Massive power user clustering algorithm based on data-physical feature joint driving

Abstract

The application discloses a kind of mass power user clustering algorithm based on data-physical feature joint driving, comprising the following steps: S1, single-user daily typical load curve clustering;S2, standardization user daily typical load data, calculate the standard value of user daily typical load;S3, multi-user daily typical load curve clustering;S4, automatic iteration, find the clustering number that satisfies contour coefficient and industry concentration standard;S5, output the user clustering result corresponding to the best clustering number.The application is based on the mass user historical power consumption data accumulated in power consumption information acquisition system, excavates and grasps the production characteristics and power demand of each industry and user, not only can improve the load prediction accuracy and dispatching management level of power grid company, but also can provide accurate data support and decision basis for electricity price setting, economic dispatching, demand response, etc.

Description

Technical Field

[0001] This invention relates to a clustering algorithm for massive electricity users, and more specifically, to a clustering algorithm for massive electricity users driven by data-physical features. Background Technology

[0002] Based on the massive historical electricity consumption data accumulated in the electricity information collection system, we can mine and understand the production characteristics and electricity demand of various industries and users. This can not only improve the load forecasting accuracy and dispatch management level of the power grid company, but also provide accurate data support and decision-making basis for electricity pricing, economic dispatch, and demand response.

[0003] Due to the massive number of users, user clustering is a prerequisite and key influencing factor for user analysis and mining. User clustering methods are divided into two categories: one is based on user industry attributes, which is simple, intuitive, and has strong physical meaning. However, because user industry attributes are often incorrect or incomplete, and even within the same industry, there are sub-sectors with different production characteristics, the clustering results are not consistent in terms of electricity consumption characteristics. On the other hand, user clustering methods based on load characteristics, while ensuring the consistency of load characteristics within clusters, lack a physical explanation for this consistency. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a massive power user clustering algorithm driven by data-physical features. This algorithm can not only improve the load forecasting accuracy and dispatch management level of power grid companies, but also provide accurate data support and decision-making basis for electricity pricing, economic dispatch, and demand response.

[0005] The technical solution adopted by this invention to solve its technical problem is: to construct a massive power user clustering algorithm driven by data-physical features, including the following steps:

[0006] S1, Clustering of typical daily load curves for a single user;

[0007] S2. Standardize the typical daily load data of users and calculate the per-unit value of the typical daily load of users;

[0008] S3, Clustering of Typical Daily Load Curves for Multiple Users;

[0009] S4. Automatic iteration to find the number of clusters that meet the profile coefficient and industry concentration standards;

[0010] S5. Output the user clustering results corresponding to the optimal number of clusters.

[0011] According to the above scheme, in step S1, the clustering of the typical daily load curve of a single user is to use hierarchical clustering to cluster the typical daily load curve of the user, and calculate the typical daily load curve of the user by the category containing the most daily load curves.

[0012] According to the above scheme, the hierarchical clustering method is as follows: adopting a top-down approach, first treating each point of all samples as a cluster, then finding the two clusters with the smallest distance and merging them, and repeating this process until the expected cluster is reached. Representative algorithms of the hierarchical clustering method include: AGNES and Agglomerative Nesting.

[0013] According to the above scheme, in step S3, the range of the number of cluster categories is selected to be N / 2 to 2N. The hierarchical clustering method is used to cluster the typical daily load curves of multiple users, and the silhouette coefficient and industry concentration of each category under different numbers of clusters are calculated at the same time.

[0014] According to the above scheme, the specific calculation method of the contour coefficient is as follows:

[0015] S101. Calculate the average distance ai from sample i to other samples in the same cluster. The smaller ai is, the more sample i should be clustered into that cluster. ai is called the intra-cluster dissimilarity of sample i. The mean ai of all samples in cluster C is called the cluster dissimilarity of cluster C.

[0016] S102. Calculate the average distance bij from sample i to all samples in another cluster Cj. This distance is called the dissimilarity between sample i and cluster Cj, and is defined as the inter-cluster dissimilarity of sample i.

[0017] bi = min{bi1,bi2,...,bik}

[0018] The larger bi is, the less likely sample i belongs to other clusters;

[0019] S103. Define the silhouette coefficient of sample i based on the intra-cluster dissimilarity ai and inter-cluster dissimilarity bi of sample i:

[0020]

[0021]

[0022] S104, Judgment:

[0023] If si is close to 1, it indicates that the clustering of sample i is reasonable;

[0024] If si is close to -1, it means that sample i should be classified into a different cluster.

[0025] If si is approximately 0, it means that sample i is on the boundary between the two clusters;

[0026] S105. The mean of si for all samples is called the silhouette coefficient of the clustering result, which is a measure of whether the clustering is reasonable and effective.

[0027] According to the above scheme, in step S4, an automatic iterative convergence algorithm is used to determine whether the silhouette coefficient and industry concentration of each cluster have reached the indicator standard and converged. If not, the number of clusters is increased and iterative calculation continues.

[0028] Implementing the massive power user clustering algorithm based on data-physical feature joint driving of the present invention has the following beneficial effects:

[0029] 1. This invention uses a data-physical feature joint driving method to cluster massive power users. Based on the profile coefficient-based hierarchical clustering method of daily typical load curves, the user load curves are clustered to obtain clustering results that meet the profile coefficient requirements and industry concentration standards.

[0030] 2. This invention is based on a massive power user clustering algorithm driven by data-physical features. The clustering results calculated by this method meet the standards of the clustering effect evaluation index and the standard of the clustering category number constraint index.

[0031] 3. The present invention is based on a massive power user clustering algorithm driven by data-physical features, which can solve the clustering problem of massive power users and ensure both the consistency of power consumption characteristics and the consistency of load characteristics. Attached Figure Description

[0032] Figure 1 This is a flowchart of the algorithm provided by the present invention;

[0033] Figure 2 This invention provides a curve showing the correlation between the average profile coefficient of a large number of users and the number of clusters.

[0034] Figure 3 This invention describes the proportion of leading industries under different contour coefficients.

[0035] Figure 4 This is an example of the number of clusters in the hierarchical clustering method of this invention. Detailed Implementation

[0036] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0037] like Figure 1-4 As shown, the massive power user clustering algorithm based on data-physical feature joint driving of the present invention includes the following steps:

[0038] S1. Cluster the typical daily load curves of a single user, as follows:

[0039] A bottom-up hierarchical clustering method is adopted. First, all single-user daily load curves are treated as a whole, and each sample is treated as a cluster. Then, the two clusters with the smallest distance are found and merged. This process is repeated until the expected cluster requirements are met. The typical daily load curve of the user is calculated based on the category containing the most daily load curves.

[0040] S2. Standardize the typical daily load data of users and calculate the per-unit value of the typical daily load of users.

[0041] S3. Use hierarchical clustering to cluster the daily typical load curves of thousands of users in a certain region, and statistically obtain the curve of the relationship between the average profile coefficient and the number of clusters.

[0042] The number of cluster categories was selected to range from 50 to 250. Hierarchical clustering was used to cluster the typical daily load curves of multiple users. At the same time, the profile coefficient and industry concentration of each category under different numbers of clusters were calculated.

[0043] The silhouette coefficient is an evaluation method for measuring the quality of clustering. The specific calculation method for the silhouette coefficient is as follows:

[0044] S101. Calculate the average distance ai from sample i to other samples in the same cluster. The smaller ai is, the more sample i should be clustered into that cluster. ai is called the intra-cluster dissimilarity of sample i. The mean ai of all samples in cluster C is called the cluster dissimilarity of cluster C.

[0045] S102. Calculate the average distance bij from sample i to all samples in another cluster Cj. This distance is called the dissimilarity between sample i and cluster Cj, and is defined as the inter-cluster dissimilarity of sample i.

[0046] bi = min{bi1,bi2,...,bik}

[0047] The larger the value of bi, the less likely sample i belongs to other clusters.

[0048] S103. Define the silhouette coefficient of sample i based on the intra-cluster dissimilarity ai and inter-cluster dissimilarity bi of sample i:

[0049]

[0050] Right now

[0051]

[0052] S104, Judgment:

[0053] If si is close to 1, it indicates that the clustering of sample i is reasonable;

[0054] If si is close to -1, it means that sample i should be classified into a different cluster.

[0055] If si is approximately 0, it means that sample i is on the boundary between the two clusters.

[0056] S1105. The mean of si for all samples is called the silhouette coefficient of the clustering result, which is a measure of whether the clustering is reasonable and effective.

[0057] As the number of clusters increases, the average silhouette coefficient shows a slow growth trend, and the curve begins to flatten out in the middle.

[0058] S4. Automatic iteration to find the number of clusters that meet the criteria of profile coefficient and industry concentration.

[0059] Using an automatic iterative convergence algorithm, determine whether the silhouette coefficient and industry concentration of each cluster meet the indicator standard and converge. If not, increase the number of clusters and continue iterative calculation.

[0060] S5. Output the user clustering results corresponding to the optimal number of clusters.

[0061] When the silhouette coefficient is close to 0.095, the top industry accounts for 55%, the top two industries account for 80%, and the top three industries account for over 90%. Furthermore, this point represents a turning point in the curve, making it suitable as the optimal cluster number for the daily typical load curve. Data analysis shows that this cluster number is 87, which is the optimal number.

[0062] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A clustering algorithm for massive electricity users based on data-physical features, characterized in that, Includes the following steps: S1, Clustering of typical daily load curves for a single user; The clustering of typical daily load curves for a single user involves using hierarchical clustering to cluster the typical daily load curves of a user, and calculating the typical daily load curve of a user by the category containing the most daily load curves. The hierarchical clustering method is as follows: it adopts a top-down approach, first treating each point of all samples as a cluster, then finding the two clusters with the smallest distance and merging them, and repeating this process until the expected cluster is reached. Representative algorithms of the hierarchical clustering method include: AGNES and Agglomerative Nesting. S2. Standardize the typical daily load data of users and calculate the per-unit value of the typical daily load of users; S3, Clustering of Typical Daily Load Curves for Multiple Users; In step S3, the range of the number of cluster categories is selected as N / 2 to 2N. The hierarchical clustering method is used to cluster the typical daily load curves of multiple users, and the silhouette coefficient and industry concentration of each category under different numbers of clusters are calculated. The specific calculation method for the contour coefficient is as follows: S101. Calculate the average distance ai from sample i to other samples in the same cluster. The smaller ai is, the more sample i should be clustered into that cluster. ai is called the intra-cluster dissimilarity of sample i. The mean ai of all samples in cluster C is called the cluster dissimilarity of cluster C. S102. Calculate the average distance bij from sample i to all samples in another cluster Cj. This distance is called the dissimilarity between sample i and cluster Cj, and is defined as the inter-cluster dissimilarity of sample i. bi = min{bi1, bi2, ..., bik} The larger bi is, the less likely sample i belongs to other clusters; S103. Define the silhouette coefficient of sample i based on the intra-cluster dissimilarity ai and inter-cluster dissimilarity bi: S104, Judgment: If si is close to 1, it indicates that the clustering of sample i is reasonable; If si is close to -1, it means that sample i should be classified into a different cluster. If si is approximately 0, it means that sample i is on the boundary between the two clusters; S105. The mean of si for all samples is called the silhouette coefficient of the clustering result, which is a measure of whether the clustering is reasonable and effective. S4. Automatic iteration to find the number of clusters that meet the profile coefficient and industry concentration standards; In step S4, an automatic iterative convergence algorithm is used to determine whether the silhouette coefficient and industry concentration of each cluster have reached the indicator standard and converged. If not, the number of clusters is increased and iterative calculation continues. S5. Output the user clustering results corresponding to the optimal number of clusters.

Images