Distributed photovoltaic cluster division method considering frequency supporting capability

Abstract

The invention discloses a distributed photovoltaic cluster division method considering frequency supporting capacity, and belongs to the field of novel power system frequency stabilization and cluster dynamic division. According to the method, a comprehensive cluster division index is formed by combining a system power flow operation result and comprehensively considering multi-dimensional indexes such as node sensitivity, frequency modulation response intensity and frequency modulation response depth. On the basis, a self-organizing mapping neural network is adopted to carry out optimization division on a photovoltaic cluster, the characteristic relation in the cluster is automatically learned, and a global optimal division result is ensured. In addition, the clustering effect is evaluated through quantization errors, neural network parameters are further optimized, and the cluster division precision and the system frequency support reliability are improved. According to the method, the defects of randomness, frequency fluctuation influence and classification precision of an existing photovoltaic cluster division method are effectively overcome, and a powerful guarantee is provided for the safety and the operation efficiency of a power system.

Description

Technical Field

[0001] The present invention relates to the field of frequency stability and cluster dynamic division of a novel power system, and in particular to a distributed photovoltaic cluster division method taking frequency support capability into account. Background Art

[0002] With the enhancement of environmental awareness and the continuous advancement of technology, distributed photovoltaic power generation has gradually become an important force in promoting the green energy revolution with its advantages of high flexibility, convenient installation, and proximity to user load centers. However, with the large-scale access of distributed photovoltaics, the system frequency has also continued to decrease. In order to maintain the safe and stable operation of the power system, distributed photovoltaics are required to have certain frequency regulation capabilities. At the same time, due to the dispersion and uncertainty of photovoltaic power generation, the analysis of a single photovoltaic contains many uncertainties, so it is necessary to cluster the distributed photovoltaics in the area and analyze their common characteristics.

[0003] At present, the classification differences of photovoltaic clusters are mainly concentrated in the selection of classification indicators and the application of classification algorithms. When selecting the classification indicators of distributed photovoltaic clusters, the intuitive and easy-to-understand electrical distance of each cluster is usually used as the main criterion, and the relevant factors affecting the photovoltaic power generation capacity such as the duration and intensity of light are combined as auxiliary indicators to construct a comprehensive photovoltaic cluster classification system through comprehensive analysis. As for the classification algorithm, simple and efficient clustering algorithms such as K-means are generally used. Some researchers will also improve the algorithm according to specific circumstances to improve the classification effect. This process involves converting the collected data into comprehensive indicators and inputting them into the corresponding classification algorithm to obtain the final classification results of the photovoltaic cluster.

[0004] However, the above-mentioned PV cluster classification method fails to fully consider the inherent randomness and intermittent characteristics of PV clusters, which may lead to fluctuations in system frequency and thus trigger a series of risks in PV cluster grid connection. In addition, the selection of classification indicators is too rigid and fails to fully consider the fluctuations in supply and demand power within PV cluster nodes and the dynamic changes in supply and demand relationships between clusters. Furthermore, the classification method is relatively simple, the K-means algorithm has limited flexibility in selecting the number of clusters, and its iterative results are only local optimal solutions, lacking global rationality, which affects the accuracy and practicality of the classification results. Summary of the invention

[0005] In view of the shortcomings of the prior art in the above background technology, the present invention proposes a method for dividing distributed photovoltaic clusters taking into account frequency modulation capability. The present invention fully considers the characteristics of system frequency fluctuation after distributed photovoltaic grid connection, incorporates frequency modulation indicators into cluster division indicators, and realizes distributed photovoltaic cluster division through self-organizing mapping neural network.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] A distributed photovoltaic cluster division method taking into account frequency support capability, the steps are as follows:

[0008] Step S1, calculating the distributed photovoltaic cluster division index; the distributed photovoltaic cluster division index includes the node sensitivity index δ, the frequency modulation response strength index K, and the frequency modulation response depth index and the comprehensive cluster partitioning index γ;

[0009] Step S2: Based on the comprehensive cluster division index γ proposed in step S1, the self-organizing map neural network algorithm is used to take the division index of each distributed photovoltaic cluster as input, and the node with the smallest distance is selected as the winning node through the similarity between the competition layer neuron and the input data. The weight is continuously updated until the iteration ends to obtain the final cluster division result;

[0010] Step S3, based on the final cluster division result obtained in step S2, the clustering precision and accuracy are calculated by calculating the quantization error of the clustering result, and the relevant parameters of the self-organizing map neural network are reset to achieve the optimal division result.

[0011] Furthermore, the node sensitivity index δ in step S1 is calculated as follows:

[0012] δ ij =S Pii -S Pij +S Qii -S Qij

[0013] Among them, δ ij is the node sensitivity index of each node in the distribution network, representing the comprehensive impact of the power change of node j on node i, δ ij The smaller it is, the greater the impact of node j’s power change on node i, which is consistent with the characteristics of distance change; S P , S Q are active sensitivity and reactive sensitivity, respectively, which are used to characterize the change rate of active power and reactive power injection when the voltage fluctuates, and S P =(N-HJ -1 L) -1 ; S Q =(L-JH -1 N) -1 ; S Pii S represents the impact of node i voltage change on the node’s own active power injection change rate; Pij S represents the impact of node j voltage change on the active power injection change rate of node i; QiiS represents the impact of node i voltage change on the node’s own reactive power injection change rate; Qij It indicates the influence of the reactive power injection change rate of node i when the voltage of node j changes. The difference between the influence of active power regulation and reactive power regulation on voltage is reflected by the size of the sensitivity coefficient, and no additional weight coefficient needs to be added;

[0014] In the above active sensitivity and reactive sensitivity, H, N, J, and L are used to simplify the partial derivatives of active power and reactive power with respect to phase angle and voltage in the Jacobian matrix, respectively. P is active power, Q is reactive power, θ is the node voltage phase angle, and V is the node voltage amplitude.

[0015] Furthermore, the frequency modulation response strength index K in step S1 is used to characterize the influence of photovoltaic output on system frequency, and is calculated as follows:

[0016]

[0017] Among them, ΔP is the change in the output of photovoltaic access to the distribution network, and Δf is the change in system frequency when the distributed photovoltaic output changes.

[0018] Furthermore, the frequency modulation response depth index in step S1 is It is used to characterize the deviation between the theoretical frequency modulation output and the actual photovoltaic output of distributed photovoltaic power generation. The calculation method is:

[0019]

[0020] Among them, P m is the rated output, P pν is the current actual output, and S is the current light intensity.

[0021] Furthermore, the comprehensive cluster division index γ in step S1 is a distributed photovoltaic cluster node sensitivity index δ, frequency modulation response strength index K, frequency modulation response depth index The calculation method of the comprehensive cluster division index γ is:

[0022]

[0023] Where λ1, λ2, and λ3 are the weight coefficients of the node sensitivity index, the frequency modulation response contribution strength index, and the frequency modulation response contribution depth index, respectively.

[0024] Furthermore, the steps of the self-organizing map neural network algorithm proposed in step S2 are as follows:

[0025] Step S2-1: normalize the comprehensive cluster division index obtained in step S1 and use it as the input of the self-organizing map neural network, and initialize the neurons at the same time to carry out the next step of data learning, as shown below:

[0026]

[0027] in, It is the input of the self-organizing map neural network, which is calculated from the distributed photovoltaic cluster division index data; The neurons of the self-organizing map neural network are initially assigned random numbers between 0 and 0.1;

[0028] Step S2-2: Calculate the connection weight matrix W between each neuron j ,from A value is randomly selected and input into the competition layer for neuron competition learning. The distance between the weight node and the weight vector of each neuron in the competition layer is calculated. The distance uses the Euclidean distance to measure the similarity. The neuron with the highest similarity distance is selected as the winning neuron, as shown below:

[0029]

[0030] Among them, X i For the input dataset The i-th element in , D is the dimension of the competition layer neural network, W j is the weight vector of the jth neuron in the competition layer. The weight vector is an attribute contained in each neuron, which is used to describe the position of the neuron in the input space;

[0031] Step S2-3, taking the winning neuron obtained in step S2-2 as the center of the circle, according to a preset neighborhood distance threshold, screen out all nodes whose distance to the node is less than the threshold, thereby forming a winning neighborhood. The winning neighborhood range formula is as follows;

[0032]

[0033] Among them, R(l) is the neighborhood radius at the lth iteration, R max , R min are the maximum and minimum neighborhood radius thresholds, respectively, and L is the total number of iterations;

[0034] Step S2-4, in the winning neighborhood, iteratively adjust the weight of each node; the principle of adjustment is: the closer the node is to the winning node, the greater the amplitude of its weight update; conversely, the farther the distance is, the smaller the update amplitude; according to this principle, set an update constraint condition for different nodes, as shown below:

[0035]

[0036] Among them, R k is the learning rate of the kth iteration, G ij is the update constraint of the jth neuron node in the winning neighborhood corresponding to the i-th sample;

[0037] Step S2-5, repeat the iterations of steps S2-2 to S2-4 until the upper limit of the number of iterations is reached, and finally output the result of the distributed photovoltaic cluster division.

[0038] Furthermore, the step S3 is specifically as follows:

[0039] Step S3-1, using the average value of the distance between the weight node and the weight vector of each neuron in the competition layer calculated in step S2-2 as a reference, the quantization error calculation formula is as follows;

[0040]

[0041] Among them, c j (x) is the quantization error value at the jth iteration, J is the total number of neurons, M j is the weight vector of all neurons at the jth iteration;

[0042] Step S3-2: The total number of neurons in the self-organizing map neural network in the final result of the distributed photovoltaic cluster partition obtained in step S2-5 and the input data set The i-th element in and the weight vector of all neurons at the j-th iteration are used as input, and the final quantization error value is calculated by the calculation formula obtained in step S3-1;

[0043] Step S3-3, change the initial self-organizing map neural network parameters, modify the dimension parameters and then repeat steps S3-1 and S3-2 to calculate the quantization errors under multiple parameters to select the optimal parameter settings.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) The present invention proposes a distributed photovoltaic cluster division method that takes frequency support capacity into account. Combined with the distribution network flow operation results, the mutual influence between the voltage and power between the distribution network nodes under the distributed photovoltaic access is analyzed to determine the node sensitivity index. At the same time, the comprehensive cluster division index is established in combination with the frequency modulation capability. The final cluster division result is obtained through the self-organizing mapping neural network. The multi-dimensional indicators such as node sensitivity, frequency modulation response strength and frequency modulation response depth are comprehensively considered, which can more comprehensively reflect the actual dynamic characteristics of the photovoltaic cluster and significantly improve the accuracy and applicability of cluster division.

[0046] (2) In order to address the problem that existing classification algorithms are not flexible enough in selecting the number of clusters and can only obtain local optimal solutions, the present invention proposes a distributed photovoltaic cluster division method that takes frequency support capability into account. The method uses a self-organizing map neural network for clustering, which can automatically learn the characteristic relationships within the cluster and provide a globally optimal division scheme, thereby effectively reducing the risk of system frequency fluctuations on photovoltaic grid connection.

[0047] (3) Compared with other methods, the distributed photovoltaic cluster division method proposed in this invention takes into account the frequency support capacity. It fully considers the scenario of distributed photovoltaic grid-connected by retaining spare capacity to participate in system frequency regulation after considering the conventional power flow results. The clustering effect is evaluated by quantization error after clustering by self-organizing map neural network, so as to achieve the purpose of optimal division.

[0048] (4) Compared with the prior art, the distributed photovoltaic cluster division method proposed in the present invention takes into account the frequency support capability, further optimizes the neural network parameters by performing quantitative error calculation on the clustering results, and ensures higher accuracy and better stability of the cluster division, thereby providing a more reliable foundation for the frequency support of the power system and improving the safety and operation efficiency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flow chart of a distributed photovoltaic cluster division method taking frequency support capability into account according to the present invention;

[0050] Figure 2 A flow chart of the self-organizing map neural network proposed by the present invention;

[0051] Figure 3 Quantization error result graph for the self-organizing map neural network;

[0052] Figure 4 Result diagram of distributed photovoltaic cluster division. DETAILED DESCRIPTION

[0053] The present invention is described below based on embodiments in conjunction with the accompanying drawings of the specification, but the present invention is not limited to these embodiments. In the following detailed description of the present invention, some specific details are described in detail. For those skilled in the art, the present invention can be fully understood without the description of these details.

[0054] A distributed photovoltaic cluster division method taking into account frequency support capability, the steps are as follows:

[0055] Step S1: In order to study the coupling relationship between the output change of distributed photovoltaic power and the distribution network data in a refined manner, a distributed photovoltaic cluster division index is proposed, including the node sensitivity index δ of the distributed photovoltaic cluster combined with the results of the power flow operation. At the same time, considering the impact of distributed photovoltaic access on the distribution network frequency, a frequency modulation response intensity index K and a frequency modulation response depth index are proposed. Finally, the comprehensive cluster division index γ is obtained by attaching different weights to the indicators;

[0056] Step S2: Based on the comprehensive cluster division index γ proposed in step S1, the self-organizing map neural network is used to take the division index of each distributed photovoltaic cluster as input, and the node with the smallest distance is selected as the winning node through the similarity between the competition layer neuron and the input data. The weight is continuously updated until the iteration ends to obtain the final cluster division result;

[0057] Step S3, based on the final cluster division result obtained in step S2, the clustering precision and accuracy are calculated by calculating the quantization error of the clustering result, and the relevant parameters of the self-organizing map neural network are reset to achieve the optimal division result.

[0058] Furthermore, the distributed photovoltaic cluster division index system in step A is constructed as follows:

[0059] A1. Due to the access of distributed photovoltaics, the voltage and power index between distribution network nodes will fluctuate. However, due to the dispersion and uncertainty of distributed photovoltaics, the influence relationship between active and reactive power between nodes under the distribution network flow is introduced, as shown in formula (1):

[0060]

[0061] Where: ΔP is the change in active power injected into the node; ΔQ is the change in reactive power injected into the node; Δθ is the change in node voltage phase angle; ΔV is the change in node voltage amplitude; J Pθ , J PV , J Qθ , J QV is the Jacobian matrix

[0062] Since the voltage amplitude has a higher coupling degree with reactive power and a lower coupling degree with active power, let ΔP = 0, we get

[0063]

[0064]

[0065] In the formula, is the part of the Jacobian matrix related to reactive power and voltage amplitude; It is the voltage / reactive power sensitivity matrix, which reflects the sensitivity of the voltage amplitude to the injected reactive power.

[0066] A2. Based on the active and reactive sensitivities of the nodes mentioned in the previous step, the simplified system decoupling equation can be obtained by numerically equivalence of the Jacobian matrix, as shown in formula (4):

[0067]

[0068] After expansion

[0069]

[0070] Eliminating Δθ, we get

[0071] ΔU=(N-HJ -1 L) -1 ΔP+(L-JH -1 N) -1 ΔQ=S P ΔP+S Q ΔQ (6)

[0072] Thus, the node active and reactive sensitivity is extracted to obtain the final node sensitivity index δ ij , specifically as formula (7)

[0073] δ ij =S Pii -S Pij +S Qii -S Qij (7)

[0074] In the formula, δ ij Characterizes the comprehensive impact of the power change of node j on node i, δ ij The smaller it is, the greater the impact of the power change of node j on node i, which is consistent with the distance change characteristics. In addition, the difference in the impact of active and reactive regulation of node injection on voltage is reflected by the size of the sensitivity coefficient, and there is no need to increase the weight coefficient.

[0075] A3. Considering the voltage and power sensitivity of the system nodes, the system frequency will fluctuate accordingly when distributed photovoltaic access is considered. Therefore, a comprehensive frequency modulation index is proposed. This index consists of the frequency modulation response contribution intensity index K and the frequency modulation response contribution depth index. The frequency modulation response contribution intensity index K analyzes the fluctuation of the existing active power and the corresponding frequency change, and proposes the influence of the photovoltaic output on the system frequency, as shown in formula (8).

[0076]

[0077] Where ΔP is the change in the output of photovoltaic access to the distribution network, and Δf is the change in system frequency when the distributed photovoltaic output changes.

[0078] In addition to the frequency modulation response contribution intensity, the frequency modulation response contribution depth is also worth considering. There must be a deviation between the theoretical frequency modulation output of distributed photovoltaic output and the actual photovoltaic output. Therefore, exploring the size of the deviation is crucial to the effect of frequency modulation, as shown in formula (9).

[0079]

[0080] Where P m is the rated output, P pν is the current actual output and S is the current light intensity.

[0081] A4. The node sensitivity index, frequency modulation response strength index and frequency modulation response depth index mentioned above are weighted to obtain the final cluster division result, as shown in formula (10).

[0082]

[0083] Where λ1, λ2, and λ3 are the weight coefficients of the node sensitivity index, the frequency modulation response contribution strength index, and the frequency modulation response contribution depth index, respectively.

[0084] Further, the process of using the self-organizing map neural network to perform distributed photovoltaic clustering in step S2 is as shown in the attached figure. Figure 2 As shown, the details are as follows:

[0085] Step S2-1: normalize the initial clustering index obtained in step S1 and use it as the input of the self-organizing map neural network. At the same time, initialize the neurons for the next step of data learning, as shown in formula (11):

[0086]

[0087] In the formula, It is the input of the self-organizing map neural network and is calculated from the distributed photovoltaic cluster division index data. The neurons of the self-organizing map neural network are initially assigned random numbers between 0 and 0.1.

[0088] Step S2-2: Calculate the connection weight matrix W between each neuron j ,from A value is randomly selected and input into the competition layer for neuron competition learning. The distance between the weight node and the weight vector of each neuron in the competition layer is calculated. Here, the Euclidean distance is usually selected to measure the similarity. The neuron with the highest similarity distance is taken as the winning neuron, as shown in formula (12).

[0089]

[0090] Among them, X i For the input dataset The i-th element in , D is the dimension of the competition layer neural network, W j is the weight vector of the jth neuron in the competition layer. The weight vector is an attribute contained in each neuron, which is used to describe the position of the neuron in the input space;

[0091] Step S2-3, taking the winning neuron obtained in step S2-2 as the center of the circle, according to the preset neighborhood distance threshold, screen out all nodes whose distance to the node is less than the threshold, thereby forming a winning neighborhood. The neighborhood range formula is shown in formula (13).

[0092]

[0093] Where R(l) is the neighborhood radius at the lth iteration, R max , R min are the maximum and minimum neighborhood radius thresholds respectively, and L is the total number of iterations.

[0094] Step S2-4, in the winning neighborhood, iteratively adjust the weight of each node. The principle of adjustment is: the closer the node is to the winning node, the greater the amplitude of its weight update; conversely, the farther the distance is, the smaller the update amplitude is. To implement this principle, it is necessary to set an update constraint condition for different nodes, which can usually be specifically calculated and determined by a Gaussian function, as shown in formula (14).

[0095]

[0096] In the formula, R k is the learning rate of the kth iteration, G ij is the update constraint of the jth neuron node in the winning neighborhood corresponding to the i-th sample.

[0097] Step S2-5, repeat the iterations of steps S2-2 to S2-4 until the upper limit of the number of iterations is reached, and finally output the result of the distributed photovoltaic cluster division.

[0098] Furthermore, the step S3 is specifically as follows:

[0099] Step S3-1, taking the average value of the distance between the weight node and the weight vector of each neuron in the competition layer calculated in step S2-2 as a reference, the quantization error calculation formula is shown in formula (15);

[0100]

[0101] Among them, c j (x) is the quantization error value at the jth iteration, J is the total number of neurons, M j is the weight vector of all neurons at the jth iteration;

[0102] Step S3-2: The total number of neurons in the self-organizing map neural network in the final result of the distributed photovoltaic cluster partition obtained in step S2-5 and the input data set The i-th element in and the weight vector of all neurons at the j-th iteration are used as input, and the final quantization error value is calculated by the calculation formula obtained in step S3-1;

[0103] Step S3-3, change the initial self-organizing map neural network parameters, modify the dimension parameters and then repeat steps S3-1 and S3-2 to calculate the quantization errors under multiple parameters to select the optimal parameter settings.

[0104] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned specific embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned specific embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed photovoltaic cluster division method taking into account frequency support capability, characterized in that: Here are the steps: Step S1, calculating the distributed photovoltaic cluster division index; the distributed photovoltaic cluster division index includes the node sensitivity index δ, the frequency modulation response strength index K, and the frequency modulation response depth index and the comprehensive cluster partitioning index γ; Step S2: Based on the comprehensive cluster division index γ proposed in step S1, the self-organizing map neural network algorithm is used to take the division index of each distributed photovoltaic cluster as input, and the node with the smallest distance is selected as the winning node through the similarity between the competition layer neuron and the input data. The weight is continuously updated until the iteration ends to obtain the final cluster division result; Step S3, based on the final cluster division result obtained in step S2, the clustering precision and accuracy are calculated by calculating the quantization error of the clustering result, and the relevant parameters of the self-organizing map neural network are reset to achieve the optimal division result.

2. A distributed photovoltaic cluster division method taking frequency support capability into account according to claim 1, characterized in that: The calculation method of the node sensitivity index δ in step S1 is: δ ij =S Pii -S Pij +S Qii -S Qij Among them, δ ij is the node sensitivity index of each node in the distribution network, representing the comprehensive impact of the power change of node j on node i, δ ij The smaller it is, the greater the impact of node j’s power change on node i, which is consistent with the characteristics of distance change; S P , S Q are active sensitivity and reactive sensitivity, respectively, which are used to characterize the change rate of active power and reactive power injection when the voltage fluctuates, and S P =(N-HJ -1 L) -1 ; S Q =(L-JH -1 N) -1 ; S Pii S represents the impact of node i voltage change on the node’s own active power injection change rate; Pij S represents the impact of node j voltage change on the active power injection change rate of node i; Qii It indicates the impact of node i voltage change on the reactive power injection change rate of the node itself; S Qij It indicates the influence of the reactive power injection change rate of node i when the voltage of node j changes. The difference in the influence of active power regulation and reactive power regulation on voltage is reflected by the size of the sensitivity coefficient, and no additional weight coefficient needs to be added. H, N, J, and L in the active sensitivity and reactive sensitivity are used to simplify the partial derivatives of active sensitivity and reactive power to phase angle and voltage in the Jacobian matrix respectively; P is active power, Q is reactive power, θ is the node voltage phase angle, and V is the node voltage amplitude.

3. A distributed photovoltaic cluster division method taking frequency support capability into account according to claim 1, characterized in that: The frequency modulation response strength index K in step S1 is used to characterize the influence of photovoltaic output on system frequency, and is calculated as follows: Among them, ΔP is the change in the output of photovoltaic access to the distribution network, and Δf is the change in system frequency when the distributed photovoltaic output changes.

4. A distributed photovoltaic cluster division method taking frequency support capability into account according to claim 1, characterized in that: The frequency modulation response depth index in step S1 It is used to characterize the deviation between the theoretical frequency modulation output and the actual photovoltaic output of distributed photovoltaic power generation. The calculation method is: Among them, P m is the rated output, P pν is the current actual output, and S is the current light intensity.

5. A distributed photovoltaic cluster division method taking frequency support capability into account according to claim 1, characterized in that: The comprehensive cluster division index γ in step S1 is a distributed photovoltaic cluster node sensitivity index δ, a frequency modulation response strength index K, and a frequency modulation response depth index. The calculation method of the comprehensive cluster division index γ is: Where λ1, λ2, and λ3 are the weight coefficients of the node sensitivity index, the frequency modulation response contribution strength index, and the frequency modulation response contribution depth index, respectively.

6. A distributed photovoltaic cluster division method taking frequency support capability into account according to claim 1, characterized in that: The steps of the self-organizing map neural network algorithm proposed in step S2 are as follows: Step S2-1: normalize the comprehensive cluster division index obtained in step S1 and use it as the input of the self-organizing map neural network, and initialize the neurons at the same time to carry out the next step of data learning, as shown below: in, It is the input of the self-organizing map neural network, which is calculated from the distributed photovoltaic cluster division index data; The neurons of the self-organizing map neural network are initially assigned random numbers between 0 and 0.1; Step S2-2: Calculate the connection weight matrix W between each neuron j ,from A value is randomly selected and input into the competition layer for neuron competition learning. The distance between the weight node and the weight vector of each neuron in the competition layer is calculated. The distance uses the Euclidean distance to measure the similarity. The neuron with the highest similarity distance is selected as the winning neuron, as shown below: Among them, X i For the input dataset The i-th element in , D is the dimension of the competition layer neural network, W j is the weight vector of the jth neuron in the competition layer. The weight vector is an attribute contained in each neuron, which is used to describe the position of the neuron in the input space; Step S2-3, taking the winning neuron obtained in step S2-2 as the center of the circle, according to a preset neighborhood distance threshold, screen out all nodes whose distance to the node is less than the threshold, thereby forming a winning neighborhood. The winning neighborhood range formula is as follows; Among them, R(l) is the neighborhood radius at the lth iteration, R max , R min are the maximum and minimum neighborhood radius thresholds, respectively, and L is the total number of iterations; Step S2-4, in the winning neighborhood, iteratively adjust the weight of each node; the principle of adjustment is: the closer the node is to the winning node, the greater the amplitude of its weight update; conversely, the farther the distance is, the smaller the update amplitude; according to this principle, set an update constraint condition for different nodes, as shown below: Among them, R k is the learning rate of the kth iteration, G ij is the update constraint of the jth neuron node in the winning neighborhood corresponding to the i-th sample; Step S2-5, repeat the iterations of steps S2-2 to S2-4 until the upper limit of the number of iterations is reached, and finally output the result of the distributed photovoltaic cluster division.

7. A distributed photovoltaic cluster division method taking frequency support capability into account according to claim 1, characterized in that: The step S3 is specifically as follows: Step S3-1, using the average value of the distance between the weight node and the weight vector of each neuron in the competition layer calculated in step S2-2 as a reference, the quantization error calculation formula is as follows; Among them, c j (x) is the quantization error value at the jth iteration, J is the total number of neurons, M j is the weight vector of all neurons at the jth iteration; Step S3-2: The total number of neurons in the self-organizing map neural network in the final result of the distributed photovoltaic cluster partition obtained in step S2-5 and the input data set The i-th element in and the weight vector of all neurons at the j-th iteration are used as input, and the final quantization error value is calculated by the calculation formula obtained in step S3-1; Step S3-3, change the initial self-organizing map neural network parameters, modify the dimension parameters and then repeat steps S3-1 and S3-2 to calculate the quantization errors under multiple parameters to select the optimal parameter settings.

Images