk level rendering calculation on the topology nodes within 50km range of the typhoon path, and update the three-dimensional color temperature layer.

[0125] When converting the power angle deviation exceeding position to physical coordinates, the registered coordinate data of the generator set in the power system geographical information database is called. The coordinate points with power angle deviation exceeding 3 degrees are mapped to the longitude and latitude positions of the generator group, and the deviation value is classified according to the exceeding limit amplitude: 3-5 degrees marked yellow warning, 5 degrees and above marked red warning. The physical coordinate conversion uses Gauss-Kruger projection coordinate system, and the longitude and latitude data are forced to align with the space reference system of the dispatching automation system topology graph.

[0126] When matching the voltage recovery delay position to the equipment identification, the circuit breaker number and bus identification are associated according to the primary wiring diagram of the substation. The position with voltage recovery delay exceeding 200 milliseconds is located to the specific bus number, and the numbering rule follows the logical node naming specification in IEC 61850 standard. The equipment identification matching operation is associated with the dispatching system topology database, and the bus number of the 500kV hub substation is additionally weighted with priority coefficient.

[0127] When rendering the color temperature gradient on the geographical wiring diagram, the margin deviation percentage interval is divided into three spectral bands: deviation value 0-5% mapped to green spectrum, 5-10% mapped to yellow gradient band, and 10% and above marked with red warning. The color temperature rendering engine superimposes digital elevation model data, and the color temperature saturation of mountainous terrain area is increased by 20%, enhancing the visual recognition.

[0128] When performing reverse mapping calculation on the power angle deviation exceeding coordinates, the mapping relationship library between the power grid physical coordinate system and the topology graph edges is established. The topology edges with fault propagation strength > 0.8 are located using the spatial nearest neighbor algorithm, and the transmission lines within 50 kilometers of the generator group are preferentially marked. The topology edge risk thermodynamic diagram is rendered according to the propagation strength value, and the edges with strength value 0.8-1.0 are marked with red flashing identification.

[0129] When performing attribute correlation analysis on the voltage recovery delay position, the node transient sensitivity attribute value in the topology risk graph is extracted. The matching operation adopts an attribute similarity calculation model, and the topology nodes with a bus number correlation sensitivity > 0.7 and a delay of more than 200 milliseconds are matched. The attribute correlation result generates a node-bus mapping relationship table, and the deviation between the support vector machine model prediction value and the actual measurement is recorded in the correlation log.

[0130] When performing risk level rendering calculation on the topology nodes within 50 kilometers of the typhoon path, the typhoon center coordinates and moving path published by the meteorological warning platform are integrated. The risk level calculation introduces a distance attenuation factor, and the risk level of the nodes within 20 kilometers of the typhoon center is set to the highest level. The three-dimensional color temperature layer dynamic update mechanism responds to the change of the typhoon path, and the color temperature saturation degree of the risk level improvement area is increased by 30% in real time.

[0131] Specifically, the power grid probability equivalence model construction method based on multi-scene clustering, the execution of the double-algorithm optimization traceability operation includes:

[0132] The genetic algorithm is used to perform feature pruning coding, and the margin deviation value is the fitness function input;

[0133] The butterfly algorithm is used to perform time window parameter optimization;

[0134] Generate a quantitative association path between the feature operation and the safety error.

[0135] When the genetic algorithm is used to perform feature pruning coding operation, the chromosome structure design includes discrete gene sites and continuous gene sites. The discrete gene sites represent the wind power sampling rate compression ratio options, and the continuous gene sites define the fault time window length parameter range. The fitness function construction operation inputs the voltage trajectory deviation coefficient and the power angle deviation coefficient of the safety margin offset thermal map, and the weighted sum value is used as the chromosome evolution evaluation benchmark. The population initialization adopts the historical optimal solution injection mechanism, and the elite preservation strategy maintains the gene integrity of high fitness individuals.

[0136] When the butterfly algorithm is used to perform time window parameter optimization operation, the butterfly position vector maps the time boundary of the feature retention window before the fault. The smell concentration calculation is based on the fitness function output value, and the high fitness individual position emits a strong smell signal. The position update process introduces quantum tunneling effect, when the local optimal solution stagnates for more than a preset number of generations, it is forced to jump to explore the unknown parameter space. The time window optimization range is limited within the engineering safety constraints, the pre-fault window length is not less than 1 second, and the post-fault window length is not more than 5 seconds.

[0137] The generation feature operation is associated with the quantization of the error path operation, and the mapping relationship between the feature trimming parameter and the margin deviation value of the direct association path is recorded. The quantization error value is calculated by a multiple linear regression model, and the voltage deviation change caused by the unit amount of sampling rate compression ratio change is recorded as the slope coefficient. The indirect association path excavates the safety impact caused by the loss of composite risk feature, and the distortion of the line breaking probability caused by the missing of the icing rate feature is converted into a conditional probability weight factor. The weight value is calibrated according to the historical fault case library.

[0138] The crossover operation of the genetic algorithm adopts a double-point crossover strategy, and a separation zone is set between the wind power sampling rate gene position and the time window length gene position. The mutation operation introduces an adaptive mutation probability mechanism, and the mutation probability decreases to 30% of the initial value in the late stage of population convergence. The evolution stop condition is set to be that the optimal fitness variation amount is less than 0.1% for 20 consecutive generations, and the Pareto frontier solution set is output for decision selection.

[0139] The smell diffusion model of the butterfly algorithm follows the inverse square decay law, and when the distance between individuals exceeds the threshold, the smell concentration decays inversely with the square of the distance. In the group coordination mechanism, the position of the optimal individual is broadcast to the whole population every generation, guiding the local search direction. The convergence criterion combines the iteration number and the position variance double index, and the position coordinate standard deviation is less than 0.01 to determine the convergence.

[0140] The quantization association path storage architecture adopts a graph database modeling, and the nodes represent feature operation types and safety indicators. The weighted edges store direct error coefficients and indirect probability weights, and the high-impact area coordinates are associated with the spatial positions in the topology risk graph. The path visualization rendering superimposes the power grid geographical wiring diagram, and the areas where the feature trimming leads to a margin deviation exceeding 10% are marked with pulse warning signs.

[0141] Specifically, the method for constructing a power grid probabilistic equivalent model based on multi-scene clustering, the generation quantization association path performs the following storage operations:

[0142] In the direct association path, the quantization error value labeling operation is performed;

[0143] In the indirect association path, the risk propagation conditional probability weight recording operation is performed;

[0144] In the high-impact area coordinates, the optimization compensation recording and identification operation is performed.

[0145] In the direct association path, a multiple linear regression model is used to establish the mapping relationship between the feature trimming parameter and the safety margin deviation. The quantization error value labeling operation records the slope coefficient of the wind power sampling rate compression operation and the voltage trajectory deviation, and the deviation change amount corresponding to a 50% reduction in the sampling rate is written into the labeling log. The labeling data is stored in the edge attribute table of the graph database, and the time series database index records the operation timestamp and the power grid operation mode snapshot, supporting historical data backtracking analysis.

[0146] In the indirect association path, the risk propagation condition probability weight record operation is performed, and the association rules of the composite risk feature missing event and the broken line failure are extracted. The risk propagation conditional probability weight calculation adopts the Bayesian network model, and the posterior probability value of the broken line probability under the missing condition of the icing rate feature is converted into a weight factor. The weight record operation integrates the historical observation data of the meteorological station and the equipment failure record, and the rules with a probability value greater than 0.8 are marked as high-risk propagation paths and written into the node attribute table of the graph database.

[0147] In the high-impact area coordinate, the optimization compensation record identification operation is performed, and the power grid geographical area with a safety margin deviation exceeding 10% is located. The optimization compensation record identification operation is associated with the double-algorithm optimization result, and the area with a deviation reduction of more than 50% after the time window parameter tuning of the butterfly algorithm is marked with a blue compensation identification. The identification data storage adopts a spatial database engine, and the geographical coordinate point is additionally provided with an optimization contribution value percentage. The 500kV hub substation area identification size is expanded by 30%.

[0148] The storage structure of the quantitative error value marking operation adopts a key-value pair database, and the key name encodes the feature operation type and the time window parameter combination. The value field stores a multiple regression coefficient matrix, and the update mechanism responds to the optimization results of the butterfly algorithm. The parameter optimization is completed within 24 hours after the coefficient recalculation. The database index establishes an associated mapping between the operating mode and the seasonal mode, and the regression coefficients in the load peak period are stored independently.

[0149] The event collection period of the risk propagation conditional probability weight record operation is set to 15 minutes, synchronized with the data update of the meteorological warning platform. The Bayesian network parameter training adopts a sliding time window mechanism, and the weight of the failure event in the last 30 days is increased by 50%. The probability weight visualization adopts a heat spectrum map rendering, and the high-risk path (probability > 0.8) is additionally provided with a pulse warning animation.

[0150] The spatial aggregation algorithm of the optimization compensation record identification operation is processed according to voltage levels, and independent identification points are used in areas above 220kV. The identification rendering engine integrates digital elevation terrain data, and the brightness of the mountain area identification is increased by 40%. The identification click interaction triggers the optimization detail display, and the margin deviation comparison curve before and after the butterfly algorithm tuning is displayed.

[0151] Specifically, the power grid probability equivalence model construction method based on multi-scene clustering executes the following judgment operation when updating the deep reinforcement learning decision rule:

[0152] When the risk propagation conditional probability weight record operation displays a broken line probability distortion, the operation of adding an icing rate feature protection item to the reward function is performed;

[0153] When the execution of the quantization error value labeling operation shows a deviation drop of more than 50%, an extended action space anti-aliasing sampling operation is performed.

[0154] When the execution of the risk propagation conditional probability weight recording operation shows a broken line probability distortion, the association log of the composite risk feature loss event and the broken line fault is analyzed. The broken line probability distortion is determined according to the posterior probability offset output by the Bayesian network, and the feature protection mechanism is triggered when the difference between the historical benchmark probability value and the current probability value is more than 20%. The icing rate feature protection item is added to the operation of the reward function, and the protection item weight coefficient is set to the inverse of the absolute value of the offset, which actively improves the sampling priority of the icing rate feature in the reinforcement learning decision process.

[0155] When the execution of the quantization error value labeling operation shows a deviation drop of more than 50%, the error slope change data of the multivariate linear regression model is analyzed. The deviation drop amplitude is calculated based on the margin deviation comparison before and after the feature pruning operation, and a drop of more than 50% is determined as an effective optimization operation. When the extended action space anti-aliasing sampling operation is performed, the action dimension definition covering the fault time window length and the sampling frequency combination parameters are added, and the action parameter range is verified by engineering safety constraints.

[0156] The event monitoring period of the risk propagation conditional probability weight recording operation is set to 15 minutes, and the real-time data of the weather station is updated synchronously. When the broken line probability distortion event occurs, the icing rate feature protection item is injected into the reward function weight item, and the original function structure is adjusted to a weighted combination of voltage deviation coefficient 0.5, power angle deviation coefficient 0.3, and feature protection item 0.2. The reward function reconstruction operation takes effect in the next decision period, and the DQN agent policy network parameters are updated synchronously.

[0157] The data backtracking window of the quantization error value labeling operation covers the last 24 hours of optimization records, and the deviation drop amplitude calculation excludes the interference factors of load fluctuations. When the anti-aliasing sampling operation is extended to the action space, the parameter combination options are preset as the Cartesian product of the fault window length [1s, 3s, 5s] and the sampling frequency [5kHz, 10kHz]. The action options whose parameters exceed the device adjustment capability range are actively shielded by using the Thevenin equivalent precalculation.

[0158] The feature protection item addition operation triggers the reinforcement learning strategy network retraining, and the target network parameter freezing period is shortened to 50% of the original value. The experience replay pool preferentially samples historical state data containing the icing rate feature, and the proportion of high-risk feature samples in the training data set is increased to 40%. The policy gradient update process adds the weight bias derivative calculation of the feature protection item, which accelerates the convergence of the protection strategy.

[0159] After the action space expansion operation, the deep reinforcement learning decision response time maintains the 200ms period constraint. The initial value of the new action dimension uses the optimization result of the butterfly algorithm, and the time window parameter before the fault is set to the mean value of the last 10 optimizations by default. The action selection strategy retains the epsilon-greedy exploration mechanism, and the exploration probability of the new action dimension is initially increased to 0.3.

[0160] In a second aspect, the present application provides a power grid probabilistic equivalence model construction device based on multi-scene clustering, which is applied to the power grid probabilistic equivalence model construction method based on multi-scene clustering, and comprises:

[0161] A data acquisition unit acquires a power grid multi-dimensional data source, and the multi-dimensional data source comprises static topology, dynamic time sequence and environmental parameters.

[0162] A data processing unit performs a space-time feature decoupling operation on the multi-dimensional data source, separates time correlation features, space correlation features and composite risk features, assigns each feature a weight identifier, and outputs a space-time feature tensor with a weight identifier.

[0163] A clustering unit performs a graph-driven reinforcement clustering operation based on the space-time feature tensor, which comprises: constructing a topological risk graph structure, outputting a reduced dimension scene feature set containing a high-risk scene label through a deep reinforcement learning decision feature compression strategy.

[0164] A heat map generation unit inputs the reduced dimension scene feature set into a hierarchical dynamic verification operation, performs differential verification according to the high-risk scene label, and generates a safety margin offset heat map.

[0165] An optimization unit performs a double-algorithm optimization tracing operation on the safety margin offset heat map, which comprises: a genetic algorithm encoding feature pruning operation, a butterfly algorithm optimization time window parameter, and an output feature operation tracing matrix.

[0166] An updating unit feeds back the feature operation tracing matrix to the graph-driven reinforcement clustering operation, and updates the deep reinforcement learning decision rule.

[0167] The present application solves the problem of dynamic safety risk quantization distortion under limited resource constraints through the following technical path: first, a risk perception type feature processing mechanism is established, a space-time feature decoupling operation separates time correlation features, space correlation features and composite risk features, and assigns them different weight identifiers; a graph-driven reinforcement clustering operation constructs a topological risk graph structure, quantifies risk distribution based on bus voltage transient sensitivity and fault propagation intensity; a deep reinforcement learning decision response real-time resource occupation rate is calculated, anti-aliasing protection sampling is performed on high-risk features to maintain millisecond-level data integrity, wavelet threshold dimension reduction compression is used on low-risk features to compress redundant data, and when the resource is over-limit, the clustering compression is performed on medium-risk features, so that the key risk data is actively protected.

[0168] Secondly, a hierarchical dynamic verification closed loop is constructed, and differential verification is triggered according to high-risk scene labels: high-risk scenes perform full dynamic time domain simulation to obtain unit regulation response time sequence data, and non-high-risk scenes use quasi-steady-state reduced-order models to calculate static parameter cumulative error; the safety margin deviation heat map maps the over-standard position of the power angle deviation to the geographical coordinates of the generator group, the voltage recovery delay is associated with the bus number of the substation, and the color temperature gradient identifies the margin deviation interval, so that the model distortion position and degree are explicitly visualized.

[0169] Finally, a double-algorithm tracing optimization mechanism is formed: genetic algorithm encoding feature pruning operation, with heat map margin deviation value as fitness function; butterfly algorithm optimizes time window parameters before failure, and quantum behavior is used to jump out of local optimum; the generated feature operation and the quantitative association path of safety error are associated, when the risk propagation conditional probability weight record shows that the probability of disconnection is distorted, the icing rate feature protection item is dynamically added to the reinforcement learning reward function; when the quantitative error label shows that the deviation has decreased by more than 50%, the action space is expanded to increase the anti-aliasing sampling operation, and the dynamic balance closed loop of resource allocation strategy and safety risk control is realized.

Claims

1. A power grid probabilistic equivalence model construction method based on multi-scene clustering, characterized in that, The method comprises the following steps: acquiring a multi-dimensional data source of a power grid, the multi-dimensional data source comprising static topology, dynamic time sequence and environmental parameters; performing a space-time feature decoupling operation on the multi-dimensional data source, separating time correlation features, space correlation features and composite risk features, assigning each feature a weight identifier, and outputting a space-time feature tensor with a weight identifier; performing a graph-driven reinforcement clustering operation based on the space-time feature tensor, comprising: constructing a topology risk graph structure, selecting a feature compression strategy through deep reinforcement learning, and outputting a reduced dimension scene feature set containing a high-risk scene label; inputting the reduced dimension scene feature set into a hierarchical dynamic verification operation, performing differential verification according to the high-risk scene label, and generating a safety margin offset heat map; performing a double-algorithm optimization tracing operation on the safety margin offset heat map, comprising: genetic algorithm coding feature pruning operation, butterfly algorithm optimization time window parameter, and outputting a feature operation tracing matrix; feeding back the feature operation tracing matrix to the graph-driven reinforcement clustering operation, and updating the deep reinforcement learning decision rule.

2. The method for constructing a power grid probability equivalent model based on multi-scenario clustering according to claim 1 is characterized in that: The assignment of each feature weight identifier comprises: adding a voltage recovery curve monotonicity identifier weight to the separated time correlation features; adding a power transfer key section weight to the separated space correlation features; adding an environmental mutation and device response coupling weight to the separated composite risk features.

3. The method of claim 1, wherein, The construction of the topology risk graph structure and the deep reinforcement learning decision-making comprises: analyzing the topology risk graph structure based on the space-time feature tensor, wherein: defining the node attribute as the bus voltage transient sensitivity; defining the edge weight as the fault propagation intensity calculation value; selecting a feature compression strategy in response to the real-time calculation resource occupancy rate based on the topology risk graph structure; outputting the reduced dimension scene feature set containing the high-risk scene label to the hierarchical dynamic verification operation.

4. The power grid probabilistic equivalence model construction method based on multi-scene clustering according to claim 3, characterized in that, The selection of the feature compression strategy comprises: if the feature type is low risk, performing wavelet threshold dimension reduction operation; if the feature contains a composite risk label, performing anti-aliasing protection sampling operation; when the calculation node dimension exceeds the preset threshold, performing clustering compression operation on the medium-risk features.

5. The method for constructing power grid probabilistic equivalence model based on multi-scene clustering according to claim 1, characterized in that, The differential verification according to the high-risk scene label is performed in the following order: According to the existence state of the high-risk scene label: if it exists, perform full dynamic time domain simulation to obtain unit regulation response time sequence data; if it does not exist, perform quasi-steady-state reduced order verification to obtain static parameter cumulative error; performing spatial mapping operation on the unit regulation response time sequence data or the static parameter cumulative error to generate a three-dimensional color temperature layer.

6. The power grid probabilistic equivalence model construction method based on multi-scene clustering according to claim 5, characterized in that, The three-dimensional color temperature layer performs the following mapping operations: convert the over-standard power angle deviation position to physical coordinates and associate it with the geographical coordinates of the generator group; match the voltage recovery delay position to the device identifier and associate it with the bus number of the substation; render the color temperature gradient on the geographical wiring diagram to identify the margin deviation percentage interval; The topology risk graph structure and the three-dimensional color temperature layer cooperatively perform the following analysis operations: performing reverse mapping calculation on the over-standard power angle deviation coordinates to locate the topology edges with fault propagation intensity > 0.8; performing attribute association analysis on the voltage recovery delay position to match the node transient sensitivity attribute; The risk level rendering calculation is performed on the topological nodes within a 50km range of the typhoon path, and a three-dimensional color temperature layer is updated.

7. The method for constructing power grid probabilistic equivalence model based on multi-scene clustering according to claim 1, characterized in that, The executing double-algorithm optimization traceability operation includes: The genetic algorithm is used to perform feature pruning coding, and the margin deviation value is the fitness function input; The butterfly algorithm is used to perform time window parameter optimization; The feature operation and safety error quantization association path is generated.

8. The method for constructing power grid probabilistic equivalence model based on multi-scene clustering according to claim 7, characterized in that, The generated quantization association path performs the following storage operations: In the direct association path, the quantization error value labeling operation is performed; In the indirect association path, the risk propagation conditional probability weight recording operation is performed; In the high-impact area coordinates, the optimization compensation recording identification operation is performed.

9. The power grid probabilistic equivalence model construction method based on multi-scene clustering according to claim 8, characterized in that, The updating deep reinforcement learning decision rule performs the following judgment operations: When the risk propagation conditional probability weight recording operation shows that the disconnection probability is distorted, the operation of adding an icing rate feature protection item to the reward function is performed; When the quantization error value labeling operation shows that the deviation decreases by more than 50%, the operation of expanding the action space to increase anti-aliasing sampling is performed.

10. A power grid probabilistic equivalence model construction device based on multi-scenario clustering, applied to the power grid probabilistic equivalence model construction method based on multi-scenario clustering in any one of claims 1-9, characterized in that, It includes: A data acquisition unit acquires a multi-dimensional data source of a power grid, and the multi-dimensional data source includes static topology, dynamic time series, and environmental parameters; The data processing unit performs a space-time feature decoupling operation on the multi-dimensional data source, separates time correlation features, spatial correlation features, and composite risk features, assigns each feature a weight identifier, and outputs a space-time feature tensor with a weight identifier; The clustering unit performs a graph-driven reinforcement clustering operation based on the space-time feature tensor, including: constructing a topological risk graph structure, compressing the decision features through a deep reinforcement learning strategy, and outputting a reduced dimension scene feature set containing high-risk scene labels; The heat map generation unit inputs the reduced dimension scene feature set into a hierarchical dynamic verification operation, performs differential verification according to the high-risk scene labels, and generates a safety margin offset heat map; The optimization unit performs a double-algorithm optimization traceability operation on the safety margin offset heat map, including: Genetic algorithm coding feature pruning operation, butterfly algorithm optimization time window parameter, output feature operation traceability matrix; The update unit feeds back the feature operation traceability matrix to the graph-driven reinforcement clustering operation, and updates the deep reinforcement learning decision rule.

Images