A power distribution network load curve clustering method based on adaptive dynamic density and related device

Abstract

The application discloses a power distribution network load curve clustering method based on adaptive dynamic density and related devices, and relates to the technical field of power system distribution networks. The method comprises the following steps: preprocessing the obtained load data of each transformer of the power distribution network to obtain corresponding normalized curves, and constructing a k-row-k-column global similarity matrix; determining the local density and density gradient of each normalized curve based on the global similarity matrix; determining the adaptive dynamic density clustering threshold of the corresponding normalized curve based on the local density and density gradient for each normalized curve; and clustering all normalized curves based on the global similarity matrix, all normalized curves and all adaptive dynamic density clustering thresholds to obtain a clustering set. The application can reduce intra-class dispersion and cross-class confusion, provide more accurate typical daily load data for improving the operation efficiency of the power distribution network, and is beneficial to improving the quality of strategies such as district planning and operation scheduling.

Description

Technical Field

[0001] This application relates to the field of power system distribution network technology, and in particular to a distribution network load curve clustering method and related apparatus based on adaptive dynamic density. Background Technology

[0002] The distribution network connects the main grid and end users, and its load patterns directly affect distribution area planning, operation scheduling, demand-side management, photovoltaic consumption, and energy storage configuration. Daily load curves reflect users' electricity consumption habits at different times; differences exist between different types of users, and users of the same type often share similar peak and valley locations and trends. By organizing historical load data through clustering, representative typical daily load patterns can be extracted, reducing the complexity of subsequent analysis and providing a data foundation for improving distribution network operating efficiency.

[0003] Existing methods for clustering load curves in power distribution networks typically employ K-means or hierarchical clustering, but these methods still suffer from insufficient adaptability to engineering data. Current clustering methods rely on fixed thresholds or preset cluster numbers, which can easily lead to intra-cluster discrepancies or inter-cluster confusion when dealing with datasets containing dense load patterns, sparse load patterns, and boundary samples. Furthermore, simply using Euclidean distance is insufficient to handle curves with slightly earlier or later peak and trough times, and it tends to underestimate the similarity between load curves with similar shapes but time misalignments. If normalization is performed directly before anomaly correction, a few spikes or missing points can alter the scale of the entire curve, reducing the reliability of subsequent clustering. In addition, some models lack post-processing mechanisms for small clusters, isolated curves, and unclassified samples, making it difficult to directly meet the requirements of engineering applications.

[0004] Furthermore, distribution network load data typically originates from multiple distribution areas or multiple distribution transformers, and the data simultaneously includes normal production and living electricity consumption curves, periodic load curves, low load curves, and abnormal or boundary curves. If clustering is performed solely based on a fixed threshold, it is easy to cause over-merging in high-density areas and under-classification in low-density areas, and it is difficult to interpret the processing results of boundary curves and noise curves. Summary of the Invention

[0005] The purpose of this application is to provide a distribution network load curve clustering method and related device based on adaptive dynamic density, which can reduce intra-class discrepancies and cross-class confusion, provide more accurate typical daily load data for improving the operating efficiency of the distribution network, and help improve the quality of strategies such as distribution area planning and operation scheduling.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a distribution network load curve clustering method based on adaptive dynamic density, including: The load data of each transformer in the distribution network is preprocessed to obtain the corresponding per-unit curves; the load data includes the multiplier and active power time series data.

[0007] Construct a global similarity matrix with k rows and k columns; the matrix element in the i-th row and j-th column of the global similarity matrix is ​​the similarity determined based on the DTW distance between the i-th and j-th standardized curves; k is the total number of standardized curves.

[0008] Based on the global similarity matrix, the local density and density gradient of each standardized curve are determined.

[0009] For each of the normalized curves, an adaptive dynamic density clustering threshold is determined based on the local density and the density gradient.

[0010] Based on the global similarity matrix, all the standardized curves, and all the adaptive dynamic density clustering thresholds, all the standardized curves are clustered to obtain one or more cluster sets; each cluster set includes one or more standardized curves.

[0011] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of the distribution network load curve clustering method based on adaptive dynamic density as described above.

[0012] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the distribution network load curve clustering method based on adaptive dynamic density described above.

[0013] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the adaptive dynamic density-based distribution network load curve clustering method described above.

[0014] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a distribution network load curve clustering method and related apparatus based on adaptive dynamic density. It uses an adaptive dynamic density clustering threshold to replace the fixed threshold of related technologies for clustering. Based on the local density, a density gradient term is introduced. The adaptive dynamic density clustering threshold of the per-unit curve is determined by the local density and the density gradient, so as to distinguish the cluster core, sparse uniform region and density transition zone. By comparing the similarity with the adaptive dynamic density clustering threshold, the corresponding cluster set is determined, so that different regions are clustered separately, which helps to reduce intra-class dispersion and cross-class confusion. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is an application environment diagram of a distribution network load curve clustering method based on adaptive dynamic density in one embodiment of this application; Figure 2 A flowchart illustrating a distribution network load curve clustering method based on adaptive dynamic density provided in an embodiment of this application; Figure 3 This is a clustering result illustration provided in an embodiment of this application, wherein, Figure 3 (a) is a schematic diagram of the scatter distribution of the per-unit curve after PCA dimensionality reduction, including unclassified noise sample labels. Figure 3 (b) is a schematic diagram of five typical daily load curves, including the category identifier of the effective load; Figure 4 This is a schematic diagram of a cluster set provided in an embodiment of this application, wherein, Figure 4 (a) is a schematic diagram of the curves for the first cluster set. Figure 4 (b) is a schematic diagram of the curve for the second cluster set. Figure 4 (c) is a schematic diagram of the curve for the third cluster set. Figure 4 (d) is a schematic diagram of the curve for the 4th cluster set. Figure 4 (e) is a schematic diagram of the curve of the 5th cluster set, with the cluster center being the center curve; Figure 5 This is a schematic diagram illustrating the effect verification of an embodiment of this application, wherein, Figure 5 (a) is a comparison chart of the DBI (Davies-Bouldin Index) of one embodiment method and four comparative methods. Figure 5(b) is a comparison chart of the changes in the DBI index under six ablation conditions in one embodiment method; Figure 6 This is a schematic diagram of parameter sensitivity analysis provided in an embodiment of this application, wherein, Figure 6 (a) is a schematic diagram of local density window index analysis. Figure 6 (b) is a schematic diagram of density sensitivity index analysis. Figure 6 (c) is a schematic diagram of gradient sensitivity index analysis. Figure 6 (d) is a schematic diagram illustrating the analysis of the minimum cluster size index. Figure 6 (e) is a schematic diagram of adaptive factor index analysis; Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application.

[0017] Attached image labels: 102 terminal, 104 server. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The distribution network load curve clustering method based on adaptive dynamic density provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the load data of each transformer in the distribution network to be processed to server 104. After receiving the load data, server 104 preprocesses the load data to obtain the corresponding per-unit curves; constructs a k-row, k-column global similarity matrix; determines the local density and density gradient of each per-unit curve based on the global similarity matrix; and clusters all the per-unit curves based on the global similarity matrix, all the per-unit curves, and all the adaptive dynamic density clustering thresholds to obtain one or more cluster sets. Server 104 can feed back the obtained cluster sets to terminal 102. Furthermore, in some embodiments, the distribution network load curve clustering method based on adaptive dynamic density can also be implemented separately by the server 104 or the terminal 102. For example, the terminal 102 can directly perform distribution network load curve clustering based on adaptive dynamic density on the load data of each transformer in the distribution network to be processed, or the server 104 can obtain the load data of each transformer in the distribution network to be processed from the data storage system and perform distribution network load curve clustering based on adaptive dynamic density on the load data of each transformer in the distribution network to be processed.

[0021] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0022] In one exemplary embodiment, such as Figure 2 As shown, a distribution network load curve clustering method based on adaptive dynamic density is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 205. Wherein: Step 201: Preprocess the load data of each transformer in the distribution network to obtain the corresponding per-unit curve; the load data includes the multiplier and active power time series data.

[0023] Step 202: Construct a k-row, k-column global similarity matrix; the matrix element in the i-th row and j-th column of the global similarity matrix is ​​the similarity determined based on the DTW distance between the i-th and j-th standardized curves; k is the total number of standardized curves. 1≤i≤k, 1≤j≤k, where i and j are both integers.

[0024] Step 203: Based on the global similarity matrix, determine the local density and density gradient of each normalized curve.

[0025] Step 204: For each of the normalized curves, determine the adaptive dynamic density clustering threshold of the corresponding normalized curve based on the local density and the density gradient.

[0026] Step 205: Based on the global similarity matrix, all the per-unit curves, and all the adaptive dynamic density clustering thresholds, cluster all the per-unit curves to obtain one or more cluster sets; each cluster set includes one or more per-unit curves.

[0027] By implementing the above steps, a density gradient term is introduced based on the local density. The adaptive dynamic density clustering threshold of the per-unit curve is determined by the local density and the density gradient, so as to distinguish the cluster core, sparse uniform region and density transition zone. The corresponding cluster set is determined by comparing the similarity with the adaptive dynamic density clustering threshold, so that different regions are clustered separately, which helps to reduce intra-class dispersion and cross-class confusion.

[0028] As an example, step 201 specifically includes: S1: Import load data from multiple transformers in the distribution network. This load data includes load factor and active power time-series data (measured values). The load data for each transformer in the distribution network needs to be corrected for load factor. Based on the load factor, the active power time-series data (measured values) is corrected to obtain the active power (actual value) reflecting the actual operating conditions. The expression is: ; In the formula: for The actual value of active power at any given time. For load factor, for The measured value of active power at any given time.

[0029] S2: The load data of each transformer in the distribution network may contain outliers, missing data, and periodic fluctuations. Directly normalizing these data may amplify outliers and affect subsequent clustering. To improve data quality, the following preprocessing is performed first: S21: Outlier correction, using... The fractional method for detecting outliers involves calculating the mean of all data points for each load curve plotted based on actual active power values. and standard deviation When data points satisfy When the value is determined to be outlier, then... The outlier detection threshold limits outliers based on their deviation direction. .

[0030] S22: Use linear interpolation to fill in missing data. If there are too many missing data, combine the neighborhood smoothing method to correct the filling result and obtain the filled load curve.

[0031] S23: Curve smoothing. The sliding window smoothing method is used to smooth the completed load curve. The smoothing formula is as follows: ; In the formula: For the smoothed first load curve The value of the moment. To adjust the sliding window size, This is the relative offset within the sliding window. This indicates rounding down to the nearest integer.

[0032] S24: Validity screening. This involves screening load curves that have at least a set threshold of valid points, are not all zero, and exhibit reasonable fluctuations, to obtain a set of clean load curves. .

[0033] S3: Set of cleaning load curves The load curves in the data are normalized to eliminate the influence of different transformers and load levels, and the load curve values ​​are normalized to a fixed value. The formula for standardizing an interval is: ; In the formula: For the first bar-marked perturbation curves in The value of the moment. For the first Cleaning load curve The actual active power at any given time , The first The minimum and maximum values ​​of the actual active power of the clean load curve.

[0034] As an example, step 202 specifically includes: S4: The load curves of each transformer in the distribution network are typical time series data, and their morphological differences are difficult to accurately represent using Euclidean distance alone. This embodiment uses the Dynamic Time Warping (DTW) algorithm to compare two per-unit curves, and uses DTW distance to handle disturbances caused by slight shifts and stretching at peak and valley times, thereby improving the accuracy and reliability of similarity evaluation.

[0035] S41: Traverse all standardized curve data and construct the DTW distance matrix pairwise for any two standardized curves. and , , and They represent the first i The first, second, and mth sampling points of the bar-standardized curve, , and They represent the first i The first, second, and m-th sampling points of the bar-standardized curve are used to construct a curve pair that corresponds only to the curve pair. (Also can be used) This indicates that the curve pair) Distance matrix When the curve pair is changed, the corresponding distance matrix also changes, and the matrix elements in the distance matrix... Represents curve The former Each sampling point and curve The former The cumulative distance of the optimal normalized path between each sampling point, where and These represent the sampling point numbers in the two curves, respectively, and their recursive calculation formula is: ; ; ; In the formula, Indicates origin from curve pair curve points Corresponding matrix elements, This represents the p-th sampling point on the i-th per-unit curve. This represents the q-th sampling point on the j-th per-unit curve. Indicates origin from curve pair curve points Corresponding matrix elements, Indicates origin from curve pair curve points Corresponding matrix elements, Indicates origin from curve pair curve points Corresponding matrix elements; , as well as These are all initial values ​​used for calculation settings.

[0036] S42: The DTW distance between the two normalized curves is ; S43: All The DTW distances obtained from pairwise calculations of the bar-labeled monotonic curves are uniformly mapped to similarity, thus constructing a... global similarity matrix The similarity matrix medium elements The formula for calculation is: ; In the formula: , The larger the value, the higher the similarity between the two load curves; This indicates that the curve is completely similar to itself.

[0037] In another exemplary embodiment of this application, steps 203 and 204 specifically include: S5: Most correlation clustering algorithms use fixed thresholds, which makes it difficult to simultaneously consider high-density cluster cores, low-density uniform regions, and transition zones between different categories, easily leading to intra-cluster dispersion or inter-cluster confusion. This embodiment first calculates the local density of the samples, then introduces a density gradient to characterize the density changes of similar samples, and generates an adaptive dynamic density clustering threshold for each per-unit curve, so that the threshold is adjusted according to the sample distribution.

[0038] S51: For a given permuted curve, its similarity matrix with any other permuted curve is... Each pair has a corresponding similarity score. These similarity scores are sorted in descending order, and the top scores are taken. The similarity is calculated, and the average of these similarities is used as the density index (unnormalized local density) of the per-unit curve. The formula is: ; ; S52: Normalize all density indices to obtain local density. : ; In the formula, This is the set of density indices for all standardized curves. For the first i Local density of bar-scaled perturbation curves For the first i Density index of bar-scaled perturbation curves This represents the minimum value of the density index for all standardized curves. This represents the maximum value of the density index for all standardized curves. It is a constant. Indicates the first i Bar-labeled perturbation curve and the Bar-labeled perturbation curve DTW distance, Let be a matrix element in the global similarity matrix, representing the _i_th i bar-marked monotonic curve and the first Similarity of bar-labeled monotonic curves, the first The bar-to-perimetric curve refers to the curve that is consistent with the first bar. i The bar-marked monotonic curve with the highest similarity bar-to-perimetric curve, To prevent the minimum value where the denominator is zero.

[0039] S53: Calculate the density gradient: For a monolithic curve, select the curve with the highest similarity to the monolithic curve. The per-unit curve is used as a set of similar curves. The density gradient is obtained by calculating the weighted average of the differences between the normalized densities of each similar curve and the normalized density of the given per-unit curve, using the similarity between the given per-unit curve and each other in the set as weights. : ; In the formula, For the first i Density gradient of bar-scaled perturbation curve, For the first i Local density of bar-scaled perturbation curves For the first j Local density of bar-scaled perturbation curves For the first i The set of similar curves to the bar-marked monotonic curve, including those similar to the first... i The bar label monotonic curve with the highest similarity Bar-labeled monotonic curve.

[0040] Density gradient is used to reflect the degree of density change in the region where a sample (a standardized curve represents a sample) is located. Smaller values ​​usually correspond to cluster cores or areas with uniform density, while larger values ​​usually correspond to cluster boundaries or transition zones.

[0041] S54: Generate an adaptive dynamic density clustering threshold. The adaptive dynamic density clustering threshold... The formula for calculation is: ; Where: benchmark threshold The set of all upper triangular off-diagonal elements in the similarity matrix The Percentile multiplied by adaptive factor To obtain, that is ; The percentile parameter is used to distinguish the sampling point numbers in S41. and .For example This indicates that the 99th percentile of the similarity distribution is taken as the benchmark reference value. This is the density sensitivity coefficient. The gradient sensitivity coefficient is... For the first i Adaptive dynamic density clustering threshold for bar-labeled permametric curves.

[0042] This formula creates three adjustment states: in high-density and low-gradient regions, the judgment strength is maintained close to the baseline threshold; in low-density and low-gradient regions, the threshold is moderately reduced to reduce the omission of valid samples; in high-gradient regions, the threshold is raised to tighten the boundary and reduce the risk of cross-cluster erroneous merging.

[0043] S55: Limit the range of the adaptive dynamic density clustering threshold to ensure ,in , These are the minimum and maximum values ​​of the adaptive dynamic density clustering threshold, respectively. This constraint clamps the threshold itself, rather than continuing to eliminate per-unit curves.

[0044] The adaptive dynamic density clustering threshold in this embodiment is determined jointly by a baseline threshold, a density sensitivity coefficient, and a gradient sensitivity coefficient, wherein the density sensitivity coefficient... Controlling the magnitude of the normalized local density correction to the threshold—when the sample is located in a low-density region ( When ), the correction item This lowers the adaptive dynamic density clustering threshold relative to the baseline threshold, relaxing the admission criteria to include valid but sparse samples; gradient sensitivity coefficient Controlling the magnitude of the density gradient correction to the threshold—when the sample is located in the density transition zone ( When the value is large, the correction term This increases the threshold relative to the baseline threshold, strictly tightening cluster boundaries and preventing erroneous cross-cluster merging.

[0045] Density sensitivity coefficient and gradient sensitivity coefficient are applied simultaneously to the same sample (i.e., the per-unit curve), jointly achieving three-state adaptive adjustment: cluster core region (high density, low gradient). Maintain normal stringency; sparse uniform region (low density, low gradient) Relaxing entry requirements; density transition zone (high gradient) Tighten the boundaries.

[0046] Related methods (such as DBSCAN) use a globally uniform threshold, which cannot adapt to the differentiated needs of the distribution network load curve in different density regions. However, the embodiments of this application introduce the density field structure (density + gradient) into the threshold generation process, so that each sample has a personalized clustering threshold that matches its local data distribution, fundamentally solving the triple contradiction of "strict threshold in high-density areas, lenient threshold in low-density areas, and prevention of false merging in transition zones".

[0047] The ablation experiment data in the subsequent embodiments directly support this conclusion: after removing the adaptive threshold, the DBI worsened by +47.2%, and after removing the density gradient, the DBI worsened by +14.9%.

[0048] In another exemplary embodiment of this application, step 205 specifically includes: S61: Construct a set of categories containing all the described per-unit curves.

[0049] Randomly select a per-unit curve from the set to be classified as the target per-unit curve.

[0050] Based on the adaptive dynamic density clustering threshold of the target permutation curve and the global similarity matrix, all target curve pairs are determined; the similarity of the target curve pairs is not lower than the adaptive dynamic density clustering threshold of the target permutation curve.

[0051] All target curves are labeled with their corresponding per-unit curves. The labeled per-unit curves are then aggregated to form a cluster set, and the labeled per-unit curves are removed from the set to be classified.

[0052] If the number of per-unit curves in the set to be classified is greater than 0, then return to the step of randomly selecting a per-unit curve from the set to be classified as the target per-unit curve.

[0053] In another exemplary embodiment of this application, after step 205, the method further includes: Step 206: Optimize the cluster set to obtain an optimized cluster set, specifically including: S62: Traverse each of the cluster sets. If the number of per-unit curves in the cluster set is greater than or equal to a preset cluster threshold, mark the corresponding cluster set as a cluster. If the number of per-unit curves in the cluster set is greater than 1 and less than the preset cluster threshold, mark the corresponding cluster set as a target cluster set. If the number of per-unit curves in the cluster set is equal to 1, mark the corresponding per-unit curve as an isolated point per-unit curve. The initial value of the noise label for all per-unit curves is a preset value. When a cluster set is marked as a cluster, the noise label for each per-unit curve in the corresponding cluster set is updated to the corresponding cluster identifier. This step completes the identification of isolated points (isolated point per-unit curves).

[0054] S63: For each target cluster set, calculate the first center distance between the target cluster set and each cluster to determine the target cluster corresponding to the target cluster set. Then, mark the noise of each per-unit curve in the target cluster set as the cluster identifier of the corresponding target cluster, and mark each per-unit curve in the target cluster set as the curve to be optimized. This step completes the integration of small clusters, where a small cluster refers to the target cluster set.

[0055] S64: For each isolated point perimeter curve, calculate the second center distance between the isolated point perimeter curve and each cluster. If the minimum second center distance is less than the noise allocation threshold, then mark the noise corresponding to the isolated point perimeter curve as the cluster identifier of the cluster with the minimum second center distance. This step completes the redistribution of noise points (isolated point perimeter curves are used as noise points).

[0056] S65: Count the number of clusters and adjust the number of clusters based on the preset target number of clusters. This step completes the cluster number optimization.

[0057] S66: Based on the DBI-guided stationary point optimization mechanism, the cluster to which each curve to be optimized belongs is adjusted, and the noise label of the curve to be optimized is updated to the cluster identifier of the adjusted cluster.

[0058] S67: Based on all the said per-unit curves and the noise labels of each of the said per-unit curves, determine the optimized cluster set.

[0059] As an example, the steps of S64 can also be placed after S65 and S66, that is, S65 and S66 are executed first, and then S64 is executed.

[0060] As an example, the first center distance is the DTW distance between the center curve of the target cluster set and the center curve of the corresponding cluster; the second center distance is the DTW distance between the isolated point permametric curve and the center curve of the corresponding cluster.

[0061] The expression for the center curve is: ; In the formula, u Indicates the target cluster set or cluster. express u The first in i bar-to-perimetric curve, express u The central curve, This means that the median is determined by statistically analyzing the sampling points of all the per-unit curves at each sampling time.

[0062] As an example, S62 involves standardizing unassigned samples (isolated points) and samples with a number less than a preset cluster threshold from step A61. Small cluster samples (perimetric curves in the target cluster set) are temporarily set as noise labels (keeping the initial value of the noise label, i.e. the preset value, which is -1), indicating that these perimetric curves are not currently accepted by any valid cluster.

[0063] As an example, in S65, if the current number of clusters is greater than the target number of clusters... Calculate the DTW distance between the center curve of each cluster and the center curves of other clusters, and merge the two clusters with the closest DTW distance; if the number of current clusters is less than the target number of clusters. The cluster with the largest number of samples (one standardized curve is one sample) is subjected to secondary clustering using the DBSCAN algorithm (Density-Based Spatial Clustering of Applications with Noise), splitting it into multiple clusters; the number of clusters is optimized through the above method until the number of clusters reaches the target number of clusters.

[0064] As an example, in S66, the DBI-oriented stationary optimization mechanism checks the candidate assignments of the curve to be optimized one by one with DBI as the target, and only accepts the reallocation that can reduce DBI.

[0065] In another example, the DBI-guided stationary point optimization mechanism can also be combined with the simulated annealing fine-tuning mechanism. Simulated annealing fine-tuning can be used to explore the boundary curve to be optimized among a small number of candidate clusters, so as to reduce the risk of getting trapped in local optima.

[0066] Specifically, perform the following operations for each curve to be optimized: Calculate the DTW distance between the curve to be optimized and the center curve of each cluster. At most, only the three closest other clusters (clusters other than the cluster to be optimized) are considered as candidate clusters, and the DTW distance between the curve to be optimized and the center curve of the candidate cluster does not exceed three times the DTW distance between the curve to be optimized and the center curve of the current cluster.

[0067] Attempt to reallocate the curve to be optimized to a candidate cluster. If the DBI decreases after reallocation, the reallocation is accepted, and the noise label of the curve to be optimized is updated to the cluster label of the corresponding candidate cluster. If the DBI increases after reallocation, the reallocation is not accepted (a simple DBI-guided stationary point optimization mechanism), or a probability-based approach is used. Accept reallocation (DBI-guided stationary optimization mechanism combined with simulated annealing fine-tuning mechanism).

[0068] in, The increase in DBI The lower the DBI increase and the higher the temperature, the greater the acceptance probability; initial temperature. After each iteration, the temperature is reduced by a decay factor of 0.8. Each round iterates a maximum of 20 times, and a total of 2 rounds of outer loop are executed. The loop terminates early when no samples move and the temperature drops below 0.0001 in a certain round.

[0069] In step A67, the noise markers of all per-unit curves are organized, and the cluster identifiers are organized into continuous ones. , The total number of clusters is given. The initial value of the noise label is a preset value (set to -1) and remains unchanged. The processed noise label is used as the optimized cluster label. When the cluster label is -1, it means that the corresponding per-unit curve is noise and does not participate in clustering. When the cluster label is other, the per-unit curves with the same cluster label belong to the same cluster set.

[0070] In another exemplary embodiment of this application, after step 205 or after step 67, the method further includes: Step 701: Summarize all the cluster sets (the clustering results of step 205 or the clustering results of step 67) and calculate the clustering evaluation index; the clustering evaluation index includes DBI, GHI (Gini-Hopkins Index) and SC.

[0071] Step 702: When the clustering evaluation index meets the preset engineering requirements, output all cluster sets.

[0072] When the clustering evaluation index does not meet the preset engineering requirements, adjust the density sensitivity coefficient and / or gradient sensitivity coefficient in the expression of the adaptive dynamic density clustering threshold, and return to step 204.

[0073] As an example, step 701 specifically includes: S71: DBI calculation, where DBI is the ratio of the degree of separation between clusters to the degree of dispersion within clusters, and the formula is: ; In the formula: The number of cluster sets, , The cluster set number, For clustering sets Center curves of all samples and their corresponding cluster sets of The average distance, For clustering sets Center curves of all samples and their corresponding cluster sets of The average distance, For clustering sets With cluster set Between the center curves distance.

[0074] S72: CHI calculation, where CHI is the ratio of the variance between clusters to the variance within clusters, and the formula is: ; In the formula: The sum of squares between cluster sets, The sum of squares within the cluster set. This represents the total number of standardized curves.

[0075] S73: SC (Shapeline Coefficient) calculation. SC is used to evaluate the relative relationship between a single standardized curve and its own cluster set and neighboring cluster sets. The formula is: ; In the formula: For the per-unit curve The average DTW distance between the curve and other standardized curves within the same cluster set. For the per-unit curve Between its nearest neighbor cluster's per-unit curve The average distance, This refers to the profile coefficient.

[0076] The method for determining the nearest neighbor cluster is as follows: For a single-unit curve (belonging to the cluster) ), calculate respectively To each other cluster ( The nearest neighbor cluster is the cluster corresponding to the minimum average DTW distance of all standardized curves in the dataset. In other words, the nearest neighbor cluster is not determined by the DTW distance between the center curves of two clusters, but rather by directly calculating... The average distance to all other cluster members – the overall distance The nearest cluster is the cluster of nearest neighbors.

[0077] As an example, step 702 specifically includes: S74: Effect Judgment: If , and If the clustering results meet the engineering requirements, the clustering results (i.e., all cluster sets) are output; otherwise, the density sensitivity coefficient and / or gradient sensitivity coefficient are adjusted, and the process returns to step 204.

[0078] The methods for determining and optimizing the density sensitivity coefficient and gradient sensitivity coefficient are as follows: First, the density sensitivity coefficient is evaluated in the pre-defined candidate set. and gradient sensitivity coefficient Perform a comprehensive scan to determine all possible density sensitivity coefficients. and gradient sensitivity coefficient The combination of .

[0079] The corresponding preset candidate set is , The corresponding preset candidate set is For each combination, a complete clustering process is performed, with DBI as the primary optimization objective and CHI and silhouette coefficient as auxiliary constraints. Under a single fast clustering iteration... (DBI=0.8649) (DBI=0.8649) represent their respective optimal single points.

[0080] Based on this, the parameter feedback adjustment mechanism is activated: adjustments are made according to the highest priority. and (Step length) Second priority adjustment of local density window (Step length) Third priority adjustment (Step length) )and (Step length) The order of clustering is adjusted, and the clusters are re-clustered and evaluated after each adjustment. If both conditions are met... , , If the condition is met, stop; otherwise, continue iterating.

[0081] After 30 rounds of random restarts, optimization using a DBI-guided stationary point optimization mechanism combined with simulated annealing fine-tuning, the final result was selected. , With this combination, the DBI reaches 0.7480 (better than the optimal value for a single scan).

[0082] The physical basis is: Too small ( A threshold of <0.03 will result in insufficient threshold reduction in low-density areas and the misclassification of sparse and effective samples; Too large ( A threshold greater than 0.08 can lead to excessively relaxed thresholds, blurred cluster boundaries, and confusion between different modes. If the value is too small, the boundary constraints of the density transition zone will be insufficient, which can easily lead to cross-cluster erroneous merging. If the size is too large, the boundaries will become excessively tight, and normal samples within the cluster may be excluded from the noise set.

[0083] In another embodiment, the engineering application search range is: , .

[0084] In another exemplary embodiment of this application, in order to verify the application effect of the above embodiments, an experiment is conducted using a power system distribution network as an example.

[0085] Step A1: Load distribution network load data for 95 transformers (tabletop transformers), with 10 days of data for each transformer, at a sampling interval of 15 minutes. The load data includes time series, multiplier series, and active power series. Calculate the actual active power using multiplier correction.

[0086] Step A2: Data cleaning, the specific steps are as follows: Step A21: Outlier removal, using... Fractional method Detect and correct outliers in the actual active power curve; Step A22: Missing value completion. Missing values ​​are completed using linear interpolation. After completion, the missing values ​​are... Smooth correction of the sliding window; Step A23: Smooth the curve and adjust the window size. The completed actual active power curve is then smoothed. Step A24: Validity screening, filtering for valid points. Non-zero, standard deviation The curve yields a clean dataset. Total number of clean samples .

[0087] Step A3: Standardization processing, cleaning the dataset. The curves in the dataset are normalized to obtain a normalized curve dataset. .

[0088] Step A4: Similarity matrix construction, for the per-unit curve dataset For each pair of samples, calculate its Distance, window size , build Similarity matrix .

[0089] Step A5: Adaptive dynamic threshold calculation, the specific steps are as follows: Step A51: Local density calculation, local density calculation window size Calculate the density index for each sample. ; Step A52: Normalize the density index to obtain the local density; Step A53: Calculate the density gradient ; Step A54: Adaptive dynamic density clustering threshold generation, setting the density sensitivity coefficient. Gradient sensitivity coefficient benchmark threshold Calculate the adaptive dynamic density clustering threshold; Step A55: The adaptive dynamic density clustering threshold range is limited to... .

[0090] Step A6: Clustering optimization and noise reduction, the specific steps are as follows: Step A61: Initial clustering. Initial clustering is performed based on the adaptive dynamic density clustering threshold to obtain the cluster set (initial cluster set). Step A62: Preset cluster threshold (minimum cluster size) Determine the isolated point standardization curve and the target cluster set (small cluster); Step A63: Merge the target cluster set with the target cluster (small cluster merging), and calculate the center curves of the small clusters and the cluster. Distance merges the target cluster set into the nearest cluster; Step A64: Cluster number optimization, target cluster number The number of clusters can be adjusted to 5 by merging or splitting clusters; Step A65: Combine the DBI-guided stationary optimization mechanism and the simulated annealing fine-tuning mechanism to attempt to redistribute the boundary samples, and retain only the redistribution schemes that improve the indicators or are acceptable according to the annealing probability; whereby, the boundary sample (or boundary curve) refers to the curve to be optimized that is far from the center curve of the initial cluster set; Step A66: Noise point redistribution, noise allocation threshold The isolated points that meet the conditions are normalized to noise labels on the curve. Modify the cluster identifier to the nearest cluster (i.e., the cluster corresponding to the minimum distance to the second center); Step A67: Determine the optimized cluster set.

[0091] Step A7: Cluster evaluation and result output, as follows: Step A71: Cluster evaluation, calculation results , And contour coefficient This indicates that the clustering results meet the requirements for engineering applications; Step A72: Output the clustering results, as shown below. Figure 3 As shown. Figure 3 The data includes the scatter distribution of 940 standardized curves after dimensionality reduction by PCA (Principal Component Analysis), the category identifiers of 5 types of effective load, the identifiers of unclassified noise samples, and various typical daily load curves, which are used to illustrate that the present invention can simultaneously output effective load categories, typical daily curves, and unclassified samples.

[0092] Step A73: Summarize similar load curves as follows Figure 4 As shown. Figure 4 The curves of the same type of load that have been effectively accepted from load type 1 to load type 5 are displayed respectively, and the median center line is used to represent the typical trend of the corresponding category, which is used to verify the consistency of load pattern within the same category and the representativeness of the center curve.

[0093] Step A74: Effect verification as follows Figure 5 As shown. Figure 5 pass Indicators for the methods and embodiments of this application Comparison of several methods (corresponding) Figure 5 (a)), and through 6 ablation conditions The changes in the metrics illustrate the contributions of modules such as dynamic density threshold, density gradient, post-processing optimization, and noise point redistribution to the clustering results (corresponding to...). Figure 5 (b)

[0094] As can be seen from this embodiment, cleaning before standardization can reduce the impact of outlier sampling points on the curve scale; after local density and density gradient jointly participate in the threshold calculation, the clustering conditions are no longer determined by a single empirical value; the DBI-guided stationary point optimization mechanism, simulated annealing fine-tuning mechanism, and noise point redistribution further preserve the opportunity to correct the boundary curve. The final clustering result has good intra-cluster consistency and inter-cluster discriminability, and can be used for distribution network planning, photovoltaic consumption, dispatch optimization, and demand-side management.

[0095] In another exemplary embodiment of this application, to further illustrate the adaptability of the embodiments of this application to different parameter values, the local density calculation window, density sensitivity coefficient, gradient sensitivity coefficient, minimum cluster size threshold, and adaptive factor are adjusted based on the data of the previous embodiment. Only one parameter is changed each time, while the remaining processing steps remain consistent, and the calculations are recalculated. , And silhouette coefficient. This method allows us to observe the impact of each parameter on the clustering results and determine the parameter range suitable for the current distribution network dataset.

[0096] When the local density calculation window When a smaller value is taken, the local density more easily reflects the clustering of a small number of similar samples near the curve, which is suitable for datasets with significant differences in load types; when Taking a larger value results in a smoother local density calculation, but may weaken the difference between the boundary curve and the core curve. Therefore, in practical applications, the appropriate value should be selected based on the sample size and the number of curve categories. .

[0097] When the minimum cluster size threshold When the load is small, the algorithm is more likely to form small-scale load categories, which is suitable for discovering a small number of special power consumption patterns; when When the number of clusters is large, smaller clusters are more likely to be labeled as noise or merged into other clusters, which can improve the stability of the output category. If the number of output clusters or noise points is too high, adjust accordingly. and noise assignment threshold .

[0098] The results of the parameter sensitivity analysis are as follows: Figure 6 As shown. Figure 6 Local density windows are given respectively (corresponding to) Figure 6 (a) of), density sensitivity (corresponding to) Figure 6 (b) of), gradient sensitivity (corresponding to) Figure 6 (c)), minimum cluster size (corresponding to Figure 6 (d) and adaptive factor (corresponding to) Figure 6 (e) in Under each value and Indicator changes; parameter values ​​that do not form effective clustering results that can be used for indicator calculation are treated as null values ​​in the figure.

[0099] The above parameter calculations show that this method does not rely on a single empirical threshold. Instead, it first characterizes the sample distribution based on the similarity between curves, then adjusts the judgment conditions around local density and boundary changes, and corrects small clusters, boundary curves, and noise curves through post-processing steps.

[0100] The embodiments of this application have the following beneficial effects: (1) Optimize the data preprocessing order by adopting a process of cleaning first and then standardizing to reduce the impact of abnormal sampling points and missing completion errors on the curve scale, and provide a more stable data foundation for subsequent clustering.

[0101] (2) An adaptive dynamic density clustering threshold is used instead of a fixed threshold, and a density gradient term is introduced based on the local density. This design can distinguish between the cluster core, sparse uniform region and density transition zone, so that different regions adopt different judgment intensities, thereby reducing intra-cluster discrepancies and cross-cluster confusion.

[0102] (3) Using DTW distance to compare the shape of the load curve can identify curves with slight shifts in peak and valley times but similar overall trends. It is more suitable for load time series than simple Euclidean distance. The cluster center is calculated using the median, which can improve the robustness and anti-interference ability of the clustering results.

[0103] (4) A system was designed to address issues ranging from outlier identification, small cluster integration, and cluster number optimization. The post-processing chain, which guides point-by-point optimization, simulated annealing fine-tuning, and noise point redistribution, ensures that boundary samples, unassigned samples, and small cluster samples all have clear processing paths.

[0104] (5) The extracted typical daily load curves can provide data support for distribution network planning, dispatch optimization, energy storage configuration, photovoltaic consumption and demand-side management.

[0105] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 7 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data from adaptive dynamic density distribution network load curve clustering. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an adaptive dynamic density-based distribution network load curve clustering method.

[0106] Those skilled in the art will understand that Figure 7 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0107] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0108] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0109] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0110] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0111] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0112] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0113] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A clustering method for distribution network load curves based on adaptive dynamic density, characterized in that, The distribution network load curve clustering method includes: The load data of each transformer in the distribution network is preprocessed to obtain the corresponding per-unit curves; the load data includes the multiplier and active power time series data; Construct a global similarity matrix with k rows and k columns; the matrix element in the i-th row and j-th column of the global similarity matrix is ​​the similarity determined based on the DTW distance between the i-th and j-th standardized curves; k is the total number of standardized curves; Based on the global similarity matrix, the local density and density gradient of each standardized curve are determined; For each of the normalized curves, an adaptive dynamic density clustering threshold is determined based on the local density and the density gradient. Based on the global similarity matrix, all the standardized curves, and all the adaptive dynamic density clustering thresholds, all the standardized curves are clustered to obtain one or more cluster sets; each cluster set includes one or more standardized curves.

2. The distribution network load curve clustering method based on adaptive dynamic density according to claim 1, characterized in that, Based on the global similarity matrix, all the standardized curves, and all the adaptive dynamic density clustering thresholds, clustering is performed on all the standardized curves to obtain one or more cluster sets, specifically including: Construct a set of objects to be classified that contains all the aforementioned standardized curves; Randomly select a per-unit curve from the set to be classified as the target per-unit curve; Based on the adaptive dynamic density clustering threshold of the target permutation curve and the global similarity matrix, all target curve pairs are determined; the similarity of the target curve pairs is not lower than the adaptive dynamic density clustering threshold of the target permutation curve. All target curves are labeled with their corresponding per-unit curves. The labeled per-unit curves are then aggregated to generate a cluster set, and the labeled per-unit curves are removed from the set to be classified. If the number of per-unit curves in the set to be classified is greater than 0, then return to the step of randomly selecting a per-unit curve from the set to be classified as the target per-unit curve.

3. The distribution network load curve clustering method based on adaptive dynamic density according to claim 1 or 2, characterized in that, The expression for the adaptive dynamic density clustering threshold is: ； In the formula, For the first i Adaptive dynamic density clustering threshold for bar-labeled perturbation curves. As the baseline threshold, This is the density sensitivity coefficient. This is the gradient sensitivity coefficient. For the first i Local density of bar-scaled perturbation curves For the first i Density gradient of bar-labeled monotonic curves.

4. The distribution network load curve clustering method based on adaptive dynamic density according to claim 1 or 2, characterized in that, The expression for the local density is: ； ； ； In the formula, For the first i Local density of bar-scaled perturbation curves For the first i Density index of bar-scaled perturbation curves This represents the minimum value of the density index for all standardized curves. This represents the maximum value of the density index for all standardized curves. It is a constant. Indicates the first i Bar-labeled perturbation curve and the Bar-labeled perturbation curve DTW distance, Let be a matrix element in the global similarity matrix, representing the _i_th i bar-marked monotonic curve and the first Similarity of bar-labeled monotonic curves, the first The bar-to-perimetric curve refers to the curve that is consistent with the first bar. i The bar-marked monotonic curve with the highest similarity bar-to-perimetric curve, To prevent the minimum value where the denominator is zero.

5. The distribution network load curve clustering method based on adaptive dynamic density according to claim 1 or 2, characterized in that, The expression for the density gradient is: ； ； In the formula, For the first i Density gradient of bar-scaled perturbation curve, For the first i Local density of bar-scaled perturbation curves For the first j Local density of bar-scaled perturbation curves For the first i The set of similar curves to the bar-marked monotonic curve, including those similar to the first... i The bar label monotonic curve with the highest similarity bar-to-perimetric curve, It is a constant. Let be a matrix element in the global similarity matrix, representing the _i_th i bar-marked monotonic curve and the first j Similarity of bar-labeled monotonic curves Indicates the first i Bar-labeled perturbation curve and the j Bar-labeled perturbation curve DTW distance.

6. The distribution network load curve clustering method based on adaptive dynamic density according to claim 1, characterized in that, After obtaining all the cluster sets, the process also includes: optimizing the cluster sets to obtain optimized cluster sets; Specifically, optimizing the cluster set to obtain an optimized cluster set includes: Each cluster set is traversed. If the number of per-unit curves in the cluster set is greater than or equal to a preset cluster threshold, the corresponding cluster set is marked as a cluster. If the number of per-unit curves in the cluster set is greater than 1 and less than the preset cluster threshold, the corresponding cluster set is marked as a target cluster set. If the number of per-unit curves in the cluster set is equal to 1, the corresponding per-unit curve is marked as an isolated point per-unit curve. The initial value of the noise label of all the per-unit curves is a preset value. When a cluster set is marked as a cluster, the noise label of each per-unit curve in the corresponding cluster set is updated to the corresponding cluster identifier. For each target cluster set, calculate the first center distance between the target cluster set and each cluster to determine the target cluster corresponding to the target cluster set. Then, mark the noise of each normalized curve in the target cluster set as the cluster identifier of the corresponding target cluster, and mark each normalized curve in the target cluster set as the curve to be optimized. For each isolated point per-unit curve, calculate the second center distance between the isolated point per-unit curve and each cluster. If the minimum second center distance is less than the noise allocation threshold, then mark the noise corresponding to the isolated point per-unit curve as the cluster identifier of the cluster with the minimum second center distance. Based on the DBI-guided stationary point optimization mechanism, the cluster to which each curve to be optimized belongs is adjusted, and the noise label of the curve to be optimized is updated to the cluster identifier of the adjusted cluster. Based on all the per-unit curves and the noise labels of each per-unit curve, the optimized cluster set is determined.

7. The distribution network load curve clustering method based on adaptive dynamic density according to claim 1 or 6, characterized in that, Also includes: Summarize all the cluster sets and calculate the cluster evaluation index; The clustering evaluation metrics include DBI, GHI, and SC; When the clustering evaluation index meets the preset engineering requirements, all cluster sets are output; When the clustering evaluation index does not meet the preset engineering requirements, adjust the density sensitivity coefficient and / or gradient sensitivity coefficient in the expression of the adaptive dynamic density clustering threshold, and return to the step of determining the adaptive dynamic density clustering threshold of the corresponding per-unit curve based on the local density and the density gradient for each per-unit curve.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement a distribution network load curve clustering method based on adaptive dynamic density as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements a distribution network load curve clustering method based on adaptive dynamic density as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the distribution network load curve clustering method based on adaptive dynamic density as described in any one of claims 1-7.

Images