Method and system for generating a peak shaving potential energy storage profile based on multi-dimensional clustering

Abstract

This invention belongs to the field of power system and automation technology, and discloses a method and system for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering. The method includes acquiring equipment power measurement data, segmenting it according to a set time interval and constructing a time-point curve matrix; calculating the standard deviation of each time point and mapping it to a time weight vector, weighting each column of the matrix to highlight the characteristics of key time periods; removing noisy time periods using a density clustering algorithm, selecting the main cluster with the most samples and calculating the typical curve of the main cluster; extracting profile feature vectors based on the curves; dividing the profile features into peak-shaving groups and basic scale groups, calculating the objective weights of each group's features using the entropy weight method, and integrating them to form a weighted feature space; scoring the profile potential of each cluster in this space using K-Means clustering to obtain the peak-shaving potential of the energy storage profile. This invention makes the peak-shaving potential assessment of energy storage equipment more realistic, stable, and reproducible.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, and in particular to a method and system for generating peak-shaving potential energy storage profiles based on multi-dimensional clustering. Background Technology

[0002] With the advancement of new power system construction, distributed energy storage, as a key flexibility and regulation resource, is being deployed on a large scale to smooth renewable energy fluctuations, participate in grid peak and frequency regulation, and perform peak shaving and valley filling on the user side. Grid operators and load aggregators need to efficiently integrate and dispatch massive, dispersed, and heterogeneous energy storage resources, especially to achieve precise and reliable output during critical periods such as evening peak hours. However, existing methods are still somewhat crude in the face of massive, heterogeneous equipment, and face the following problems:

[0003] 1) Over-reliance on static nameplate parameters. Common evaluation methods rely solely on static indicators such as rated capacity and rated power to roughly judge equipment capabilities, which fails to reflect key issues such as "how much energy the equipment can actually release during operation within a specific peak period, how long it can continue, and whether it occurs stably every day."

[0004] 2) Poor adaptability to varying operating conditions and noise levels. Actual operating data not only includes "abnormal days" such as shutdowns and maintenance, but also shows significant differences in the density of equipment operating modes, such as high density at full capacity and low density at partial output. Existing methods, such as direct statistical analysis or traditional clustering using a single threshold, struggle to adaptively identify effective patterns at different densities, leading to inaccurate extraction of the main operating patterns.

[0005] 3) Lack of a unified high-dimensional profile and automatic classification mechanism. Existing methods often only provide a single indicator (such as daily average charging and discharging energy), which cannot simultaneously characterize multi-dimensional features such as scale capacity, available energy and average power during critical periods, operational stability and regularity. Furthermore, there is a lack of a general method to automatically generate "potential level labels" from these features, which means that dispatchers still need a lot of manual experience to make judgments.

[0006] 4) The empowerment mechanism lacks business orientation and is subjective in its parameters. Existing profile clustering mostly uses equal-weighted Euclidean distance, ignoring the weight difference between "peak shaving capacity" and "basic scale" in terms of scheduling value, resulting in the classification results failing to accurately identify high-potential equipment. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies, such as excessive reliance on static nameplate parameters for peak shaving and valley filling, poor adaptability to varying operating conditions and noise, lack of business-oriented multi-dimensional profiling and automatic grading mechanisms, and subjective weighting, which lead to rough, inaccurate assessments of the peak shaving potential of energy storage devices and reliance on human experience. This invention provides a method and system for generating peak shaving potential energy storage profiling based on multi-dimensional clustering. The method involves constructing a time-point curve matrix based on energy storage device measurement data, calculating the standard deviation of each time point and mapping it to time weights for weighting, selecting the main cluster to calculate typical curves after noise removal through density clustering, extracting multi-dimensional profiling features and dividing them into peak shaving groups and basic scale groups, using entropy weighting to calculate objective weights within each group and constructing a weighted feature space, and scoring the profiling potential of each cluster using K-Means clustering. This process achieves the goal of generating multi-dimensional peak shaving potential energy storage profiling and accurately assessing the peak shaving potential of energy storage devices.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] The method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering includes the following steps:

[0010] Step 1: Obtain the power measurement data of the device, divide the power measurement data into sets of time intervals, and construct a time point curve matrix;

[0011] Step 2: Calculate the standard deviation of each time point based on the time point curve matrix, map the standard deviation vector to the time weight vector, multiply each column of the time point curve matrix by the corresponding time weight, weight the features of the time point curve matrix, perform clustering operation on the weighted time point curve matrix through density clustering algorithm to remove noise time, select the main cluster with the most samples and calculate the typical curve of the main cluster.

[0012] Step 3: Extract the feature vectors that constitute the portrait based on the typical curves of the main cluster;

[0013] Step 4: Divide the profile features into peak shaving group and basic scale group. Use the entropy weight method to calculate the objective weight of the profile features in each group. After integrating the objective weight into the profile features, a weighted feature space is formed. In the weighted feature space, K-Means clustering is used to score the profile potential in each cluster to obtain the peak shaving potential of the energy storage profile.

[0014] Preferably, in step 1, the power measurement data is further denoised. If there is no valid power measurement data at a certain time point, the power measurement data at that time point is deleted from the time point curve matrix.

[0015] Preferably, in step 2, the time weight is set as follows: the time point with a larger standard deviation has a higher weight, and the time point with a smaller standard deviation has a lower weight.

[0016] Preferably, in step 2, the stability of the equipment operation mode and the reliability of the main cluster extraction are judged by statistically analyzing the main cluster ratio, noise ratio, and contour coefficient of the density clustering algorithm in the time point curve matrix. If the stability of the equipment operation mode is lower than the stability threshold or the reliability of the main cluster extraction is lower than the reliability threshold, the parameters of the density clustering algorithm are adjusted for re-clustering until the stability of the equipment operation mode and the reliability of the main cluster extraction are above the set threshold.

[0017] Preferably, in step 3, the profile feature vector includes basic scale features, peak-shaving potential features, stability features, and regularity features. The profile feature vectors of all devices are summarized into a unified feature table for data input in step 4.

[0018] Preferably, core features are extracted through the profile feature vector. The core features include peak available energy, peak average power, peak duration, peak CV, scale power, scale energy, and operational regularity. The core features are Z-score standardized to obtain a standardized feature matrix. In step 4, the profile features are divided into a peak shaving group and a basic scale group. The core features of the peak shaving group include peak available energy, peak average power, peak duration, and peak CV. The core features of the basic scale group include scale power, scale energy, and operational regularity.

[0019] Preferably, in step 4, the objective weights of the portrait features in each group are calculated using the entropy weight method. These objective weights are then incorporated into the portrait features to form a weighted feature space. Specifically:

[0020] First, the data within the peak shaving group and the basic scale group are normalized and the information entropy is calculated. The profile feature with greater variability receives a higher weight. The information entropy weight is multiplied by the business weight between groups to obtain the hierarchical mixed weight of each profile feature. In K-Means clustering, the weighted feature space is obtained by multiplying each dimension of the standardized feature matrix by the square root of the corresponding hierarchical mixed weight.

[0021] As a preferred method, the generation method for peak-shaving potential energy storage profiles based on multi-dimensional clustering further optimizes the number of clusters in K-Means clustering, specifically as follows:

[0022] K-Means clustering is performed in the weighted feature space to calculate the corresponding DBI index and SI (Silhouette) index. Then, the DBI index is sorted in ascending order and the SI index is sorted in descending order. The two rankings are added together to obtain a comprehensive index. The optimal number of clusters is selected as the optimal number of clusters based on the comprehensive index with the smallest sum of rankings.

[0023] Preferably, in step 4, the potential of the energy storage profile is scored in each cluster using K-Means clustering in the weighted feature space to obtain the peak-shaving potential of the energy storage profile. Specifically:

[0024] For each cluster center, a peak comprehensive score is constructed. The standardized values ​​of each cluster center in four dimensions—peak available energy, peak average power, peak duration, and peak CV—are calculated. The comprehensive score is constructed in a linear combination form, positively rewarding high energy, high power, and long duration, and negatively penalizing high volatility. Each cluster is sorted from largest to smallest based on the comprehensive score, and the cluster label and potential level of each device are written into the energy storage profile list to realize the peak shaving potential classification of the device's energy storage profile.

[0025] A system for generating peak-shaving potential energy storage profiles based on multi-dimensional clustering is applicable to methods for generating peak-shaving potential energy storage profiles based on multi-dimensional clustering, including:

[0026] The module includes: data acquisition and matrix construction, weight calculation and data weighting, density clustering and typical curve generation, portrait feature extraction, feature grouping and mixed weighting, weighted feature space construction, and cluster scoring and potential classification.

[0027] The data acquisition and matrix construction module is used to acquire the power measurement data of the device, divide the data according to the set time interval, and construct a time point curve matrix.

[0028] The weight calculation and data weighting module is used to calculate the standard deviation of each time point in the time point curve matrix, map the standard deviation vector into a time weight vector, and multiply each column of the time point curve matrix by the corresponding time weight to achieve feature weighting.

[0029] The density clustering and typical curve generation module is used to cluster the weighted time point curve matrix using a density clustering algorithm, remove noisy time data, select the main cluster with the largest number of samples, and calculate the typical curve of the main cluster.

[0030] The portrait feature extraction module is used to extract portrait feature vectors based on the typical curves of the main cluster.

[0031] The feature grouping and hybrid weighting module is used to divide the portrait feature vector into peak-shaving group and basic scale group, calculate the objective weight of the portrait features within each group using the entropy weight method, and obtain the hierarchical hybrid weight by combining the business weights between groups.

[0032] The weighted feature space construction module is used to integrate hierarchical mixed weights into the profile features to construct a weighted feature space;

[0033] The clustering scoring and potential grading module is used to score the potential of each cluster profile in the weighted feature space through K-Means clustering, and output the peak-shaving potential level of the energy storage profile.

[0034] The beneficial effects of this invention are: 1. The evaluation results are accurate and reliable, solving the problem of dual interference from noisy periods and low-information periods. This invention introduces adaptive density clustering with automatic variance weighting, automatically calculating variance weights according to the time dimension, assigning higher clustering weights to key periods with high variability and significant differences, and suppressing low-information background periods such as late-night standby. Combined with the adaptive density clustering mechanism of the density clustering algorithm, it can automatically identify and eliminate noisy periods such as downtime, maintenance, and communication anomalies, constructing typical curves using only the daily curves of the stable main clusters, fundamentally avoiding the risk of traditional curves being diluted by outliers and low-value periods, and significantly improving the authenticity and stability of subsequent profile features.

[0035] 2. A multi-dimensional profile feature system strongly coupled with peak shaving operations has been constructed. This invention breaks away from the crude reliance on "nameplate capacity" or a single total power indicator. Instead, based on the "typical curves of the main clusters" selected by the variance-weighted density clustering algorithm, a multi-dimensional profile feature system covering "scale, peak capacity, stability, and regularity" has been systematically constructed. This system not only includes features that directly characterize peak shaving capabilities, such as available energy during the evening peak, average power during peak periods, and duration of high load, but also explicitly introduces reliability indicators such as the peak power variation coefficient and the regularity of daily operation. This achieves an integrated quantitative description of "energy availability, power output, output stability, and regularity of operation patterns," enabling a more comprehensive and detailed characterization of the true dispatchable potential of energy storage equipment during critical peak periods.

[0036] 3. This invention achieves business-oriented automated profiling and horizontal comparison. It introduces hierarchical hybrid weighting combined with weighted K-Means clustering to automatically classify standardized fused profiles of all devices. First, features are divided into "peak shaving groups" and "basic scale groups." Within each group, entropy weighting is used to objectively characterize the information content of each indicator, emphasizing the importance of peak shaving-related features. Based on this, K-Means clustering is run using weighted Euclidean distance, and a comprehensive ranking index is constructed using DBI and SI to adaptively optimize the number of candidate clusters, overcoming the shortcomings of traditional K-value selection which relies on experience and is difficult to reproduce programmatically. Furthermore, this invention constructs a comprehensive score based on peak-related features, comprehensively considering peak energy, average power, long duration, and output stability, and sorts and names each cluster, naturally embedding the potential level as a profile label. This achieves unified, objective, and business-oriented horizontal comparative evaluation of massive heterogeneous devices. Attached Figure Description

[0037] Figure 1 This is a flowchart of the present invention;

[0038] Figure 2 This is a schematic diagram of the portrait clustering and potential classification process of this invention. Detailed Implementation

[0039] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0040] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0041] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0042] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0043] Example: A method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering, such as... Figure 1 As shown, it includes the following steps:

[0044] Step 1: Obtain the power measurement data of the device, divide the power measurement data into segments according to the set time intervals, and construct a time point curve matrix; also, denoise the power measurement data, and if there is no valid power measurement data within a certain time point, delete the power measurement data of that time point in the time point curve matrix.

[0045] Step 2: Calculate the standard deviation of each time point based on the time-point curve matrix. Map the standard deviation vector to a time weight vector. Multiply each column of the time-point curve matrix by its corresponding time weight to weight the features of the time-point curve matrix. Perform clustering operations on the weighted time-point curve matrix using the OPTICS density clustering algorithm to remove noisy times, select the main cluster with the most samples, and calculate the typical curve of the main cluster. The time weights are set as follows: the larger the standard deviation, the higher the weight; the smaller the standard deviation, the lower the weight. Furthermore, by statistically analyzing the proportion of main clusters, the noise proportion, and the contour coefficient of the density clustering algorithm in the time-point curve matrix, the stability of the equipment operation mode and the reliability of the main cluster extraction are judged. If the stability of the equipment operation mode is lower than the stability threshold or the reliability of the main cluster extraction is lower than the reliability threshold, the parameters of the density clustering algorithm are adjusted, and re-clustering is performed until the stability of the equipment operation mode and the reliability of the main cluster extraction are above the set thresholds.

[0046] Step 3: Extract the profile feature vector based on the typical curve of the main cluster; the profile feature vector includes basic scale features, peak shaving potential features, stability features and regularity features. The profile feature vectors of all devices are summarized into a unified feature table for data input in step 4.

[0047] Step 4: Divide the profile features into peak shaving group and basic scale group. Use the entropy weight method to calculate the objective weight of the profile features in each group. After integrating the objective weight into the profile features, a weighted feature space is formed. In the weighted feature space, K-Means clustering is used to score the profile potential in each cluster to obtain the peak shaving potential of the energy storage profile.

[0048] This embodiment is based on 15 minutes of particle size measurement data from energy storage devices and adopts an overall framework of "two-stage + hybrid weighting":

[0049] The first phase addresses the "high noise and variability" characteristics of massive time-series data by introducing a combined strategy of "automatic standard deviation weighting + OPTICS adaptive density clustering." Through a data-driven weighting mechanism, it focuses on key working periods, adaptively identifies and locks the main operating mode of equipment to obtain typical daily curves and multi-dimensional profile features. The second phase addresses the pain point of "unclear guidance" in traditional evaluation by constructing a hybrid weighting mechanism of "business layering + entropy weighting." It also automatically selects the number of clusters K by combining the DBI index and SI index. Based on this, a "peak comprehensive score" is constructed, and each cluster is sorted and named according to high / medium / low potential, realizing full automation from raw data to potential level labels.

[0050] In terms of data organization, the smallest modeling unit is "user-asset number". The system automatically traverses the raw data files in the specified directory, identifies the asset number column in the file, and processes them separately by device to avoid interference from multiple assets. The raw time series is divided by natural day and aligned to a standard 15-minute grid to construct a daily curve matrix of 96 points per day. The power vector on day i is denoted as To ensure input integrity, as long as there are valid measurement points on a given day, the system uses linear interpolation to fill in 96 points within that day's range. Only when there is no valid data for the entire day is the day removed. Then, the daily curves are standardized using a point-by-point global Z-score. Based on this, the OPTICS algorithm with automatic variance weighting is used to automatically identify and isolate "noise days" such as downtime, maintenance, and communication anomalies. The main operating mode is adaptively locked, and the typical curve of the main cluster is calculated accordingly for subsequent portrait feature extraction.

[0051] Specifically, the first stage, based on standard deviation weighted OPTICS, aims to identify the main operating mode and construct typical curves from the daily curve matrix, which includes abnormal operating conditions such as downtime, maintenance, and strategy changes. From this, a reliable "typical cluster curve" that can represent the stable operating conditions of the equipment was extracted.

[0052] To reduce the influence of dimensions and preserve the "amplitude difference" between different operating days, this embodiment modifies the matrix. Each column (i.e., the same time of day) is subjected to point-by-point global Z-score normalization to obtain a normalized matrix. Unlike daily normalization, this "time-by-time global normalization" will not stretch "high-output days" and "low-output days" into the exact same shape, so that "full-load" and "shallow-load" operation days can still be distinguished in subsequent clustering.

[0053] In this embodiment, the standard deviation of each time point in the historical samples is calculated to construct a set of time weight vectors: the larger the standard deviation, the higher the weight of the time (usually corresponding to high information intervals such as daytime working hours and evening peak hours), and the lower the weight of the time with extremely small standard deviation (such as late-night standby and periods with almost no change). By multiplying each column of the normalized matrix by the corresponding weight, an automatic weighted variance representation that "amplifies key time periods and suppresses background time periods" is formed in the feature space.

[0054] In the weighted feature space, the OPTICS density clustering algorithm is used to identify the main operating mode. OPTICS constructs a reachability map through core distance and reachability distance, and can identify operating modes at different density levels simultaneously without relying on a single density threshold. The algorithm automatically divides the system into clusters based on the "valley structure" of the reachability map, identifying outlier operating days such as downtime, maintenance, and communication anomalies as noise points (denoted as cluster −1) and removing them from subsequent analysis. Then, the cluster with the largest number of samples among all non-noise clusters is selected as the set of main operating modes for the equipment. The daily curves within this set are averaged over time in the original power dimension to generate the "typical curve of the main cluster".

[0055] Meanwhile, the system records statistics such as the percentage of primary clusters, the percentage of noise, and the OPTICS profile coefficient to reflect the stability of the equipment's operating mode and the reliability of primary cluster extraction.

[0056] The specific calculation process is as follows:

[0057] First, in the original daily curve matrix The standard deviation of each time point is calculated column-wise to characterize the degree of variation among all sample days at different time points. Let the first... The mean of each time point over all sample days is The standard deviation is:

[0058] ;

[0059] in, Indicates the number of sample days. This represents the element in the i-th row and t-th column of the original daily curve matrix X.

[0060] Then, the standard deviation vector Mapped to time weight vector For example, by linear scaling to a range In this embodiment, we take This ensures that low-information periods are not completely ignored, while significantly amplifying the impact of high-variability peaks:

[0061] ;

[0062] in These represent the minimum and maximum standard deviations for all time points, respectively. To prevent small values ​​from being divided by zero, peak times with high variability (significant differences) are automatically given higher weights, while times with low variability (such as those in standby mode at night) are given relatively lower weights.

[0063] To reduce the influence of dimensions and ensure that clustering focuses primarily on "shape + relative amplitude," after constructing time weights, point-by-point global Z-score standardization is performed on the daily curves (96 points per day): This involves applying a matrix... Each column (i.e., each time point) ), calculate its mean over all sample days. with standard deviation Standardization process to obtain :

[0064] ;

[0065] in To prevent zero-value items, this preprocessing method, while preserving macroscopic differences such as "full placement and shallow placement," eliminates dimensional differences at different time points, making subsequent distance measurements more stable.

[0066] The time weights are explicitly injected into the standardized matrix in the form of a "weighted Euclidean distance":

[0067] ;

[0068] Then the daily curves for any two days The distance between them is:

[0069] ;

[0070] and similar, and Similarly, the only difference lies in the different expressions for "i days" and "j days".

[0071] This means that in OPTICS, a time-weighted Euclidean distance naturally emerges where "times with higher weights contribute more." This is reflected in the weighted feature space. Execute the OPTICS clustering algorithm on the above:

[0072] 1) Core distance and reachability distance calculation: OPTICS does not explicitly generate clusters, but calculates the core distance and reachability distance for each sample to generate a reachability graph that reflects the data density structure. This mechanism allows the algorithm to identify clusters of different densities under a weighted distance metric, and is especially able to adapt to the variable operating conditions of energy storage under the premise of emphasizing the differences during the evening peak period.

[0073] 2) Adaptive cluster extraction: Employing a slope-based extraction method ( Using either a threshold-based or threshold-based method, automatically identify "valley" structures as clusters on the reachability graph, and label samples that do not belong to any cluster as "noise days" (label=-1) and remove them.

[0074] 3) Main Cluster Selection: Count the number of samples in all non-noise clusters, and select the cluster with the largest number of samples as the main operating mode set for this device. .

[0075] This is based on the main running mode set. Calculate the typical curve of the principal cluster as a benchmark for subsequent feature extraction and hierarchical evaluation. Let the first... The daily power curve (96-point alignment) is The typical curve of the principal cluster is then defined as:

[0076] ;

[0077] The curve is calculated in the original dimension (kW) to maintain a one-to-one correspondence with the original time series, and is used for subsequent calculation and evaluation of profiling features such as scale, peak capacity, stability, and regularity. In the main operating mode recognition, the minimum number of samples for the main cluster of OPTICS is set to... That is, it should account for at least about 15% of the historical sample days, and no less than 5 days.

[0078] Simultaneously record denoising statistics, such as the proportion of OPTICS noise:

[0079] ;

[0080] Percentage of main cluster:

[0081] ;

[0082] Used as a reference for image quality, where N represents the number of calendar days. Image features are based on... Extracting and supplementing with necessary cross-day statistics, an indicator system of "scale—peak capacity—stability—regularity" is formed, with "discharge as positive and charging as negative" defined. Typical daily discharge / charge amounts are:

[0083] ;

[0084] in, h, peak power To reduce peak traffic during the evening rush hour, peak-hour windows... As a configurable option (default 17:00–21:00, corresponding to a 15-minute granularity index). According to the statistical methods used in this plan, peak available energy and peak-period average power are calculated as net values:

[0085] ;

[0086] in, This represents the average power during the evening rush hour. This indicates the number of time points included in the peak window W.

[0087] To characterize the stability of the peak output, the coefficient of variation (CV) is used:

[0088] ;

[0089] in, This represents the standard deviation of power during peak hours. This represents the power variation coefficient during peak hours. To reflect the "high-intensity-long-lasting" supply capacity, it statistically measures the duration of peak periods exceeding a certain percentage (80%, with the 80% threshold representing a trade-off between filtering short-term spikes and maintaining near-full power output).

[0090] ;

[0091] The daily curve within the main cluster is used to predict the cross-day pattern. With the mean of the main cluster Invert the normalized average relative error:

[0092] ;

[0093] This forms the device profile vector. :

[0094] ;

[0095] in, Indicates scale power.

[0096] Therefore, device profile vectors can be mainly divided into:

[0097] 1) Basic scale characteristics: such as typical daily peak power (approximately representing the maximum output that the equipment can achieve under stable operating conditions), typical daily total discharge energy (characterizing the total energy that can be released in a single day), scale power approximation and scale energy approximation, etc., are used to characterize the overall scale and capacity limit of the equipment;

[0098] 2) Peak shaving potential characteristics: For the evening peak period (corresponding to 17:00–21:00), calculate the peak available energy (peak power integral), peak average power, and the duration when the peak power is not lower than a certain percentage of the peak value, reflecting the equipment's peak shaving capability and continuous supply capability within the target window.

[0099] 3) Stability characteristics: such as the power variation coefficient CV (the ratio of standard deviation to mean, the smaller the value, the more stable the output) during the evening peak period, the duration of high load intervals, etc., are used to characterize the smoothness and dispatchability of the equipment's output during peak periods;

[0100] 4) Regularity characteristics: such as "daily regularity average", which measures the similarity between the daily curves within the main cluster and the typical curves of the main cluster, and quantifies the consistency and predictability of the equipment's cross-day operation mode.

[0101] Finally, the multi-dimensional profile features of all devices are summarized into a unified feature table, which serves as the input for the second stage.

[0102] The second stage involves profile clustering and potential classification based on hierarchical hybrid weighting and weighted K-Means, such as... Figure 2 As shown, this step performs automated and scientific horizontal comparison and potential classification of all device profiles. The core idea is to explicitly elevate the status of peak-shaving features in the evening rush hour through "hierarchical hybrid weighting," run K-Means clustering in the weighted feature space, and use DBI and SI to analyze the cluster numbers. Objective selection is performed, and finally, a peak comprehensive score is constructed to rank and name each cluster as "high / medium / low potential".

[0103] First, to eliminate the dimensional differences between energy-related characteristics and "fluctuation coefficients," the system reads the characteristic summary table generated in step three. The seven selected core features (including peak available energy, peak average power, peak duration, peak CV, scale power, scale energy, and operational regularity) are Z-score standardized to obtain a standardized feature matrix. These features are then divided into two business levels:

[0104] Peak shaving group: including peak available energy, peak average power, peak duration, peak power CV, etc., is given a high inter-group weight (set to 0.7) to reflect its core position in evening peak shaving business;

[0105] Basic scale group: including scale power, scale energy, and daily regularity, etc., is assigned a low inter-group weight (set to 0.3) to supplement the characterization of the basic scale and long-term operating characteristics of the equipment.

[0106] Within each feature group, this embodiment uses the entropy weight method to automatically calculate the objective weights of each feature within the group: by normalizing the data within the group and calculating the information entropy, features with greater variability receive greater weights. The entropy weights within the group are normalized and then multiplied by the inter-group business weights to obtain the hierarchical mixed weights for each feature. In the K-Means clustering implementation, by multiplying each dimension of the standardized feature matrix by the square root of the corresponding weight, these weights are "embedded" into the Euclidean distance, which is equivalent to performing K-Means clustering in the weighted feature space.

[0107] To avoid cluster numbers Relying entirely on human experience, this embodiment operates within a given integer range. Internal differences Test one by one, here the integer range is taken K-Means is run in the weighted feature space to calculate the corresponding Davies-Bouldin index (DBI, smaller is better) and Silhouette index (SI, larger is better). Then, the DBI is sorted in ascending order and the SI in descending order, and the two rankings are summed to obtain a comprehensive ranking index. The index with the best comprehensive performance is selected based on the principle of "minimum sum of rankings". As a recommended clustering number, it realizes adaptive selection of the clustering number based on objective indicators.

[0108] To obtain the optimal After clustering results, this embodiment no longer simply sorts by "peak energy," but instead constructs a peak comprehensive score for the center of each cluster: In the original feature space, the standardized values ​​of each cluster center in four dimensions—peak available energy, peak average power, peak duration, and peak CV—are calculated, and a comprehensive score is constructed in a linear combination form. High energy, high power, and long duration are positively rewarded, while high volatility is negatively penalized. The comprehensive score coefficient is set to meet the guiding requirement of "peak available energy ≥ peak average power ≥ duration ≈ stability," for example, values ​​of (0.4, 0.3, 0.2, -0.1). Clusters are sorted from highest to lowest comprehensive score and named "high potential," "medium potential," and "low potential" (if there are many clusters, they can be further subdivided into "medium potential-1," "medium potential-2," etc.). The cluster label and potential level of each device are written into the final profile list, achieving unified potential classification of massive energy storage assets.

[0109] The specific calculation process is as follows:

[0110] Z-score normalization is applied to the fused profile matrix of all devices to eliminate dimensional differences and ensure balanced contributions across dimensions. Let the... The population mean and standard deviation of dimension are ,equipment The standardization characteristics are:

[0111] ;

[0112] in For the first The equipment in the first The values ​​of each feature These represent the mean and standard deviation of this feature across all devices. To prevent zero terms, a standardized matrix is ​​obtained. .

[0113] Considering the core role of peak-shaving capacity in business operations, this embodiment further divides the profile features into two categories: "Peak Shaving Group" and "Basic Scale Group." The former includes peak-available energy, peak average power, peak high load duration, and peak power fluctuation coefficient, among other peak-shaving related features. The latter includes scale power, scale energy, and operational regularity, reflecting the overall capacity and stability of the equipment. Let PEAK group contain a set of feature indexes. The SCALE group contains ,and , .

[0114] Based on this, a hybrid weighting mechanism of "business layering + entropy weighting" is constructed: first, the entropy weighting method is used to characterize the dispersion of indicators within each group, and then business preferences are reflected through inter-group weights. Specifically, for any group... (Can be PEAK or SCALE), record the number of the group. The non-negative normalized value of on all devices is Then the sample In terms of indicators The proportion of the above is:

[0115] ;

[0116] in, Represents the total number of energy storage devices, defined as a constant. The first in the group The entropy value of each indicator is:

[0117] ;

[0118] The coefficient of difference is:

[0119] ;

[0120] The entropy weights are obtained by normalizing the indices within the group:

[0121] ;

[0122] in, This reflects the "information content" of the feature within the group: the greater the dispersion (the more obvious the difference), the higher the weight. This represents the characteristic difference coefficient.

[0123] To reflect the business principle of "peak shaving taking precedence over base scale," this embodiment sets a total weight for peak groups. Total weight of the scale group Between groups satisfy The final mixed weight vector Defined as:

[0124] ;

[0125] That is, within a group, the entropy weight method objectively characterizes "which type of peak feature is more discriminative," and between groups, it uses... Explicitly enhance the overall contribution of peak-shaving related features.

[0126] In terms of distance metric, this embodiment injects the hybrid weights into the normalized feature space in the form of "weighted Euclidean distance". For any device Standardization characteristics Perform a linear transformation:

[0127] ;

[0128] Where M represents the total number of profile features participating in clustering; then any two devices The distance between them is:

[0129] ;

[0130] In other words, in K-Means, a weighted Euclidean distance naturally forms in which "peak group features contribute more and basic scale features contribute less".

[0131] Based on DBI and SI, this method automatically optimizes the number of clusters. To address the subjectivity issue in cluster number selection, this method considers the candidate cluster number set... The above adopts a multi-index cross-evaluation. In this embodiment, it is based on actual needs ( Take 2, Take 8): For each given In the aforementioned weighted feature space K-Means clustering is performed, and the corresponding DBI index and SI coefficient are calculated. Given that a smaller DBI index and a larger SI index may lead to different recommendations regarding the optimal number of clusters, this embodiment employs a comprehensive ranking method for robust decision-making: Rank in descending order ,right Rank in ascending order ,definition:

[0132] ;

[0133] Take smallest As the optimal number of clusters, and retaining the evaluation curve for future review. To clarify the definitions, standard definitions of DBI and SI are given (using weighted Euclidean distance as an example). Let the th... The sample set of each cluster is Its center of mass is (exist (Spatial calculation), then the intra-class divergence (the average distance from the sample within a cluster to the centroid) and the inter-cluster center distance are respectively ,thus:

[0134] ;

[0135] The smaller the value, the better the intra-cluster compactness and inter-cluster separation of K-Means clustering, and the more reliable the clustering results. For any sample (Its cluster number is) ), define the average distance within a cluster and the minimum average distance to other clusters:

[0136] ;

[0137] in, This represents the average intra-cluster distance of the i-th device. Let i represent the sample set of the cluster to which the i-th device belongs. This represents the minimum average inter-cluster distance for the i-th device. Indicates the first A sample set of each cluster;

[0138] The sample contour value and the global SI are then:

[0139] ;

[0140] The closer the number is to 1, the better the intra-cluster aggregation and inter-cluster separation of all energy storage device samples in K-Means clustering. This results in a clearer potential classification for each device and stronger guidance for energy storage grading. Determining the optimal number of clusters... Then, the weighted standardized profile matrix Perform K-Means analysis to find the solution:

[0141] .

[0142] On the other hand, this embodiment provides a system for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering, applicable to methods for generating peak-shaving potential energy storage profiles based on multi-dimensional clustering, including:

[0143] The module includes: data acquisition and matrix construction, weight calculation and data weighting, density clustering and typical curve generation, portrait feature extraction, feature grouping and mixed weighting, weighted feature space construction, and cluster scoring and potential classification.

[0144] The data acquisition and matrix construction module is used to acquire the power measurement data of the device, divide the data according to the set time interval, and construct a time point curve matrix.

[0145] The weight calculation and data weighting module is used to calculate the standard deviation of each time point in the time point curve matrix, map the standard deviation vector into a time weight vector, and multiply each column of the time point curve matrix by the corresponding time weight to achieve feature weighting.

[0146] The density clustering and typical curve generation module is used to cluster the weighted time point curve matrix using a density clustering algorithm, remove noisy time data, select the main cluster with the largest number of samples, and calculate the typical curve of the main cluster.

[0147] The portrait feature extraction module is used to extract portrait feature vectors based on the typical curves of the main cluster.

[0148] The feature grouping and hybrid weighting module is used to divide the portrait feature vector into peak-shaving group and basic scale group, calculate the objective weight of the portrait features within each group using the entropy weight method, and obtain the hierarchical hybrid weight by combining the business weights between groups.

[0149] The weighted feature space construction module is used to integrate hierarchical mixed weights into the profile features to construct a weighted feature space;

[0150] The clustering scoring and potential grading module is used to score the potential of each cluster profile in the weighted feature space through K-Means clustering, and output the peak-shaving potential level of the energy storage profile.

[0151] In engineering practice, this embodiment also conducted a simple sensitivity analysis on the inter-group weights: the weights of the peak group were adjusted within the range of [0.6, 0.8] and the weights of the scale group were adjusted within the range of [0.4, 0.2] for testing. The results showed that the "high potential" set remained basically stable under different weight settings, with only a few boundary devices migrating, thus verifying the robustness of this method to weight settings.

[0152] In summary, this embodiment also achieves the following beneficial effects: The data processing flow of this embodiment is rigorous, adaptable to complex engineering scenarios, and has reproducible parameters. This embodiment features a rigorous design in its data preprocessing and modeling processes, fully considering multiple data sources and complex working conditions on the engineering side: It supports automatic splitting of multiple assets within the same file through intelligent header recognition and asset number parsing; it segments by natural day and aligns to a standard 96-point time grid, interpolating only within the current day's range to avoid cross-day extrapolation; in the main operating mode recognition stage, it employs "global Z-score + variance automatic weighting OPTICS" processing, taking into account both amplitude and morphological differences; in the portrait clustering stage, it uses StandardScaler standardization and applies hybrid feature weights on this basis, effectively alleviating the problem of inconsistent feature dimensions and scales. The core parameters and criteria in the method (such as variance weight generation rules, minimum sample ratio of OPTICS, peak segment window settings, stratification weights of 0.7 / 0.3, DBI / SI evaluation intervals, and peak comprehensive score coefficients) are all managed uniformly through configuration files, ensuring traceability and reproducibility when applied across regions, seasons, and projects, and making it suitable for large-scale automated deployment and long-term operation and maintenance in actual engineering projects.

[0153] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0154] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0155] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for generating peak-shaving potential energy storage profiles based on multi-dimensional clustering, characterized by: Includes the following steps: Step 1: Obtain the power measurement data of the device, divide the power measurement data into sets of time intervals, and construct a time point curve matrix; Step 2: Calculate the standard deviation of each time point based on the time point curve matrix, map the standard deviation vector to the time weight vector, multiply each column of the time point curve matrix by the corresponding time weight, weight the features of the time point curve matrix, perform clustering operation on the weighted time point curve matrix through density clustering algorithm to remove noise time, select the main cluster with the most samples and calculate the typical curve of the main cluster. Step 3: Extract the feature vectors that constitute the portrait based on the typical curves of the main cluster; Step 4: Divide the profile features into peak shaving group and basic scale group. Use the entropy weight method to calculate the objective weight of the profile features in each group. After integrating the objective weight into the profile features, a weighted feature space is formed. In the weighted feature space, K-Means clustering is used to score the profile potential in each cluster to obtain the peak shaving potential of the energy storage profile. The objective weights of the portrait features in each group are calculated using the entropy weight method. These objective weights are then incorporated into the portrait features to form a weighted feature space, specifically: First, the data within the peak shaving group and the basic scale group are normalized and the information entropy is calculated. The profile feature with greater variability receives a higher weight. The information entropy weight is multiplied by the inter-group business weight to obtain the hierarchical mixed weight of each profile feature. In K-Means clustering, the weighted feature space is obtained by multiplying each dimension of the standardized feature matrix by the square root of the corresponding hierarchical mixed weight.

2. The method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering according to claim 1, characterized in that, In step 1, the power measurement data is also denoised. If there is no valid power measurement data at a certain time point, the power measurement data at that time point is deleted from the time point curve matrix.

3. The method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering according to claim 1, characterized in that, In step 2, the method for setting the time weight is as follows: the time point with a larger standard deviation has a higher weight, and the time point with a smaller standard deviation has a lower weight.

4. The method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering according to claim 1, characterized in that, In step 2, the stability of the equipment operation mode and the reliability of the main cluster extraction are judged by statistically analyzing the main cluster ratio, noise ratio, and contour coefficient of the density clustering algorithm in the time point curve matrix. If the stability of the equipment operation mode is lower than the stability threshold or the reliability of the main cluster extraction is lower than the reliability threshold, the parameters of the density clustering algorithm are adjusted for re-clustering until the stability of the equipment operation mode and the reliability of the main cluster extraction are above the set threshold.

5. The method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering according to claim 1, characterized in that, In step 3, the profile feature vector includes basic scale features, peak shaving potential features, stability features, and regularity features. The profile feature vectors of all devices are summarized into a unified feature table for data input in step 4.

6. The method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering according to claim 5, characterized in that, Core features are extracted from the profile feature vector. These core features include peak available energy, peak average power, peak duration, peak CV, scale power, scale energy, and operational regularity. The core features are then Z-score standardized to obtain a standardized feature matrix. In step 4, the profile features are divided into a peak shaving group and a basic scale group. The core features of the peak shaving group include peak available energy, peak average power, peak duration, and peak CV. The core features of the basic scale group include scale power, scale energy, and operational regularity.

7. The method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering according to claim 1, characterized in that, Furthermore, the number of clusters in K-Means clustering was optimized, specifically as follows: K-Means clustering is performed in the weighted feature space to calculate the corresponding DBI index and SI index. Then, the DBI index is sorted in ascending order and the SI index is sorted in descending order. The two rankings are added together to obtain a comprehensive index. The optimal number of clusters is selected as the optimal number of clusters based on the comprehensive index with the smallest sum of rankings.

8. The method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering according to claim 1, characterized in that, In step 4, the potential of the energy storage profile is scored in each cluster using K-Means clustering in the weighted feature space to obtain the peak-shaving potential of the energy storage profile. Specifically: For each cluster center, a peak comprehensive score is constructed. The standardized values ​​of each cluster center in four dimensions—peak available energy, peak average power, peak duration, and peak CV—are calculated. The comprehensive score is constructed in a linear combination form, positively rewarding high energy, high power, and long duration, and negatively penalizing high volatility. Each cluster is sorted from largest to smallest based on the comprehensive score, and the cluster label and potential level of each device are written into the energy storage profile list to realize the peak shaving potential classification of the device's energy storage profile.

9. A system for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering, applicable to the method for generating a peak-shaving potential energy storage profile based on multi-dimensional clustering as described in any one of claims 1-8, characterized in that, include: The module includes: data acquisition and matrix construction, weight calculation and data weighting, density clustering and typical curve generation, portrait feature extraction, feature grouping and mixed weighting, weighted feature space construction, and cluster scoring and potential classification. The data acquisition and matrix construction module is used to acquire the power measurement data of the device, divide the data according to the set time interval, and construct a time point curve matrix. The weight calculation and data weighting module is used to calculate the standard deviation of each time point in the time point curve matrix, map the standard deviation vector into a time weight vector, and multiply each column of the time point curve matrix by the corresponding time weight to achieve feature weighting. The density clustering and typical curve generation module is used to cluster the weighted time point curve matrix using a density clustering algorithm, remove noisy time data, select the main cluster with the largest number of samples, and calculate the typical curve of the main cluster. The portrait feature extraction module is used to extract portrait feature vectors based on the typical curves of the main cluster. The feature grouping and hybrid weighting module is used to divide the portrait feature vector into peak-shaving group and basic scale group, calculate the objective weight of the portrait features within each group using the entropy weight method, and obtain the hierarchical hybrid weight by combining the business weights between groups. The weighted feature space construction module is used to integrate hierarchical mixed weights into the profile features to construct a weighted feature space, specifically as follows: First, the data within the peak shaving group and the basic scale group are normalized and the information entropy is calculated. The profile feature with greater variability receives a higher weight. The information entropy weight is multiplied by the inter-group business weight to obtain the hierarchical mixed weight of each profile feature. In K-Means clustering, the weighted feature space is obtained by multiplying each dimension of the standardized feature matrix by the square root of the corresponding hierarchical mixed weight. The clustering scoring and potential grading module is used to score the potential of each cluster profile in the weighted feature space through K-Means clustering, and output the peak-shaving potential level of the energy storage profile.

Images