An lstm-based distributed photovoltaic cluster frequency modulation equivalent modeling method

Abstract

This invention discloses an LSTM-based equivalent modeling method for frequency regulation of distributed photovoltaic (PV) clusters. The steps are as follows: extracting clustering indices for distributed PV systems and normalizing the collected clustering index data; using an improved K-means clustering algorithm to cluster the distributed PV systems based on the obtained clustering index data, resulting in K distributed PV clusters; obtaining the corresponding input dataset through experiments on detailed models of the distributed PV clusters, and selecting the input and output variables of the LSTM neural network; constructing the neural network using the LSTM training algorithm and training it with the experimental data results; and constructing an equivalent model of the distributed PV clusters based on the training parameters of the trained LSTM neural network. This invention rapidly clusters and groups distributed PV systems when facing a large number of grid-connected distributed PV systems, leveraging the frequency regulation capabilities of the distributed PV clusters and significantly improving the stability of the power system.

Description

Technical Field

[0001] This invention relates to the field of frequency regulation participation in distributed photovoltaic (PV) clusters, and in particular to an LSTM-based equivalent modeling method for frequency regulation of distributed PV clusters. Background Technology

[0002] Since 2018, the installed capacity of distributed photovoltaic (PV) power has been developing rapidly. In the first half of 2022, the newly installed PV capacity reached 30.88 GW, a year-on-year increase of 130%, of which distributed PV accounted for 64%, and the cumulative installed capacity of distributed PV exceeded 100 million kilowatts. With the rapid development of distributed PV installed capacity, the impact of large-scale distributed PV grid connection on the power system is also becoming increasingly significant.

[0003] However, with the increasing proportion of distributed photovoltaic (PV) systems connected to the power grid, their complex and highly variable dynamic operating characteristics have significantly impacted the dynamic stability of the system after disturbances, making frequency stability issues increasingly prominent. Distributed PV power generation systems typically operate in maximum power point tracking (MPPT) mode, utilizing power electronic components to connect to the grid. Their power supply side is decoupled from the grid, making it difficult to provide inertia, leading to reduced power system inertia and decreased frequency response capability. Conducting equivalent modeling for distributed PV grid-connected frequency regulation is fundamental to studying these impacts and is also an effective tool for ensuring the safe and stable operation of the distribution network. Currently, methods for single-unit modeling of individual PV power plants are quite mature, while research on equivalent modeling for large-scale distributed PV integration into the distribution network is relatively limited. Multi-unit equivalent modeling is based on the idea of ​​clustering. Clustering indices are selected to divide distributed PV systems into clusters. Then, distributed PV systems with similar characteristics are equivalently represented by a single PV unit, while the entire distributed PV cluster is equivalently represented by several PV units with significant differences. Commonly used clustering methods include K-means, FCM, and hierarchical clustering algorithms, each with its own advantages and disadvantages. Existing research on equivalent modeling for frequency regulation of distributed photovoltaic (PV) systems is relatively limited. For distribution networks with large-scale distributed PV integration, modeling each PV system individually leads to excessively high order and low efficiency. Therefore, conducting equivalent modeling for frequency regulation of distributed PV clusters is of great significance for ensuring the safe and stable operation of the power grid. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of the prior art by providing an LSTM-based equivalent modeling method for frequency regulation of distributed photovoltaic (PV) clusters. This invention rapidly performs clustering when dealing with a large number of distributed PV grid-connected systems, providing a model foundation for fully utilizing the frequency regulation capabilities of distributed PV.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A distributed photovoltaic cluster frequency regulation equivalent modeling method based on LSTM, the specific steps of which are as follows:

[0007] A. Extract clustering indicators from distributed photovoltaic data and normalize the collected clustering indicator data.

[0008] B. Based on the normalized clustering index data obtained in step A, the distributed photovoltaic system is clustered using the improved K-means clustering algorithm to obtain multiple distributed photovoltaic clusters.

[0009] C. Obtain the corresponding input dataset through experiments on the detailed model of the distributed photovoltaic cluster, and select the input and output variables of the LSTM neural network;

[0010] D. Use the LSTM training algorithm to build a neural network and train it with the experimental data obtained in step C;

[0011] E. Construct an equivalent model of a distributed photovoltaic cluster based on the training parameters of the LSTM neural network obtained from the training.

[0012] Furthermore, step A involves normalizing the collected distributed photovoltaic clustering index data to obtain the time series of each distributed photovoltaic clustering index. The specific steps are as follows:

[0013] A1. The collected distributed photovoltaic clustering index data includes, but is not limited to, the output voltage, output current, and output power of the photovoltaic array;

[0014] A2. Normalize the collected clustering indicators to obtain the time series of the clustering indicators. The specific formula is as follows:

[0015]

[0016] x' represents the normalized value of each clustering index data, and x represents the actual value of each clustering index data. min x represents the minimum value of each clustering index. max This represents the maximum value of each clustering index.

[0017] Furthermore, the specific process of clustering distributed photovoltaic systems using the improved K-means clustering algorithm in step B is as follows:

[0018] B1. Calculate the Euclidean distance d between the clustering indices (x1, y1) and (x2, y2) of any two photovoltaic power generation units, and generate the distance matrix D. The specific formula is as follows:

[0019]

[0020] x and y are the actual values ​​of the clustering index data;

[0021] B2. Group the two photovoltaic power generation units with the smallest distance in the distance matrix D into one group, and take the mean point of the two clustering indices as the first initial cluster center point.

[0022] B3. Set a distance threshold a, find the cluster indicators whose distances from the first cluster center are both greater than a through the distance matrix D, and group the two cluster indicators with the smallest distance into one group, and use their mean point as the second initial cluster center point.

[0023] B4. Repeat the above steps to determine K initial cluster centers.

[0024] Furthermore, in step C, the corresponding input dataset is obtained through experiments using a detailed model of the distributed photovoltaic cluster, and the input and output variables of the LSTM neural network are selected. The specific steps are as follows:

[0025] C1, Light intensity from (100kJ + 1000) W / m 2 Change to (100k2+1000)W / m 2 Where k1, k2 = [0, 2, 4, ..., 20], and k1 ≠ k2, with other conditions remaining unchanged, a total of 110 sets of data are generated;

[0026] C2. The ambient temperature changes from (n1-10)℃ to (n2-10)℃, where n1, n2 = [0, 2, 4, ..., 50], and n1 ≠ n2. Other conditions remain unchanged, generating a total of 650 sets of data.

[0027] The voltage fluctuation at points C3 and PCC changes from 0.1q1p.u. to 0.1q2p.u., where q1, q2 = [0, 1, 2, ..., 11] and q1 ≠ q2. Other conditions remain unchanged, generating a total of 132 sets of data.

[0028] The above three scenarios generated a total of 892 sets of data. 300 sets of data were selected as the test dataset and 592 sets of data were selected as the training dataset. The irradiance of the distributed photovoltaic system, the ambient temperature, and the output voltage of the photovoltaic system were selected as the input variables of the LSTM network, and the active power and reactive power of the PCC point were selected as the output variables.

[0029] Furthermore, in step D, the neural network is constructed using the LSTM training algorithm, and the experimental data obtained in step C is used for training.

[0030] Furthermore, in step E, an equivalent model of a distributed photovoltaic cluster is constructed based on the training parameters of the trained LSTM neural network.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) The LSTM-based equivalent modeling method for frequency regulation of distributed photovoltaic clusters proposed in this invention focuses on the distributed photovoltaic side while taking into account the frequency regulation capability of the photovoltaic power generation system, and establishes an equivalent model for distributed photovoltaic clusters. It more intuitively reflects the cluster modeling process of distributed photovoltaics, and the improved K-means algorithm can more efficiently cluster distributed photovoltaics.

[0033] (2) Compared with traditional distributed photovoltaic cluster modeling, this invention is based on long short-time neural network (LSTM), which can quickly and effectively construct the equivalent model of distributed photovoltaic cluster according to the disturbance of the input variables of distributed photovoltaic, which helps distributed photovoltaic to participate in grid frequency regulation quickly. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the principle of a distributed photovoltaic cluster frequency modulation equivalent modeling method based on LSTM according to the present invention;

[0035] Figure 2 This is a flowchart of an LSTM-based distributed photovoltaic cluster frequency modulation equivalent modeling method according to the present invention. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0037] A distributed photovoltaic cluster frequency regulation equivalent modeling method based on LSTM, such as... Figure 1 As shown, the steps are as follows:

[0038] A. Extract clustering indicators from distributed photovoltaic data and normalize the collected clustering indicator data.

[0039] B. Based on the normalized clustering index data obtained in step A, the distributed photovoltaic system is clustered using the improved K-means clustering algorithm to obtain multiple distributed photovoltaic clusters.

[0040] C. Obtain the corresponding input dataset through experiments on the detailed model of the distributed photovoltaic cluster, and select the input and output variables of the LSTM neural network;

[0041] D. Use the LSTM training algorithm to build a neural network and train it with the experimental data obtained in step C;

[0042] E. Construct an equivalent model of a distributed photovoltaic cluster based on the training parameters of the LSTM neural network obtained from the training.

[0043] Step A involves normalizing the collected distributed photovoltaic clustering index data to obtain the time series of each distributed photovoltaic clustering index. The specific steps are as follows:

[0044] A1. The collected distributed photovoltaic clustering index data includes, but is not limited to, the output voltage, output current, and output power of the photovoltaic array;

[0045] A2. Normalize the collected clustering indicators to obtain the time series of the clustering indicators. The specific formula is as follows:

[0046]

[0047] x' represents the normalized value of each clustering index data, and x represents the actual value of each clustering index data. min x represents the minimum value of each clustering index. max This represents the maximum value of each clustering index.

[0048] like Figure 2 As shown, the specific process of clustering distributed photovoltaics using the improved K-means clustering algorithm in step B is as follows:

[0049] Considering that the traditional K-means clustering algorithm is sensitive to the initial cluster centers, an inappropriate selection of K can lead the algorithm into local optima. The improved K-means clustering algorithm selects the initial K cluster centers using the following steps:

[0050] B1. Calculate the Euclidean distance d between the clustering indices (x1, y1) and (x2, y2) of any two photovoltaic power generation units, and generate the distance matrix D. The specific formula is as follows:

[0051]

[0052] x and y are the actual values ​​of the clustering index data;

[0053] B2. Group the two photovoltaic power generation units with the smallest distance in the distance matrix D into one group, and take the mean point of the two clustering indices as the first initial cluster center point.

[0054] B3. Set a distance threshold a, find the cluster indicators whose distances from the first cluster center are both greater than a through the distance matrix D, and group the two cluster indicators with the smallest distance into one group, and use their mean point as the second initial cluster center point.

[0055] B4. Repeat the above steps to determine K initial cluster centers.

[0056] In step C, the corresponding input dataset is obtained through experiments using a detailed model of a distributed photovoltaic cluster. The input and output variables of the LSTM neural network are then selected. The specific steps are as follows:

[0057] C1, Light intensity from (100kJ + 1000) W / m 2 Change to (100k2+1000)W / m 2 Where k1, k2 = [0, 2, 4, ..., 20], and k1 ≠ k2, with other conditions remaining unchanged, a total of 110 sets of data are generated;

[0058] C2. The ambient temperature changes from (n1-10)℃ to (n2-10)℃, where n1, n2 = [0, 2, 4, ..., 50], and n1 ≠ n2. Other conditions remain unchanged, generating a total of 650 sets of data.

[0059] The voltage fluctuation at points C3 and PCC changes from 0.1q1p.u. to 0.1q2p.u., where q1, q2 = [0, 1, 2, ..., 11] and q1 ≠ q2. Other conditions remain unchanged, generating a total of 132 sets of data.

[0060] The above three scenarios generated a total of 892 sets of data. 300 sets of data were selected as the test dataset and 592 sets of data were selected as the training dataset. The irradiance of the distributed photovoltaic system, the ambient temperature, and the output voltage of the photovoltaic system were selected as the input variables of the LSTM network, and the active power and reactive power of the PCC point were selected as the output variables.

[0061] In step D, the neural network is constructed using the LSTM training algorithm, and the experimental data obtained in step C is used for training.

[0062] In step E, an equivalent model of a distributed photovoltaic cluster is constructed based on the training parameters of the trained LSTM neural network.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 frequency regulation equivalent modeling method based on LSTM, the specific steps of which are as follows: A. Extract clustering indicators for distributed photovoltaic (PV) power and normalize the collected clustering indicator data; the specific steps for normalizing the collected distributed PV clustering indicator data to obtain the time series of each distributed PV clustering indicator are as follows: A1. The collected distributed photovoltaic clustering index data includes, but is not limited to, the output voltage, output current, and output power of the photovoltaic array; A2. Normalize the collected clustering indicators to obtain the time series of the clustering indicators. The specific formula is as follows: These are the normalized values ​​for each clustering index. These are the actual values ​​of each clustering index. The minimum value of each clustering index data, This represents the maximum value of each clustering index. B. Based on the normalized clustering index data obtained in step A, the distributed photovoltaic system is clustered using the improved K-means clustering algorithm to obtain multiple distributed photovoltaic clusters. The specific process of clustering distributed photovoltaic systems using the improved K-means clustering algorithm is as follows: B1. Calculate the Euclidean distance d between any two photovoltaic power generation unit clustering indices (x1, y1) and (x2, y2), and generate the distance matrix D. The specific formula is as follows: , These are the actual values ​​of the clustering index data; B2. Group the two photovoltaic power generation units with the smallest distance in the distance matrix D into one group, and take the mean point of the two clustering indices as the first initial cluster center point. B3. Set a distance threshold a, find the cluster indicators whose distances from the first cluster center are both greater than a through the distance matrix D, and group the two cluster indicators with the smallest distance into one group, and use their mean point as the second initial cluster center point. B4. Repeat the above steps to determine K initial cluster centers; C. Obtain the corresponding input dataset through experiments on the detailed model of the distributed photovoltaic cluster, and select the input and output variables of the LSTM neural network; The input dataset was obtained through experiments on a detailed model of a distributed photovoltaic cluster. The input and output variables of the LSTM neural network were then selected, and the specific steps are as follows: C1, light intensity from Change to ,in , ,and With other conditions remaining unchanged, a total of 110 sets of data were generated; C2, Ambient temperature from Change to ,in ,and With other conditions remaining unchanged, a total of 650 sets of data were generated; Voltage fluctuations at C3 and PCC points from Change to ,in ,and With other conditions remaining unchanged, a total of 132 sets of data were generated; The above three scenarios generated a total of 892 sets of data. 300 sets of data were selected as the test dataset and 592 sets of data were selected as the training dataset. The irradiance of the distributed photovoltaic system, the ambient temperature, and the output voltage of the photovoltaic system were selected as the input variables of the LSTM network, and the active power and reactive power of the PCC point were selected as the output variables. D. Use the LSTM training algorithm to build a neural network and train it with the experimental data obtained in step C; E. Construct an equivalent model of a distributed photovoltaic cluster based on the training parameters of the LSTM neural network obtained from the training.

2. The method for equivalent modeling of frequency regulation in a distributed photovoltaic cluster based on LSTM according to claim 1, characterized in that, In step D, the neural network is constructed using the LSTM training algorithm, and the experimental data obtained in step C is used for training.

3. The LSTM-based distributed photovoltaic cluster frequency modulation equivalent modeling method according to claim 1, characterized in that, In step E, an equivalent model of a distributed photovoltaic cluster is constructed based on the training parameters of the trained LSTM neural network.

Images