Method and device for constructing power consumption load feature signature model based on attention

Abstract

The application provides a construction method and device of an electricity load feature signature model based on attention, S1, electricity data of each user is collected according to a first set time interval; S2, the electricity data of each user is processed based on attention to obtain an electricity load feature model of each user; and S3, the electricity load feature models of all users are clustered to obtain a global electricity load feature signature module. The application obtains better performance improvement with a small time cost based on a double-layer feature extraction architecture of an attention mechanism, reduces the operation amount, avoids feature loss in a feature area, and thus guarantees the accuracy of the constructed electricity load feature signature model and the value of subsequent analysis and application.

Description

Technical Field

[0001] This invention relates to the field of power system analysis technology, and in particular to a method and apparatus for constructing an attention-based electrical load characteristic information model. Background Technology

[0002] With the continuous expansion of the power grid and the increasing coverage of smart terminal installations, power grid data is experiencing explosive growth. Extracting valuable information from this data has become a challenging problem in the current power system. Therefore, understanding the electricity consumption behavior of different users is of profound significance for how power departments operate the power system. With the development of the internet, numerous smart terminals have been deployed throughout cities to collect user electricity consumption data. This allows for the calculation of user electricity consumption behavior characteristics from massive amounts of user electricity consumption data, providing the possibility of distinguishing user electricity consumption behavior patterns.

[0003] Currently, the reciprocal dynamic time normalization (DDTW) algorithm is commonly used to cluster user electricity consumption data. The DDTW algorithm takes into account the local shape information of the time series. That is, after differentiating the time series, it considers the neighboring nodes on both sides in the local area and calculates the distance between corresponding points of two time series using Euclidean distance. This better describes the local information and has better robustness to outliers. Therefore, it can effectively solve the singularity problem.

[0004] Because the distance calculated by the DDTW algorithm has a high time complexity, data smoothing is needed during data processing to remove featureless small fluctuations in order to improve computational efficiency. However, during the data smoothing process, if the values ​​of the two left and right points differ significantly from the center data point when calculating the slope k through the window, the feature region will be incorrectly smoothed, resulting in feature loss and inaccurate subsequent analysis of user behavior features. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a method and apparatus for constructing an attention-based electricity consumption characteristic information model, so as to solve the problem of inaccurate user characteristic analysis in the prior art.

[0006] To achieve the above and other related objectives, the present invention provides a method for constructing an attention-based electrical load characteristic token model, comprising at least the following steps:

[0007] S1, collect electricity consumption data of each user according to the first set time interval;

[0008] S2, Based on attention, the electricity consumption data of each user is processed to obtain the electricity load characteristic model of each user;

[0009] S3 performs clustering on the electricity load characteristic models of all users to obtain the global electricity load characteristic information module.

[0010] Preferably, in step 2, a first load curve and a second load curve are obtained based on the electricity consumption data; then, the electricity load characteristic model is obtained based on the first load curve and the second load curve.

[0011] Preferably, a fine-sampled power load curve is obtained based on the power consumption data, and the fine-sampled power load curve is processed to obtain a first load curve; and a coarse-sampled power load curve is obtained based on the power consumption data, and the coarse-sampled power load curve is processed to obtain a second load curve.

[0012] Preferably, the step of obtaining the coarse-sampled power load curve based on the power consumption data is to downsample the power consumption data according to a second set time interval to obtain the coarse-sampled power load curve; the second set time interval is greater than the first set time interval.

[0013] Preferably, the process of processing the coarse-sampled electrical load curve to obtain the second load curve includes:

[0014] The coarse sampling electrical load curve is smoothed to obtain a second smooth curve;

[0015] The second smooth curve is normalized to obtain the second normalized curve;

[0016] The second load curve is obtained by discretizing the second normalized curve.

[0017] Preferably, the step of smoothing the coarse-sampled electrical load curve to obtain a second smoothed curve includes:

[0018] Distinguish between the characteristic region and the stable region in the coarse sampling power load curve according to the set requirements;

[0019] The stable region is smoothed to obtain a smooth curve portion of the stable region, while maintaining the original curve portion of the feature region;

[0020] The smoothed curve portion and the original curve portion are combined to form the second smoothed curve.

[0021] Preferably, in step 3, hierarchical clustering is used for clustering.

[0022] Preferably, the module for clustering the electricity load characteristic models of all users to obtain the global electricity load characteristic token module includes:

[0023] Divide all user load curves according to certain rules to obtain user load curves in at least two regions; each region includes at least two users.

[0024] Clustering the user load curves of each region yields the edge power load characteristic information module;

[0025] The global electricity load characteristic model is obtained by clustering the edge electricity load characteristic models of all regions.

[0026] Preferably, it also includes application analysis of the global electricity load characteristic information model; the application analysis includes user electricity consumption behavior, energy saving and electricity pricing.

[0027] To achieve the above and other related objectives, the present invention also provides an apparatus for constructing an attention-based electrical load characteristic object model, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the steps of the above-described attention-based electrical load characteristic object model construction method.

[0028] As described above, the method and apparatus for constructing the attention-based electricity consumption characteristic token model of the present invention have the following beneficial effects:

[0029] This invention achieves significant performance improvements with minimal time overhead through a two-layer feature extraction architecture based on an attention mechanism. This reduces computational load and avoids feature loss in feature regions, thereby ensuring the accuracy of the constructed electrical load feature information model and its value for subsequent application analysis. Attached Figure Description

[0030] Figure 1 The diagram shows the curves before and after smoothing in the prior art.

[0031] Figure 2 The diagram shown is a schematic representation of the power system structure based on the user in this invention.

[0032] Figure 3 The diagram shows the construction method of the attention-based electrical load characteristic token model of the present invention.

[0033] Figure 4 The diagram shown is a structural schematic of the device for constructing an attention-based electrical characteristic information object feature model according to the present invention. Detailed Implementation

[0034] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0035] Please see Figure 1-4 It should be noted that the illustrations provided in this embodiment are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0036] With the rapid increase in the number of students on campus, the existing dormitories are no longer sufficient, leading to the addition of new dormitories near the school. Consequently, schools now have different dormitory areas. Students today often use high-powered appliances to improve their comfort and convenience, but this can easily lead to safety issues or excessive electricity waste if not addressed promptly. Therefore, monitoring electricity usage in school dormitories is essential. This necessitates feature extraction and analysis of student electricity consumption patterns in different dormitory areas to better manage school electricity usage.

[0037] Based on the above phenomena, the present invention proposes as follows: Figure 2 The diagram illustrates a user-based power system architecture. The system includes numerous users at the front end, and monitoring terminals and monitoring endpoints at the edge. Each user at the front end is equipped with a smart meter to collect electricity consumption data and send it to the monitoring endpoint. The monitoring endpoints process the data and send it to the control terminal, which then manages the electricity usage for safety. In this power system architecture, the monitoring terminal represents the school, the edge monitoring endpoints represent the control terminals for different dormitory areas, and the users are individual dormitories within each dormitory area.

[0038] Considering that current smoothing methods, in the actual processing of electricity consumption data, may encounter situations where the current point has a high electricity consumption while the two preceding and following points have low and equal electricity consumption, the local slope of the least squares fitting may be low. If the local slope is below a threshold, it can lead to the removal of feature points, resulting in feature loss. Figure 1As shown, the spikes appearing in the left image lose their feature values ​​in the right image after smoothing. To address the feature loss problem, this paper proposes an attention-based two-layer feature extraction structure, which restores lost features with fewer points through two-layer feature concatenation. Specifically, in conjunction with this power system structure, this invention proposes a method and apparatus for constructing an attention-based electricity consumption feature model. This model enables the assessment of user electricity usage; if an anomaly is detected, an alarm is immediately triggered, power is cut off, or professional personnel are dispatched to the site for inspection and maintenance to ensure the safety of user electricity consumption.

[0039] like Figure 3 The diagram shows a flowchart of the method for constructing an attention-based electrical load characteristic information model according to the present invention; based on different dormitory areas of the aforementioned school as application examples, the present invention combines... Figure 2 A detailed description is provided of the method for constructing an attention-based electrical load characteristic information model, which includes at least the following steps:

[0040] S1, collect electricity consumption data of each user according to the first set time interval;

[0041] Smart meters are installed in each dormitory. The smart meters collect electricity consumption data of the dormitory at set time intervals and send the electricity consumption data of each user (each dormitory) out through the smart meters. The transmission method can include wired communication or wireless communication.

[0042] In this embodiment of the invention, the electricity consumption data is described using daily data as an example. The first set time interval is half an hour, that is, the electricity consumption data is collected once every half hour, with a total of 48 sampling points, that is, there are 48 data points of electricity consumption data. The 48 data points are connected to obtain the fine-sampled electricity load curve.

[0043] In another embodiment, if it is necessary to analyze the electricity consumption data for one year, the electricity consumption data includes the electricity consumption data corresponding to the sampling points for 365 days of the year, and the electricity consumption data for the 365 days are connected in chronological order to obtain the load curve corresponding to the year.

[0044] S2, Based on attention, the electricity consumption data of each user is processed to obtain the electricity load characteristic model of each user;

[0045] This step aims to process each user's electricity consumption data based on attention. Attention involves processing characteristic and stable data within the electricity consumption data separately to obtain a new user load characteristic module; it includes at least the following steps:

[0046] S21, obtain the first load curve and the second load curve based on the power consumption data;

[0047] This step aims to analyze the electricity consumption data X = {x} raw,1 ,x raw,2 ,…,x raw,T The first load curve is obtained by processing the data, and the second load curve is obtained by downsampling the data according to the second set time interval; the second set time interval is greater than the first set time interval.

[0048] S211, A fine-sampled power load curve is obtained based on the power consumption data, and the fine-sampled power load curve is processed to obtain a first load curve;

[0049] In this invention, the electricity consumption data is obtained by fine sampling at a first predetermined time interval, and the electricity consumption data is concatenated to obtain the finely sampled electricity load curve; wherein, the processing of the finely sampled electricity load curve includes smoothing, normalization, and discretization. Specifically, the processing of the finely sampled electricity load curve to obtain the first load curve includes at least the following steps:

[0050] S2111, the fine-sampling electrical load curve is smoothed to obtain a first smoothed curve.

[0051] Considering that appliances such as water heaters can cause slight fluctuations in the load curve, and that the amplitude of these fluctuations is uncertain and varies greatly, the present invention uses a method of sampling local trends to detect them. Since the first time interval of the user data of the finely sampled power load curve is small, the feature region corresponding to the fluctuation is also considered a stable region. Therefore, when smoothing the finely sampled power load curve, the feature region corresponding to the fluctuation is also directly smoothed as a stable region.

[0052] Smoothing is performed based on n data points within a stable region. X S Smoothing in,

[0053] After smoothing all stable regions in X, a load curve with smoothed stable regions can be obtained.

[0054] S2112, normalize the first smooth curve to obtain the first normalized curve;

[0055] This step aims to adjust the data to a similar range, facilitating subsequent comparison and processing. In other words, normalization maps the original data to a uniform range of [0,1] for easier comparison and effectively avoids issues such as different load baselines or amplitude variations, thus enabling better comparison of the similarity in shape between load curves.

[0056] In this embodiment of the invention, the first smooth curve is processed using max-min normalization, and the resulting first normalized curve is expressed as follows:

[0057]

[0058] in,

[0059] S2113, Discretize the first normalized curve to obtain the first load curve.

[0060] In this invention, the equal width method is used to define the load curve. Discretization can effectively improve both speed and classification accuracy.

[0061] Specifically, [0,1] is divided into L equal-width intervals {(0,α1],(α1,α2],… ... L-1 ,1],},will Each point x in i Mapping to the median of the interval, as follows:

[0062]

[0063] The first load curve after discretization is:

[0064] S212, obtain a coarse sampling power load curve based on the power consumption data, and process the coarse sampling power load curve to obtain a second load curve;

[0065] This step aims to downsample the electricity consumption data according to a second set time interval to obtain a coarse-sampled electricity load curve, and then obtain a second load curve based on the coarse-sampled electricity load curve; the second set time interval is greater than the first set time interval.

[0066] S2121, Downsample the power consumption data to obtain coarse sample data, and obtain the coarse sample power load curve based on the coarse sample data;

[0067] In this invention, the electricity consumption data is downsampled to obtain coarse sampled data; the coarse sampled data is then concatenated to obtain the coarse sampled electricity load curve.

[0068] The downsampling involves extracting the electricity consumption data according to a second set time interval, where the second set time interval is greater than a first set time interval. Coarse-sampled data is then extracted from the fine-sampled electricity consumption data obtained in step S11 according to the second set time interval; finally, the coarse-sampled data is concatenated to obtain a coarse-sampled electricity load curve.

[0069] In this embodiment, the second set time interval is one hour, so there are 24 sampling points for daily electricity consumption data, i.e., 24 data points. These 24 data points are connected to obtain the coarse-sampled electricity load curve. That is, the coarse-sampled data is Y = {y...} raw,1 ,y raw,2 ,…,y raw,T / 2}

[0070] S2122, The coarse sampling electrical load curve is smoothed to obtain a second smooth curve;

[0071] The coarse sampling power load curve has a small number of sampled data points. In order to improve the smoothing efficiency, the region corresponding to the fluctuation is no longer smoothed. That is, the data point values ​​of the region corresponding to the fluctuation are retained, and other stable regions are smoothed to obtain the second smooth curve.

[0072] In this step, it is necessary to first identify the characteristic region. Therefore, before smoothing the coarse sampled power load curve, the characteristic region corresponding to the fluctuation will be screened out to distinguish the characteristic region from the stable region. The region corresponding to the fluctuation is called the characteristic region.

[0073] Considering the slope of the power load curve The absolute value of the slope is small and close to 0 in the stable region, but the slope will change drastically when various electrical appliances are used. Therefore, the region corresponding to the set of coarse sampled data with a slope greater than the set threshold in the coarse sampled power load curve is called the feature region.

[0074] Let the threshold be δ, and assume there exists a region with n data points.

[0075] When satisfied Such a Y S For a stable region; when the following conditions are met Such a Y S For the feature region; where,

[0076] Specifically, the user's electricity consumption data is formed into a set Y = {y raw,1 ,y raw,2 ,…,y raw,T The set Y is partitioned into sliding windows, and then the slope of the sliding window is calculated using the least squares method. Each slope These are all local trends of finely sampled power load curves;

[0077] In this embodiment of the invention, the sliding window size is 3, and the sliding window is represented as k. t ={y raw,t-1 y raw,t yraw,t+1}

[0078] The slope of each sliding window is calculated using the least squares method. The slopes of the fine-sampled electrical load curves are listed below. The expression for the slope is:

[0079] Where k and b are constants.

[0080] Using the aforementioned method for distinguishing between characteristic regions and stable regions, the slope corresponding to the data points is... The results are compared with a set threshold δ to distinguish between stable and characteristic regions in the coarse-sampled power load curve.

[0081] The stable region is smoothed using the method in step S2111 to obtain the smoothed curve portion of the stable region; the characteristic region is preserved by retaining the original curve portion of the characteristic region in the coarse sampling electrical load curve, and the smoothed curve portion and the original curve portion form the second smoothed curve.

[0082] S2123, Normalize the second smooth curve to obtain the second normalized curve;

[0083] S2124, Discretize the second normalized curve to obtain the second load curve.

[0084] In this step, the normalization and discretization of the second smooth curve are performed in the same way as in step S211, and will not be repeated here.

[0085] The discretized second load curve is as follows:

[0086] S22, the electrical load characteristic model is obtained based on the first load curve and the second load curve;

[0087] S221, record the characteristic regions and their coarse sampling data in the second load curve;

[0088] The coarse-sampled data corresponding to the feature region is dict_Y a ={y a,1 :y a,1 ,…,y a,n :y a,n}

[0089] S222, Process the coarsely sampled data of the feature region to obtain a new feature region;

[0090] Specifically, the coarsely sampled data of the feature region is sorted, and the top k local trends are selected. a,iThe coarsely sampled data yields feature regions with better characteristics:

[0091] Y opt ={y opt,1 ,…,y opt,k}; where 0≤k≤n

[0092] The better feature region Y opt ={y opt,1 ,…,y opt,k Each point in} is mapped to according to the sampling interval. Above, a new feature region is obtained: X opt ={x opt,1 ,…,x opt,k / 2}

[0093] S223, superimpose the data points of the new feature region onto the feature region corresponding to the first load curve to obtain the user's electricity load feature model;

[0094] New feature region X opt ={x opt,1 ,…,x opt,k / 2 The data points are superimposed on the first load curve. The user's final electricity load curve is obtained from the corresponding feature region:

[0095]

[0096] Then, using the user's final electricity consumption characteristic load curve as input, the DDTW distance is calculated as follows:

[0097]

[0098] Where ATT_FDDTW(Y1,Y2) represents the user's electricity load characteristic model. The load curve represents the user's electricity consumption characteristics.

[0099] Therefore, this invention proposes a construction method based on the ATT-FDDTW (Attenion Fast Derivative Dynamic Time Warping) algorithm, which can effectively capture characteristic segments on the load curve and smooth other segments with fewer features. Through a two-layer feature concatenation method, lost features are restored with fewer data points.

[0100] In a preferred embodiment, step S212 is first executed to downsample the user data to obtain a second load curve. Then, based on the characteristic regions in the second load curve, only the data points in the user data corresponding to the characteristic regions are smoothed to obtain a first local smoothed curve. The first local smoothed curve is then mapped onto the characteristic region portion of the second load curve to obtain the user's electricity load curve. This implementation method can accelerate the construction efficiency.

[0101] S3 clusters all user load curves to obtain a global power load characteristic information module.

[0102] In this embodiment, clustering the electricity load curves of all users to obtain a global electricity load feature model includes:

[0103] S31, divide all user load curves according to certain rules to obtain user load curves in at least two regions; each region includes at least two users;

[0104] In this embodiment of the invention, dormitories can be divided according to their location within the dormitory area. Alternatively, they can be categorized by grade level, depending on the user's situation. This invention does not impose any limitations on the rules for categorization.

[0105] S32, cluster the user load curves of each region to obtain the edge power load characteristic information module;

[0106] In the construction of LSD (Low-Sum Distributed Load) clusters, to ensure the quality of their representation of load characteristic curves, the number of load characteristic curves within the same cluster should be small and relatively close together. On the other hand, due to the differences in electricity load behavior among different users, the shapes of electricity load curves vary, resulting in different numbers of load characteristics for different users. This makes methods such as k-means and k-mediods, which require pre-specifying the number of cluster centers, less effective in constructing LSDs.

[0107] This invention uses hierarchical clustering, which can automatically control the number of final clusters by adjusting parameters. Hierarchical clustering initially treats each data point as a separate cluster. In each iteration, it finds the two most similar data points and merges them into a single cluster, until no more clusters can be merged. The similarity between two clusters is determined based on their distance, which can be quantified by minimum inter-cluster distance, maximum inter-cluster distance, or average inter-cluster distance.

[0108] This invention specifically uses the UPGMA (umweighted pair-group method of average) metric to define the distance between two clusters. Cluster Ω u With cluster Ω vThe distance between two clusters is the average distance between all pairs of elements:

[0109]

[0110] Where, N u With N v They are respectively clusters Ω u With cluster Ω v The total number of elements.

[0111] Compared to other clustering methods that require calculating cluster centers, UPGMA-based hierarchical clustering avoids the complex time warping calculations required to obtain cluster centers in each iteration, effectively saving time. For a given threshold θ, the cluster Ω... u With cluster Ω v The following condition must be met during the merging process: dist(Ω) u ,Ω v )≤θ.

[0112] S33, cluster the edge power load characteristic information models of all regions to obtain the global power load characteristic information model;

[0113] In this step, the clustering method is the same as in step S32. That is, the edge power load characteristic information model is clustered again to obtain a more complete and broader global power load characteristic information module. This facilitates subsequent application analysis from a more global perspective.

[0114] The user electricity load feature token module of this invention uses the stable region of the first load curve and the feature region of the second load curve. Based on the edge electricity load feature model and the global electricity load feature module obtained by the user electricity load feature token module, the computational load is reduced during clustering, which makes the algorithm well applied to the massive data in the real world. It also avoids mistaking some feature segments as featureless segments and smoothing them when extracting feature segments, which would lead to the loss of features. The feature extraction architecture of this invention based on the attention mechanism achieves a good performance improvement with a small time cost.

[0115] The method of the present invention also includes application analysis of the global electrical load characteristic object model; the application analysis based on the global electrical load characteristic object model includes:

[0116] By analyzing the data from the fused electricity load characteristic information model, the characteristics of user electricity consumption can be effectively extracted. This allows for effective application in the following scenarios:

[0117] Segmenting user electricity consumption behavior: By extracting user electricity consumption characteristics, we can obtain user electricity consumption habits, preferences, and thinking patterns, and segment user electricity consumption behavior to distinguish different user groups, such as residential electricity consumption, industrial electricity consumption, etc. At the same time, we can also analyze the impact of external factors on user electricity consumption.

[0118] Energy saving: By acquiring users' electricity consumption behavior, it is possible to determine when users consume excessive amounts of electricity, thus reminding users to save electricity and reduce pollution.

[0119] Electricity pricing: By extracting the characteristics of user electricity consumption, peak and off-peak periods can be segmented for electricity pricing. Pricing based on peak and off-peak periods encourages users to stagger their electricity usage, making full use of equipment and energy, and also facilitates the regulation of grid load.

[0120] Device Example:

[0121] like Figure 4 The diagram shows a schematic of the structure of the attention-based electricity load characteristic object feature model construction device of the present invention. The construction device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described attention-based electricity load characteristic object model construction method.

[0122] The steps and principles of the method for constructing the attention-based electrical load characteristic token model are described in detail in the method implementation examples, and will not be repeated here.

[0123] The electricity load feature object model constructed using the attention-based electricity load feature object model of this invention first collects electricity consumption data from different dormitory areas of the school, uploads it to several edge computing points for calculation, and extracts the corresponding electricity load feature object models. This allows for the determination of electricity consumption in different dormitory areas and students' electricity consumption behavior. Then, through clustering by the control terminal, a global electricity load feature object model for the entire school dormitory area is obtained, thereby detecting the electricity consumption situation in the school dormitory area. The edge electricity load feature object model is used to judge the students' electricity consumption in the dormitory area. If an anomaly is detected, an anomaly alarm needs to be issued, power cut off, or professional personnel need to be dispatched to the site to ensure the safety of students' electricity consumption behavior.

[0124] In summary, this invention achieves significant performance improvements with minimal time overhead through a two-layer feature extraction architecture based on an attention mechanism. This reduces computational load while preventing feature loss in feature regions, thus ensuring the accuracy of the constructed electricity consumption feature information model and its value for subsequent analysis and applications. Therefore, this invention effectively overcomes the various shortcomings of existing technologies and possesses high industrial applicability.

[0125] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for constructing an attention-based electrical load characteristic information model, characterized in that, At least the following steps are included: S1, collect electricity consumption data of each user according to the first set time interval; S2, Based on attention, the electricity consumption data of each user is processed to obtain the electricity load characteristic model of each user; In step 2, a first load curve and a second load curve are obtained based on the power consumption data; a fine-sampled power load curve is obtained based on the power consumption data, and the fine-sampled power load curve is processed to obtain the first load curve; The process involves obtaining a coarse-sampled load curve based on electricity consumption data, processing the coarse-sampled load curve to obtain a second load curve, and then obtaining the load characteristic model based on the first and second load curves. The process of obtaining the load characteristic model based on the first and second load curves includes: recording the characteristic regions in the second load curve and their coarse-sampled data; processing the coarse-sampled data of the characteristic regions to obtain new characteristic regions; sorting the coarse-sampled data of the characteristic regions and selecting the top k local trends. The coarsely sampled data is used to obtain a new feature region; the data points of the new feature region are superimposed on the feature region corresponding to the first load curve to obtain the user's electricity load feature model; wherein, the region corresponding to the set of coarsely sampled data with a slope greater than a set threshold in the coarsely sampled electricity load curve is called the feature region; S3 performs clustering on the electricity load characteristic models of all users to obtain the global electricity load characteristic information module.

2. The method for constructing an attention-based electrical load characteristic information model according to claim 1, characterized in that, The process of obtaining the coarse-sampled power load curve based on the power consumption data involves downsampling the power consumption data according to a second set time interval; the second set time interval is greater than the first set time interval.

3. The method for constructing an attention-based electrical load characteristic information model according to claim 1, characterized in that, The process of processing the coarse sampled load curve to obtain the second load curve includes: The coarse sampling electrical load curve is smoothed to obtain a second smooth curve; The second smooth curve is normalized to obtain the second normalized curve; The second load curve is obtained by discretizing the second normalized curve.

4. The method for constructing an attention-based electrical load characteristic information model according to claim 3, characterized in that, The process of smoothing the coarse-sampled electrical load curve to obtain the second smoothed curve includes: Distinguish between the characteristic region and the stable region in the coarse sampling power load curve according to the set requirements; The stable region is smoothed to obtain a smooth curve portion of the stable region, while maintaining the original curve portion of the feature region; The smoothed curve portion and the original curve portion are combined to form the second smoothed curve.

5. The method for constructing an attention-based electrical load characteristic information model according to claim 1, characterized in that, In step 3, hierarchical clustering is used.

6. The method for constructing an attention-based electrical load characteristic information model according to claim 1, characterized in that, The module for clustering the electricity load characteristic models of all users to obtain the global electricity load characteristic information module includes: Divide all user load curves according to certain rules to obtain user load curves in at least two regions; each region includes at least two users. Clustering the user load curves of each region yields the edge power load characteristic information module; The global electricity load characteristic model is obtained by clustering the edge electricity load characteristic models of all regions.

7. The method for constructing an attention-based electrical load characteristic information model according to claim 1, characterized in that, It also includes the application analysis of a global electricity load characteristic information model; the application analysis includes user electricity consumption behavior, energy saving and electricity pricing.

8. An apparatus for constructing an attention-based electrical load characteristic token model, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for constructing an attention-based electrical load characteristic token model as described in claims 1-7.

Images