Load clustering and classification method comprehensively considering load response characteristics

Abstract

A load clustering and classification method comprehensively considering load response characteristics comprises the following steps: defining a performance evaluation index of controllable load resources participating in frequency regulation and a performance evaluation index of controllable load participating in peak regulation auxiliary service; based on the defined evaluation indexes, a two-stage classification method is provided, and controllable load resources are divided into frequency response loads and peak regulation loads; and according to the classification of controllable load resources, the controllable loads are aggregated based on curve dual-scale distance features. According to the method, the load resources are accurately classified and aggregated, the clustering precision and the load adaptability are effectively improved, the defects of a traditional method in the aspects of load aggregation, classification precision and market adaptability are overcome, and the resource scheduling efficiency and the market competitiveness of the virtual power plant are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of virtual power plant load dispatching, and particularly relates to a load clustering and classification method that comprehensively considers load response characteristics. Background Art

[0002] With the rapid development of the intelligentization of the power system and distributed energy resources, the virtual power plant (VPP), as an important platform for integrating and managing distributed energy resources, is gradually playing an increasingly important role in the power market. Through advanced communication technologies and software systems, the virtual power plant organically integrates various types of controllable loads, distributed energy, and other resources to form a virtual power supply system, which can participate in various ancillary services in the power market, such as peak shaving and frequency modulation.

[0003] However, due to the diverse types of load resources in the virtual power plant and the significant differences in their response characteristics, when aggregating these load resources, it is necessary to consider the timeliness response characteristics of various loads and their matching with different ancillary services. In particular, frequency regulation and peak shaving services have different requirements for characteristics such as the response time, response duration, and response frequency of loads. Therefore, how to effectively classify and aggregate controllable loads based on these response characteristics has become a key research topic.

[0004] The current load clustering research mainly focuses on clustering methods based on user load characteristics, but these methods do not fully consider the attributes of the virtual power plant participating in the market and its requirements for load response characteristics. Summary of the Invention

[0005] Aiming at the problem of diverse types of load resources in the virtual power plant and significant differences in their response characteristics, the present invention provides a load clustering and classification method that comprehensively considers load response characteristics. This method accurately classifies and aggregates load resources, effectively improves the clustering accuracy and load adaptability, solves the deficiencies of traditional methods in load aggregation, classification accuracy, and market adaptability, and improves the resource dispatching efficiency and market competitiveness of the virtual power plant.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A load clustering and classification method that comprehensively considers load response characteristics, comprising the following steps:

[0008] Step 1: Define performance evaluation indicators for controllable load resources to participate in frequency regulation, and performance evaluation indicators for controllable load resources to participate in peak shaving ancillary services;

[0009] Step 2: Based on the evaluation indicators defined in Step 1, propose a two-stage classification method to classify controllable load resources into frequency-responsive loads and peak shaving loads;

[0010] Step 3: Based on the classification of controllable load resources in Step 2, aggregate the controllable loads based on the curve double-scale distance feature.

[0011] In the above Step 1, the evaluation index for the controllable load resources to participate in frequency regulation is shown in Equation (1):

[0012]

[0013] In Equation (1), A AGC represents the evaluation index for the controllable load resources to participate in frequency regulation; and respectively represent the availability frequency, response time, and daily load volatility of the controllable load resources.

[0014] The performance evaluation index for the controllable load resources to participate in peak shaving ancillary services is shown in Equation (2):

[0015]

[0016] In Equation (2), A reg represents the performance index of peak adjustment response, and respectively represent the response duration and response capacity.

[0017] In addition, the parameters of the controllable load resources need to meet the requirements for participating in frequency modulation services and peak shaving services; the controllable load resources participating in frequency modulation services meet the following constraints:

[0018]

[0019] Among them, τ i,res represents the response time in the frequency modulation service; f i,res represents the minimum response frequency in the frequency modulation service; ν i,res represents the load volatility in the frequency modulation service; respectively represent the upper limit of the response time, the lower limit of the frequency response, and the upper limit of the daily load volatility in the frequency modulation service.

[0020] The controllable load resources participating in peak shaving services meet the following constraints:

[0021]

[0022] Among them, T i,last represents the response duration in the peak shaving service; S i,res represents the response capacity in the peak shaving service; respectively represent the upper limit of the response time, the lower limit of the response duration, and the lower limit of the response capacity in the peak shaving service. In the above Step 2, in the two-stage classification method:

[0023] Stage 1: The controllable load resources participating in frequency regulation service and peak shaving service first meet the basic constraints of the corresponding ancillary services. If the controllable load resources satisfy Equations (3), (4), and (5), they will participate in the frequency regulation service;

[0024] If the controllable load resources satisfy Equations (6), (7), and (8), they will participate in the peak shaving service;

[0025] If the controllable load resources simultaneously meet the requirements of Equations (3), (4), (5), (6), (7), and (8), the controllable load resources need to be further segmented according to the proposed ancillary service performance indicators. Specifically as follows:

[0026] The frequency regulation performance evaluation index A for all loads AGC and the peak shaving performance evaluation index A reg are sorted in ascending order respectively to obtain the frequency regulation ranking w of each load AGCi and the peak shaving ranking If indicates that the frequency regulation performance of the load is better than the peak shaving performance, it is classified as a frequency regulation load. If indicates that its peak shaving performance is better, it is classified as a peak shaving load. If (i.e., the rankings are the same): If w AGCi is in the top 50% percentile of all loads, it is classified as a frequency regulation load; otherwise, it is classified as a peak shaving load. The controllable load resources that cannot meet the basic constraints of the frequency regulation service and the peak shaving service are no longer considered.

[0027] Stage 2: Due to the different parameter selections of the performance evaluation indicators for peak shaving and frequency regulation, the numerical values of a single controllable load resource compared with the evaluation indicators are meaningless. In this stage, a ranking-based classification method is used for the controllable load resources that simultaneously meet the requirements of Equations (3), (4), (5), (6), (7), and (8).

[0028] First, the performance evaluation indicator values of the peak shaving and frequency regulation of the controllable load are sorted in ascending order;

[0029] Let w AGCi represent the ranking of the frequency regulation performance evaluation indicator value of load i among all controllable load resources, and w regi represent the ranking of the peak shaving performance evaluation indicator value of load i.

[0030] If w AGCi < w regi , it indicates that from the perspective of the overall virtual power plant VPP, the frequency regulation performance of load i is better than the peak shaving performance. Therefore, this controllable load resource is classified as a frequency response load.

[0031] On the contrary, if w regi>w AGCi , it indicates that the load regulation performance of load i is better than its frequency regulation performance. Therefore, this controllable load resource is classified as a peak shaving load.

[0032] In the case where the two metrics have the same ranking, the following rule is applied: If w AGCi is within the first 50%, the controllable load resource is classified as a frequency response load; otherwise, it is classified as a peak shaving load.

[0033] After the two-stage classification process, the loads of the virtual power plant VPP are assigned to the corresponding types to participate in the corresponding ancillary services.

[0034] Step 3 includes the following steps:

[0035] S3.1: Use the Euclidean distance and DTW distance as dual-scale features to measure the similarity between load curves;

[0036] S3.2: Use the entropy weighting method to determine the weights of the Euclidean distance and DTW distance metrics;

[0037] S3.3: Use the NJW spectral clustering algorithm to achieve controllable load clustering.

[0038] In the above S3.1:

[0039] 1) The Euclidean distance is expressed as follows:

[0040] For time series X E and Y E , the Euclidean distance d ED (X E , Y E ) can be calculated using the following formula:

[0041]

[0042] In Equation (9), x t and y t represent the values of X E and Y E at time t respectively; T represents the length of the time series;

[0043] 2) The DTW distance is expressed as follows:

[0044] For any two time series, X D = [x D1 , x D2 , …, x Dn and Y D = [y D1 , y D2 , ……, y Dn ,

[0045] Among them, x D1 , x D2 , …, x Dn represents the observations of the time series X D at consecutive time points. y D1 , y D2 , ……, y Dn represents the observations of the time series Y D at consecutive time points.

[0046] The calculation formula of the DTW distance is as follows:

[0047]

[0048] In formula (10), g s represents the coordinates of the s-th point on the warping path, that is, g s =(i, j); D(g s ) represents the cumulative distance of the warping path; k represents the length of the warping path G; G represents the set of all possible warping paths;

[0049] A two-scale similarity metric is adopted, denoted as d total (X, Y), which simultaneously considers the Euclidean distance and the DTW distance of the energy curve. The formula for d total (X, Y) is shown in formula (11):

[0050] d total (X, Y)=αd ED (X, Y)+βd DTW (X, Y) (11);

[0051] In formula (10), the parameters α and β are weighting parameters used to adjust the importance of different similarity metrics; by calculating d total (X, Y), the similarity value between the load curves X and Y can be obtained; the smaller the value of d total (X, Y), the higher the similarity between the load curves.

[0052] In the above S3.2: The entropy weighting method is used to determine the weights of the Euclidean distance and the DTW distance metrics in the two-scale similarity metric. The entropy weighting method measures the differences between different indicators according to the entropy value and determines the weights accordingly, specifically as follows:

[0053] a: Calculate the entropy value ξ m :

[0054] Suppose there are M metrics and n objects to be evaluated. The formula for calculating the entropy value of the MTH metric is as follows:

[0055]

[0056] Where: m represents the mth evaluation indicator; n represents the nth evaluated object; r mn represents the actual measured value of the nth evaluated object under the mth metric, and λ mn This is the value assigned to r for this particular metric. mn The weight of; M represents the total number of evaluation indicators; N represents the total number of evaluated objects;

[0057] Calculate the information utility value ζ m and entropy weight w m :

[0058] ζ m =1-ξ m (14);

[0059]

[0060] In S3.3: the NJW spectral clustering algorithm is combined with the dual-scale distance feature of the curve to aggregate the controllable loads, including the following steps:

[0061] 1) Calculate the similarity matrix:

[0062] Standardize the controllable load output data.

[0063]

[0064] In formula (16), x i ′ represents the standardized value of the output data of the ith controllable load; x i represents the original value of the output data of the ith controllable load; x1,…,x n They represent the original values ​​of the 1st to the pth controllable load output data respectively; p represents the total number of controllable load output data;

[0065] The Euclidean distance, DTW distance and entropy weight between the load curves are calculated to obtain a two-scale distance metric matrix, called D total . D total The elements in are represented by d total (i, j) represents the dual-scale distance metric between curve i and curve j.

[0066] 2) Use Gaussian kernel function to construct adjacency matrix:

[0067]

[0068] In formula (17), G(i,j) represents the similarity between the i-th node and the j-th node in the adjacency matrix; μ represents the bandwidth parameter in the Gaussian kernel function;

[0069] 3) Calculate the degree matrix H, where the degree matrix H represents the total weight of the edges connected to each vertex in the adjacency matrix;

[0070]

[0071] In Equation (18), H(i,i) represents the degree of the i-th node in the degree matrix H, that is, the total weight of the edges connected to the i-th node. a represents an index variable used to traverse the columns in the adjacency matrix G. a varies from 1 to b for calculating the degree of each vertex; b represents the total number of vertices; i represents the i-th node.

[0072] 4) Calculate the normalized Laplacian matrix:

[0073] R = H - D (19);

[0074] R std = H -1 / 2 RH 1 / 2 (20);

[0075] In the above formula, R represents the Laplacian matrix of the graph; H represents the degree matrix, which is a diagonal matrix where each diagonal element H(i,i) represents the degree of the i-th node, that is, the total weight of the edges connected to the i-th node; D represents the adjacency matrix, which is used to describe the connection relationship between vertices in the graph. The element D(i,j) in the adjacency matrix represents the connection weight between the i-th node and the j-th node; R std represents the normalized Laplacian matrix.

[0076] 5) Calculate the eigenvalues and eigenvectors of R std and sort the eigenvalues in descending order. Select the first p eigenvalues and their corresponding eigenvectors to construct the clustering matrix Q, whose dimension is n×q.

[0077] Q = [l1, l2, …, l q (21);

[0078] In Equation (21), l1, l2, …, l q respectively represent the first q eigenvectors of the normalized Laplacian matrix R std ;

[0079] 6) Use the K-means algorithm to cluster Q. Q is a matrix composed of the first q eigenvectors of the normalized Laplacian matrix R std Q = [l1, l2, …, l q . The steps of the K-means algorithm are as follows:

[0080] ① Initialization: Randomly select K initial centroids.

[0081] ② Assign data points: Each row of Q represents a load and is assigned to the cluster to which the centroid with the closest Euclidean distance belongs. The specific formula is as follows:

[0082] argmin j' ‖q i' -c j' ‖ 2

[0083] where q i' is the i'-th row of matrix Q, and c j' is the j'-th centroid.

[0084] ③ Update centroids: Recalculate the centroid of each cluster, which is the mean of all rows within the cluster. The specific formula is as follows:

[0085]

[0086] where S j' is the data set of the j'-th cluster.

[0087] ④ Iteration: Repeat the assignment and update steps until the centroids converge.

[0088] Through the above steps, the loads with similar energy curve characteristics are divided into the same cluster, obtaining multiple load clusters;

[0089] For the loads in each cluster, by accumulating the power consumption of the loads within the same cluster, several equivalent loads are obtained. The specific formula is as follows:

[0090]

[0091] where represents the maximum total AGC power of cluster U m at frequency f; represents the maximum power of a single load q in cluster U m at frequency f; q represents the load index within the cluster; represents the total number of loads contained in cluster U m ;

[0092] Its performance index is the average of the performance indices of all controllable loads. The specific formula is as follows:

[0093]

[0094] where represents the average frequency modulation performance index of cluster U m ; is the frequency modulation performance index of a single load;

[0095] Calculate the sum of its power consumption:

[0096]

[0097] Among them, K is the total number of clusters; Denote the cluster U m The maximum total AGC power at frequency f;

[0098] Accumulate the power of each equivalent load in time series, plot it as an energy curve, and obtain the energy curve of the virtual load to achieve the aggregation effect.

[0099] A load clustering and classification method that comprehensively considers load response characteristics according to the present invention has the following beneficial effects:

[0100] 1) The present invention considers the different requirements of frequency regulation service and peak shaving service, and designs corresponding performance evaluation indexes. The frequency regulation service requires that the load can respond quickly, have high availability, and have stable daily load characteristics to ensure the stability of the power system frequency. The peak shaving service requires that the load can provide additional power generation capacity during high demand periods, or increase load consumption during low demand periods to achieve supply-demand balance.

[0101] 2) The present invention proposes a two-stage classification method. In the first stage, controllable load resources that meet the requirements of frequency regulation service and peak shaving service are screened according to basic constraints. In the second stage, based on the ranking of performance indexes, the load is further subdivided into frequency response load and peak shaving load.

[0102] 3) The present invention adopts an NJW spectral clustering algorithm based on dual-scale distance features in the aggregation stage. This algorithm combines Euclidean distance and dynamic time warping (DTW) distance, and can more comprehensively evaluate the similarity between load energy distribution curves. By using the entropy weight method to determine the weights of the two distance metrics, the present invention further optimizes the clustering process and improves the quality and reliability of the clustering results.

[0103] 4) Through the NJW spectral clustering algorithm, the present invention aggregates controllable load resources into virtual loads, forms several equivalent load resources, and provides convenience for virtual power plants to participate in subsequent bidding processes..

[0104] 5) By comprehensively considering load response characteristics, the present invention not only improves the accuracy of aggregating controllable load resources, but also enhances the effectiveness and adaptability of virtual power plants in frequency regulation and peak shaving services, significantly improves the resource scheduling efficiency of virtual power plants, and provides strong support for the stable operation of the power system and the better utilization of renewable energy.

[0105] 6) The advantages of the present invention can be summarized as follows: ① improving the accuracy of load aggregation in the virtual power plant; ② enhancing the effectiveness and adaptability of controllable load resources in frequency regulation and peak shaving services; ③ improving the market adaptability and resource scheduling efficiency of the virtual power plant. Brief Description of the Drawings

[0106] The present invention will be further described below in conjunction with the drawings and embodiments:

[0107] Figure 1 It is a flowchart of the method of the present invention.

[0108] Figure 2 It is the energy curve of the virtual load.

[0109] Figure 3 It is the result of frequency modulation load aggregation.

[0110] Figure 4 It is the result of peak shaving load aggregation. Detailed Embodiment

[0111] A load clustering and classification method that comprehensively considers load response characteristics. First, performance evaluation indexes for controllable load resources to participate in frequency regulation are designed, providing a quantitative standard for classification. The two-stage classification method divides controllable load resources into different categories based on these standards. Secondly, the aggregation method based on double-scale distance utilizes the classification results and aggregates the controllable load resources of the same category into virtual loads through a clustering algorithm, ultimately achieving the purpose of classifying and aggregating virtual power plants based on controllable load resources considering integrated response characteristics.

[0112] A load clustering and classification method that comprehensively considers load response characteristics, characterized by including the following steps:

[0113] Step 1: Define performance evaluation indexes for controllable load resources to participate in frequency regulation and performance evaluation indexes for controllable load resources to participate in peak shaving ancillary services;

[0114] Step 2: Based on the evaluation indexes defined in Step 1, propose a two-stage classification method to divide controllable load resources into frequency response loads and peak shaving loads;

[0115] Step 3: According to the classification of controllable load resources in Step 2, aggregate the controllable loads based on the curve double-scale distance characteristics.

[0116] In the said Step 1, the evaluation index for controllable load resources to participate in frequency regulation is as shown in Equation (1):

[0117]

[0118] In Equation (1), A AGC represents the evaluation index for controllable load resources to participate in frequency regulation; and respectively represent the availability frequency, response time, and daily load volatility of controllable load resources.

[0119] The performance evaluation index for controllable load resources participating in peak shaving ancillary services is shown in Equation (2):

[0120]

[0121] In Equation (2), A reg represents the performance index of peak adjustment response, and respectively represent the response duration and response capacity.

[0122] In addition, the parameters of controllable load resources need to meet the requirements for participating in frequency regulation services and peak shaving services; the controllable load resources participating in frequency regulation services meet the following constraints:

[0123]

[0124] Among them, τ i,res represents the response time in frequency regulation services; f i,res represents the minimum response frequency in frequency regulation services; ν i,res represents the load volatility in frequency regulation services; respectively represent the upper limit of the response time, the lower limit of the frequency response, and the upper limit of the daily load volatility in frequency regulation services.

[0125] The controllable load resources participating in peak shaving services meet the following constraints:

[0126]

[0127] Among them, T i,last represents the response duration in peak shaving services; S i,res represents the response capacity in peak shaving services; respectively represent the upper limit of the response time, the lower limit of the response duration, and the lower limit of the response capacity in peak shaving services. In step 2, in the two-stage classification method:

[0128] The first stage: The controllable load resources participating in frequency regulation services and peak shaving services first meet the basic constraints of the corresponding ancillary services. If the controllable load resources meet Equation (3), Equation (4), and Equation (5), they will participate in frequency regulation services;

[0129] If the controllable load resources meet Equation (6), Equation (7), and Equation (8), they will participate in peak shaving services;

[0130] If the controllable load resources simultaneously meet the requirements of Equations (3), (4), (5), (6), (7), and (8), the controllable load resources need to be further segmented according to the proposed auxiliary service performance indicators. Specifically as follows:

[0131] The frequency modulation performance evaluation index A for all loads AGC and the peak shaving performance evaluation index A reg are sorted in ascending order respectively to obtain the frequency modulation ranking w of each load AGCi and the peak shaving ranking If indicates that the frequency modulation performance of the load is better than the peak shaving performance, it is classified as a frequency modulation load. If indicates that its peak shaving performance is better, it is classified as a peak shaving load. If (i.e., the rankings are the same): If w AGCi is in the top 50% of all loads, it is classified as a frequency modulation load; otherwise, it is classified as a peak shaving load. Controllable load resources that cannot meet the basic constraints of frequency modulation services and peak shaving services are no longer considered.

[0132] Phase 2: Since the parameter selections of the performance evaluation indicators for peak shaving and frequency modulation are different, the numerical values of a single controllable load resource compared with the evaluation indicators are meaningless. In this phase, a classification method based on rankings is used for controllable load resources that simultaneously meet the requirements of Equations (3), (4), (5), (6), (7), and (8).

[0133] First, the performance evaluation indicator values of peak shaving and frequency modulation of the controllable load are arranged in ascending order;

[0134] Let w AGCi represent the ranking of the frequency modulation performance evaluation indicator value of load i among all controllable load resources, and w regi represent the ranking of the peak shaving performance evaluation indicator value of load i.

[0135] If w AGCi < w regi , it indicates that from the perspective of the overall virtual power plant VPP, the frequency modulation performance of load i is better than the peak shaving performance. Therefore, this controllable load resource is classified as a frequency response load.

[0136] On the contrary, if w regi > w AGCi , it indicates that the peak shaving performance of load i is better than its frequency modulation performance. Therefore, this controllable load resource is classified as a peak shaving load.

[0137] In the case where the two indicators have the same ranking, the following rule is applied: If w AGCi is within the first 50%, the controllable load resource is classified as a frequency response load; otherwise, it is classified as a peak shaving load.

[0138] After the two-stage classification process, the loads of the virtual power plant (VPP) are assigned to the corresponding types to participate in the corresponding ancillary services.

[0139] Step 3 includes the following steps:

[0140] S3.1: Use the Euclidean distance and the DTW distance as dual-scale features to measure the similarity between load curves;

[0141] S3.2: Use the entropy weighting method to determine the weights of the Euclidean distance and the DTW distance metrics;

[0142] S3.3: Use the NJW spectral clustering algorithm to achieve controllable load clustering.

[0143] In the above S3.1:

[0144] 1) The Euclidean distance is expressed as follows:

[0145] For time series X E and Y E , the Euclidean distance d can be calculated using the following formula ED (X E , Y E ):

[0146]

[0147] In Equation (9), x t and y t represent the values of X E and Y E at time t, respectively; T represents the length of the time series;

[0148] 2) The DTW distance is expressed as follows:

[0149] For any two time series, X D = [x D1 , x D2 , …, x Dn and Y D = [y D1 , y D2 , ……, y Dn ,

[0150] where x D1 , x D2 , …, x Dn represent the observed values of the time series X D at consecutive time points. y D1 , y D2 , ……, y Dn represent the observed values of the time series Y DObservation values at consecutive time points.

[0151] The calculation formula of DTW distance is:

[0152]

[0153] In formula (10), g s represents the coordinates of the s-th point on the warping path, that is, g s =(i, j); D(g s ) represents the cumulative distance of the warping path; k represents the length of the warping path G; G represents the set of all possible warping paths;

[0154] A dual-scale similarity metric is adopted, denoted as d total (X, Y), which takes into account both the Euclidean distance and the DTW distance of the energy curve. The formula for d total (X, Y) is shown in formula (11):

[0155] d total (X, Y)=αd ED (X, Y)+βd DTW (X, Y) (11);

[0156] In formula (10), the parameters α and β are weighting parameters used to adjust the importance of different similarity metrics; by calculating d total (X, Y), the similarity value between the load curves X and Y can be obtained; the smaller the value of d total (X, Y), the higher the similarity between the load curves.

[0157] In S3.2: The entropy weighting method is used to determine the weights of the Euclidean distance and the DTW distance metrics in the dual-scale similarity metric. The entropy weighting method measures the differences between different indicators according to the entropy value and determines the weights accordingly, as follows:

[0158] a: Calculate the entropy value ξ m :

[0159] Suppose there are M evaluation metrics and n objects. The formula for calculating the entropy value of the MTH metric is as follows:

[0160]

[0161] where: m represents the m-th evaluation metric; n represents the n-th evaluated object; r mn represents the actual measured value of the n-th evaluated object under the m-th metric, and λ mn represents the weight assigned to r mn ; M represents the total number of evaluation metrics; N represents the total number of evaluated objects;

[0162] Calculate the information utility value ζ m and the entropy weight w m :

[0163] ζ m = 1 - ξ m (14);

[0164]

[0165] In the above S3.3: The NJW spectral clustering algorithm is combined with the dual-scale distance feature of the curve to aggregate the controllable loads, including the following steps:

[0166] 1) Calculate the similarity matrix:

[0167] Since the outputs of different loads vary significantly, the first step is to standardize the output data of the controllable loads.

[0168]

[0169] In Equation (16), x i ′ represents the value of the i-th controllable load output data after standardization; x i represents the original value of the i-th controllable load output data; x1, …, x n respectively represent the original values of the 1st to the p-th controllable load output data; p represents the total number of controllable load output data;

[0170] Calculate the Euclidean distance, DTW distance, and entropy weight between the load curves to obtain a dual-scale distance metric matrix, denoted as D total . The elements in D total are denoted as d total (i, j), representing the dual-scale distance metric value between curve i and curve j.

[0171] 2) Use the Gaussian kernel function to construct the adjacency matrix:

[0172]

[0173] In Equation (17), G(i, j) represents the similarity between the i-th node and the j-th node in the adjacency matrix; μ represents the bandwidth parameter in the Gaussian kernel function;

[0174] 3) Calculate the degree matrix H, where the degree matrix H represents the sum of the weights of the edges connected to each vertex in the adjacency matrix;

[0175]

[0176] In Equation (18), H(i, i) represents the degree of the i-th node in the degree matrix H, that is, the total weight of the edges connected to the i-th node. a represents an index variable used to traverse the columns in the adjacency matrix G. a varies from 1 to b and is used to calculate the degree of each vertex; b represents the total number of vertices; i represents the i-th node.

[0177] 4) Calculate the normalized Laplacian matrix:

[0178] R = H - D (19);

[0179] R std = H -1 / 2 RH 1 / 2 (20);

[0180] In the above formula, R represents the Laplacian matrix of the graph; H represents the degree matrix, which is a diagonal matrix, where each diagonal element H(i, i) represents the degree of the i-th node, that is, the total weight of the edges connected to the i-th node; D represents the adjacency matrix, which is used to describe the connection relationship between vertices in the graph. The element D(i, j) in the adjacency matrix represents the connection weight between the i-th node and the j-th node; R std represents the normalized Laplacian matrix.

[0181] 5) Calculate the eigenvalues and eigenvectors of R std and sort the eigenvalues in descending order. Select the first p eigenvalues and their corresponding eigenvectors to construct the clustering matrix Q, whose dimension is n×q.

[0182] Q = [l1, l2, …, l q (21);

[0183] In Equation (21), l1, l2, …, l q respectively represent the first q eigenvectors of the normalized Laplacian matrix R std ;

[0184] 6) Use the K-means algorithm to cluster Q. Q is a matrix composed of the first q eigenvectors of the normalized Laplacian matrix R std Q = [l1, l2, …, l q . The steps of the K-means algorithm are as follows:

[0185] ① Initialization: Randomly select K initial centroids (cluster centers).

[0186] ② Assign data points: Assign each row of Q (representing a load) to the cluster to which the centroid with the closest Euclidean distance belongs.

[0187] The specific formula is as follows

[0188] argminj' ‖q i' -c j' ‖ 2

[0189] where q i' is the i'-th row of matrix Q, and c j' is the j'-th centroid.

[0190] ③ Update the centroid: Recalculate the centroid of each cluster, that is, the mean of all rows within the cluster. The specific formula is as follows:

[0191]

[0192] where S j' is the data set of the j'-th cluster.

[0193] ④ Iteration: Repeat the assignment and update steps until the centroid converges.

[0194] Through the above steps, the loads with similar energy curve characteristics are divided into the same cluster, obtaining multiple load clusters;

[0195] Specifically, the matrix Q is clustered by the above K-means algorithm, and the loads are divided into two types of load clusters: frequency regulation and peak shaving.

[0196] For the loads in each cluster, by accumulating the power consumption of the loads within the same cluster, several equivalent loads are obtained. The specific formula is as follows:

[0197]

[0198] where represents the maximum total AGC power of cluster U m at frequency f; represents the maximum power of a single load q in cluster U m at frequency f; q represents the load index within the cluster (i.e., traversing the q-th load in the cluster; represents the total number of loads contained in cluster U m ;

[0199] Its performance index is the average of the performance indices of all controllable loads. The specific formula is as follows:

[0200]

[0201] where represents the average frequency regulation performance index of cluster U m ; is the frequency regulation performance index of a single load;

[0202] Calculate the sum of its power consumption:

[0203]

[0204] Among them, K is the total number of clusters; denotes the cluster U m the maximum total AGC power at frequency f;

[0205] Accumulate the power of each equivalent load in time series and plot it as an energy curve, and the energy curve of the virtual load is obtained as Figure 2 shown, achieving the aggregation effect.

[0206] Verification example:

[0207] Figure 3 is the result of frequency modulation load aggregation, from Figure 3 it can be seen that the frequency modulation loads are aggregated into 4 virtual loads, and their curves represent the average power of each type of virtual load. Since frequency modulation services require fast response and frequent adjustment, the energy consumption curves of frequency modulation loads have two major characteristics: high volatility and strong consistency. The DTW distance weight (0.545) of frequency modulation loads is slightly higher than the Euclidean distance (0.455), indicating that more attention is paid to the similarity of curve shapes (such as peak-valley time series matching) during clustering, rather than just numerical differences. This helps to group loads with similar dynamic response patterns into one category, improving the overall efficiency of frequency modulation services. The aggregated virtual loads can participate in power grid frequency regulation as a whole, reducing the need for fine control of individual loads, while ensuring the fast coordination ability of frequency modulation resources.

[0208] Figure 4 is the result of peak shaving load aggregation, from Figure 4 it can be seen that the peak shaving loads are aggregated into 6 virtual loads, and their curves represent the average power of each type of virtual load. Peak shaving services require long-term and large-capacity power adjustment, so the energy consumption curves of peak shaving loads have two major characteristics: significant persistence and large capacity differences. The DTW distance weight (0.549) of peak shaving loads is also slightly higher than the Euclidean distance (0.451). This indicates that during clustering, not only the total energy consumption difference is considered, but also the temporal alignment of energy consumption patterns is concerned. Through aggregation, peak shaving resources can be divided into virtual load groups with different capacities and durations as needed, facilitating the flexible bidding of virtual power plants in the electricity market and optimizing the peak shaving and valley filling strategies at the same time.

[0209] Through the two-stage classification and double-scale clustering method, controllable load resources are effectively divided into frequency modulation and peak shaving types and aggregated into a small number of equivalent loads. The energy curves of frequency modulation loads are smooth and respond quickly, and the power of peak shaving loads is concentrated during peak hours. The two respectively adapt to different auxiliary service requirements. This method significantly improves the resource utilization efficiency of virtual power plants in frequency modulation and peak shaving scenarios, providing technical support for flexible bidding in the electricity market.

Claims

1. A load clustering and classification method that comprehensively considers load response characteristics, characterized in that It includes the following steps: Step 1: Define the performance evaluation indicators for controllable load resources to participate in frequency regulation, and the performance evaluation indicators for controllable load resources to participate in peak shaving ancillary services; Step 2: Based on the evaluation indicators defined in Step 1, propose a two-stage classification method to divide controllable load resources into frequency response loads and peak shaving loads; Step 3: According to the classification of controllable load resources in Step 2, aggregate the controllable loads based on the curve double-scale distance feature.

2. The load clustering and classification method considering load response characteristics according to claim 1, characterized in that: In Step 1, the evaluation indicator for controllable load resources to participate in frequency regulation is shown in Equation (1): In formula (1), A AGC represents the evaluation index for controllable load resources to participate in frequency regulation; and respectively represent the availability frequency, response time, and daily load volatility of the controllable load resources; The performance evaluation indicator for controllable load resources to participate in peak shaving ancillary services is shown in Equation (2): In formula (2), A reg represents the performance index of the peak adjustment response, and represent the response duration and the response capacity respectively.

3. The load clustering and classification method considering load response characteristics according to claim 2, characterized in that: The parameters of the controllable load resources meet the requirements for participating in frequency modulation services and peak shaving services; the controllable load resources participating in frequency modulation services meet the following constraints: Among them, τ i,res represents the response time in the frequency regulation service; f i,res represents the lowest response frequency in the frequency regulation service; ν i,res represents the load volatility in the frequency regulation service; respectively represent the upper limit of the response time, the lower limit of the frequency response, and the upper limit of the daily load volatility of the frequency regulation service.

4. A load clustering and classification method that comprehensively considers load response characteristics according to claim 3, characterized in that: The controllable load resources participating in peak shaving services meet the following constraints: Among them, T i,last represents the response duration in the peak shaving service; S i,res represents the response capacity in the peak shaving service; respectively represent the upper limit of the response time, the lower limit of the response duration, and the lower limit of the response capacity of the peak shaving service.

5. The load clustering and classification method considering load response characteristics according to claim 4, characterized in that: In Step 2, in the two-stage classification method: The first stage: The controllable load resources participating in frequency modulation services and peak shaving services first meet the basic constraints of the corresponding ancillary services; if the controllable load resources meet Equations (3), (4), and (5), they will participate in frequency modulation services; If the controllable load resources meet Equations (6), (7), and (8), they will participate in peak shaving services; If the controllable load resources simultaneously meet the requirements of Equations (3), (4), (5), (6), (7), and (8), the controllable load resources need to be further subdivided according to the proposed ancillary service performance indicators; specifically as follows: Frequency modulation performance evaluation index A for all loads AGC and peak shaving performance evaluation index A reg are sorted in ascending order respectively to obtain the frequency modulation ranking w of each load AGCi and the peak shaving ranking If indicates that the frequency modulation performance of this load is better than the peak shaving performance, and it is classified as a frequency modulation load; if indicates that its peak shaving performance is better, and it is classified as a peak shaving load; if (i.e., the rankings are the same): if w AGCi is in the top 50% percentile of all loads, it is classified as a frequency modulation load; otherwise, it is classified as a peak shaving load; controllable load resources that cannot meet the basic constraints of frequency modulation services and peak shaving services are no longer considered; The second stage: Since the parameter selections of the performance evaluation indicators for peak shaving and frequency modulation are different, the numerical values of a single controllable load resource compared with the evaluation indicators are meaningless; in this stage, a ranking-based classification method is used for the controllable load resources that meet the requirements of Equations (3), (4), (5), (6), (7), and (8); First, arrange the performance evaluation indicator values of the controllable load for peak shaving and frequency modulation in ascending order; Let w AGCi represent the ranking of the frequency regulation performance evaluation index value of load i among all controllable load resources, and w regi represent the ranking of the peak shaving performance evaluation index value of load i; If w AGCi <w regi , it indicates that from the perspective of the overall virtual power plant VPP, the frequency regulation performance of load i is better than its peak shaving performance; therefore, this controllable load resource is classified as a frequency response load; On the contrary, if w regi > w AGCi , it indicates that the load regulation performance of load i is better than its frequency regulation performance; therefore, this controllable load resource is classified as a peak shaving load; In the case where two metrics have the same ranking, the following rules apply: If w AGCi is within the first 50%, the controllable load resource is classified as a frequency response load; otherwise, it is classified as a peak shaving load. After the two-stage classification process, the load of the virtual power plant VPP is assigned to the corresponding type to participate in the corresponding ancillary services.

6. The method for load clustering and classification that comprehensively considers the load response characteristics according to claim 1, wherein: Step 3 includes the following steps: S3.1: Use the Euclidean distance and DTW distance as double-scale features to measure the similarity between load curves; S3.2: Use the entropy weighting method to determine the weights of the Euclidean distance and DTW distance metrics; S3.3: Use the NJW spectral clustering algorithm to achieve controllable load clustering.

7. A load clustering and classification method that comprehensively considers load response characteristics according to claim 6, characterized in that: In S3.1: 1) The Euclidean distance is expressed as follows: For time series X E and Y E , the Euclidean distance d can be calculated using the following formula ED (X E , Y E ): In formula (9), x t and y t respectively represent the values of X E and Y E at time t; T represents the length of the time series; 2) The DTW distance is expressed as follows: For any two time series, X D =[x D1 , x D2 , …, x Dn and Y D =[y D1 , y D2 , ……, y Dn , where x D1 , x D2 , …, x Dn represent the observations of the time series X D at consecutive time points; y D1 , y D2 , ……, y Dn represent the observations of the time series Y D at consecutive time points; The calculation formula of the DTW distance is: In formula (10), g s represents the coordinates of the sth point on the warping path, that is, g s =(i, j); D(g s ) represents the cumulative distance of the warping path; k represents the length of the warping path G; G represents the set of all possible warping paths; A dual-scale similarity metric, denoted as d, is adopted. total (X, Y), which takes into account both the Euclidean distance and the DTW distance of the energy curves. The formula for d total (X, Y) is shown in Equation (11) as follows: d total (X,Y) = αd ED (X,Y) + βd DTW (X,Y) (11); In Equation (10), the parameters α and β are weighting parameters used to adjust the importance of different similarity measures; by calculating d total (X, Y), the similarity value between the load curves X and Y can be obtained; the smaller the value of d total (X, Y), the higher the similarity between the load curves.

8. The method for load clustering and classification considering load response characteristics according to claim 6, characterized in that: In the above S3.2: The entropy weighting method is adopted to determine the weights of the Euclidean distance and the DTW distance metrics in the dual-scale similarity metric; the entropy weighting method measures the differences between different indicators according to the entropy value and determines the weights accordingly, specifically as follows: a: Calculate the entropy value ξ m : Suppose there are M metrics and n objects to be evaluated; the formula for calculating the entropy value of the MTH metric is as follows: Where: m represents the m-th evaluation index; n represents the n-th object to be evaluated; r mn represents the actual measured value of the n-th object to be evaluated under the m-th metric, while λ mn represents the weight assigned to r mn for this specific metric; M represents the total number of evaluation indices; N represents the total number of objects to be evaluated; Calculate the information utility value ζ m and the entropy weight w m : ζ m = 1 - ξ m (14); 9. A load clustering and classification method that comprehensively considers load response characteristics according to claim 6, characterized in that: In S3.3: The NJW spectral clustering algorithm is combined with the curve double-scale distance feature to aggregate the controllable loads, including the following steps: 1) Calculate the similarity matrix: Perform normalization processing on the output data of the controllable load; In formula (16), x i ′ represents the value of the output data of the i-th controllable load after standardization; x i represents the original value of the output data of the i-th controllable load; x1, …, x n respectively represent the original values of the output data of the 1st to the p-th controllable loads; p represents the total number of the output data of the controllable loads; Calculate the Euclidean distance, DTW distance, and entropy weight between the computational load curves to obtain a two-scale distance metric matrix, called D total ; D total The elements in are denoted as d total (i, j), representing the two-scale distance metric value between curve i and curve j; 2) Use the Gaussian kernel function to construct the adjacency matrix: In Equation (17), G(i,j) represents the similarity between the i-th node and the j-th node in the adjacency matrix; μ represents the bandwidth parameter in the Gaussian kernel function; 3) Calculate the degree matrix H, where the degree matrix H represents the sum of the weights of the edges connected to each vertex in the adjacency matrix; In Equation (18), H(i, i) represents the degree of the i-th node in the degree matrix H, that is, the sum of the weights of the edges connected to the i-th node; a represents an index variable used to traverse the columns in the adjacency matrix G, and a varies from 1 to b to calculate the degree of each vertex; b represents the total number of vertices; i represents the i-th node; 4) Calculate the normalized Laplacian matrix: R = H - D (19); R std = H -1 / 2 RH 1 / 2 (20); In the above formula, R represents the Laplacian matrix of the graph; H represents the degree matrix, which is a diagonal matrix, where each diagonal element H(i, i) represents the degree of the i-th node, that is, the sum of the weights of the edges connected to the i-th node; D represents the adjacency matrix, which is used to describe the connection relationship between vertices in the graph, and the element D(i, j) in the adjacency matrix represents the connection weight between the i-th node and the j-th node; R std represents the normalized Laplacian matrix; 5) Calculate the eigenvalues and eigenvectors of R std and sort the eigenvalues in descending order; select the first p eigenvalues and their corresponding eigenvectors to construct the clustering matrix Q, whose dimension is n×q; Q＝[l1,l2,…,l q ] (21)； In formula (21), l1, l2, …, l q respectively represent the first q eigenvectors of the normalized Laplacian matrix R std ; 6) Use the K-means algorithm to cluster Q, where Q is a matrix composed of the first q eigenvectors of the normalized Laplacian matrix R std Q = [l1, l2, …, l q , and the steps of the K-means algorithm are as follows: ① Initialization: Randomly select K initial centroids; ② Assign data points: Each row of Q represents a load and is assigned to the cluster to which the centroid with the closest Euclidean distance belongs; the specific formula is as follows: argmin j' ‖q i' -c j' ‖ 2 where q i' is the i'-th row of matrix Q, and c j' is the j'-th centroid; ③ Update centroids: Recalculate the centroid of each cluster, that is, the mean of all rows within the cluster; the specific formula is as follows: Among them, S j' is the data set of the j'-th cluster; ④ Iteration: Repeat the assignment and update steps until the centroids converge. Through the above steps, the loads with similar energy curve characteristics are divided into the same cluster, obtaining multiple load clusters.

10. The load clustering and classification method considering load response characteristics according to claim 9, characterized in that: For the loads in each cluster, by accumulating the power consumption of the loads within the same cluster, several equivalent loads are obtained, and the specific formula is as follows: Among them, represents the maximum total AGC power of cluster U m at frequency f; represents the maximum power of a single load q in cluster U m at frequency f; q represents the load index within the cluster; represents the total number of loads contained in cluster U m ; Its performance metric is the average of the performance metrics of all controllable loads, and the specific formula is as follows: Among them, represents the average frequency regulation performance index of cluster U m ; is the frequency regulation performance index of a single load; Calculate the sum of its power consumption: where K is the total number of clusters; represents the cluster U m the maximum total AGC power at the frequency f; Accumulate the power of each equivalent load in time series and plot it as an energy curve, and the energy curve of the virtual load reaches the aggregation effect.

Images