A combined supervised and unsupervised method for extracting typical scenarios of complex power loads in power systems

Abstract

This invention relates to a method for extracting typical scenarios of complex power loads in a power system using a combination of supervised and unsupervised learning. The method includes the following steps: acquiring reactive power load data of power system nodes; preprocessing the reactive power load data; calculating the weighted similarity between each complex power load using a multi-scale similarity index; optimizing and pre-classifying the number of classes of complex power loads across all scenarios using unsupervised learning based on the weighted similarity between the complex power loads, with the classification results serving as labels; smoothing the classification boundaries of the complex power loads using supervised learning based on the labels, obtaining the final classification results; and extracting typical operating scenarios based on the final classification results. Compared with existing technologies, this invention has advantages such as high classification efficiency and the ability to achieve more accurate classification.

Description

Technical Field

[0001] This invention belongs to the field of power system optimization operation technology, and relates to a load characteristic extraction method, particularly a method for extracting typical scenarios of complex power loads in power systems using a combination of supervised and unsupervised methods. Background Technology

[0002] Currently, with the large-scale development of power systems, the loads on these systems have become more complex and variable. The randomness and diversity of loads pose significant challenges to the safety and reliability of power systems. Optimizing each of the massive load scenarios individually is inefficient and computationally burdensome. Extracting typical scenarios can effectively reduce the number of scenarios and improve the efficiency of power system optimization under the influence of random factors.

[0003] In recent years, both domestic and international researchers have accumulated certain research results in the field of typical scenario extraction. The literature "A method for generating multi-dimensional typical scenarios of distribution networks based on load clustering and network equivalence" (Li Bo, Sun Jianjun, Yu Pan, Zha Xiaoming, Wang Chaoliang, Xu Feng. Proceedings of the CSEE, 2021, 41(08):2661-2671.) is based on the vector K-medoids clustering method. It divides the load into n categories according to the power fluctuation characteristics and divides distributed generation into two categories: photovoltaic and wind power. It performs n+1-dimensional and 1-dimensional K-means clustering on load / photovoltaic and wind power respectively, and calculates the initial network node typical scenarios based on the cluster centers. The literature "Distributed Clustering Algorithm for Sensing Electricity Consumption Characteristics of Massive Users" (Zhu Wenjun, Wang Yi, Luo Min, Lin Guoying, Cheng Jiangnan, Kang Chongqing. Automation of Electric Power Systems, 2016, 40(12):21-27.) targets massive and dispersed electricity consumption data. It uses an adaptive k-means clustering algorithm to perform local clustering analysis on electricity consumption data distributed in various regions, and then uses a traditional clustering algorithm to perform secondary clustering analysis on the obtained local model to obtain the typical load curves of the whole. Chinese patent application CN118589595A discloses a method and system for extracting typical scenarios of power systems, which combines the simulation of AC power flow unit combination to extract phase scenarios and evaluate indicators from the simulation results. Chinese patent CN112467803B discloses a method and system for generating typical scenarios of distributed power sources and new loads. It performs clustering processing on the optimized distributed power source output and new load demand to obtain a dataset, and then classifies the scenarios by maximizing the change in scenario information entropy based on the distributed power source output and new load cluster dataset to generate a variety of typical scenarios. Chinese patent application CN117060398A discloses a method and system for extracting and modeling typical key scenarios of power grids under massive load data. Based on clustering algorithms and global search algorithms, it calculates typical key scenarios of load and power flow data.

[0004] However, existing methods still have the following inherent drawbacks:

[0005] 1. Similarity measurement relies on traditional Euclidean distance to evaluate the similarity of curves. It only considers the distance between points and ignores the correlation between different dimensions of samples, making it difficult to distinguish the shape differences of curves and unable to analyze the temporal nature and active and reactive coupling of load scenarios.

[0006] 2. Unsupervised classification is simple in principle, relying on a specific distance as a similarity measure, making it difficult to describe complex samples. Supervised learning, on the other hand, is complex in principle and its classification depends on labeled data, leading to overfitting. Existing classification methods often combine unsupervised learning with a small number of samples for classification, and then use supervised learning to feed a large number of samples into a fixed classification, failing to combine the advantages of both types of learning.

[0007] 3. Currently, the evaluation of classification effectiveness mostly focuses on the overall scenario, without highlighting extreme scenarios that are important for power system analysis. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for extracting typical scenarios of complex power loads in power systems that combines supervised and unsupervised methods, which has high classification efficiency and can achieve more accurate classification.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] A method for extracting typical scenarios of complex power loads in power systems using a combination of supervised and unsupervised methods includes the following steps:

[0011] Acquire reactive load data of power system nodes, preprocess the reactive load data to convert it into complex power load samples, and use multi-scale similarity index to calculate the weighted similarity between each complex power load;

[0012] Based on the weighted similarity between each complex power load, unsupervised learning is used to optimize the number of classes and pre-classify complex power loads in the whole scene, and the classification results are used as labels;

[0013] Based on the labels, supervised learning is used to smooth the classification boundary of complex power loads and obtain the final classification result.

[0014] Typical operational scenarios are extracted based on the final classification results.

[0015] Furthermore, the preprocessing includes vectorization of complex power loads over continuous time periods and normalization of complex power loads across the entire scenario.

[0016] Furthermore, the multi-scale similarity index is a similarity metric that takes into account both amplitude and power factor.

[0017] Furthermore, the multi-scale similarity index is a weighted fusion index of Euclidean distance and cosine similarity distance.

[0018] Furthermore, the class number optimization and pre-classification specifically include:

[0019] The complex power load is classified multiple times, with a different number of categories selected for each classification. The group with the best classification effect is selected as the optimal number of categories based on the clustering reward index.

[0020] Based on the optimal number of classifications and the weighted similarity between each complex power load, unsupervised learning is used to pre-classify complex power loads in the entire scenario.

[0021] Furthermore, the smooth correction of the complex power load classification boundary using supervised learning specifically includes:

[0022] The labels generated by unsupervised classification are used to train a kernel trick-based support vector machine model to formally classify the complex power load in the entire scene and obtain the final classification result.

[0023] Furthermore, the extraction of typical operating scenarios includes:

[0024] Typical load segments are extracted based on the final classification results to form typical operating scenarios;

[0025] Calculate the power factor range and / or probability for each typical operating scenario.

[0026] Furthermore, before extracting typical operating scenarios, the final classification results are evaluated in multiple dimensions. The evaluation indicators include inter-group difference evaluation indicators, intra-group similarity evaluation indicators, and edge scenario classification evaluation indicators.

[0027] Furthermore, the inter-group difference assessment index is characterized by the silhouette coefficient, the intra-group similarity assessment index is characterized by the sum of squared errors combining distance and shape, and the expression for the edge scene classification assessment index is:

[0028]

[0029] Among them, C b Let m represent the set of all boundary scenes. i It is the center of the i-th type of load, D(x,m) i ) represents a multi-scale similarity measure from the boundary scene to the classification center, n b Indicates the number of boundary scenes.

[0030] The present invention also provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the method described above for extracting typical scenarios of complex power loads in a power system with and without supervision.

[0031] Compared with existing technologies, this invention can accurately extract typical load scenarios with active and reactive power coupling, and has the following beneficial effects:

[0032] 1. This invention employs a two-stage classification method for the active and reactive power curves of complex power loads: unsupervised learning pre-classification followed by supervised learning correction classification. This method combines the low computational complexity and high classification efficiency of unsupervised algorithms with the ability of supervised algorithms to describe the complex features of samples, thereby improving the accuracy of load classification and typical scenario extraction.

[0033] 2. This invention uses a multi-scale similarity index to measure the similarity of complex power loads, taking into account both amplitude and power factor. It can balance the proximity between load curves and the similarity of load curve shapes, effectively reflecting the temporal nature of load curves and the coupling between active and reactive power, thereby improving the accuracy of classification and typical scenario extraction.

[0034] 3. This invention designs classification effect evaluation rules for all operating scenarios, which can cover extreme scenarios that have a significant impact on the system and can comprehensively evaluate the classification effect. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the process of the present invention;

[0036] Figure 2 This is a schematic diagram illustrating the specific steps of the present invention. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0038] like Figure 1 and Figure 2 As shown, this embodiment provides a method for extracting typical scenarios of complex power loads in a power system using a combination of supervised and unsupervised methods, including the following steps:

[0039] Step S100: Similarity measurement of complex power loads that takes into account both amplitude and power factor.

[0040] In this step, reactive power load data of power system nodes is acquired, and the reactive power load data is preprocessed, including vectorization of complex power loads over continuous time periods and normalization of complex power loads across the entire scenario. A multi-scale similarity index is used to calculate the weighted similarity between each complex power load. This multi-scale similarity index is a similarity metric that considers both amplitude and power factor. In this embodiment, a cosine similarity metric is added to the traditional Euclidean distance; that is, the multi-scale similarity index uses a weighted fusion of Euclidean distance and cosine similarity distance.

[0041] This step specifically includes:

[0042] 1.1 Vectorization of Complex Power Loads in Continuous Time Periods

[0043] Data preprocessing requires transforming a scenario into a directly processable sample, while considering the active and reactive loads of a specific node. The collected data often has different time resolutions, determined by the monitoring equipment. Therefore, the collected data needs to be adjusted to an hourly resolution using averaging and interpolation methods. Then, the hourly active and reactive load data are used as the first and last 24 dimensions, resulting in 48 dimensions of data.

[0044] 1.2 Normalization of Complex Power Loads in All Scenarios

[0045] In this embodiment, normalization is performed. Given that the values ​​of active and reactive loads at a node typically differ, they need to be normalized separately to ensure that the numerical proportions of each dimension fall within the range [0,1]. After this process, corresponding weights can be assigned to different dimensions. Specifically, this involves normalizing the active and reactive loads separately:

[0046] Active load normalization p t ′=(p t -p min ) / (p max -p min ),

[0047] Reactive load normalization q t ′=(q t -q min ) / (q max -q min ).

[0048] Where, p t q t p represents the active and reactive power of the complex power load at time t. max ,p min ,q max ,q minThese represent the maximum and minimum active power and reactive power values ​​at all times for all complex power loads.

[0049] 1.3 Comprehensive Distance Calculation for Complex Power Loads in All Scenarios

[0050] In this embodiment, a multi-scale similarity index is used to evaluate the similarity between complex power loads. Euclidean distance is typically used to measure the similarity between load curves. This metric primarily focuses on the distance between points, i.e., their spatial proximity, but ignores the similarity of curve shapes. However, in power systems, load curves are time-series, and there is a correlation between active and reactive power. Therefore, a cosine similarity index is used to measure the similarity of curve shapes.

[0051] This embodiment defines a multi-scale similarity measure d. ij It uses weighted Euclidean distance cosine similarity distance Given that the former measures point-to-point distance and the latter measures shape similarity, and k represents the dimensions of the complex power load, then:

[0052] d ij =ε×a ij +(1-ε)×b ij

[0053] Among them, a ij Denotes Euclidean distance, b ij ε represents the cosine similarity distance, and ε is the weighting coefficient of the multi-scale similarity measure.

[0054] Step S200: Class number optimization and pre-classification of full-scenario complex power load based on unsupervised learning.

[0055] In this step, based on the weighted similarity between each complex power load, unsupervised learning is used to optimize the number of classes and pre-classify the complex power loads in the whole scene, and the classification results are used as labels.

[0056] This embodiment uses the unsupervised k-means clustering algorithm to divide a portion of the complex power load data and train labels for subsequent supervised classification.

[0057] This step includes:

[0058] 2.1 Class Optimization of Complex Power Loads in All Scenarios

[0059] In this embodiment, the inflection point method based on the clustering effectiveness index is used to determine the optimal number of clusters. Specifically, the largest possible number of clusters k is first selected. max Then select [2,k] max Each integer within the range is used as the cluster number, and k is performed.max Perform clustering at level -1 and calculate the clustering reward index. The clustering result is optimal when the clustering reward index is maximized, and the number of clusters in this case is the optimal number of clusters.

[0060] Clustering performance metrics need to comprehensively consider both inter-cluster separation and intra-cluster cohesion to describe the differences in complex power loads across different categories and the similarity within the same category. To quantify intra-cluster cohesion, the sum of squares due to error (SSE) is generally used, defined as: Where, m i C is the center of the i-th type of complex power load. i Let k be the current cluster number, representing the i-th cluster of complex power loads. SSE (Solution-Oriented Sequence) calculates the Euclidean distance from all complex power loads to their respective cluster centers. SSE is a reward function that decreases with increasing cluster number; the decrease is significant when k is less than the optimal cluster number, but slows down considerably when k reaches the optimal cluster number. To select the optimal cluster number, an error reduction coefficient β needs to be defined. SSE To describe the change in the magnitude of the SSE decline:

[0061] On the other hand, to describe the inter-cluster separation after clustering, the silhouette coefficient (SC) needs to be introduced. For a given sample, its silhouette coefficient is defined as: Where, b(x) i ) and sample x i The average Euclidean distance to samples from other clusters, a(x i ) is the sample x i The average Euclidean distance from other samples in the cluster. The closer this value is to 1, the better the clustering effect. The average silhouette coefficient of all samples is used as the clustering reward metric. n is the number of samples.

[0062] Combining the above two indicators, namely α SC +β SSE To select the optimal number of clusters.

[0063] 2.2 Complex Power Load Preclassification Based on Unsupervised Learning

[0064] In this embodiment, after determining the optimal number of classifications, the k-means algorithm is used for pre-classification, including:

[0065] Randomly initialize K data points as initial cluster centers (centroids);

[0066] Distance calculation: For each data point, calculate its distance to all cluster centers. Use the weighted similarity obtained in step S100 as the distance metric.

[0067] Assign to the nearest cluster: Assign each data point to the cluster corresponding to the nearest cluster center.

[0068] Recalculate cluster centers: For each cluster, calculate the mean of all data points within the cluster and use this mean as the new cluster center.

[0069] Repeat the above steps iteratively until the stopping condition is met, that is, the change in the cluster center is less than a certain threshold or the preset number of iterations is reached.

[0070] Step S300: Smooth correction of the complex power load classification boundary based on supervised learning.

[0071] In this step, based on the labels, supervised learning is used to smooth the classification boundary of complex power loads and obtain the final classification result.

[0072] In this embodiment, the pre-classification result of the load curve by k-means is regarded as the label of supervised clustering, and the final classification is implemented by the supervised learning algorithm of support vector machine (SVM).

[0073] This step includes:

[0074] 3.1 Support Vector Machine Binary Classification Model Based on Kernel Tricks

[0075] In practical complex power load classification problems where the loads are linearly inseparable, kernel tricks are generally used to map the samples to a high-dimensional space, where the classification is then performed. In this implementation example, a radial basis function kernel is used to achieve the low-dimensional to high-dimensional mapping: k(x,x′)=exp(-γ||xx′|| 2 ); Select parameter γ, which controls the range of influence of a single training sample.

[0076] Next, the SVM classification algorithm seeks a hyperplane to divide the mapped samples into two classes, maximizing the margin of the division while maximizing the classification accuracy. The optimization problem is:

[0077]

[0078] sty i (w·x i +b)≥1-ξ i

[0079] The objective function's first term minimizes the reciprocal of the classification margin, while the second term is the penalty function for misclassification. Choosing an appropriate parameter C controls the severity of the penalty for misclassification; a larger C value results in a greater penalty for misclassification. The optimization problem is solved using the interior-point method.

[0080] 3.2 Multi-classification model of complex power load based on combined strategy

[0081] In practice, the classification of complex power loads is often a multi-class problem. SVM is a binary classification algorithm, while the classification of system operation modes based on complex power loads is a multi-class problem. Therefore, it is necessary to use appropriate strategies to combine multiple binary classification problems in a reasonable way.

[0082] Currently, there are two combined strategies: a one-to-many strategy and a one-to-one strategy. Assuming the optimal number of classes is k, the one-to-many strategy requires building k binary SVMs. The i-th SVM divides the load into the i-th class and the remaining (k-1) classes, directly determining the class of the complex power load. This strategy is relatively simple but prone to contradictions, where the same complex power load is assigned to different classes by multiple binary classifiers. The one-to-one strategy, on the other hand, builds k(k-1) / 2 binary classifiers, comparing each class pairwise to determine which class the complex power load is more likely to belong to, and then using a voting process to determine the class of the complex power load. Therefore, this technique adopts the one-to-one strategy.

[0083] 3.3 Boundary Smoothing Correction for Formal Classification of Complex Power Loads

[0084] n complex power loads {x1, x2, ... x} n The labels {y1, y2, ... y} are pre-classified into k groups. n The final classification is carried out through the following steps:

[0085] (1) Initialization: Construct k binary SVM classifiers, that is, generate k hyperplanes and their corresponding parameters {w,b}.

[0086] (2) Network training: using complex power load {x1,x2,…x} n} and the corresponding labels {y1, y2, ... y n Solve the optimization problem to obtain the parameters {w,b}.

[0087] (3) Grouping: All complex power loads are put into the SVM classifier and the group corresponding to each complex power load is calculated.

[0088] Step S400: Multidimensional evaluation of the classification effect of complex power loads. Specifically, the evaluation indicators include inter-group difference evaluation indicators, intra-group similarity evaluation indicators, and edge scene classification evaluation indicators.

[0089] In this implementation example, the sum of squared errors (SSE) and silhouette coefficient (SC) are selected as effectiveness indicators to achieve the dual goal of highlighting the similarity within groups and the differences between groups.

[0090] The evaluation rules for each evaluation indicator include:

[0091] 4.1 Rules for assessing inter-group differences in complex power loads

[0092] The differences between groups are characterized by SC. SC can be applied directly, where the Euclidean distance is replaced by a multi-scale similarity measure d. ij , Where b′(x i ) is the sample x i The average multi-scale similarity to other cluster samples, a′(x i ) is the sample x i The average multi-scale similarity of other samples in this cluster.

[0093] 4.2 Intra-group similarity assessment rules for complex power loads

[0094] Intra-group similarity is characterized by SSE, which also needs to be improved from using Euclidean distance to a multi-scale similarity index that can describe the shape of the curve. The improved SSE consists of two parts. d SSE measures proximity. s Measuring shape similarity:

[0095]

[0096] Where, m i It is the center of the i-th type of complex power load, and ε is the weighting coefficient of the multi-scale similarity measure.

[0097] 4.3 Classification and Evaluation Rules for Extreme Scenarios of Complex Power Loads

[0098] In this implementation example, traditional evaluation metrics such as SSE and SC primarily focus on overall classification performance; more specifically, these metrics are calculated based on the average of all scenarios. However, in this implementation case, the goal is to reclassify scenarios on the classification edge. Although these scenarios represent a small proportion of the overall scenario set, they are critical and complex situations within the power system. Therefore, a new metric—Boundary Scenario SSE (bSSE)—is defined to specifically emphasize the clustering effect of these edge scenarios: Among them, C b Let m represent the set of all boundary scenes. i It is the center of the i-th type of scene, D(x,m) i ) represents a multi-scale similarity measure from the boundary scene to the classification center, n bThis represents the number of boundary scenarios (x∈C). b This accounts for 5% of all scenes.

[0099] Step S500: Extract and generate typical operating scenarios of actual systems using complex power load classification.

[0100] This step includes: extracting typical load segments based on the final classification results to form typical operating scenarios; and calculating the power factor range and / or probability for each type of typical operating scenario.

[0101] Specifically, in this embodiment, a combined supervised and unsupervised classification algorithm is applied to the extraction and characteristic analysis of typical scenarios of complex power loads, as follows:

[0102] 5.1 Extraction of Typical Load Segments Based on Complex Power Load Classification

[0103] Clustering techniques divide the load curves of all scenarios into several groups, each representing a typical operating scenario. Each typical scenario can be represented by a typical curve, which is the center point of that curve group. N i This represents the number of scenes in this curve group.

[0104] These typical curves illustrate the typical electricity consumption patterns of nodes over a period of time. For power systems, typical load curves not only provide information on the timing and magnitude of user electricity consumption, but also the coupling characteristics of active and reactive power, which are crucial for system operation and planning.

[0105] 5.2 Feature Analysis and Probabilistic Description of Typical Load Operation Scenarios

[0106] A wealth of load information can be extracted from typical load scenarios. A classification algorithm combining supervised and unsupervised learning is used to divide the massive load data into k classes, each with a different peak active power P in typical load scenarios. i,max and peak reactive power Q i,max Furthermore, the power factor (PF) can be calculated for typical load scenarios at different times. i,t Calculate the power factor range (PF) for typical scenarios of each type of load. i .

[0107] The probability of each typical scenario can also be approximated by frequency. Assume there are N historical complex power load scenarios, which are divided into K groups using a combination of supervised and unsupervised classification algorithms. Assume that each typical load in the k-th group has N... i Given several scenarios, the probability of each typical load scenario is:

[0108] The above method addresses two major shortcomings of existing load analysis: it often neglects reactive power and the classification methods used do not leverage the respective advantages of supervised and unsupervised machine learning. It proposes a two-stage classification algorithm for complex power loads, which includes unsupervised pre-classification and supervised smoothing of class boundaries, thus providing core technical support for the extraction of power system operating characteristics and typical scenarios.

[0109] If the above methods are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0111] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0112] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for extracting typical scenarios of complex power loads in a power system using a combination of supervised and unsupervised methods, characterized in that, Includes the following steps: Acquire reactive load data of power system nodes, preprocess the reactive load data to convert it into complex power load samples, and use multi-scale similarity index to calculate the weighted similarity between each complex power load; Based on the weighted similarity between each complex power load, unsupervised learning is used to optimize the number of classes and pre-classify complex power loads in the whole scene, and the classification results are used as labels; Based on the labels, supervised learning is used to smooth the classification boundary of complex power loads and obtain the final classification result. Typical operational scenarios are extracted based on the final classification results; The multi-scale similarity index is a similarity metric that takes into account both the magnitude and power factor of the complex power load. The multi-scale similarity index is a weighted fusion index of Euclidean distance and cosine similarity distance; The class number optimization and pre-classification specifically include: The complex power load is classified multiple times, with a different number of categories selected for each classification. The group with the best classification effect is selected as the optimal number of categories based on the clustering reward index. Based on the optimal number of classifications and the weighted similarity between each complex power load, unsupervised learning is used to pre-classify complex power loads in the entire scenario. The method of using supervised learning to smooth the boundary of complex power load classification specifically includes: The labels generated by unsupervised classification are used to train a binary classification model based on kernel trick support vector machine to formally classify the complex power load of the entire scene and obtain the final classification result. In the process of formally classifying the complex power loads in the entire scenario using a binary classification model, each category is compared pairwise to determine which category the complex power load is more inclined to, and then the category of the complex power load is determined by voting.

2. The method for extracting typical scenarios of complex power loads in a power system using a combination of supervised and unsupervised methods according to claim 1, characterized in that, The preprocessing includes vectorization of complex power loads over continuous time periods and normalization of complex power loads across the entire scenario.

3. The method for extracting typical scenarios of complex power loads in a power system using a combination of supervised and unsupervised methods according to claim 1, characterized in that, The typical operating scenarios extracted include: Typical load segments are extracted based on the final classification results to form typical operating scenarios; Calculate the power factor characteristics and probability for each typical operating scenario.

4. The method for extracting typical scenarios of complex power loads in a power system using a combination of supervised and unsupervised methods according to claim 1, characterized in that, Before extracting typical operating scenarios, the final classification results are evaluated in multiple dimensions. The evaluation indicators include inter-group difference evaluation indicators, intra-group similarity evaluation indicators, and edge scenario classification evaluation indicators.

5. The method for extracting typical scenarios of complex power loads in a power system using a combination of supervised and unsupervised methods according to claim 4, characterized in that, The inter-group difference assessment index is characterized by the silhouette coefficient, the intra-group similarity assessment index is characterized by the sum of squared errors combining distance and shape, and the expression for the edge scene classification assessment index is: in, Represents the set consisting of all boundary scenes. It is the center of the i-th type of scene. This represents a multi-scale similarity measure from the boundary scene to the classification center. Indicates the number of boundary scenes.

6. A computer-readable storage medium, characterized in that, Includes one or more programs executed by one or more processors of an electronic device, said one or more programs including instructions for performing the supervised and unsupervised combined method for extracting typical scenarios of complex power loads in a power system as described in any one of claims 1-5.

Images