A power grid intelligent safety and stability early warning and control method

Abstract

The application discloses a power grid intelligent safety and stability early warning and control method, and relates to the technical field of power grid safety and stability control. The method comprises the following steps: acquiring a historical operation data set of each transformer substation in a target power grid area; performing transient energy gradient analysis processing on a voltage phase angle fluctuation sequence to obtain a phase angle energy gradient feature vector, performing oscillation mode decomposition processing on an active power disturbance record sequence to obtain a power oscillation mode feature vector, and performing nonlinear dynamics feature coupling processing on the two feature vectors to generate a coupled stability feature tensor; calling a pre-trained graph attention deep network model to perform stability risk propagation reasoning processing on the coupled stability feature tensor to generate a stability risk propagation path set; and performing collaborative stability boundary deduction processing on the voltage phase angle fluctuation sequence and the protection device action time sequence record based on the stability risk propagation path set to generate a stability out-of-control boundary prediction result of the target power grid area.

Description

Technical Field

[0001] This invention relates to the field of power grid safety and stability control technology, specifically to a method for intelligent power grid safety and stability early warning and control. Background Technology

[0002] Early warning and control of power grid safety and stability are key technological areas for ensuring the reliable operation of power systems. Existing early warning methods for power grid safety and stability mainly rely on time-domain simulation calculations based on physical mechanisms, generating preset stability control strategies through offline analysis of the power grid's operating state. These methods typically assess system stability margins based on single-type electrical quantity parameters, such as over-limit monitoring of voltage amplitude or phase angle, or based on spectral analysis of active power oscillations. These techniques have limitations when processing multi-source heterogeneous data, making it difficult to effectively integrate the time-series variation characteristics of voltage phase angle with information on active power oscillation modes and the response characteristics of protection devices. Existing technologies in power grid stability risk assessment often employ linear model assumptions, neglecting the deep nonlinear coupling relationship between transient energy changes in voltage phase angle and power oscillation modes under different operating states of the power grid, resulting in inaccurate characterization of the dynamic evolution trend of power grid stability. Furthermore, conventional early warning methods struggle to effectively infer the propagation path and diffusion direction of stability risks between substations, failing to accurately identify the initiating substation of stability failure and the propagation path of failure within the power grid. When the power grid faces complex disturbances, early warning results based on a single parameter often lag behind the actual physical process, resulting in insufficient targeting and timeliness of control. Summary of the Invention

[0003] The purpose of this invention is to provide a smart power grid safety and stability early warning and control method to solve the problems in the prior art, which is difficult to effectively integrate the transient energy characteristics of voltage phase angle and the active power oscillation mode characteristics, resulting in inaccurate characterization of the nonlinear dynamic evolution process of the power grid stability state, and the difficulty in accurately reasoning and predicting the propagation path of stability risks between substations.

[0004] The objective of this invention can be achieved through the following technical solutions:

[0005] This invention provides a method for intelligent safety and stability early warning and control of power grids, aiming to improve the accuracy of identifying transient stability risks, the ability to infer propagation paths, and the timeliness of adaptive control responses during power grid operation. This method achieves fully automated processing from risk perception to control decision-making through collaborative analysis of multi-dimensional historical operating data and coupling with nonlinear dynamic characteristics, combined with graph attention deep networks and stability boundary deduction mechanisms.

[0006] As a technical solution of the present invention, the method includes: acquiring a set of historical operating data from various substations in a target power grid area, including voltage phase angle fluctuation sequences, active power disturbance record sequences, and protection device action timing records; performing transient energy gradient analysis on the voltage phase angle fluctuation sequences to obtain phase angle energy gradient feature vectors; performing oscillation mode decomposition on the active power disturbance record sequences to obtain power oscillation mode feature vectors; performing nonlinear dynamic feature coupling processing on the phase angle energy gradient feature vectors and the power oscillation mode feature vectors to generate coupled stable feature tensors; calling a pre-trained graph attention deep network model to perform stability risk propagation inference processing on the coupled stable feature tensors to generate a set of stability risk propagation paths corresponding to each substation; and performing collaborative stability boundary deduction processing on the voltage phase angle fluctuation sequences and protection device action timing records of each substation based on the set of stability risk propagation paths to generate a stability runaway boundary prediction result for the target power grid area.

[0007] This invention, through the coordinated processing of transient energy gradient analysis and oscillation mode decomposition, can extract energy accumulation features and oscillation mode features of the power grid transient process from two dimensions: voltage phase angle and active power. Preferably, transient energy gradient analysis is performed on the voltage phase angle fluctuation sequence to obtain a phase angle energy gradient feature vector. This includes: performing differential operations on adjacent time points of the voltage phase angle fluctuation sequence to obtain a phase angle change rate sequence; performing sliding window energy integration on the phase angle change rate sequence to obtain the transient energy accumulation value within each time window; calculating the gradient direction of the transient energy accumulation value to obtain an energy gradient direction sequence; and concatenating the energy gradient direction sequence with the phase angle values ​​at corresponding time points in the voltage phase angle fluctuation sequence to obtain the phase angle energy gradient feature vector. This processing method fully preserves the directionality and amplitude information of energy changes during the transient process, providing high-resolution input for subsequent coupled feature extraction. As a further optimization, the sliding window length used in the transient energy gradient analysis process is positively correlated with the period of the dominant oscillation mode of the active power disturbance recording sequence, so that the energy integration window can adaptively match the current dominant oscillation frequency of the power grid and avoid energy integration distortion caused by window size mismatch.

[0008] In terms of power oscillation feature extraction, this invention performs oscillation mode decomposition on the active power disturbance recording sequence to obtain a power oscillation mode feature vector. This includes: performing empirical mode decomposition on the active power disturbance recording sequence to obtain multiple intrinsic mode function (IMF) components; selecting the dominant oscillation mode component from the multiple IMF components, wherein the dominant oscillation mode component is the IMF component whose energy proportion exceeds a preset energy proportion threshold; extracting the instantaneous frequency and instantaneous amplitude of the dominant oscillation mode component to obtain an oscillation frequency sequence and an oscillation amplitude sequence; and alternately arranging and combining the oscillation frequency sequence and the oscillation amplitude sequence to generate the power oscillation mode feature vector. This process can effectively filter out noise interference and secondary modes, retaining the oscillation characteristics most sensitive to the grid stability state, thereby improving the targeting of subsequent coupling analysis.

[0009] To achieve a high-dimensional characterization of the power grid's stable state, this invention performs nonlinear dynamic feature coupling processing on the phase angle energy gradient feature vector and the power oscillation mode feature vector to generate a coupled stable feature tensor. As a specific technical solution of this invention, the coupling processing includes: performing phase space reconstruction processing on the phase angle energy gradient feature vector to obtain a phase space trajectory matrix, where the rows of the phase space trajectory matrix correspond to the phase space dimension and the columns correspond to the time steps; performing Lyapunov exponential spectrum mapping processing on the power oscillation mode feature vector to obtain an exponential spectrum feature vector; and constructing an energy-oscillation cross-coupling function, performing element-wise nonlinear multiplication of the phase space trajectory matrix and the exponential spectrum feature vector to obtain a cross-coupling matrix, wherein the energy-oscillation cross-coupling function is expressed as:

[0010] ;

[0011] in: Represents the first in the cross-coupling matrix Line 1 The element values ​​of the column, The first in the phase space trajectory matrix Line 1 Column elements, The first eigenvector in the exponential spectrum One element, This is a coupling strength adjustment factor. The nonlinear suppression coefficient is... Denotes the Euclidean norm. The hyperbolic tangent function is used; the cross-coupling matrix and the phase space trajectory matrix are deeply concatenated and stacked to generate the coupled stable feature tensor. This coupling function, by introducing a nonlinear suppression term and hyperbolic tangent activation, can adaptively adjust the interaction strength between the energy trajectory and the oscillation mode. When the difference between the two is large, the coupling weight is automatically reduced, thereby avoiding noise introduced by spurious correlations and improving the tensor's ability to express true stability risks. Preferably, the Lyapunov exponent of the chaotic system is introduced as a coupling strength adjustment factor in the nonlinear dynamic feature coupling processing, further enhancing the sensitivity of the feature tensor to the chaotic edge states of the system.

[0012] In terms of stable risk propagation inference, this invention calls a pre-trained graph attention deep network model to perform stable risk propagation inference processing on the coupled stable feature tensor, generating a set of stable risk propagation paths corresponding to each substation. Specifically, the set of stable risk propagation paths includes the influence direction and intensity distribution of stable risks between each substation; the coupled stable feature tensor is input into the node feature encoding layer of the graph attention deep network model for node embedding encoding processing to obtain the stable state latent feature vector corresponding to each substation; the attention coefficient calculation layer of the graph attention deep network model is called to calculate the risk propagation attention coefficient between any two substations based on the vector similarity and topological adjacency relationship between the stable state latent feature vectors; the adjacency matrix of the graph attention deep network model is updated with edge weights based on the risk propagation attention coefficient to obtain a weighted risk propagation graph; multi-hop risk diffusion simulation processing is performed on the weighted risk propagation graph to generate the set of stable risk propagation paths starting from each substation. Through the graph attention mechanism, the model can automatically learn the nonlinear dependency relationship of risk propagation between different substations, avoiding the defect of traditional graph convolutional networks that treat each neighbor node with equal weight. Preferably, the graph attention deep network model adopts a multi-head attention mechanism, where each attention head independently calculates the risk propagation attention coefficient between substations and takes the average value as the final attention coefficient, thereby improving the stability and generalization ability of the attention coefficient.

[0013] Based on the aforementioned propagation path set, this invention further performs collaborative stability boundary extrapolation processing on the voltage phase angle fluctuation sequence and protection device action timing records of each substation, generating a stability runaway boundary prediction result for the target power grid area. The stability runaway boundary prediction result is used to indicate the starting substation where stability runaway is expected to occur and the runaway propagation path. The specific extrapolation process includes: constructing a stability risk directed graph based on the influence direction and influence intensity distribution, where the nodes of the stability risk directed graph are substations, the direction of the directed edges is the risk propagation direction, and the weight of the directed edges is the corresponding intensity value in the influence intensity distribution; mapping the voltage phase angle fluctuation sequence to the nodes of the stability risk directed graph as the initial node state value, and marking the protection device action timing records as boundary trigger conditions on the corresponding nodes; starting from the node with an out-degree of zero in the stability risk directed graph, backpropagation extrapolation is performed, and the stability margin decay trajectory of each node is calculated sequentially; the node that first crosses the preset stability margin lower limit in the stability margin decay trajectory and its upstream propagation path are determined as the stability runaway boundary prediction result.

[0014] As a further technical solution of the present invention, backward propagation is performed starting from the node with zero out-degree in the directed graph of stability risk, and the stability margin decay trajectory of each node is calculated sequentially. This includes: determining the node with zero out-degree in the directed graph of stability risk as the set of starting nodes for propagation; for each starting node in the set of starting nodes, traversing all its predecessor nodes in the opposite direction of the directed edges; for each traversed predecessor node, calculating the stability margin decay start time and decay slope of the predecessor node based on the weight of the directed edge between it and the successor node, the stability margin decay rate of the successor node, and the voltage phase angle fluctuation amplitude of the predecessor node. The formula for calculating the stability margin decay slope of the predecessor node is as follows:

[0015] ;

[0016] in: Represents the predecessor node The slope of the stability margin decay. Indicates the successor node The stability margin decay rate, For the node To the node The weight of the directed edge. The propagation attenuation coefficient, This is the voltage disturbance amplification factor. Predecessor node The voltage phase angle fluctuation amplitude, Risk from nodes propagation to nodes The estimated propagation delay is calculated; the stability margin decay start time and decay slope of each node are sequentially concatenated to generate the stability margin decay trajectory of that node. This backpropagation deduction method fully utilizes the directionality of the risk propagation path. By recursively extrapolating from the farthest node, it can effectively avoid the path divergence problem caused by forward propagation and improve the accuracy of boundary prediction.

[0017] After generating the stability runaway boundary prediction result, the present invention further includes generating an adaptive generator and load shedding control strategy based on the stability runaway boundary prediction result; and distributing the adaptive generator and load shedding control strategy to the stability control device corresponding to the starting substation to trigger dynamic tuning of control parameters. Specifically, generating the adaptive generator and load shedding control strategy based on the stability runaway boundary prediction result includes: parsing the starting substation identifier and the runaway propagation path node sequence in the stability runaway boundary prediction result; obtaining the real-time total output power of the starting substation corresponding to the starting substation identifier, and the load importance level coefficient of each node substation in the runaway propagation path node sequence; determining the total generator capacity to be shedding based on the real-time total output power and a preset power reduction target value; sorting each node substation from low to high according to the load importance level coefficient, and accumulating the load shedding amount sequentially according to the sorting result until the total generator capacity to be shedding is reached, forming a load shedding node list; allocating the total generator capacity to each generator unit of the starting substation, generating a generator shedding execution sequence, and combining the load shedding node list and the generator shedding execution sequence into the adaptive generator and load shedding control strategy. The strategy generation process takes into account both the tripping location (starting substation) and the load shedding location (propagation path node), achieving precise control over the instability source and propagation path, and avoiding the problems of unnecessary or insufficient load shedding in traditional unified load reduction strategies.

[0018] As a refined scheme for generating the generator tripping execution sequence, the total tripping capacity is allocated to each generator unit in the starting substation to generate the tripping execution sequence. This includes: obtaining the primary frequency regulation response speed and moment of inertia coefficient of each generator unit in the starting substation; sorting the generator units from slowest to fastest according to their primary frequency regulation response speed to obtain a frequency regulation response ranking list; for each generator unit in the ranking list, calculating the initial tripping allocation capacity corresponding to that generator unit based on the ratio of its moment of inertia coefficient to the sum of the moment of inertia coefficients of all generator units; and sequentially accumulating the initial tripping allocation capacity corresponding to each generator unit according to the order of the ranking list until the accumulated total capacity reaches the total tripping capacity. The selected generator units and their corresponding allocation capacities are then used as the tripping execution sequence. This scheme prioritizes the tripping of units with slow frequency regulation response speeds, maximizing the retention of fast-regulating units to maintain frequency stability. Simultaneously, by allocating capacity proportionally based on the moment of inertia coefficient, the allocation of tripped power is more consistent with the dynamic characteristics of the units, reducing the secondary impact of tripping on the power grid.

[0019] In terms of strategy delivery and execution, the adaptive generator tripping and load shedding control strategy is delivered to the stability control device corresponding to the starting substation to trigger dynamic adjustment of control parameters. This includes: encapsulating the adaptive generator tripping and load shedding control strategy into a strategy frame conforming to the stability control device's communication protocol; the strategy frame containing the starting substation identifier, the generator tripping execution sequence, and the load shedding node list; delivering the strategy frame to the stability control device corresponding to the starting substation via a dedicated stability control communication channel; in the stability control device, updating the generator tripping setpoint parameter table according to the generator tripping execution sequence in the strategy frame, and updating the low-frequency and low-voltage load shedding setpoint parameter table according to the load shedding node list; triggering the stability control device to perform a hot-loading operation of the setpoint parameters, so that the updated generator tripping setpoint parameter table and the low-frequency and low-voltage load shedding setpoint parameter table take effect without restarting the device. This hot-loading mechanism ensures the immediate update of control parameters, avoids the control window period caused by the need to restart the device for traditional setpoint modifications, and significantly improves the real-time performance of emergency control response.

[0020] The beneficial effects of this invention are:

[0021] A transient energy gradient analysis process is performed on the voltage phase angle fluctuation sequence to obtain the phase angle energy gradient feature vector. An oscillation mode decomposition process is performed on the active power disturbance record sequence to obtain the power oscillation mode feature vector. The phase angle energy gradient feature vector and the power oscillation mode feature vector are then coupled using nonlinear dynamics to generate a coupled stable feature tensor. This process effectively integrates the gradient features characterizing the transient energy change trend of the voltage phase angle with the mode features characterizing the frequency and amplitude of power oscillations at the dynamic level. By constructing an energy-oscillation cross-coupling function and introducing a nonlinear suppression coefficient and a hyperbolic tangent function, this process can uncover the interaction relationships between different physical quantities under nonlinear dynamic mechanisms. This results in the generated coupled stable feature tensor containing the co-evolutionary information between the phase space trajectory and the power oscillation exponential spectrum. This coupling method avoids simply splicing or linearly weighting the two features, overcoming the shortcomings of conventional methods that ignore the dynamic coupling relationship between voltage energy changes and power oscillation modes. Therefore, it more fundamentally characterizes the nonlinear evolution features of the grid's stable state at the feature level. A pre-trained graph attention deep network model is used to perform stable risk propagation inference on the coupled stable feature tensor, generating a set of stable risk propagation paths for each substation. This set includes the influence direction and intensity distribution of stable risks between substations. The graph attention deep network model transforms the coupled stable feature tensor into latent stable-state features of the substations through a node feature encoding layer, and adaptively calculates the risk propagation attention coefficients based on the vector similarity and topological adjacency relationships between latent features. This mechanism can automatically adjust the weights of risk propagation between substations according to the dynamic changes in the real-time operating status of the power grid, thereby generating a weighted risk propagation graph that more closely reflects the actual physical process. Based on this graph, multi-hop risk diffusion simulation processing can generate a set of risk propagation paths starting from each substation, clearly identifying the chain structure and influence intensity of risk propagation from a certain node to other nodes. Compared with traditional propagation path estimation based on fixed topology or preset rules, this risk propagation inference method based on graph attention mechanism achieves a more accurate characterization of the dynamic diffusion direction of risks. Based on this, a collaborative stability boundary extrapolation process is performed on the voltage phase angle fluctuation sequences and protection device operation timing records of each substation, according to the distribution of influence direction and intensity. By constructing a directed graph of stability risk, the voltage phase angle fluctuation sequences are mapped to the initial values ​​of node states. The protection device operation timing records are used as boundary trigger conditions, and the extrapolation is performed backward from the node with zero out-degree, gradually calculating the stability margin decay trajectory of each node. This collaborative extrapolation process combines the phase angle fluctuation information reflecting changes in the electrical state of the power grid with the timing information reflecting the operation behavior of protection equipment, jointly constraining the decay process of the stability margin.By identifying the node that first crosses the preset lower limit of stability margin in the stability margin decay trajectory and its upstream path as the stability runaway boundary prediction result, this method can accurately locate the starting substation of stability runaway and the spread path of runaway in the power grid, providing clear spatiotemporal boundary information for subsequent control. Attached Figure Description

[0022] The invention will now be further described with reference to the accompanying drawings.

[0023] Figure 1 This is a flowchart of the intelligent power grid safety and stability early warning and control method described in this invention;

[0024] Figure 2 This is a flowchart for extracting the features of the dominant oscillation mode of the active power perturbation sequence;

[0025] Figure 3 This is a flowchart of the construction of stable feature tensors for phase space reconstruction and Lyapunov exponential mapping. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See Figure 1 This invention provides a method for intelligent safety and stability early warning and control of power grids. The method includes: acquiring historical operating data sets from various substations in a target power grid area, including voltage phase angle fluctuation sequences, active power disturbance record sequences, and protection device action timing records; performing transient energy gradient analysis on the voltage phase angle fluctuation sequences to obtain phase angle energy gradient feature vectors; performing oscillation mode decomposition on the active power disturbance record sequences to obtain power oscillation mode feature vectors; coupling the phase angle energy gradient feature vectors and the power oscillation mode feature vectors with nonlinear dynamic features to generate a coupled stability feature tensor; calling a pre-trained graph attention deep network model to perform stability risk propagation inference processing on the coupled stability feature tensor to generate a set of stability risk propagation paths corresponding to each substation; and performing collaborative stability boundary deduction processing on the voltage phase angle fluctuation sequences and protection device action timing records of each substation based on the stability risk propagation path set to generate a stability runaway boundary prediction result for the target power grid area.

[0028] Example 1: In a specific implementation, historical operating data sets, including voltage phase angle fluctuation sequences, active power disturbance record sequences, and protection device action timing records, are obtained from various substations in the target power grid area. Transient energy gradient analysis processing is performed on the voltage phase angle fluctuation sequences to obtain the phase angle energy gradient feature vector. The specific implementation method of the transient energy gradient analysis processing is as follows:

[0029] The voltage phase angle fluctuation sequence is subjected to a difference operation between adjacent time points to obtain a phase angle change rate sequence. The voltage phase angle fluctuation sequence is represented as follows: ,in Let be the sampling time point, and let be the time interval between adjacent time points. The first phase angle change rate sequence element The calculation formula is: ,in , This represents the total length of the voltage phase angle fluctuation sequence. The phase angle change rate sequence includes... Each element.

[0030] The phase angle change rate sequence is subjected to sliding window energy integration to obtain the cumulative transient energy value within each time window. The length of the sliding window is denoted as . , is a positive integer representing the number of data points contained in each window. For the phase angle change rate sequence from the , A window starting with an element, the cumulative transient energy value within this window is expressed as... ,in The starting time point of each window corresponds to a time point in the original voltage phase angle fluctuation sequence. Therefore, the sequence of transient energy accumulation values ​​is denoted as The sequence length is .

[0031] The gradient direction is calculated for the transient energy accumulation value to obtain an energy gradient direction sequence. For the transient energy accumulation values ​​corresponding to two adjacent time windows... and Calculate the difference ,in The difference As elements in the energy gradient direction sequence, this sequence is denoted as The sequence length is Positive values ​​in the energy gradient direction sequence indicate an increase in the transient energy accumulation, while negative values ​​indicate a decrease in the transient energy accumulation.

[0032] The phase angle energy gradient feature vector is obtained by concatenating the energy gradient direction sequence with the phase angle values ​​at corresponding time points in the voltage phase angle fluctuation sequence. This concatenation of the energy gradient direction sequence and the voltage phase angle fluctuation sequence results in the phase angle energy gradient feature vector. Correspondingly, in specific implementations, the first of the energy gradient direction sequences is selected. element The voltage phase angle fluctuation sequence of the first Phase angle values ​​at each time point Perform concatenation. The concatenated vector is... ,in All splicing results are combined in chronological order to obtain the phase angle energy gradient feature vector.

[0033] In specific implementation, the sliding window length used in the transient energy gradient analytical processing is... With respect to the dominant oscillation mode period of the active power disturbance recording sequence A positive correlation exists. The period of the dominant oscillation mode. The acquisition method is as follows: Empirical mode decomposition (EMD) is performed on the active power disturbance recording sequence to obtain multiple intrinsic mode function (IMF) components. The IMF component with the largest energy proportion is selected from these IMF components as the dominant oscillation mode component. The average period of this dominant oscillation mode component is the dominant oscillation mode period. The length of the sliding window The value of is determined in the following way: the period of the dominant oscillation mode is denoted as . (Unit: seconds), the sampling time interval is denoted as... (Unit: seconds) ,in This indicates rounding to the nearest integer.

[0034] Example 2: In specific implementation, refer to Figure 2 The active power disturbance recording sequence is subjected to oscillation mode decomposition (EMD) to obtain power oscillation mode feature vectors. The specific implementation method of the EMD is as follows:

[0035] The active power disturbance recording sequence is subjected to empirical mode decomposition (EMD) to obtain multiple intrinsic mode function (IMF) components. The active power disturbance recording sequence is denoted as... ,in For the sampling time point, , Where is the sequence length. The specific process of empirical mode decomposition is as follows: Each oscillation mode in the active power disturbance recording sequence is extracted through a sieving process, and one eigenmode function component obtained from each sieving is denoted as . ,in , This represents the total number of components obtained from the decomposition. Each intrinsic mode function component... The length of the active power disturbance recording sequence is related to the length of the active power disturbance recording sequence. The components are identical, and each component satisfies the basic definition of the intrinsic mode function, that is, the number of extreme points and the number of zero crossings are equal or differ by no more than one in the whole sequence, and the local mean is zero.

[0036] The dominant oscillation mode component is selected from the plurality of intrinsic mode function components. The dominant oscillation mode component is the intrinsic mode function component whose energy proportion exceeds a preset energy proportion threshold. The specific selection method is as follows: for each intrinsic mode function component... Calculate its energy value Calculate the sum of the energies of all intrinsic mode function components. For each eigenmode function component, calculate its energy percentage. The preset energy percentage threshold is denoted as... The value ranges from 0.3 to 0.5, and in specific implementation... Set to 0.4. Filter out all that meet the criteria. The intrinsic mode function components are taken as the dominant oscillation mode components. If multiple intrinsic mode function components satisfy the conditions, the one with the largest energy proportion is selected as the unique dominant oscillation mode component. The dominant oscillation mode component is denoted as... ,in .

[0037] The dominant oscillation mode component is processed by extracting its instantaneous frequency and amplitude to obtain an oscillation frequency sequence and an oscillation amplitude sequence. Specifically, the extraction method involves processing the dominant oscillation mode component... Perform a Hilbert transform to obtain the analytic signal. ,in The imaginary unit, This represents the Hilbert transform. The instantaneous amplitude of the analytic signal is... The instantaneous phase is Discrete difference operations are performed on the instantaneous phase sequence to obtain the instantaneous angular frequency sequence. ,in This is the sampling time interval for the active power disturbance recording sequence. The instantaneous angular frequency is converted to the instantaneous frequency. This yields the oscillation frequency sequence. The instantaneous amplitude sequence is then used. As a sequence of oscillation amplitudes.

[0038] The power oscillation mode feature vector is generated by alternately combining the oscillation frequency sequence and the oscillation amplitude sequence. Specifically, starting from the first time point, one element from the oscillation frequency sequence and one element from the oscillation amplitude sequence are sequentially and alternately selected to form a vector. The power oscillation mode feature vector is represented as follows: ,in For the first The oscillation frequency at each time point For the first The oscillation amplitude at each time point, the total length of the vector is The elements in the power oscillation mode feature vector are arranged strictly according to the alternation of frequency and amplitude, and correspond to the same time point.

[0039] Example 3: In specific implementation, refer to Figure 3 The phase angle energy gradient feature vector and the power oscillation mode feature vector are coupled nonlinearly with each other to generate a coupled stable feature tensor. The specific implementation method of the nonlinear dynamic feature coupling process is as follows:

[0040] Step 1: Perform phase space reconstruction processing on the phase angle energy gradient eigenvector to obtain a phase space trajectory matrix. The rows of the phase space trajectory matrix correspond to the phase space dimension, and the columns correspond to the time steps. The phase angle energy gradient eigenvector is denoted as... ,in , The length of the vector. Indicates the first The element corresponding to each time point Phase space reconstruction is performed using the delayed coordinate reconstruction method: the embedding dimension is set. and delay time Then the first phase space The state vectors are represented as follows: ,in The dimension of the phase space trajectory matrix is... ,in Let be the time step, and be the in the matrix. Line 1 Column elements Corresponding to the The th state vector of the th state vector Each component. In specific implementation, the embedding dimension. The pseudo-nearest neighbor ratio is determined by taking the smallest integer that makes the proportion of pseudo-nearest neighbors lower than a preset threshold, where the preset threshold is set to 0.05; delay time. Determined using the autocorrelation function method, the autocorrelation function is taken to first decrease to... The corresponding time delay.

[0041] Step 2: Perform Lyapunov exponential spectrum mapping on the power oscillation mode feature vector to obtain the exponential spectrum feature vector. The power oscillation mode feature vector is denoted as... Its length is The specific method for Lyapunov exponential spectrum mapping is as follows: the power oscillation mode feature vector is used as a time series input, and its Lyapunov exponential spectrum is calculated using the standard Jacobian method. The calculation process includes: reconstructing the phase space of the power oscillation mode feature vector (embedding dimension). and delay time Setup and Step 1 and The determination method is the same, but for (Independent calculation) to obtain the reconstructed phase space trajectory; calculate the Jacobian matrix in the tangent space for each state point; extract the Lyapunov exponents of all dimensions through QR decomposition iteration to obtain the exponent set. ,in To determine the embedding dimension, the set of indices is sorted from largest to smallest to form the feature vector of the index spectrum. Vector length The number of rows of the phase space trajectory matrix To obtain the same value, that is, to adjust the reconstruction parameters to make .

[0042] Step 3: Construct the energy-oscillation cross-coupling function to convert the phase space trajectory matrix. With the exponential spectrum feature vector Element-by-element nonlinear multiplication is performed to obtain the cross-coupling matrix. The energy-oscillation cross-coupling function is expressed as:

[0043]

[0044] in: Represents the first in the cross-coupling matrix Line 1 The element values ​​of the column, , . The first in the phase space trajectory matrix Line 1 Column elements, The first eigenvector in the exponential spectrum One element, This is a coupling strength adjustment factor. The nonlinear suppression coefficient is... Denotes the Euclidean norm. It is a hyperbolic tangent function. In specific implementations, the Lyapunov exponent of the chaotic system is introduced as a coupling strength adjustment factor in the nonlinear dynamic characteristic coupling processing. ,Right now The value of is equal to the exponential spectrum feature vector. The largest Lyapunov index in . This is the nonlinear suppression coefficient, ranging from 0.1 to 1.0, and is set to 0.2 in practice. Element-by-element operation refers to processing each element of the matrix... The corresponding row index of the vector elements The energy-oscillation cross-coupling function was calculated as follows: Thus forming a dimension of The cross-coupling matrix. The value range of each element in the cross-coupling matrix is... .

[0045] Step 4: The cross-coupling matrix With the phase space trajectory matrix Deep concatenation and stacking are performed to generate the coupled stable feature tensor. Deep concatenation and stacking refers to concatenating two matrices along a new dimension, increasing the dimension of the tensor. The phase space trajectory matrix... The dimension is The cross-coupling matrix The dimension is Stacking the two matrices along the third dimension yields the coupled stable feature tensor. Its dimensions are This refers to a three-dimensional tensor with a depth of 2. Specifically, the first channel in the tensor... The phase space trajectory matrix Second Channel The cross-coupling matrix .

[0046] Example 4: In specific implementation, a pre-trained graph attention deep network model is invoked to perform stability risk propagation inference processing on the coupled stable feature tensor, generating a set of stability risk propagation paths corresponding to each substation. The set of stability risk propagation paths includes the influence direction and intensity distribution of stability risks between each substation. The input to the graph attention deep network model is the coupled stable feature tensor. The output is the set of stable risk propagation paths. The graph attention deep network model employs a multi-head attention mechanism.

[0047] The graph attention deep network model comprises a node feature encoding layer and an attention coefficient calculation layer. The node feature encoding layer receives the coupled stable feature tensor. The dimension of the coupled stable feature tensor is The node feature encoding layer performs node embedding encoding on the coupled stable feature tensor to obtain the stable state latent feature vector corresponding to each substation. In a specific implementation, it is assumed that the target power grid area contains For each substation, each substation corresponds to a subtensor in the coupled stable feature tensor, and the subtensor dimension is... ,in This represents the time step corresponding to the substation. The node feature encoding layer employs a two-layer fully connected neural network. The first fully connected layer flattens each subtensor into a one-dimensional vector and then performs a linear transformation. The second fully connected layer maps the output of the first layer to the latent space, obtaining the steady-state latent feature vector corresponding to each substation. ,in , The dimension is , As a hidden feature dimension, in specific implementation Set it to 64.

[0048] The attention coefficient calculation layer calculates the risk propagation attention coefficient between any two substations based on the vector similarity and topological adjacency relationship between their steady-state latent feature vectors. Specifically, the calculation method is as follows: for each substation... and substation First, the topological adjacency relationship is obtained. This topological adjacency relationship is determined by the power grid topology. If the substation With substation If there is a direct power transmission line connection, then ,otherwise The attention coefficient calculation is based on a multi-head attention mechanism, where each attention head independently calculates its risk propagation attention coefficient, and the average value is taken as the final attention coefficient. In practice, the number of attention heads... Set to 8. For the first One point of attention, Risk transmission attention coefficient The calculation formula is:

[0049]

[0050] in: and Substations and substation The stable state latent feature vector. For the first The learnable weight matrix for each attention head has dimensions of . , As for the attention head output dimension, in specific implementation Set it to 64. For the first The learnable weight vectors for each attention head have a dimension of [missing information]. . This indicates a vector concatenation operation. The activation function is set to a negative slope of 0.2. Then, softmax normalization is applied to all adjacent nodes to obtain the final risk propagation attention coefficients.

[0051]

[0052] in Indicates substation The set of topologically adjacent nodes. For non-adjacent substation pairs, the attention coefficient is set to 0. The final risk propagation attention coefficient is obtained by averaging the outputs of all attention heads. .

[0053] Based on the risk propagation attention coefficients, the adjacency matrix of the graph attention deep network model is updated with edge weights to obtain a weighted risk propagation graph. The adjacency matrix... The primitive elements represent topological connections ( (or 0), elements in the updated adjacency matrix That is, only the attention coefficients between adjacent nodes are retained as edge weights, while the weights of edges between non-adjacent nodes remain at 0. The weighted risk propagation graph uses all substations as nodes and the updated adjacency matrix as the edge weight matrix.

[0054] A multi-hop risk propagation simulation is performed on the weighted risk propagation graph to generate a set of stable risk propagation paths starting from each substation. The multi-hop risk propagation simulation is implemented as follows: starting from each substation, the propagation proceeds hop-by-hop outwards, with each hop selecting the propagation direction according to the outgoing edge weight of the current node. Specifically, for the starting substation... The set of candidate nodes reached by the first hop is... Adjacent nodes are ranked according to edge weight. Sort the values ​​in descending order and take the first one with the largest weight. Each node serves as the first hop in the actual propagation path, in specific implementation... Set to 3. For each node in the first hop, continue propagating to the second hop using the same strategy until the propagation hop count reaches the preset maximum hop count. In specific implementation Set to 5. Each path records the sequence of substation nodes traversed sequentially, as well as the edge weights between adjacent nodes. The set of paths generated from all starting substations constitutes the set of stable risk propagation paths. This set contains the influence direction (node ​​sequence order) and influence strength (weights of each edge on the path) of each path. The pre-training process of the graph attention deep network model is as follows: the training data consists of historical power grid fault cases, each case containing a historical coupled stable feature tensor as input and a manually labeled set of stable risk propagation paths as labels. The loss function is cross-entropy loss, the optimizer is Adam, the learning rate is set to 0.001, the training epochs are 200, and training stops when the validation set loss does not decrease for 10 consecutive epochs.

[0055] Example 5: In specific implementation, based on the aforementioned set of stable risk propagation paths Voltage phase angle fluctuation sequence for each of the substations The stability boundary is calculated by coordinating the timing records of the protection device's actions with the stability boundary prediction results for the target power grid area. These results indicate the expected initiation substation of the stability runaway and the runaway propagation path. The specific implementation method is as follows:

[0056] A stable risk directed graph is constructed based on the influence direction and influence intensity distribution. The nodes in this graph represent substations, the directions of the directed edges represent the risk propagation directions, and the weights of the directed edges are the corresponding intensity values ​​in the influence intensity distribution. The set of stable risk propagation paths is also described. Each path includes the starting substation, the sequence of substation nodes it passes through, and the influence strength (i.e., edge weight) between adjacent nodes. All unique substation nodes are extracted from all paths as nodes in the directed graph of the stability risk, denoted as the total number of nodes. For each path, two adjacent nodes and (Direction from) arrive Add a node from the stable risk directed graph. Pointing to node The weight of the directed edge. Take from this path arrive The influence strength value. If multiple paths contain the same node pair, the average of all influence strength values ​​is taken as the final weight.

[0057] The voltage phase angle fluctuation sequence The initial values ​​of the node states are mapped onto the nodes of the directed graph of stable risks, and the action sequence records of the protection devices are marked on the corresponding nodes as boundary trigger conditions. For each substation corresponding to a node, the phase angle value of the last time point in the voltage phase angle fluctuation sequence of that substation is used. As the initial value of the node state ,in Simultaneously, for each node, check the substation's protection device action sequence record for any action events (such as protection device trip signals). If an action event exists, mark the node as a boundary trigger node and record the time of the action event as... Otherwise, mark it as a non-trigger node.

[0058] The backpropagation process begins with the nodes in the directed graph of stable risk that have an out-degree of zero, calculating the stability margin decay trajectory for each node in turn. A node with an out-degree of zero refers to a node that has no outgoing edges, meaning it does not propagate risk to other nodes. The nodes in the directed graph of stable risk that have an out-degree of zero are determined as the starting set for the backpropagation, denoted as […]. For each starting node in the set of starting nodes for the deduction... Traverse all predecessor nodes along the opposite direction of the directed edges (i.e., from the successor node to the predecessor node). Specifically, for each starting node... traverse all For a directed edge to a successor node, there exists a predecessor node for that directed edge. point to nodes For each predecessor node that is traversed According to its relationship with the successor node Weight of directed edges between The stability margin decay rate of the successor node and the voltage phase angle fluctuation amplitude of the predecessor node are used to calculate the stability margin decay start time and decay slope of the predecessor node.

[0059] The formula for calculating the stability margin decay slope of the predecessor node is:

[0060]

[0061] in: Represents the predecessor node The stability margin decay slope, in degrees per second; Indicates the successor node The stability margin decay rate, in degrees per second; For the node To the node The weight of the directed edge is dimensionless and ranges from 0 to 1; The propagation attenuation coefficient is dimensionless and ranges from 0.1 to 1.0. In practical implementation... Set to 0.5; This is the voltage disturbance amplification factor, dimensionless, with a value ranging from 0.01 to 0.1. In specific implementations... Set to 0.05; Predecessor node The voltage phase angle fluctuation amplitude, in degrees, is obtained by calculating the difference between the maximum and minimum values ​​of the voltage phase angle fluctuation sequence of the predecessor node. Risk from nodes propagation to nodes The estimated propagation delay is in seconds, and the actual propagation time is determined by the node in the specific implementation. With nodes The normalized value of the electrical distance between the lines is multiplied by a preset reference time, which is set to 0.1 seconds. ,in This refers to the per-unit value of the electrical distance of the line.

[0062] The predecessor node stability margin decay start time The calculation method is as follows: if the successor node As the starting node for the deduction (i.e., with an out-degree of zero), then take ,in The current system time (or simulation start time); if subsequent nodes If it has been marked as a boundary trigger node, then take... Otherwise, take Initial set of starting nodes for the initial deduction The node in the middle, its stability margin decay start time Set to current system time Its stability margin decay rate Based on the voltage phase angle fluctuation amplitude of the node itself Calculation, the calculation formula is: ,in The time base is set to 1 second. For non-starting nodes, the time of each predecessor node is calculated sequentially according to the backpropagation order. and .

[0063] Decrease the stability margin of each node starting time and attenuation slope Time-series stitching is performed to generate the stability margin decay trajectory of this node. Specifically, from... Starting from a moment, with a time step (In specific implementation) (Set to 0.01 seconds) Generate a series of time points until the preset maximum simulation time (In specific implementation) (Set to 10 seconds). The stability margin value at each time point. Calculated using a linear decay model: ,in This represents the initial stability margin of the node, which is based on the initial values ​​of the node's state. Compared with the preset stability margin benchmark value The difference is determined. Set to 180 degrees. Time series With the corresponding stability margin sequence The splicing forms a stable margin decay trajectory.

[0064] Step 4: The first time the stability margin decay trajectory crosses the preset lower limit of stability margin. The nodes and their upstream propagation paths are determined as the predicted stable runaway boundary. The preset stability margin lower limit In practice, it is set to 30 degrees. For each node Search for the condition that satisfies the stability margin decay trajectory. The earliest time point If such a time point exists, then that node is determined to be a stable out-of-control node. Among all stable out-of-control nodes, the node that first experienced a crossover (i.e., the smallest) is selected. The node is designated as the starting substation. Then, from that starting substation Starting from the beginning, backtracking along the opposite direction (i.e., the outgoing edge direction) of the stable risk directed graph, extracting all upstream node sequences that can reach the starting substation, forming the runaway propagation path. The stable runaway boundary prediction result includes the identifier of the starting substation. And the sequence of nodes along the runaway diffusion path.

[0065] In specific implementation, after generating the stability runaway boundary prediction result, an adaptive load shedding control strategy is generated based on the stability runaway boundary prediction result. The starting substation identifier and runaway propagation path node sequence in the stability runaway boundary prediction result are parsed. The real-time total output power of the starting substation corresponding to the starting substation identifier is obtained. And the load importance level coefficient of each substation in the uncontrolled propagation path node sequence. The load importance level coefficient The load importance level is obtained from a pre-set load importance table by the power dispatch center, with values ​​ranging from 0 to 1; the smaller the value, the lower the load importance. The total capacity to be tripped is determined based on the real-time total output power and the preset power reduction target value. ,in In practice, it is set to 0.20. This is based on the load importance level coefficient of each substation node. Sort the loads from lowest to highest, and accumulate the load cut-off amounts sequentially according to the sorting results until the total capacity of the cut-off machine is reached. This creates a list of load shedding nodes. Specifically, for each substation node along the runaway propagation path... Get its current total load. Starting with the substation with the lowest load importance coefficient, the total load of that substation is sequentially calculated. Add to the cut list until the total accumulated load is greater than or equal to At this point, the last substation to join may only disconnect a portion of the load (i.e., the disconnected amount is the remaining required amount), ultimately forming a list of load disconnection nodes. Each entry in the list contains the substation identifier and the amount of load shelved.

[0066] The total capacity of the cutting machine Each generator set assigned to the starting substation is assigned a generator tripping sequence. The primary frequency regulation response speed of each generator set in the starting substation is obtained. and rotational inertia coefficient Primary frequency modulation response speed Provided by the generator set's technical parameters, the unit is seconds; a smaller value indicates a faster response. This is based on the primary frequency regulation response speed of each generator set. The generator sets are sorted from slowest to fastest to obtain a frequency regulation response sorting list. For each generator set in the frequency regulation response sorting list, based on the generator set's moment of inertia coefficient... The initial offloading capacity corresponding to a generator set is calculated by using the ratio of its moment of inertia coefficient to the sum of the moment of inertia coefficients of all generator sets. Following the order of the frequency regulation response sorting list (from slowest to fastest), the initial generator tripping capacity corresponding to each generator set is allocated sequentially. Accumulate until the total accumulated capacity reaches the total capacity of the cutting machine. Specifically, let the generator sets in the sorted list be, in order: ,from Start accumulating; if the sum is less than... Then continue accumulating until the sum is greater than or equal to the sum for the first time. Until then. At this point, the allocated capacity of the last accumulated generator set is adjusted to... The difference between the selected generator set and the previous cumulative sum. The selected generator set and its corresponding allocated capacity are used as the generator switching execution sequence. .

[0067] List of load-cutting nodes and the cutting machine execution sequence This is combined into the adaptive generator tripping and load shedding control strategy. In specific implementation, the adaptive generator tripping and load shedding control strategy is sent to the stability control device corresponding to the starting substation to trigger dynamic adjustment of control parameters. The adaptive generator tripping and load shedding control strategy is encapsulated into a strategy frame conforming to the communication protocol of the stability control device. The strategy frame includes the starting substation identifier and the tripping execution sequence. and load cut-off node list The format of the strategy frame is defined according to the communication protocol of the stability control device (such as IEC61850 or a proprietary protocol), the starting substation identifier is a string or integer code, and the tripping execution sequence is... This is an ordered list, where each element contains a generator set identifier and allocated capacity value, and a list of load shedding nodes. The list is ordered, with each element containing a substation identifier and the amount of load shelving. The strategy frame is sent to the stability control device corresponding to the starting substation via a dedicated stability control communication channel. This dedicated stability control communication channel can be a fiber optic channel or a power line carrier channel. Within the stability control device, the load shelving execution sequence in the strategy frame is executed. Update the setpoint parameter table for the load shedding machine according to the load shedding node list. The low-frequency and low-voltage load shedding setting parameter table is updated. This table contains the tripping settings for each generator unit (e.g., over-frequency and low-frequency settings). The allocated capacity in the tripping execution sequence is converted into the corresponding power setting value, and then merged or replaced with the original setting value. The low-frequency and low-voltage load shedding setting parameter table contains the load shedding cycle and load shedding amount for each substation. The shedding amount from the load shedding node list is added to the load shedding setting value for the corresponding cycle. The stability control device is triggered to perform a hot-load operation on the setting parameters, so that the updated tripping setting parameter table and the low-frequency and low-voltage load shedding setting parameter table take effect without restarting the device. The hot-load operation is achieved by writing parameters to the parameter storage area in the device's memory and triggering the parameter reload flag; the device automatically applies the new parameters.

[0068] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for intelligent safety and stability early warning and control of power grids, characterized in that, The method includes: Acquire historical operational data sets from each substation in the target power grid area, including voltage phase angle fluctuation sequences, active power disturbance record sequences, and protection device action timing records; The voltage phase angle fluctuation sequence is subjected to transient energy gradient analysis to obtain the phase angle energy gradient feature vector. The active power disturbance record sequence is subjected to oscillation mode decomposition to obtain the power oscillation mode feature vector. The phase angle energy gradient feature vector and the power oscillation mode feature vector are coupled nonlinearly to generate a coupled stable feature tensor. A pre-trained graph attention deep network model is invoked to perform stability risk propagation inference processing on the coupled stable feature tensor, generating a set of stability risk propagation paths corresponding to each substation; Based on the set of stable risk propagation paths, the voltage phase angle fluctuation sequence and protection device action timing records of each substation are processed collaboratively to perform stability boundary extrapolation, thereby generating the stability runaway boundary prediction result of the target power grid area.

2. The method for intelligent safety and stability early warning and control of power grids according to claim 1, characterized in that, Perform transient energy gradient analysis on the voltage phase angle fluctuation sequence to obtain the phase angle energy gradient feature vector, including: Perform adjacent time point difference operations on the voltage phase angle fluctuation sequence to obtain the phase angle change rate sequence; The phase angle change rate sequence is subjected to sliding window energy integration to obtain the transient energy accumulation value within each time window; The gradient direction is calculated for the transient energy accumulation value to obtain an energy gradient direction sequence; The phase angle energy gradient feature vector is obtained by concatenating the energy gradient direction sequence with the phase angle values ​​at corresponding time points in the voltage phase angle fluctuation sequence.

3. The method for intelligent safety and stability early warning and control of power grids according to claim 2, characterized in that, The sliding window length used in the transient energy gradient analysis is positively correlated with the period of the dominant oscillation mode of the active power perturbation recording sequence.

4. The method for intelligent safety and stability early warning and control of power grids according to claim 1, characterized in that, The active power disturbance recording sequence is subjected to oscillation mode decomposition to obtain a power oscillation mode feature vector, including: The active power disturbance recording sequence is subjected to empirical mode decomposition to obtain multiple intrinsic mode function components; The dominant oscillation mode component is selected from the plurality of intrinsic mode function components. The dominant oscillation mode component is the intrinsic mode function component whose energy ratio exceeds a preset energy ratio threshold. The instantaneous frequency and instantaneous amplitude of the dominant oscillation mode components are extracted to obtain the oscillation frequency sequence and the oscillation amplitude sequence; The power oscillation mode feature vector is generated by alternating and combining the oscillation frequency sequence and the oscillation amplitude sequence.

5. The method for intelligent safety and stability early warning and control of power grids according to claim 1, characterized in that, The phase angle energy gradient feature vector and the power oscillation mode feature vector are coupled nonlinearly to generate a coupled stable feature tensor, including: The phase angle energy gradient feature vector is subjected to phase space reconstruction processing to obtain a phase space trajectory matrix. The rows of the phase space trajectory matrix correspond to the phase space dimension, and the columns correspond to the time steps. The power oscillation mode feature vector is subjected to Lyapunov exponential spectrum mapping to obtain the exponential spectrum feature vector. Construct an energy-oscillation cross-coupling function by performing an element-by-element nonlinear multiplication operation between the phase space trajectory matrix and the exponential spectrum eigenvector to obtain the cross-coupling matrix; The cross-coupling matrix and the phase space trajectory matrix are deeply concatenated and stacked to generate the coupled stable feature tensor.

6. The method for intelligent safety and stability early warning and control of power grids according to claim 5, characterized in that, The nonlinear dynamics feature coupling process introduces the Lyapunov exponent of chaotic systems as a coupling strength adjustment factor.

7. The method for intelligent safety and stability early warning and control of power grids according to claim 1, characterized in that, The method involves calling a pre-trained graph attention deep network model to perform stability risk propagation inference processing on the coupled stable feature tensor, generating a set of stability risk propagation paths corresponding to each substation, including: The set of stable risk propagation paths includes the distribution of the direction and intensity of the impact of stable risks among various substations. The coupled stable feature tensor is input into the node feature encoding layer of the graph attention deep network model for node embedding encoding to obtain the stable state latent feature vector corresponding to each substation. The attention coefficient calculation layer of the graph attention deep network model is invoked to calculate the risk propagation attention coefficient between any two substations based on the vector similarity and topological adjacency relationship between the steady-state latent feature vectors of any two substations. Based on the risk propagation attention coefficient, the adjacency matrix of the graph attention deep network model is updated with edge weights to obtain a weighted risk propagation graph; Perform multi-hop risk diffusion simulation processing on the weighted risk propagation graph to generate a set of stable risk propagation paths starting from each substation.

8. The method for intelligent safety and stability early warning and control of power grids according to claim 7, characterized in that, The graph attention deep network model adopts a multi-head attention mechanism, where each attention head independently calculates the risk propagation attention coefficient between substations and takes the average value as the final attention coefficient.

9. The method for intelligent safety and stability early warning and control of power grids according to claim 7, characterized in that, Based on the influence direction and intensity distribution, a collaborative stability boundary extrapolation process is performed on the voltage phase angle fluctuation sequence and protection device operation timing records of each substation to generate a stability runaway boundary prediction result for the target power grid area, including: The stability runaway boundary prediction results are used to indicate the starting substation where stability runaway is expected to occur and the runaway propagation path; A stable risk directed graph is constructed based on the influence direction and influence intensity distribution, wherein the nodes of the stable risk directed graph are substations, the direction of the directed edges is the risk propagation direction, and the weight of the directed edges is the corresponding intensity value in the influence intensity distribution. The voltage phase angle fluctuation sequence is mapped to the nodes of the stable risk directed graph as the initial value of the node state, and the action sequence record of the protection device is marked on the corresponding node as the boundary trigger condition. Starting from the node with zero out-degree in the directed graph of stability risk, backpropagation is performed to calculate the stability margin decay trajectory of each node in turn. The node that first crosses the preset lower limit of stability margin in the stability margin decay trajectory and its upstream propagation path are determined as the stability runaway boundary prediction result.

10. The method for intelligent safety and stability early warning and control of power grids according to claim 9, characterized in that, Starting from the node with zero out-degree in the directed graph of stability risk, backpropagation is performed to calculate the stability margin decay trajectory of each node in turn, including: The nodes with zero out-degree in the directed graph of the stable risk are determined as the set of starting nodes for the deduction; For each starting node in the set of starting nodes for the deduction, traverse all its predecessor nodes in the opposite direction of the directed edges; For each traversed predecessor node, the stability margin decay start time and decay slope of the predecessor node are calculated based on the weight of the directed edge between it and the successor node, the stability margin decay rate of the successor node, and the voltage phase angle fluctuation amplitude of the predecessor node. The stability margin decay start time and decay slope of each node are spliced ​​together in time sequence to generate the stability margin decay trajectory of that node.

Images