New energy black-start partition optimization method considering unconventional risk

Abstract

The invention relates to the technical field of power system stability control, and provides a new energy black-start partition optimization method considering unconventional risks, which comprises the following steps: simulating unconventional risks possibly encountered by a power system by adopting Monte Carlo sampling; extracting the output characteristics of wind power under the unconventional risk by adopting variational mode decomposition, and performing output prediction in combination with a long-short-term memory neural network; a power system recovery partition is divided based on an LPA algorithm, and the LPA algorithm is improved by considering a partition size balance principle, so that the partition area is more balanced, and the tag oscillation effect is reduced. According to the method, the black-start partition optimization model fusing new energy output uncertainty modeling and unconventional risk scene generation is designed, and the partition scheme is cooperatively solved by adopting the improved label propagation algorithm, so that the robustness and the recovery efficiency of the black-start partition are effectively improved; and rapid and stable power supply of a power system under complex disturbance can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system stability control, and particularly relates to a new energy black start partition optimization method considering unconventional risks. BACKGROUND

[0002] With the promotion of the "double carbon" goal, the energy structure of the power system is accelerating towards clean and low-carbon direction, and the installed capacity of new energy units is significantly improved. Under this background, black start is essential to ensure the safe and stable operation of the power system, and its main function is to gradually restore system operation through starting self-starting generator units after large-area power failure, and ultimately realize the process of re-powering the entire power grid. Therefore, in order to improve the efficiency of black start and better integrate new energy resources, it is necessary to divide the partition to ensure that each region can quickly and stably restore power supply after failure.

[0003] However, the current research has not fully considered the potential impact of unconventional risks in the black start process, and its suddenness and uncertainty have brought great challenges to the division of black start regions. Traditional black start region division is often based on the static structure of the system and the preset fault scenario, while the occurrence of unconventional risks may cause the above preset scenario to fail, so that the originally divided black start region cannot be started independently. At the same time, the highly destructive unconventional risks may directly affect the availability of black start power sources, causing some regions to lose starting ability.

[0004] The present application considers unconventional risks in the consideration category, deeply analyzes their specific impact on the black start process and develops a corresponding partition optimization method, thereby improving the robustness and response ability of the black start partition. SUMMARY

[0005] The present application aims to provide a new energy black start partition optimization method considering unconventional risks, which can meet the demand of improving the response ability of black start partition.

[0006] The present application is implemented as follows: a new energy black start partition optimization method considering unconventional risks, which comprises:

[0007] The Monte Carlo sampling algorithm is used to simulate the unconventional risks that the power system may encounter, including the fault lines, fault locations and fault duration caused by unconventional risks, to quantify the impact of unconventional risks on wind power output and make risk assessment;

[0008] The wind speed data of the wind power cluster is preprocessed, the variational mode decomposition is used to extract the wind power output characteristics under unconventional risks, and the long short-term memory neural network is used for output prediction to predict potential risks in advance and obtain prediction information;

[0009] The label propagation algorithm is improved, and the power system restoration partition is divided based on the improved label propagation algorithm based on the risk assessment and prediction information, so that the partition area is more balanced and the label shock effect is reduced.

[0010] Preferably, the Monte Carlo sampling algorithm is used to simulate the unconventional risks that the power system may encounter, including fault lines, fault locations and fault durations, specifically including:

[0011] A probability model of fault lines is established:

[0012] ,

[0013] ,

[0014] Wherein, p(L i ) is the probability of failure of the i-th line in the power grid; L is the fault line; L i is the length of the i-th line, i∈[1,N] and i∈Z; u1 is a random number, u1~(0,1), satisfying uniform distribution from 0 to 1; N is the total number of lines;

[0015] A probability model of fault location is established:

[0016] Three fault points with equal distance are set on each fault line, and it is ensured that the probability of failure of each fault point in a fault line is the same and independent of each other;

[0017] ,

[0018] Wherein, LF is the fault point; u2 is a random number, u2~(0,1), satisfying uniform distribution from 0 to 1;

[0019] A probability model of fault duration is established:

[0020] The fault is divided in time in the power system, including permanent fault and transient fault;

[0021] Rayleigh distribution is used to model the probability of fault time:

[0022] ,

[0023] ,

[0024] Wherein, f(t) is the probability density function; M is the scale parameter for probability modeling; m is the scale parameter; LT is the fault duration; u3 is a random number, u3~(0,1), satisfying uniform distribution from 0 to 1.

[0025] Preferably, the wind speed data of the wind power cluster is preprocessed, the wind power output features under unconventional risks are extracted by using variational mode decomposition, and long short-term memory neural network is combined for output prediction, specifically including:

[0026] Firstly, the t-SNE algorithm is used to preprocess the wind speed data of the wind power cluster, and the dimension of the time series data is reduced;

[0027] The two-dimensional coordinates after dimension reduction are decomposed by using variational mode decomposition to extract wind power features of different time scales; the wind power time series is decomposed into K components, and a variational optimization model is established:

[0028] ,

[0029] Wherein, u k is each modal function; ω k is the center frequency of each mode; is the partial derivative of time t; δ(t) is the impulse function; j is the imaginary number; u k (t) is the function relationship of u k and time; f(t) is the original signal;

[0030] Finally, the extracted features u k and ω k are input into the long short-term memory neural network for training, and a wind power prediction model suitable for unconventional risk scenarios is constructed.

[0031] Preferably, the extracted features are input into the long short-term memory neural network for training, and a wind power prediction model suitable for unconventional risk scenarios is constructed, specifically including:

[0032] ,

[0033] ,

[0034] ,

[0035] Wherein, r t , w t , l t are the gate structures of the input gate, the forget gate and the output gate respectively; σ is the activation function; W r , b r are the weight matrix and bias vector of the input gate respectively; W w , b w are the weight matrix and bias vector of the forget gate respectively; W l , b l are the weight matrix and bias vector of the output gate respectively; h t-1 ​is the hidden state input at time t-1; x t is the information input at time t;

[0036] The calculation and update of temporary memory:

[0037] ,

[0038] ,

[0039] wherein, is the temporary memory variable; W c , b c are the weight matrix and bias vector of temporary memory respectively; c t , c t-1 are the memory information at time t and time t-1 respectively;

[0040] The output of the hidden state:

[0041] ,

[0042] wherein, h t is the hidden state output at time t.

[0043] Preferably, the improved label propagation algorithm is used to divide the power system restoration partition, specifically including:

[0044] Abstract the grid topology to obtain the network topology graph G=(V, E), wherein V is the vertex set of the graph, and E is the branch set of the graph;

[0045] Calculate the branch weight based on the branch restoration time and effective power:

[0046] ,

[0047] wherein, W ij is the branch weight between node i and node j; |P ij | is the absolute value of the active power flowing through the branch between node i and node j before power failure; t ij is the restoration time of the branch between node i and node j; * is the normalization of the value;

[0048] Determine the location of the black start power supply and initialize the label classification matrix F;

[0049] Calculate the historical influence coefficient H and the community size matrix M:

[0050] ,

[0051] ,

[0052] Wherein, i, j are nodes of community size matrix M; m c is community size factor of label category c; d c is the sum of degrees of all nodes whose label category is c; l is the total number of branches in the grid;

[0053] Calculate the propagation probability matrix T:

[0054] ,

[0055] ,

[0056] Wherein, w ij is the branch weight between node i and node j; u is the range of index variable; k is the index variable of branch weight summation; is a binary indicator variable, when node i belongs to c label at t-1 round iteration, then =1, otherwise =0; is the historical influence increment of node i to c label at t round iteration; is c label of node i at t moment;

[0057] Continue to update the label classification matrix F based on the label influence matrix Y:

[0058] ;

[0059] Restore the black start power node label information;

[0060] Judge whether the label classification matrix F converges, if yes, output the partition result; if not, recalculate the geographical influence factor H and the community size matrix M.

[0061] Preferably, the label propagation algorithm is improved, and specifically includes:

[0062] ,

[0063] Wherein, f(i) is the label information of node i; N(i) is the set of adjacent nodes of node i; w i (f j ) is the sum of edge weights of all adjacent nodes of target node i and label information f j ; m(f j ) is the community size influence factor with label information f j .

[0064] Compared with the prior art, the present application has the beneficial effects that:

[0065] Compared with the traditional unimproved LPA algorithm, the new energy power grid black start partition optimization method considering unconventional risks in the present application fully combines the Monte Carlo sampling algorithm and the improved label propagation algorithm, and when dealing with the uncertainty brought by unconventional risks, not only the balance of node distribution in each partition and the consistency of the proportion of key nodes are realized, but also the starting time of the generator set is significantly shortened and the utilization efficiency of the available active power of the system is improved, thereby effectively enhancing the stability, reliability and rapid response capability of the black start process. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0067] Figure 1 is a step flow chart of the present application; Figure 2 is a basic structure schematic diagram of the LSTM unit in the embodiment of the present application; Figure 3 is a label propagation algorithm schematic diagram in the embodiment of the present application; Figure 4 is a power system restoration partition flow chart schematic diagram based on the improved LPA in the embodiment of the present application; Figure 5 is a wiring schematic diagram of the IEEE 39 system containing new energy in the embodiment of the present application; Figure 6 is a power prediction curve schematic diagram of the wind power prediction model in the embodiment of the present application; Figure 7 is a total available active power curve schematic diagram of the system before and after the improved optimization algorithm in the embodiment of the present application. DETAILED DESCRIPTION

[0068] In order to more fully understand the technical content of the present application, the technical solutions of the present application will be further introduced and described below in combination with specific embodiments, but not limited to this. The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are involved in the scope of protection of the present application.

[0069] Reference Figure 1 The new energy black start partition optimization method considering unconventional risks comprises:

[0070] Step A, Monte Carlo sampling is used to simulate the unconventional risks that the power system may encounter, including the caused fault line, location and duration, and analyze the influence of unconventional risks on wind power output;

[0071] Step B, the characteristics of wind power output under unconventional risks are extracted by using variational mode decomposition (VMD), and long short-term memory (LSTM) neural network is combined to carry out output prediction, so as to early warn potential risks and effectively reduce the influence of new energy output uncertainty on black start partitioning;

[0072] Step C, the power system restoration partition is divided based on label propagation algorithm (LPA), the LPA algorithm is improved considering the balance principle of partition size, so that the partition area is more balanced and the label shock effect is effectively reduced, and the restoration efficiency of the black start process of the power grid is accelerated.

[0073] In the embodiment of the present application, the specific operation of step A is as follows:

[0074] Monte Carlo algorithm is a numerical simulation method based on probability statistics, and its core idea is to solve complex problems through random sampling and statistical analysis. First of all, constructing a probability model is the basis of Monte Carlo method, and random variables and deterministic variables in the problem need to be handled respectively. Random variables are described by constructing the corresponding probability distribution model, which ensures that the model can effectively reflect its randomness; while deterministic variables are analyzed by mathematical methods such as calculus, linear solution, etc. Secondly, random sampling is the key to the implementation of Monte Carlo method, and random numbers are generated according to the known probability distribution model by using random number generator, usually by using pseudo-random number method to generate random numbers with the same distribution rule as the actual quantity, realizing the random sampling of zero-one uniform distribution. Finally, through multiple Monte Carlo sampling simulation, the simulation values of various random variables are established, and the sampling results are statistically analyzed, so as to obtain the fault characteristics and other statistical estimates of unconventional risks.

[0075] Since the random process of Monte Carlo method contains multiple variables, it is necessary to determine the basic variables involved in the unconventional risk process and establish its probability distribution function. The probability model of the fault location, fault type and duration caused by unconventional risks is taken as the representation, and the specific expression is as follows:

[0076] (1) Probability model of fault line

[0077] It is assumed that the fault probability of each line in the power grid is only related to the length of the line:

[0078] The probability model of fault line is established:

[0079] ,

[0080] ,

[0081] Wherein, p(L i ) is the probability of fault occurred in the i line of power grid;L is the fault line;L i The length of the i line, i∈[1,N] and i∈Z;u1 is random number, u1~(0,1), satisfy uniform distribution from 0 to 1;N is the total number of lines.

[0082] (2) the probability model of fault location

[0083] In order to facilitate calculation, the present application sets three fault points with equal distance interval in each fault line, and ensures that the probability of each fault point in a fault line is the same and independent of each other, and LF represents that a fault point has occurred, and set random number u2~(0,1), satisfy uniform distribution from 0 to 1, then the probability model of fault location is:

[0084] .

[0085] (3) the probability model of fault duration

[0086] If the fault is divided in time in power system, it can be divided into two kinds: permanent fault and transient fault, and in actual production, transient fault is more common, and Relayigh distribution is used to model the probability of fault time, which is generally suitable for describing the probability distribution of reliability and communication clutter, evolved from exponential distribution, and only contains a parameter M, which is usually 10:

[0087] ,

[0088] Set random number u3~(0,1), satisfy uniform distribution from 0 to 1;f(t) is the probability density function;M is the scale parameter for probability modeling;m is the scale parameter;

[0089] The probability model of fault duration LT is as follows:

[0090] ,

[0091] LT is the fault duration;u3 is random number, u3~(0,1), satisfy uniform distribution from 0 to 1.

[0092] The specific operation of step B of the method of the present application is as follows:

[0093] The wind power cluster data as typical time series data not only contains rich feature information, but also exhibits significant correlation between different time series segments. In the unconventional risk scenario, such as extreme weather, equipment failure and other sudden situations, higher requirements are put forward for the accuracy and robustness of wind power prediction. Therefore, the wind speed data of the wind power cluster is pretreated by using t-SNE, and the dimensionality of the time series data is reduced; then, the two-dimensional coordinates after dimensionality reduction are decomposed by using the VMD method, and the wind power features of different time scales are effectively extracted, so that the multi-level change law in the wind power signal is captured; finally, the extracted features are input into the LSTM neural network for training, and a wind power prediction model suitable for unconventional risk scenarios is constructed.

[0094] (1) Data preprocessing based on t-SNE

[0095] The wind power cluster data as typical time series data not only contains rich feature information, but also exhibits significant correlation between different time series segments. In the unconventional risk scenario, such as extreme weather, equipment failure and other sudden situations, higher requirements are put forward for the accuracy and robustness of wind power prediction. Therefore, the wind speed data of the wind power cluster is pretreated by using t-SNE, and the dimensionality of the time series data is reduced; then, the two-dimensional coordinates after dimensionality reduction are decomposed by using the VMD method, and the wind power features of different time scales are effectively extracted, so that the multi-level change law in the wind power signal is captured; finally, the extracted features are input into the LSTM neural network for training, and a wind power prediction model suitable for unconventional risk scenarios is constructed.

[0096] The visualization processing of high-dimensional data is to reduce the data segment higher than 3D to 2D or 3D, and retain the feature information contained in the high-dimensional state. The t-distributed stochastic neighbor embedding (t-Distributed Stochastic Neighbor Embedding, t-SNE) method is to replace the Gaussian distribution in the low-dimensional space with the t-distribution, and construct a probability distribution function similar to the high-dimensional space.

[0097] Let x g , x h (g, h = 1, 2, 3, …) be two points in the high-dimensional space, and x g be the center to construct a Gaussian distribution with variance σ g , p h|g be the conditional probability of x h in the neighborhood of x g :

[0098] ,

[0099] where x zFor x g is a point in the neighborhood of y

[0100] Such conditional probability can also be used to define distance in low dimensional space.

[0101] Let x g , x h be mapped into low dimensional space respectively corresponding to y g , y h , then y g is the conditional probability q h of y h|g neighborhood. The expression of q h|g is:

[0102] ,

[0103] where y z is a point in the neighborhood of y g .

[0104] In high dimensional space, p h|g ≠ p g|h , in low dimensional space q h|g ≠ q g|h . By using the joint probability distribution instead of the original conditional probability, p h|g = p g|h , q h|g = q g|h , then the expression of p h|g and q h|g is:

[0105] ,

[0106] ,

[0107] where x l , y l are the best simulation points in high dimensional space and low dimensional space respectively.

[0108] The expression of high dimensional joint probability distribution p gh is:

[0109] ,

[0110] After introducing the t distribution with 1 degree of freedom, the expression of low dimensional joint probability distribution q gh is:

[0111] ,

[0112] The t-SNE method ensures that the local and global structures of the data remain consistent before and after dimensionality reduction, preserving key feature information. Simultaneously, this method effectively reduces redundant information in the data, improves the overall data quality, and provides more reliable input for subsequent prediction models.

[0113] (2) Time series processing based on variational mode decomposition

[0114] The essence of Variational Mode Decomposition (VMD) is to transform the decomposition process into an optimization process, i.e., the solution process of a variational problem. The wind power time series is decomposed into K components, ensuring that the decomposed mode components have a center frequency and finite bandwidth; simultaneously, the sum of the estimated bandwidths of each mode is minimized, and constraints are used to ensure that the sum of all modes is equal to the original signal. The variational optimization model is shown below:

[0115] ,

[0116] Among them, u k For each mode function; ω k The center frequencies of each mode; δ(t) is the partial derivative with respect to time t; δ(t) is the impulse function; j is the imaginary number; u k (t) is u k The relationship between time and the signal; f(t) is the original signal.

[0117] (3) Wind power prediction based on LSTM neural network

[0118] In short-term wind power output prediction, the power at each predicted time point is closely related to the power data from multiple previous time points. LSTM neural networks perform well on periodic continuous time series and can better address the limitations of recurrent neural networks in remembering long-term data, as well as issues like vanishing and exploding gradients. LSTM neural networks are a variant of recurrent neural networks, utilizing the logic control of gate units to manipulate data. The basic structure of an LSTM neural network unit includes a forget gate, input gate, output gate, and memory unit, as shown in the diagram. Figure 2 As shown in Figure c. t c t-1 These are the memory information at time t and time t-1, respectively; w t r t l t These correspond to the gate structures for the forget gate, input gate, and output gate, respectively; c t x For supplementary information; σ is the activation function; x t For the information input at time t; h t h t-1is the hidden state output at time t and the hidden state input at time t-1, respectively.

[0119] The input gate is responsible for deciding which feature information in the current input and the hidden state at the last time step should be preserved, as shown in the following equation:

[0120] ,

[0121] where W r and b r are the weight matrix and bias vector of the input gate, respectively.

[0122] The forget gate is responsible for processing information in the memory cell at the last time step that should be forgotten, as shown in the following equation:

[0123] ,

[0124] where W w and b w are the weight matrix and bias vector of the forget gate, respectively.

[0125] The forget gate is responsible for processing information in the memory cell at the last time step that should be forgotten, as shown in the following equation:

[0126] ,

[0127] where W l and b l are the weight matrix and bias vector of the output gate, respectively.

[0128] The expression of the temporary memory is as follows:

[0129] ,

[0130] ,

[0131] where c t r is the temporary memory variable; W c and b c are the weight matrix and bias vector of the temporary memory, respectively; c t and c t-1 are the memory information at time t and time t-1, respectively.

[0132] The expression of the memory update is as follows:

[0133] ,

[0134] The expression of the hidden state is as follows:

[0135] .

[0136] In the embodiments of the present application, the specific operation of step C is as follows:

[0137] The present application adopts a label propagation algorithm (LPA) to optimize the partition of the black start recovery area. The LPA is a semi-supervised learning method based on graph theory, which has the characteristics of rapidness, high efficiency, no need for any prior parameters and good classification effect, and can quickly propagate labels in large-scale networks, thereby realizing the rapid division of the black start area. Meanwhile, the algorithm has good adaptability and robustness, and can adapt to different topological structures and dynamic changes of the power grid environment, ensuring the stability and reliability of the partition result. In addition, the LPA algorithm updates the label based on local information, effectively reducing the dependence on global information, so that the algorithm is more efficient in actual black start partition application.

[0138] The basic idea of the LPA algorithm is to propagate label information between nodes. Each node not only propagates its own label information, but also is affected by the continuous updating of the label information of all adjacent nodes. Label information often communicates and propagates between closely related nodes. The algorithm predicts the labels of unmarked nodes according to the labels of marked nodes and iterates in a loop until the labels of each node no longer change, and nodes with the same label information are classified into the same community. The process of label propagation between nodes is shown in Figure 3 . Initially, each node may have no label or a unique label. After iterative propagation, all nodes in the community will have the same label information after stabilization.

[0139] The specific steps of the LPA algorithm are as follows:

[0140] First, the black start power grid contains l initial marked nodes and u unmarked nodes, and there are C marked categories. The weight of the edge is calculated according to the following formula. The smaller the distance between two points, the greater the weight corresponding to the edge, and the higher the similarity between nodes:

[0141] ,

[0142] where W ij is the weight of the edge between node i and node j; d ij is the Euclidean distance from node i to node j; the parameter σ affects the size of the weight; D is the dimension number of node features; and are the dth dimension feature value of node i and the dth dimension feature value of node j, respectively;

[0143] The propagation probability matrix T is calculated as shown in the following formula. The T matrix measures the probability of label information propagation between nodes and is determined by the size of the edge weight.

[0144] ,

[0145] where T ij is the probability of the label information of node j propagating to node i; W ij is the branch weight between node i and node j; u is the range of the index variable; k is the index variable of the branch weight summation; is a binary indicator variable, when node i belongs to the c-th label at the t-1th iteration, then = 1, otherwise = 0; is the historical influence increment of node i on the c-th label at the tth iteration; is the c-th label of node i at the tth iteration.

[0146] The label classification matrix F of (l+u) rows and C columns is defined as follows, the F matrix records the label classification information of all nodes in the network, F ic is an element in the classification matrix F, when node i belongs to the C-th label, F ic equals 1, otherwise 0. The label classification matrix F is calculated according to the initial node label.

[0147] .

[0148] The label information of each node starts the propagation process according to the propagation probability, as shown in the following formula:

[0149] ,

[0150] where Y is the label influence matrix, representing the influence caused by the label update of each node in this round of propagation; Y t is the label influence matrix of the tth round of label propagation process; F t-1 is the label classification matrix at the t-1th iteration;

[0151] The label classification matrix F is updated, and the label propagation process influence matrix Y.

[0152] The label classification matrix F is updated according to the label Figure 3 in the figure, the numbers represent the label categories of the nodes, which are used to show the update and convergence process of the node labels in the label propagation algorithm. The label propagation process influence matrix Y. If the element of the i-th row and the c-th column of Y t has the maximum value, then the F ic element in the corresponding label classification matrix is set to 1, and the other elements of the i-th row of the F matrix are set to 0, so that the label information of each node can be updated.

[0153] If the change degree of the label classification matrix in the adjacent two iterations is less than a preset threshold, it is determined that the label classification matrix converges, and the partition result is output; if not, the geographical influence factor H and the community size matrix M are recalculated.

[0154] The above updating process is repeatedly performed until the label classification matrix F converges, and finally the final label information of each node and the community classification result are obtained.

[0155] After the accident of the power system, serial and parallel recovery can be performed. In actual large-scale power grids with many nodes and complex topologies, multiple black-start power sources need to be configured for the reliability of the power system. After power failure, multiple black-start power sources with self-starting capability provide initial recovery power, and parallel partition recovery is performed, and each partition is independently recovered before synchronous interconnection operation, thereby improving the recovery speed and efficiency of the entire complex large-scale power grid.

[0156] The parallel partition recovery of the power system consists of recovery partition division, partition self-recovery and partition grid connection operation, and the basic principles of the partition considered by the application are as follows:

[0157] (1) The partition has self-recovery capability;

[0158] (2) All partitions are centered on black-start power sources as much as possible;

[0159] (3) The network within the partition is tight, and the connection between the partitions is sparse;

[0160] (4) The number of node lines in each partition is as equal as possible.

[0161] In actual parallel partition recovery of the power system, the speed and reliability of recovery are crucial. The recovery sub-partition method adopted by the application not only considers the reasonable division of the network topology, but also considers the recovery time within the partition, thereby accelerating the independent recovery of the partition as much as possible. At the same time, combined with the system operation condition, the available active power of the system in the black-start recovery process is improved to improve the reliability of parallel connection.

[0162] First, the grid topology is abstracted to obtain a network topology graph G=(V,E), where V represents the vertex set of the graph, and E represents the branch set of the graph. In the application, the branch weight is set by considering the recovery time of the branch and the active power flowing through each branch before power failure, as shown in the following formula. The shorter the branch recovery time, the greater the weight, and the two nodes connected by the branch are more likely to be classified into the same partition. At the same time, the tidal data before power failure intuitively reflects the tightness of the connection of each node in the power system. The greater the active power value transmitted on the branch, the tighter the connection between the two nodes connected by the branch.

[0163] ,

[0164] wherein W ij is the branch weight between node i and node j; |P ij | is the absolute value of the active power flowing through the branch between node i and node j before the outage; t ij is the recovery time of the branch between node i and node j; and * is the normalization of the value.

[0165] Suppose that there are n nodes in the power grid and k black-start power sources (located at nodes x1, x2, …, x k The node where the black-start power source is located is used as the initial marked node, and the label information remains unchanged during the entire propagation process. A value is assigned to the initial phase label category matrix F, as shown in the following formula. During the propagation process, the label information of the remaining power sources in the power grid that are not used for black-start (non-black-start power sources) and loads needs to be updated and obtained, so that all elements are zero except the elements involved in the following formula.

[0166] .

[0167] The label propagation strategy of the traditional LPA algorithm is shown in the following formula:

[0168] ,

[0169] In the formula, f(i) represents the label information of node i; C represents the type of label information; N(i) represents the set of adjacent nodes of node i; w i (f j ) represents the sum of edge weights of all adjacent nodes of target node i and label information f j .

[0170] Due to the small number of nodes containing label information in the initial network, the early propagation process has high randomness and instability. Therefore, part of the label information is easily propagated along the path endlessly, leading to a high proportion of a certain label information in the network, thereby causing the emergence of a large community. In addition, if there is a small community near a large community, the small community is easily annexed, leading to further expansion of the large community and causing a vicious cycle, which does not meet the principle that the size of each subzone required for power system restoration partitioning is roughly balanced. At the same time, the traditional propagation strategy may cause label oscillation, leading to difficulty in algorithm convergence and affecting the stability of the partitioning method. The present application improves the traditional LPA algorithm by introducing a community size matrix and a historical influence factor into the propagation strategy of power system restoration subzone partitioning to solve the above problems.

[0171] The community size coefficient is defined as shown in the following formula:

[0172] ,

[0173] where m c denotes the community size factor of label class c; d c denotes the sum of degrees of all nodes with label class c; l is the total number of branches in the grid.

[0174] The community size matrix M is shown in the following formula, and the diagonal elements are the corresponding community size factor values.

[0175] ,

[0176] The M matrix represents the size of the existing community in the system; the larger the community size of the label, the smaller the corresponding size factor. Before each round of label propagation, the community size matrix of the current grid is calculated and incorporated into the node label update. For the label information with a large existing community size, the propagation ability of the label is limited by the factor value, thereby ensuring that the size of each recovered sub-region obtained from the final division is roughly equal.

[0177] The improved label propagation strategy is shown in the following formula:

[0178] ,

[0179] where m(f j ) denotes the community size influence factor of the node with label information f j .

[0180] In addition, the traditional LPA algorithm is prone to label oscillation because the diagonal elements of the propagation probability matrix are all zero, indicating that the node does not consider its own historical label information when performing label update, which affects the stability and partition effect of the algorithm. Therefore, the historical influence coefficient H is defined as shown in the following formula:

[0181] ,

[0182] where denotes the c-class label at node i at time t; is the historical influence factor.

[0183] The propagation probability matrix is improved by considering the historical influence factor of the node, as shown in the following formula:

[0184] ,

[0185] Considering the community size matrix and the historical influence factor, the node propagation process is shown in the following formula:

[0186] ;

[0187] In summary, the flow chart of the power system restoration partitioning method based on the improved LPA algorithm is shown in Figure 4 .

[0188] To verify the effect of the application, the following examples illustrate the advancement of the proposed new energy black start partition optimization method considering unconventional risks.

[0189] This section takes the IEEE 10-machine 39-node system as a simulation model to verify the proposed black start partition optimization model with new energy units. The topological structure of the system is shown in Figure 5 , which consists of 10 generator units, 46 transmission lines, 12 transformers, and 19 load nodes. The specific simulation settings and parameter settings are as follows: the generator unit number is consistent with the node number; the unit startup and bus restoration time are synchronized; the transient frequency response coefficient of thermal power units is 0.105, and the detailed parameters are shown in Table 1. The wind farms are connected at nodes 36 and 37, each with an installed capacity of 400 MW, equipped with 1.5 MW DFIG wind turbines, and the specific parameters are shown in Table 2. In addition, the restoration time of all transmission lines is set to 4 minutes. During the entire simulation process, all units will strictly follow the output curve for power generation.

[0190] Table 1 Simulation parameters of thermal power units

[0191]

[0192] Table 2 DFIG parameters

[0193]

[0194] To verify the effectiveness of the proposed wind power prediction algorithm, the predicted wind power curve considering unconventional risks is compared with the actual wind power curve, as shown in Figure 6 . By comparing the curves in Figure 6 , it can be seen that the prediction results of the proposed algorithm have high consistency with the actual wind power trend. In local details, the proposed algorithm can also accurately predict the peak and valley values of wind power, which is of great significance for the scheduling and operation of wind farms. Further analysis shows that the proposed algorithm can effectively handle abnormal data and reduce the impact of unconventional risks on prediction results. In addition, the robustness of the algorithm ensures stable prediction performance in the face of uncertainty factors.

[0195] Taking the failure of the key node as the consequence of unconventional risks, the time when all generator units are started is taken as the black start restoration time of the entire system, and the simulation results under different failures are shown in the following table.

[0196] Table 3 System restoration time under unconventional risks

[0197]

[0198] From the above table, it can be seen that the recovery time of the system is taken as the evaluation index. When the key node 16 fails due to the unconventional risk, the recovery time is the longest, and the effect of the black start of the power system is the largest. Therefore, this scenario is selected as the main research object for subsequent analysis.

[0199] To verify the superiority of the proposed algorithm, the black start partition results of the LPA algorithm before and after improvement are compared, as shown in Tables 4 and 5.

[0200] Table 4 LPA algorithm without improvement

[0201]

[0202] Table 5 LPA algorithm improved in this paper

[0203]

[0204] From the comparison in Tables 4 and 5, it can be seen that in the partition optimization results of the LPA algorithm without improvement, regions 1, 2 and 3 contain 8, 12 and 9 nodes respectively, and the number of key nodes in each region is 4. Due to the uneven size of the partition, the proportion of key nodes in some regions is relatively high, which leads to unreasonable resource allocation and scheduling in the black start process. The improved LPA algorithm shows more balanced results in partition optimization. Regions 1, 2 and 3 contain 9, 10 and 10 nodes respectively, and the number of key nodes in each region remains 4. The proportion of key nodes in different regions is closer, which improves the stability and reliability of the black start process.

[0205] Taking the start-up time of all generator units as the evaluation index, the partition optimization performance of the LPA algorithm before and after improvement is compared, as shown in Table 6.

[0206] Table 6 Unit recovery time under different algorithms

[0207]

[0208] As can be seen from the comparison data in Table 6, the improved LPA algorithm exhibits a more efficient node allocation strategy when dealing with the problem of black-start partitioning under unconventional risks, making the start-up time of each generator unit more balanced. Specifically, the start-up time of node 38 in region 1 is shortened from 21 min to 16 min, node 34 in region 2 is reduced from 30 min to 25 min, and node 36 in region 3 is decreased from 70 min to 62 min, alleviating the situation of some units waiting for start-up for too long a time and improving the overall start-up efficiency of the black-start unit. At the same time, the improved algorithm improves the energy utilization efficiency of the black-start system through more reasonable partitioning, which is of great significance to the long-term operation and economy of the power system.

[0209] At the initial stage of black-start, the system faces the problems of limited resources and high load uncertainty, and the quality of the partitioning strategy directly affects the efficiency and stability of the system recovery. In addition, the results at the initial stage of black-start have significant visibility, and any improvement or defect of the partitioning strategy will be directly reflected in the system recovery speed and stability. Therefore, to further verify the superiority of the proposed algorithm, the total available active power of the system at the initial stage of black-start is taken as the evaluation index, and the partitioning optimization performance before and after the improvement of the LPA algorithm is compared, as shown in Figure 7

[0210] From the comparison curve of Figure 7 , through accurate partitioning optimization, the improved LPA algorithm can more effectively utilize the active resources in the system, and the total available active power is increased by 156 MW compared with the non-optimized partitioning strategy at 24 min, ensuring the system to quickly and smoothly transit to the normal operating state, verifying the superior performance of the proposed algorithm in the black-start recovery process.

[0211] It should be noted that: the electrical components appearing in the present application are all connected with the main controller and 220V mains, and the main controller can be a conventional known device for controlling servo motors, touch sensors, processors, alarm modules and driving modules, the standard parts used in the present application can be purchased from the market, the specific connection mode of each part is connected by conventional means such as bolts, rivets and welding in the prior art, and the mechanical parts and devices are of conventional types in the prior art, and the circuit connection adopts the conventional connection mode in the prior art, which will not be described in detail here.

[0212] The above-described embodiments are only a part of the embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes within the technical range disclosed in the present application according to the technical solution and inventive concept of the present application, which should be covered in the protection scope of the present application. The parts not involved in the present application are the same as or can be realized by the prior art.​

Claims

1. A new energy black-start partitioning optimization method considering unconventional risks, characterized in that, The method includes: The Monte Carlo sampling algorithm is used to simulate unconventional risks that the power system may encounter, including fault lines, fault locations and fault durations caused by unconventional risks, to quantify the impact of unconventional risks on wind power output and make a risk assessment. The wind speed data of the wind power cluster is preprocessed, and the wind power output characteristics under unconventional risks are extracted by variational mode decomposition and combined with long short-term memory neural network to predict the output, so as to predict potential risks in advance and obtain predictive information. An improved tag propagation algorithm is used to divide the power system recovery zones based on the aforementioned risk assessment and prediction information, making the zones more balanced and reducing tag oscillation effects.

2. The new energy black-start partition optimization method considering unconventional risks as described in claim 1, characterized in that, The Monte Carlo sampling algorithm is used to simulate unconventional risks that power systems may encounter, including faulty lines, fault locations, and fault durations, specifically including: Establish a probabilistic model for the faulty circuit: ， ， Among them, p(L i Let be the probability of a fault occurring on the i-th line in the power grid; L be the faulty line; L i Let be the length of the i-th line, i∈[1,N] and i∈Z; u1 is a random number, u1~(0,1), which satisfies a uniform distribution from 0 to 1; N is the total number of lines; Establish a probabilistic model of the fault location: Three fault points with equal distances are set up on each faulty line, and it is ensured that the probability of each fault point failing is the same and that they are independent of each other within a faulty line. ， Where LF is the fault point; u2 is a random number, u2~(0,1), which satisfies a uniform distribution from 0 to 1; Establish a probabilistic model for the duration of the failure: In power systems, faults are classified according to time, including permanent faults and transient faults; The Rayleigh distribution is used to probabilistically model the failure time: ， ， Where f(t) is the probability density function; M is the scale parameter used for probability modeling; m is the scale parameter; LT is the fault duration; u3 is a random number, u3~(0,1), which satisfies a uniform distribution from 0 to 1.

3. The new energy black-start partition optimization method considering unconventional risks as described in claim 1, characterized in that, The process of preprocessing wind speed data from wind power clusters, extracting wind power output characteristics under unconventional risks using variational mode decomposition, and then combining this with a long short-term memory neural network for output prediction specifically includes: First, the t-SNE algorithm is used to preprocess the wind speed data of the wind power cluster to reduce the dimensionality of the time series data; Variational mode decomposition is used to decompose the dimensionality-reduced two-dimensional coordinates to extract wind power characteristics at different time scales; the wind power time series is decomposed into K components, and a variational optimization model is established. ， Among them, u k For each mode function; ω k The center frequencies of each mode; δ(t) is the partial derivative with respect to time t; δ(t) is the impulse function; j is the imaginary number; u k (t) is u k The functional relationship with time; f(t) is the original signal; Finally, the extracted features u k ω k The data is fed into a long short-term memory neural network for training to build a wind power prediction model adapted to unconventional risk scenarios.

4. The new energy black-start partition optimization method considering unconventional risks as described in claim 1, characterized in that, The process of inputting the extracted features into a long short-term memory neural network for training to construct a wind power prediction model adapted to unconventional risk scenarios specifically includes: ， ， ， Where, r t w t l t These are the gating structures for the input gate, forget gate, and output gate, respectively; σ is the activation function; W r b r These are the weight matrix and bias vector of the input gate, respectively; W w b w These are the weight matrix and bias vector of the forget gate, respectively; W l b l These are the weight matrix and bias vector of the output gate, respectively; h t-1 The hidden state input at time t-1; x t Input information at time t; Calculation and updating of temporary memory: ， ， in, W is a temporary memory variable. c b c These are the weight matrix and bias vector for temporary memory, respectively; c t c t-1 These are the memory information at time t and time t-1, respectively. Output in hidden state: ， Among them, h t Output the hidden state at time t.

5. The new energy black-start partition optimization method considering unconventional risks as described in claim 1, characterized in that, The improved label propagation algorithm is used to divide the power system recovery zones, specifically including: Abstract the grid topology to obtain the network topology graph G=(V,E), where V is the set of vertices of the graph and E is the set of branches of the graph; Calculation of branch weights based on branch recovery time and effective power: ， Among them, W ij Let |P| be the branch weight between node i and node j; ij | represents the absolute value of the active power flowing through the branch between node i and node j before the power outage; t ij The recovery time of the branch between node i and node j; * represents the normalization of the value; Determine the location of the black-start power source and initialize the label classification matrix F; Calculate the historical influence coefficient H and the community size matrix M: ， ， Where i and j are nodes of the community size matrix M; m c Community size factor for tag category c; d c is the sum of the degrees of all nodes with label category c; l is the total number of branches in the grid; Calculate the propagation probability matrix T: ， ， Among them, w ij Let be the branch weight between node i and node j; u is the range of the index variable; k is the index variable for summing the branch weights; Let be a binary indicator variable. If node i belongs to label class c at iteration t-1, then... =1, otherwise =0; This represents the historical impact increment of node i on the class c label during iteration t. Let the label be of class c at node i at time t; The label classification matrix F is further updated based on the label influence matrix Y: ； Restore the black boot power node label information; Determine whether the label classification matrix F has converged. If it has, output the partitioning results; otherwise, recalculate the geographical influence factor H and the community size matrix M.

6. The new energy black-start partition optimization method considering unconventional risks as described in claim 5, characterized in that, The improvements to the label propagation algorithm specifically include: ， Where f(i) is the label information of node i; N(i) is the set of neighboring nodes of node i; w i (f j ) represents the target node i and the label information f. j The sum of the edge weights of all adjacent nodes; m(f j ) is for those with tag information f j Community size is an influencing factor.

Images