Rapid LDP-MST lightweight method based on novel load and distributed resource clustering

Abstract

The invention relates to a power system control technology, and discloses a rapid LDP-MST lightweight method based on novel load and distributed resource clustering, which comprises the following steps: collecting novel load data and distributed resource data, and preprocessing the collected data; key features are extracted, the weight of each feature is calculated through an entropy method, and a multi-dimensional feature vector is constructed; performing privacy protection processing on the features based on an LDP mechanism; constructing a lightweight minimum spanning tree (MST) based on the features after privacy protection processing; performing clustering division based on a minimum spanning tree (MST) to obtain clusters; dynamically updating a cluster result; standardizing the parameters to generate power grid adaptive parameters; and efficiency optimization is carried out on the LDP-MST lightweight algorithm. The method has the advantages that rapid, accurate and dynamic clustering of mass novel loads and distributed resources is achieved, and therefore efficient and reliable technical support is provided for real-time scheduling and planning of a power grid.

Description

Technical Field

[0001] This invention relates to the field of power system control technology, and in particular to a fast LDP-MST lightweight method based on novel load and distributed resource clustering. Background Technology

[0002] As energy systems transition towards decentralization and multi-energy complementarity, the number of distributed resources and new types of loads is growing exponentially. Analyzing the characteristics of each resource / load individually would lead to a dramatic increase in computational complexity, far exceeding the time requirements for real-time grid dispatch (typically requiring a response time in seconds). Traditional clustering algorithms such as K-Means, DBSCAN, and traditional MST clustering suffer from low clustering accuracy and lack of privacy protection. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a fast LDP-MST lightweight method based on novel load and distributed resource clustering.

[0004] The objective of this invention is achieved through the following technical solution: a fast LDP-MST lightweight method based on novel load and distributed resource clustering, the method comprising,

[0005] Data acquisition and preprocessing for new loads and distributed resources;

[0006] Extract key features, calculate the weights of each feature using the entropy method, and construct a multidimensional feature vector.

[0007] Privacy protection processing based on LDP mechanism;

[0008] Construct a lightweight minimum spanning tree (MST);

[0009] Clusters are obtained by clustering based on the minimum spanning tree (MST).

[0010] Dynamically update cluster results;

[0011] Standardized parameters are used to generate power grid adaptation parameters;

[0012] Efficiency optimization of the LDP-MST lightweight algorithm.

[0013] Specifically, the acquisition and preprocessing of novel load and distributed resource data includes acquiring novel load data and distributed resource data, cleaning the acquired data to remove outliers, and filling in missing values ​​using interpolation.

[0014] ;

[0015] In the formula, , These are the valid points before and after the missing point, respectively; , These represent the time intervals between missing points and the preceding and following valid points, respectively.

[0016] Finally, the extreme value method is used to process the data:

[0017] ;

[0018] In the formula, This is the standardized power value; This is the original power value; , These are the minimum and maximum power in the sample set, respectively;

[0019] After preprocessing, the data forms a sample set to be clustered.

[0020] ;

[0021] In the formula, n is the total number of samples.

[0022] Specifically, the extracted features include temporal features, fluctuation features, and controllability features;

[0023] The time-series characteristics include the times when peaks and troughs occur, and the average daily load factor.

[0024] ;

[0025] In the formula, Rated power of the equipment; Let be the power value at time t;

[0026] The fluctuation characteristics include short-time fluctuation frequency and power fluctuation coefficient:

[0027] ;

[0028] In the formula, This represents the average power. ;

[0029] The controllability features include adjustable response speed and a scalable capacity ratio:

[0030] ;

[0031] In the formula, This represents the maximum power that can be reduced.

[0032] Specifically, the weights of each feature are calculated using the entropy method as follows:

[0033] ;

[0034] In the formula, For the first In the nth sample The normalized percentage of each feature; For feature dimensions; For the first The weights of each feature; Let be the information entropy of the j-th feature.

[0035] Specifically, the privacy protection processing employs a layered noise addition strategy, where highly sensitive features are subject to strong privacy protection:

[0036] ;

[0037] In the formula, These are highly sensitive feature values ​​after differential privacy protection; Represents the original true feature values; ( ) represents the Laplace distribution, used to generate random noise that satisfies differential privacy requirements;

[0038] Low-sensitivity features employ weak privacy protection:

[0039] ;

[0040] In the formula, These are low-sensitivity eigenvalues ​​after differential privacy protection; the probability density function of Laplace noise is... ; For scale parameters; For feature sensitivity; , Privacy budgets for high / low sensitivity features, respectively.

[0041] Specifically, constructing the lightweight minimum spanning tree (MST) includes:

[0042] First, Principal Component Analysis (PCA) is used to analyze the privacy-preserving feature set. Dimensionality reduction, calculation of the covariance matrix:

[0043] ;

[0044] In the formula, Represent the covariance matrix; Let be the feature vector of the i-th sample after privacy protection; This is the mean vector of the feature vectors of all samples after privacy protection.

[0045] The eigenvalue decomposition of the covariance matrix yields the eigenvector matrix. The first k principal components are selected, and the mapping formula is:

[0046] ;

[0047] In the formula, The eigenvector matrix; A feature set after privacy protection;

[0048] Then, the weighted Euclidean distance is defined, with the formula:

[0049] ;

[0050] In the formula, The weighted Euclidean distance between the i-th sample and the j-th sample; The weight of the l-th feature; Let be the projection value of the i-th sample onto the l-th principal component; Let be the projection value of the j-th sample onto the l-th principal component;

[0051] Set distance threshold Only retain Sample pairs;

[0052] The distance matrix calculation is accelerated by FPGA hardware parallelism, and the minimum spanning tree (MST) is constructed by combining Kruskal's algorithm and disjoint set.

[0053] Specifically, the clustering partitioning based on the Minimum Spanning Tree (MST) topology includes:

[0054] Calculate the median weight of all edges in the minimum spanning tree (MST), and set a cutting threshold:

[0055] ;

[0056] In the formula, The median of all edge weights in the minimum spanning tree (MST); This is the adjustment coefficient;

[0057] When the target number of clusters is C, the number of clusters is converged to C by adjusting α using the bisection method. If the number of clusters Then increase ;like Then decrease The process continues until the number of clusters converges to C; after splitting, C connected components are formed, each component corresponding to a cluster G1, G2, ..., Gc, satisfying that the distance between samples within the same category is less than 1 / 2. The distance between out-of-class samples is greater than .

[0058] Specifically, the dynamic updating of cluster results includes,

[0059] Collect daily new data and calculate the deviation of its features from the average features of its cluster:

[0060] ;

[0061] In the formula, The feature deviation between the newly added sample i and its cluster; Let be the average value of the j-th feature of cluster c; Let i be the j-th feature value of the newly added sample i; For the first The weights of each feature; For feature dimensions;

[0062] Based on historical data verification, the following settings are established. ,when When a sample is identified as having significant feature changes, its distance from other samples is recalculated, the local topology of the minimum spanning tree (MST) is updated, and its cluster affiliation is adjusted.

[0063] Specifically, the standardized parameter output to generate power grid adaptation parameters includes:

[0064] Typical curves for clusters:

[0065] ;

[0066] In the formula, For sample weights; Let be the typical power value of the c-th cluster at time t; The actual power value of the i-th sample at time t; Regulation potential index:

[0067] ;

[0068] In the formula, The maximum adjustable power of the i-th sample after it has been cut;

[0069] The average response time is:

[0070] ;

[0071] In the formula, This represents the average response time of cluster c. Let be the individual response time of the i-th sample; For the first One cluster;

[0072] Clustering confidence:

[0073] ;

[0074] In the formula, It is a weighted Euclidean distance.

[0075] Specifically, the efficiency optimization of the LDP-MST lightweight algorithm includes:

[0076] S1. Perform wavelet transform compression, compressing and segmenting the adaptive data, specifically including:

[0077] S11. The time-series feature data are decomposed into three levels using the db4 wavelet basis.

[0078] The first layer retains high-frequency coefficients;

[0079] The second layer retains the intermediate frequency coefficients;

[0080] Third-level quantization low-frequency coefficients;

[0081] The compression formula is:

[0082] ;

[0083] In the formula, W is the wavelet transform matrix; This is for compression error; This is the sample set after wavelet transform compression; This is the preprocessed sample set to be clustered.

[0084] S12. Perform block clustering, dividing the data into sub-blocks according to voltage level. Each sub-block independently constructs a minimum spanning tree (MST). Calculate the inter-block distance using core samples, and then perform global cluster fusion.

[0085] ;

[0086] In the formula, This represents the global clustering result; The core sample of sub-block b; This is the inter-block fusion threshold; This is the core sample of sub-block d; The core sample distance between the c-th sub-block and the d-th sub-block;

[0087] S2, heterogeneous hardware collaborative acceleration, calculates sample distance in parallel using multiple GPUs, the calculation formula is:

[0088] ;

[0089] In the formula, The distance between the i-th sample and the j-th sample is calculated using GPU acceleration. This is a parallel computation function, executed collaboratively by multiple GPUs; Let be the vector of the i-th sample in the reduced feature space; Let be the vector of the j-th sample in the reduced-dimensional feature space; The feature weight vector;

[0090] The system calculates and outputs local cluster centers, and then performs global clustering in the cloud based on the center samples.

[0091] S13, Minimum Spanning Tree (MST) Construction and Pruning Optimization;

[0092] S131. Randomly sample s samples to construct an initial minimum spanning tree (MST), and then perform a k-nearest neighbor search. After correcting the critical edges, the approximation error is ≤3%, and the time complexity is optimized to... ;

[0093] S132, Dynamic threshold pruning, adjusting the distance threshold based on sample density:

[0094] ;

[0095] In the formula, For sample density; The baseline threshold;

[0096] S14. Improved time-aware distance measurement;

[0097] S141. Dynamic time warping, replacing Euclidean distance for handling time offsets:

[0098] ;

[0099] In the formula, This is a time alignment function; Let be the dynamic time-normalized distance between the i-th sample and the j-th sample; Let be the power value of the i-th sample at time t; Let be the power value of the j-th sample at the corresponding time after time alignment; To find the optimal path that minimizes the total distance among all possible time-aligned paths;

[0100] S142. Physical constraint weighting, combining geographical distance and temporal distance, the total distance is:

[0101] ;

[0102] In the formula, Geographical distance; The maximum geographical span;

[0103] S15, Feature engineering for embedding domain knowledge;

[0104] S151, Construction of Cross Features:

[0105] ;

[0106] In the formula, For sample i, the peak load period; For the j-th sample, the peak load period is; Pearson correlation coefficient; Peak load overlap; Source-load matching degree; Let be the power generation of the i-th sample at time t; Let be the load power of the j-th sample at time t;

[0107] S152, Median standardization, employing anti-extreme value processing for fluctuation characteristics:

[0108] ;

[0109] In the formula, These are the original eigenvalues; Let x be the median; The difference in interquartile ranges of x; These are the standardized eigenvalues; These are the 25th and 75th percentiles of the data, respectively.

[0110] S16. Noise adjustment for data distribution perception: reducing noise in high-density feature regions, with the following scale parameter:

[0111] ;

[0112] In the formula, For feature probability density;

[0113] S17. Differential privacy aggregation: District-level samples are aggregated first, then noise is added; the aggregated noise is individual noise. The calculation is as follows:

[0114] ;

[0115] In the formula, Let be the original feature value of the i-th sample; N is the total number of samples; The aggregated feature values ​​after adding Laplacian noise are obtained, i.e., differential privacy protection is introduced on the basis of the mean.

[0116] S18, Post-clustering processing correction;

[0117] S181, Median smoothing, using the cluster median instead of the mean to eliminate noise:

[0118] ;

[0119] In the formula, is the median smoothed typical characteristic value of the c-th cluster;

[0120] S182, Boundary Sample Validation: For cross-cluster boundary samples, secondary classification is performed using physical labels.

[0121] The present invention has the following advantages:

[0122] This invention provides a fast, lightweight LDP-MST algorithm based on novel load and distributed resource clustering. Through lightweight design and optimization of the topological characteristics of the Minimum Spanning Tree (MST), it simplifies redundant computational steps in traditional clustering, reducing computational complexity from O(N²) to O(NlogN). It avoids the inefficient process of traversing all sample pairs in traditional density clustering, enabling rapid identification of similar sample groups and supporting real-time analysis and decision-making in large-scale scenarios. Furthermore, by addressing four core dimensions—computational efficiency, clustering accuracy, privacy-accuracy balance, and dynamic adaptability—and considering the characteristics of source-load data (temporal sequence, high dimensionality, and dynamism) and power grid application scenarios, the efficiency and accuracy of the model are optimized. Attached Figure Description

[0123] Figure 1 This is a schematic diagram of the lightweighting method of the present invention. Detailed Implementation

[0124] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0125] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0126] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0127] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0128] like Figure 1 As shown, a fast LDP-MST lightweight method based on novel load and distributed resource clustering is presented. This method includes,

[0129] Data acquisition and preprocessing for new loads and distributed resources;

[0130] A two-tier edge-cloud architecture is constructed, with edge terminals deployed in 10kV distribution areas to collect new load data and distributed resource data in real time.

[0131] The new load data includes charging power sequences of electric vehicle charging stations. Sampling interval 15 minutes; Time-sharing load of smart buildings Real-time power consumption of 5G base stations ;

[0132] Distributed resource data includes residential solar power output. Distributed wind power output Energy storage charging and discharging power Positive values ​​indicate charging, while negative values ​​indicate discharging.

[0133] The collected data were cleaned using the 3σ criterion. For the power sequence of each sample, the mean μ and standard deviation σ were calculated with a 7-day window. Sample values ​​deviating from μ ± 3σ were marked as outliers; outliers were then removed. Samples were marked as anomalies; for newly deployed new loads or distributed resources, if the sample data is less than 7 days old, historical data from the same type of transformer area was used instead, and missing values ​​were filled using interpolation.

[0134] ;

[0135] In the formula, , These are the valid points before and after the missing point, respectively; , These represent the time interval between the missing point and the preceding and following valid points; if more than 3 consecutive missing values ​​are found, they are filled using historical data from the same period.

[0136] Finally, the standardization process uses the extreme value method to map to the [0, 1] interval:

[0137] ;

[0138] In the formula, This is the standardized power value; This is the original power value; , These are the minimum and maximum power in the sample set, respectively;

[0139] After preprocessing, the data forms a sample set to be clustered.

[0140] ;

[0141] In the formula, n is the total number of samples.

[0142] Key features are extracted, and the weights of each feature are calculated using the entropy method to construct a multidimensional feature vector. The extracted features include time-series features, fluctuation features, and controllability features.

[0143] The time-series characteristics include the times when peaks and troughs occur, and the peak times of the power curve are identified by a sliding window (window size 2 hours). Valley value time ;

[0144] And the average daily load factor:

[0145] ;

[0146] In the formula, The rated power of the equipment; 96 represents the number of sampling points at 15-minute intervals per day; Let be the power value at time t;

[0147] The fluctuation characteristics include short-time fluctuation frequency, which is the number of times the power change exceeds 10% of the rated power per unit time. ;

[0148] And power fluctuation coefficient:

[0149] ;

[0150] In the formula, This represents the average power. ;

[0151] The controllability feature includes the adjustment response speed, which is the time τ (minutes) from receiving the control command to the power adjustment being in place.

[0152] And the ratio of capacity that can be reduced:

[0153] ;

[0154] In the formula, This represents the maximum power that can be reduced.

[0155] The feature weights are calculated using the entropy method. When calculating the feature weights using the entropy method, the first... The first sample Normalized proportion of each feature It refers to the standardized feature value of the sample. In the The proportion of each feature in the sum of all standardized values ​​of all samples:

[0156] ;

[0157] In the formula, n is the total number of samples, and the denominator is the sum of the standardized values ​​of all samples for the j-th feature, ensuring... This satisfies the basic requirements of normalization;

[0158] Feature weights are calculated using the entropy method:

[0159] ;

[0160] In the formula, For the first In the nth sample The normalized percentage of each feature; For feature dimensions; For the first The weights of each feature; Let be the information entropy of the j-th feature.

[0161] Privacy protection is implemented based on the LDP mechanism; the privacy protection process employs a layered noise addition strategy, specifically, for highly sensitive features (such as charging periods)... Strong privacy protection is adopted:

[0162] ;

[0163] In the formula, These are highly sensitive feature values ​​after differential privacy protection; These are the original true feature values; ( ) represents a Laplace distribution used to generate random noise that satisfies differential privacy requirements;

[0164] Low-sensitivity features (such as the volatility coefficient δ) employ weak privacy protection:

[0165] ;

[0166] In the formula, These are low-sensitivity eigenvalues ​​after differential privacy protection; the probability density function of Laplace noise is... ; For scale parameters; For feature sensitivity (maximum possible change); , Privacy budgets for high-sensitivity and low-sensitivity features are set based on the power data security standard GB / T36952, with a privacy budget for high-sensitivity features. Low sensitivity characteristics .

[0167] Constructing a lightweight minimum spanning tree (MST) specifically includes:

[0168] Dimensionality reduction optimization is performed using Principal Component Analysis (PCA) to reduce the dimensionality of the privacy-preserving feature set, and the covariance matrix is ​​calculated.

[0169] ;

[0170] In the formula, Represent the covariance matrix; Let be the feature vector of the i-th sample after privacy protection; This is the mean vector of the feature vectors of all samples after privacy protection.

[0171] The eigenvalue decomposition of the covariance matrix yields the eigenvector matrix V. The first k principal components (cumulative contribution rate ≥ 90%) are selected, and the mapping formula is as follows:

[0172] ;

[0173] In the formula, The eigenvector matrix; A feature set after privacy protection;

[0174] Distance calculation is simplified by defining a weighted Euclidean distance, with the following formula:

[0175] ;

[0176] In the formula, This represents the weighted Euclidean distance between the i-th sample and the j-th sample; represents the weight of the l-th feature; represents the projection value of the i-th sample onto the l-th principal component; This represents the projection value of the j-th sample onto the l-th principal component;

[0177] Set distance threshold Only retain Sample pairs;

[0178] Parallel computing units are built using FPGA hardware to process the distance matrix (size: Divide by block (each block) The sample pairs are distributed across 16 independent processing cores, enabling distance calculation for 8 sample pairs to be completed within a single clock cycle. Combining the Kruskal algorithm (a greedy algorithm for finding the minimum sum of edge weights and acyclic elements in an undirected weighted graph) with a disjoint-set data structure to construct a minimum spanning tree (MST), the time complexity is optimized from... ( (For the inverse Ackermann function) further reduced to an approximation .

[0179] Clustering based on the Minimum Spanning Tree (MST) yields clusters, including:

[0180] Calculate the median weight of all edges in the minimum spanning tree (MST), and set a cutting threshold:

[0181] ;

[0182] In the formula, The median of all edge weights in the minimum spanning tree (MST); This is the adjustment coefficient;

[0183] When the target number of clusters is C, the number of clusters is converged to C by adjusting α using the bisection method. If the number of clusters Then increase (Step size 0.1); if Then decrease The process continues until the number of clusters converges to C; after splitting, C connected components are formed, each component corresponding to a cluster G1, G2, ..., Gc, satisfying that the distance between samples within the same category is less than 1 / 2. The distance between out-of-class samples is greater than .

[0184] The cluster results are dynamically updated; the dynamic update of cluster results adopts a feature monitoring-local adjustment mechanism, which collects newly added data daily and calculates the feature deviation between its features and the average feature deviation of its respective cluster.

[0185] ;

[0186] In the formula, The feature deviation between the newly added sample i and its cluster; Let be the average value of the j-th feature of cluster c; Let i be the j-th feature value of the newly added sample i; For the first The weights of each feature; For feature dimensions;

[0187] Based on historical data validation (when the deviation exceeds 10%, the sample regulation potential error is ≥20%), the following settings are established: ,when When a sample is identified as having significant feature changes, only the distance between that sample and other samples is recalculated, the local topology of the minimum spanning tree (MST) is updated, and the clustering affiliation is adjusted. There is no need to reconstruct the entire tree structure. The update cycle is adaptively set according to the intensity of source load fluctuations. When photovoltaic output fluctuates greatly in summer, the update cycle is set to 1 day; when the load is stable in winter, the update cycle is set to 7 days.

[0188] Standardized parameters are used to generate power grid adaptation parameters;

[0189] The standardized parameter output generates three types of power grid adaptation parameters, including:

[0190] Typical curves for clusters:

[0191] ;

[0192] In the formula, Assign sample weights (based on data quality); Let be the typical power value of the c-th cluster at time t; This represents the actual power value of the i-th sample at time t.

[0193] Indicators of regulatory potential:

[0194] Total adjustable capacity of the cluster:

[0195] ;

[0196] In the formula, The maximum adjustable power of the i-th sample after it has been cut;

[0197] The average response time is:

[0198] ;

[0199] In the formula, This represents the average response time of cluster c. Let be the individual response time of the i-th sample; For the first One cluster;

[0200] Clustering confidence:

[0201] ;

[0202] In the formula, The weighted Euclidean distance is used; the cluster confidence value ranges from [0, 1], and the closer it is to 1, the better the clustering effect.

[0203] The above three types of grid adaptation parameters are converted into a format recognizable by the dispatching system through the IEC61850 protocol, directly connecting to the distribution network energy management system (EMS) to support source-load coordinated dispatching decisions.

[0204] Efficiency optimizations were made to the LDP-MST lightweight algorithm, including:

[0205] S1. Compress and perform block-based computation on the adaptive data, specifically including:

[0206] S11. Implement wavelet transform compression by performing a three-level decomposition of the power time series characteristic data using the db4 wavelet basis.

[0207] The first layer retains high-frequency coefficients, reflecting fluctuations at the 15-minute level;

[0208] The second layer retains the mid-frequency coefficients, reflecting hourly trends;

[0209] The third layer quantizes low-frequency coefficients, reflecting the daily baseline;

[0210] The compressed data size is reduced by 65%, and low-frequency coefficients are quantified. The compression formula is as follows:

[0211] ;

[0212] In the formula, W is the wavelet transform matrix; To reduce errors, the data volume was reduced by 65%; This is the sample set after wavelet transform compression; This is the preprocessed sample set to be clustered.

[0213] S12. Perform block clustering, dividing the area into sub-blocks according to voltage level (e.g., 10kV transformer sub-blocks). Construct a minimum spanning tree (MST) independently for each sub-block. Calculate the inter-block distance using core samples (sub-block cluster centers) and perform global cluster fusion.

[0214] ;

[0215] In the formula, This represents the global clustering result; The core sample of sub-block b; This is the inter-block fusion threshold; This is the core sample of sub-block d; The core sample distance between the c-th sub-block and the d-th sub-block;

[0216] S2. Heterogeneous hardware collaborative acceleration, multiple GPUs perform parallel computing, deploying the sample distance matrix calculation on GPUs and utilizing CUDA kernel functions for parallelization, with a single kernel function handling the process. Level 1 sample pairs improve distance calculation efficiency by 5-8 times. The calculation formula is as follows:

[0217] ;

[0218] In the formula, The distance between the i-th sample and the j-th sample is calculated using GPU acceleration. This is a parallel computation function, executed collaboratively by multiple GPUs; Let be the vector of the i-th sample in the reduced feature space; Let be the vector of the j-th sample in the reduced-dimensional feature space; The feature weight vector;

[0219] Edge-cloud collaboration, edge terminal processing of area-level samples (≤10) 4 The cloud-based system outputs local cluster centers based on the central samples, reducing the amount of communication data to 5% of the total transmission volume.

[0220] S13, Minimum Spanning Tree (MST) Construction and Pruning Optimization;

[0221] S131, Approximate MST algorithm, when At that time, random sampling s Construct an initial minimum spanning tree (MST) from the samples, and then perform a k-nearest neighbor search. After correcting the critical edges, the approximation error is ≤3%, and the time complexity is optimized to... ;

[0222] S132, Dynamic threshold pruning, adjusting the distance threshold based on sample density:

[0223] ;

[0224] In the formula, For sample density; As a baseline threshold, high-density areas (such as urban load) Reduce the size of the computation, thereby minimizing unnecessary calculations.

[0225] S14. Improved time-aware distance measurement;

[0226] S141, Dynamic Time Warping (DTW), replacing Euclidean distance for handling time offsets:

[0227] ;

[0228] In the formula, Using time alignment functions, cluster purity is improved by 15%-20% compared to Euclidean distance; Let be the dynamic time-normalized distance between the i-th sample and the j-th sample; Let be the power value of the i-th sample at time t; Let be the power value of the j-th sample at the corresponding time after time alignment; To find the optimal path that minimizes the total distance among all possible time-aligned paths;

[0229] S142. Physical constraint weighting, combining geographical distance and temporal distance, the total distance is:

[0230] ;

[0231] In the formula, Geographical distance; To maximize geographical span, enhance the physical consistency of photovoltaic / wind power clusters.

[0232] S15, Feature engineering for embedding domain knowledge;

[0233] S151. Construction of cross-features, adding peak load overlap. matching degree with source and load :

[0234] ;

[0235] In the formula, For sample i, the peak load period is... This represents the peak load period for the j-th sample; Pearson correlation coefficient; Peak load overlap; Source-load matching degree; Let be the power generation of the i-th sample at time t; Let be the load power of the j-th sample at time t;

[0236] S152, Median standardization, employing anti-extreme value processing for fluctuation characteristics:

[0237] ;

[0238] In the formula, These are the original eigenvalues; Let x be the median; The difference in interquartile ranges of x; These are the standardized eigenvalues; These are the 25th and 75th percentiles of the data, respectively.

[0239] S16. Noise adjustment for data distribution perception: reducing noise in high-density feature regions, with the following scale parameter:

[0240] ;

[0241] In the formula, The feature probability density is used; after adjustment, the noise intensity is reduced by 30%, and the cluster center error is ≤5%.

[0242] S17. Differential privacy aggregation: District-level samples are aggregated first, then noise is added; the aggregated noise is individual noise. The calculation is as follows:

[0243] ;

[0244] In the formula, Let be the original feature value of the i-th sample; N is the total number of samples; The aggregated feature values ​​after adding Laplacian noise are obtained, i.e., differential privacy protection is introduced on the basis of the mean.

[0245] S18, Post-clustering processing correction;

[0246] S181, Median smoothing, using the cluster median instead of the mean to eliminate noise:

[0247] ;

[0248] In the formula, is the median smoothed typical characteristic value of the c-th cluster;

[0249] S182. Boundary sample verification: For cross-cluster boundary samples, secondary classification is performed using physical labels (such as charging pile type), which reduces the misclassification rate by 10%.

[0250] This invention reduces the traditional O(N²) similarity matrix construction to O(NlogN) through a triple compression of principal component dimensionality reduction, hierarchical LDP noise, and distance threshold pruning, achieving a memory footprint of ≤15% of the original size. Utilizing FPGA-GPU heterogeneous parallelism and an approximate Kruskal algorithm, the MST construction time is compressed from hours to seconds, meeting the second-level scheduling requirements of the power grid. While maintaining clustering purity ≥95%, the amount of communication data is reduced by 80%, achieving lightweight real-time clustering under an edge-cloud collaborative framework. Through multi-dimensional optimization, the fast LDP-MST lightweight algorithm constructed in this invention can adapt to the ultra-large-scale, highly time-varying, and highly sensitive source-load clustering requirements of new power systems, providing core technical support for distribution network planning and source-load collaborative scheduling.

[0251] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.

Claims

1. A fast, lightweight LDP-MST method based on novel load and distributed resource clustering, characterized by: The method includes the following steps: Collect new types of load data and distributed resource data, and preprocess the collected data; Extract key features, calculate the weights of each feature using the entropy method, and construct a multidimensional feature vector. Privacy protection processing of features is performed based on the LDP mechanism; A lightweight minimum spanning tree (MST) is constructed based on privacy-preserving features. Clusters are obtained by clustering based on the minimum spanning tree (MST). Dynamically update cluster results; Standardized parameters are used to generate power grid adaptation parameters; Efficiency optimization of the LDP-MST lightweight algorithm.

2. The fast LDP-MST lightweight method based on novel load and distributed resource clustering as described in claim 1, characterized in that: Preprocessing the collected data includes cleaning the data and removing outliers; and imputing missing values ​​using interpolation. ； In the formula, , These are the valid points before and after the missing point, respectively; , These represent the time intervals between missing points and the preceding and following valid points, respectively. Finally, the extreme value method is used to process the data: ； In the formula, This is the standardized power value; This is the original power value; , These are the minimum and maximum power in the sample set, respectively; After preprocessing, the data forms a sample set to be clustered. ； In the formula, n is the total number of samples.

3. The fast LDP-MST lightweight method based on novel load and distributed resource clustering as described in claim 1, characterized in that: The extracted features include time-series features, fluctuation features, and controllability features; The time-series characteristics include the times when peaks and troughs occur, and the average daily load factor. ； In the formula, Rated power of the equipment; Let be the power value at time t; The fluctuation characteristics include short-time fluctuation frequency and power fluctuation coefficient: ； In the formula, This represents the average power. ; The controllability features include adjustable response speed and a scalable capacity ratio: ； In the formula, This represents the maximum power that can be reduced.

4. The fast LDP-MST lightweight method based on novel load and distributed resource clustering as described in claim 3, characterized in that: The weights of each feature are calculated using the entropy method: ； In the formula, For the first In the nth sample The normalized percentage of each feature; For feature dimensions; For the first The weights of each feature; Let be the information entropy of the j-th feature.

5. The fast LDP-MST lightweight method based on novel load and distributed resource clustering as described in claim 1, characterized in that: The privacy protection processing employs a layered noise addition strategy, specifically, highly sensitive features are protected with strong privacy measures: ； In the formula, These are highly sensitive feature values ​​after differential privacy protection; The initial true eigenvalues; ( ) represents the Laplace distribution; Low-sensitivity features employ weak privacy protection: ； In the formula, These are low-sensitivity feature values ​​after differential privacy protection; For feature sensitivity; , Privacy budgets for high / low sensitivity features, respectively.

6. The fast LDP-MST lightweight method based on novel load and distributed resource clustering as described in claim 1, characterized in that: The construction of the lightweight minimum spanning tree (MST) includes... First, principal component analysis (PCA) is used to reduce the dimensionality of the privacy-preserving feature set, and the covariance matrix is ​​calculated: ； In the formula, It is the variance matrix; Let be the feature vector of the i-th sample after privacy protection; This is the mean vector of the feature vectors of all samples after privacy protection. The eigenvalue decomposition of the covariance matrix yields the eigenvector matrix. The first k principal components are selected, and the mapping formula is: ； In the formula, The eigenvector matrix; A feature set after privacy protection; Then, the weighted Euclidean distance is defined, with the formula: ； In the formula, This represents the weighted Euclidean distance between the i-th sample and the j-th sample; This represents the weight of the l-th feature; This represents the projection value of the i-th sample onto the l-th principal component; This represents the projection value of the j-th sample onto the l-th principal component; a distance threshold is set. Only retain Sample pairs; The distance matrix calculation is accelerated by FPGA hardware parallelism, and the minimum spanning tree (MST) is constructed by combining Kruskal's algorithm and disjoint set.

7. The fast LDP-MST lightweight method based on novel load and distributed resource clustering as described in claim 6, characterized in that: The clustering partitioning based on the Minimum Spanning Tree (MST) topology includes... Calculate the median weight of all edges in the minimum spanning tree (MST), and set a cutting threshold: ； In the formula, The median of all edge weights in the minimum spanning tree (MST); This is the adjustment coefficient; When the target number of clusters is C, the number of clusters is converged to C by adjusting α using the bisection method. If the number of clusters Then increase ;like Then decrease The process continues until the number of clusters converges to C; after splitting, C connected components are formed, each component corresponding to a cluster G1, G2, ..., Gc, satisfying that the distance between samples within the same category is less than 1 / 2. The distance between out-of-class samples is greater than .

8. The fast LDP-MST lightweight method based on novel load and distributed resource clustering as described in claim 7, characterized in that: The dynamic updating of cluster results includes, Collect daily new data and calculate the deviation of its features from the average features of its cluster: ； In the formula, The feature deviation between the newly added sample i and its cluster; Let be the average value of the j-th feature of cluster c; Let i be the j-th feature value of the newly added sample i; For the first The weights of each feature; For feature dimensions; Based on historical data verification, the following settings are established. ,when When a sample is identified as having significant feature changes, its distance from other samples is recalculated, the local topology of the minimum spanning tree (MST) is updated, and its cluster affiliation is adjusted.

9. The fast LDP-MST lightweight method based on novel load and distributed resource clustering as described in claim 7, characterized in that: The standardized parameter output generates power grid adaptation parameters including: Typical curves for clusters: ； In the formula, For sample weights; Let be the typical power value of the c-th cluster at time t; This represents the actual power value of the i-th sample at time t. Indicators of controllability potential include total adjustable capacity of the cluster and average response time; The total adjustable capacity of the cluster is: ； In the formula, The maximum adjustable power of the i-th sample after it has been cut; The average response time is: ； In the formula, This represents the average response time of cluster c. Let be the individual response time of the i-th sample; For the first One cluster; Clustering confidence: ； In the formula, It is a weighted Euclidean distance.

10. The fast LDP-MST lightweight method based on novel load and distributed resource clustering according to claim 7, characterized in that: The efficiency optimization of the LDP-MST lightweight algorithm includes: S1. Perform wavelet transform compression, compressing and segmenting the adaptive data, specifically including: S11. The time-series feature data are decomposed into three levels using the db4 wavelet basis. The first layer retains high-frequency coefficients; The second layer retains the intermediate frequency coefficients; Third-level quantization low-frequency coefficients; The compression formula is: ； In the formula, W is the wavelet transform matrix; This is for compression error; This is the sample set after wavelet transform compression; This is the preprocessed sample set to be clustered. S12. Perform block clustering, dividing the data into sub-blocks according to voltage level. Each sub-block independently constructs a minimum spanning tree (MST). Calculate the inter-block distance using core samples, and then perform global cluster fusion. ； In the formula, This represents the global clustering result; This is the core sample of sub-block d; This is the core sample of sub-block d; This is the inter-block fusion threshold; The core sample distance between the c-th sub-block and the d-th sub-block; S2, heterogeneous hardware collaborative acceleration, calculates sample distance in parallel using multiple GPUs, the calculation formula is: ； In the formula, The distance between the i-th sample and the j-th sample is calculated using GPU acceleration. This is a parallel computation function, executed collaboratively by multiple GPUs; Let be the vector of the i-th sample in the reduced feature space; Let be the vector of the j-th sample in the reduced-dimensional feature space; The feature weight vector; The system calculates and outputs local cluster centers, and then performs global clustering in the cloud based on the center samples. S13, Minimum Spanning Tree (MST) Construction and Pruning Optimization; S131. Randomly sample s samples to construct an initial minimum spanning tree (MST), and then perform a k-nearest neighbor search. After correcting the critical edges, the approximation error is ≤3%, and the time complexity is optimized to... ; S132, Dynamic threshold pruning, adjusting the distance threshold based on sample density: ； In the formula, For sample density; The baseline threshold; S14. Improved time-aware distance measurement; S141. Dynamic time warping, replacing Euclidean distance for handling time offsets: ； In the formula, This is a time alignment function; Let be the dynamic time-normalized distance between the i-th sample and the j-th sample; Let be the power value of the i-th sample at time t; Let be the power value of the j-th sample at the corresponding time after time alignment; To find the optimal path that minimizes the total distance among all possible time-aligned paths; S142. Physical constraint weighting, combining geographical distance and temporal distance, the total distance is: ； In the formula, Geographical distance; The maximum geographical span; S15, Feature engineering for embedding domain knowledge; S151, Construction of Cross Features: ； In the formula, For sample i, the peak load period; This represents the peak load period for the j-th sample; Pearson correlation coefficient; Peak load overlap; Source-load matching degree; Let be the power generation of the i-th sample at time t; Let be the load power of the j-th sample at time t; S152, Median standardization, employing anti-extreme value processing for fluctuation characteristics: ； In the formula, These are the original eigenvalues; Let x be the median; The difference in interquartile ranges of x; These are the standardized eigenvalues; These are the 25th and 75th percentiles of the data, respectively. S16. Noise adjustment for data distribution perception: reducing noise in high-density feature regions, with the following scale parameter: ； In the formula, For feature probability density; S17. Differential privacy aggregation: District-level samples are aggregated first, then noise is added; the aggregated noise is individual noise. The calculation is as follows: ； In the formula, Let be the original feature value of the i-th sample; N is the total number of samples; These are the aggregated feature values ​​after adding Laplacian noise; S18. Post-clustering processing correction; S181, Median smoothing, using the cluster median instead of the mean to eliminate noise: ； In the formula, is the median smoothed typical characteristic value of the c-th cluster; S182, Boundary Sample Validation: For cross-cluster boundary samples, secondary classification is performed using physical labels.

Images