A user power consumption load classification method based on data feature analysis

Abstract

The present application relates to the technical field of data classification, and particularly relates to a user power load classification method based on data feature analysis, which comprises the following steps: obtaining multi-dimensional power load data of a plurality of users, determining a category center probability of each user, taking a user corresponding to the maximum category center probability as an initial classification center, constantly screening a new classification center according to distances between other users and the determined classification center and category center probabilities of the other users, until a preset stop condition is reached, and finally classifying power load data according to each obtained classification center. In addition to considering the power load data of the same day, the present application also considers power load data of other dimensions of the user, so that the classification accuracy of the power load data of the user can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data classification, and particularly relates to a user electricity load classification method based on data feature analysis. BACKGROUND

[0002] User electricity load data refers to the power consumption record of a power user within a certain time. By classifying these data, the user can be divided into groups with similar electricity characteristics, thereby helping the power company to more accurately perform load forecasting, demand response management and power resource optimization allocation.

[0003] Based on the similarity between the electricity load data of multiple users, multiple different categories are formed. However, for users with the same electricity load, it does not mean that they belong to the same category, because the electricity load of the user is affected by multiple factors such as environmental temperature, environmental humidity, user type and other variables. For example, users with the same daily electricity consumption in winter and summer may have completely different electricity mode and demand characteristics, and cannot be simply classified into the same category. Therefore, the accuracy of classifying only by load size is low. SUMMARY

[0004] In order to solve the technical problem of low accuracy of the existing user electricity load classification method, the purpose of the present application is to provide a user electricity load classification method based on data feature analysis, and the technical solution adopted is as follows:

[0005] The present application provides a user electricity load classification method based on data feature analysis, comprising:

[0006] Obtaining multi-dimensional electricity load data of multiple users, the multi-dimensional electricity load data comprising daily electricity load data;

[0007] Determining the category center probability of each user, the category center probability being obtained from the association of each dimension of electricity load data with daily electricity load data and the isolated condition of each dimension of electricity load data of each user;

[0008] Taking the user corresponding to the maximum category center probability as the initial classification center, and constantly screening new classification centers according to the distance between each user and the determined classification center and the category center probability of each user until a preset stop condition is reached; the determined classification center comprises the initial classification center;

[0009] Classifying the electricity load data of all users according to the final obtained category center.

[0010] In an exemplary embodiment, the category center probability of each user is obtained by the central constructivity of the electricity load data in each dimension and the category centrality of the electricity load data in each dimension of each user. The central constructivity represents the association between the electricity load data in each dimension and the electricity load data of the day, and the category centrality represents the isolation of the electricity load data in each dimension of each user.

[0011] In an exemplary embodiment, the process of obtaining the central constructivity includes:

[0012] The correlation coefficient between the power load data of each dimension and the power load data of the current day is obtained, and the central constructivity of each dimension is obtained from the correlation coefficient.

[0013] In an exemplary embodiment, the process of obtaining the category centrality includes:

[0014] Based on the electricity load data of the same dimension of each user, the local outlier factor LOF value of the electricity load data of each dimension of each user is obtained;

[0015] The local outlier factor (LOF) value of the power load data of each dimension of each user is negatively correlated to obtain the category centrality of the power load data of each dimension of each user.

[0016] In an exemplary embodiment, the class center probability is calculated as follows:

[0017] The centrality weight of each dimension is obtained according to the centrality constructivity of each dimension; the centrality weight is proportional to the centrality constructivity;

[0018] According to the centrality weight of each dimension, the category centrality of the power load data of each dimension of the target user is weighted and summed to obtain the category centrality probability of the target user; the target user is any user.

[0019] In an exemplary embodiment, continuously selecting new classification centers based on the distances of other users from the determined classification center and the classification center probabilities of other users includes: continuously iterating according to the following cyclic process to continuously select new classification centers:

[0020] Based on the central constructivity of each dimension and the distance between the power load data of each dimension of the remaining users and the power load data of each dimension of the confirmed classification centers, the target distance of each remaining user with respect to each confirmed classification center is obtained;

[0021] The target distances and category center probabilities corresponding to the remaining users are integrated to obtain the target probabilities of the remaining users belonging to the new category center.

[0022] The user corresponding to the largest target probability is used as another classification center.

[0023] In an exemplary embodiment, the process of acquiring the target distance includes:

[0024] According to the centrality weight of each dimension, the distances between the power load data of each dimension of the reference user and the power load data of each dimension of each confirmed classification center are weighted and summed to obtain the target distance of the reference user; the reference user is any one of the remaining other users; the centrality weight of each dimension is obtained by the central construct of each dimension.

[0025] In an exemplary embodiment, the preset cutoff condition includes:

[0026] Classify the electricity load data of all users according to the classification centers obtained in any two adjacent iterative processes, obtain two classification results, and obtain the classification effects of the two classification results;

[0027] If the classification effects of the two classification results satisfy a preset classification effect degradation condition, the iteration is stopped, and the previous iteration process of any two adjacent iteration processes is used as the last iteration process.

[0028] In an exemplary embodiment, the classification effects of the two classification results satisfy a preset classification effect degradation condition, including:

[0029] The difference between the classification effect corresponding to the latter iterative process and the classification effect corresponding to the previous iterative process in any two adjacent iterative processes is calculated, and the difference is less than a preset threshold.

[0030] In an exemplary embodiment, classifying the electricity load data of all users according to the finally obtained classification centers includes:

[0031] Calculate the distance between each user's electricity load data and each classification center;

[0032] The electricity load data of each user is classified into the category corresponding to the classification center with the smallest distance.

[0033] The present invention has the following beneficial effects: in addition to considering the electricity load data of the day, the present invention also considers the electricity load data of multiple other dimensions of the user, and obtains the category center probability of each user based on the correlation between the electricity load data of each dimension and the electricity load data of the day, as well as the isolation of the electricity load data of each dimension of each user. The category center probability is the probability of the user as the classification center, so that the user corresponding to the largest category center probability is first used as the initial classification center, and then according to the distance of each other user from the determined classification center and the category center probability of each other user, continuous iteration is performed to screen out new classification centers, and finally the electricity load data of all users are classified according to the finally obtained classification centers, which can improve the classification accuracy of user electricity load data. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of a method for classifying user power loads based on data feature analysis provided by one embodiment of the present invention;

[0035] Figure 2 is a flowchart for obtaining category centrality provided by one embodiment of the present invention;

[0036] Figure 3 is a flow chart for obtaining the category center probability provided by one embodiment of the present invention;

[0037] Figure 4 is a flow chart of a cyclic iterative process provided by one embodiment of the present invention;

[0038] Figure 5 This is a specific implementation flow chart of the preset cutoff condition provided by one embodiment of the present invention;

[0039] Figure 6 is a flow chart of a classification process provided by one embodiment of the present invention;

[0040] Figure 7 It is a schematic diagram of the structure of a decision tree provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] To further illustrate the technical means and effects employed by the present invention to achieve its intended objectives, the following detailed description of the specific embodiments, structures, features, and effects of the present invention is provided in conjunction with the accompanying drawings and preferred embodiments. In the following description, references to "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this invention pertains. The data and information collected in this application were obtained with full consent and authorization.

[0043] This embodiment provides a method for classifying user electricity loads based on data feature analysis. When classifying users based on their electricity load data, it is not enough to rely solely on the current day's electricity load data; electricity load data from other dimensions is also needed to improve the accuracy of user classification. It should be understood that user classification essentially involves classifying users' electricity load data.

[0044] like Figure 1 As shown, this embodiment provides a user power load classification method based on data feature analysis, including:

[0045] Step S1: Acquire multi-dimensional electricity load data of multiple users, where the multi-dimensional electricity load data includes electricity load data for the current day;

[0046] Step S2: Determine the category center probability of each user. The category center probability is obtained by the correlation between the power load data of each dimension and the power load data of the day, as well as the isolation of the power load data of each dimension of each user.

[0047] Step S3: The user with the highest probability of being the classification center is selected as the initial classification center. New classification centers are continuously selected based on the distances of other users from the determined classification center and the classification center probabilities of other users until a preset cutoff condition is reached. The determined classification centers include the initial classification center.

[0048] Step S4: Classify the electricity load data of all users according to the finally obtained classification centers.

[0049] The implementation process of each step is described in detail below with reference to the accompanying drawings.

[0050] Step S1: Acquire multi-dimensional electricity load data of multiple users, where the multi-dimensional electricity load data includes electricity load data for the current day.

[0051] Obtain multi-dimensional electricity load data for multiple users, where each user is an object that needs to be classified. The number of users is set according to the actual classification needs, such as all users in a certain area.

[0052] The multidimensional electricity load data includes electricity load data of multiple different dimensions, one of which is the current day's electricity load data (i.e., the current day's electricity load data). In addition to the current day's electricity load data, the electricity load data of the other dimensions is set according to actual conditions. In an exemplary embodiment, the electricity load data of the other dimensions includes: user type, the temperature of the user's environment on the current day, the humidity of the user's environment on the current day, and electricity load data for several historical days. The user type represents the type of user's electricity usage, such as: office worker families, single-elderly families, night shift workers' families, freelance families, etc. To facilitate data processing, each user type needs to be quantified, with different numerical values ​​representing different user types. For example, the quantization value for office worker families is 1, the quantization value for single-elderly families is 2, and so on. The electricity load data for several historical days can specifically be: electricity load data for the 10 days before the current day, i.e., electricity load data for the past 10 days.

[0053] It should be understood that for power load data representing the temperature and humidity of the user's environment on that day, the average of multiple data points for that day can be used as the power load data for that day. Specifically, for the temperature of the user's environment on that day, the temperatures of the user's environment at multiple different times on that day are obtained according to a preset sampling period or random sampling, and then the average of the temperatures for that day is calculated to serve as the temperature of the user's environment on that day. The humidity of the user's environment on that day is obtained in a similar manner. For any day in the power load data for several historical days and the power load data for that day, the power consumption of the user on that day is obtained as the power load data for that day.

[0054] In order to achieve accurate classification of user electricity consumption data, a multi-dimensional feature group is constructed based on the various dimensions contained in the multi-dimensional electricity load data of each user, as shown in Table 1. Table 1 is a multi-dimensional feature group.

[0055] Table 1

[0056]

[0057] Where M represents the number of dimensions, Represents the power load data of the first dimension, specifically the quantitative value corresponding to the user type, Represents the second dimension of power load data, specifically the temperature of the user's environment on that day. Represents the third dimension of power load data, specifically the humidity of the user's environment on that day. Represents the fourth dimension of electricity load data, specifically the user load on the 10th day in the past, and so on. Represents the electricity load data of the M-1th dimension, specifically the user load on the first day in the past. The Mth dimension represents the power load data, specifically the user load data for that day. Based on the above multi-dimensional user load data division method, the multi-dimensional power load data of each user is obtained.

[0058] Step S2: Determine the category center probability of each user. The category center probability is obtained by the association between the power load data of each dimension and the power load data of the day, as well as the isolation of the power load data of each dimension of each user.

[0059] It should be understood that many factors influence the classification of user electricity load data. In addition to daily electricity load data, this includes electricity load data from various other dimensions, such as the temperature and humidity of the user's environment on that day, as well as electricity load data from several days of history. Each characteristic factor may have a different impact on the classification results and its importance may vary (i.e., different central structural factors).

[0060] Therefore, the category centrality probability of each user is derived from the central constructivity of the power load data in each dimension and the category centrality of the power load data in each dimension of each user. The central constructivity represents the correlation between the power load data in each dimension and the power load data of the day, while the category centrality represents the isolation of the power load data in each dimension of each user.

[0061] In classification, if the electricity load data of a certain dimension shows a similar trend of change to the electricity load data of the day (for example, when the temperature rises, the user's electricity load increases), then it can be considered that this characteristic factor has a higher contribution to the classification task. In this case, the correlation between the electricity load data of this dimension and the electricity load data of the day is strong, which plays a more critical role in classification. Specifically, this shows that the "central structurality" of the electricity load data of this dimension is strong, which can effectively help classification and improve the accuracy of the classification results. Accordingly, a specific process for obtaining the central structurality is given as follows: obtain the correlation coefficient between the electricity load data of each dimension and the electricity load data of the day, and obtain the central structurality of each dimension from the correlation coefficient. Specifically: set any dimension as the Ath dimension, then the quantitative formula for the central structurality of the Ath dimension is as follows:

[0062] ;

[0063] in, represents the central constructivity of the Ath dimension, Represents the A-th dimension electricity load data of the first user, Represents the power load data of the Ath dimension of the Nth user, where N represents the number of users and M represents the number of dimensions. represents the A-th dimension electricity load data set for all users, Represents the electricity load data of the Mth dimension of the first user, that is, the electricity load data of the first user on that day, Represents the electricity load data of the Mth dimension of the Nth user, that is, the electricity load data of the Nth user on that day, It represents the M-th dimension electricity load data set for all users, that is, the electricity load data set for the day.

[0064] express and Pearson correlation coefficient.

[0065] represents the maximum-minimum normalization method. Here, the maximum and minimum values ​​of the Pearson correlation coefficients corresponding to each dimension are obtained. Then, the maximum-minimum normalization method is used to normalize the Pearson correlation coefficient corresponding to the Ath dimension. The result is the central structural property of the Ath dimension. Using this method, the central structural property of the electricity load data for each dimension is obtained. It should be understood that the Pearson correlation coefficient between the electricity load data set for that day and itself needs to be obtained, which is a value of 1.

[0066] The multidimensional electricity load data of each user is composed of electricity load data in M ​​dimensions. Now we need to calculate the category centrality of each user's electricity load data in each dimension. Specifically, if the data points of the electricity load data are clustered more closely in a certain dimension and can form a clear center, it means that the category centrality of the electricity load data in this dimension is high. On the contrary, if the distribution of data points in a dimension is relatively isolated, it means that the category centrality of the electricity load data in this dimension is weak. Accordingly, if Figure 2 As shown, a specific process of obtaining category centrality is given as follows:

[0067] Step S2-1: Based on the electricity load data of the same dimension of each user, obtain the local outlier factor LOF value of the electricity load data of each dimension of each user.

[0068] Obtain the electricity load data of each user in the same dimension, for example, obtain the electricity load data of the Ath dimension of each user, so as to obtain the electricity load data set corresponding to the Ath dimension, which includes the electricity load data of the Ath dimension of each user.

[0069] The electricity load data set corresponding to the Ath dimension is then processed using a local outlier factor (LOF) algorithm to obtain the local outlier factor value (i.e., LOF value) for each user's electricity load data in the Ath dimension. Since processing a data set using the LOF algorithm to obtain the LOF value for each data point is a conventional technique, it will not be described in detail here. The LOF value for each user's electricity load data in each dimension is also called an isolated value (or local outlier). A higher LOF value indicates that the corresponding electricity load data is more isolated and has a higher degree of anomaly. Electricity load data with a higher LOF value is farther away from other electricity load data in the feature space and has a lower probability of belonging to the category center. Conversely, electricity load data with a lower LOF value exhibits more pronounced clustering and is more likely to belong to the category center. Therefore, the category centrality of electricity load data is inversely proportional to its LOF value.

[0070] In an exemplary embodiment, a two-dimensional plane of the Ath dimension can also be constructed. The horizontal and vertical axes of the two-dimensional plane both represent the Ath dimension. Then, the electricity load data of each user in the Ath dimension is mapped to the two-dimensional plane. It should be understood that at the coordinate points of the electricity load data of each user in the Ath dimension in the two-dimensional plane, the horizontal and vertical coordinate values ​​are the same, both of which are the values ​​of the electricity load data in the Ath dimension. The electricity load data of each user in the Ath dimension mapped to the two-dimensional plane is then processed using the LOF algorithm to obtain the LOF value of each data.

[0071] Step S2-2: negatively correlate the local outlier factor (LOF) value of the power load data of each dimension of each user to obtain the category centrality of the power load data of each dimension of each user.

[0072] After obtaining the LOF value of the A-th dimension of the power load data of each user, the LOF value of the A-th dimension of the power load data of each user is negatively correlated. The result of the negative correlation is the category centrality of the A-th dimension of the power load data of each user. In an exemplary embodiment, a specific implementation method of the negative correlation is given as follows:

[0073] ;

[0074] in, represents the category centrality of the A-th dimension of the electricity load data of the i-th user, represents the A-th dimension electricity load data of the i-th user, represents the LOF value of the A-th dimension of the electricity load data of the i-th user, and exp represents an exponential function with the natural constant e as the base.

[0075] By adopting the above method, the category centrality of the A-th dimension of the power load data of each user is obtained, thereby obtaining the category centrality of the power load data of each dimension of each user.

[0076] The central structure of the power load data in different dimensions is usually different. The power load data corresponding to the dimension with stronger central structure has a greater impact on the data category center, while the dimension with weaker central structure has a smaller impact on the data category center. Therefore, the category center probability of each user is obtained by combining the central structure of the power load data in each dimension and the category centrality of the power load data in each dimension of each user. In an exemplary embodiment, Figure 3 As shown, a specific calculation method of the category center probability is given as follows:

[0077] Step S2-3: Obtain the centrality weight of each dimension according to the central constructivity of each dimension.

[0078] The centrality weight of each dimension is obtained based on the centrality constructivity of each dimension, wherein the centrality weight is proportional to the centrality constructivity. The greater the centrality constructivity, the greater the centrality weight. In an exemplary embodiment, the sum of the centrality constructivity of each dimension is calculated, and then the ratio of the centrality constructivity of each dimension to the sum is calculated, and the result obtained is the centrality weight of each dimension. In this way, not only does the logic of the greater the centrality constructivity, the greater the centrality weight is satisfied, but also the sum of the centrality weights of each dimension can be made to be 1.

[0079] Step S2-4: Based on the centrality weights of each dimension, perform weighted summation on the category centrality of the target user's power load data in each dimension to obtain the category centrality probability of the target user.

[0080] For ease of explanation, the target user is set to any user, and the target user is represented by the i-th user. According to the centrality weight of each dimension, the category centrality of the power load data of each dimension of the i-th user is weighted and summed to obtain the category centrality probability of the i-th user. The calculation formula is as follows:

[0081] ;

[0082] in, represents the category center probability of the i-th user, represents the centrality weight of the A-th dimension.

[0083] The above method is used to obtain the category center probability of each user. The larger the category center probability, the greater the possibility that the corresponding user is the category center.

[0084] Step S3: The user corresponding to the largest category center probability is used as the initial category center. New category centers are continuously screened out based on the distances of other users from the determined category center and the category center probabilities of other users until the preset cutoff condition is reached; the determined category centers include the initial category center.

[0085] In step S2, the class center probability for each user is obtained. Next, the electricity load data for all users needs to be classified. Since the larger the class center probability, the more likely the corresponding user is to be the class center, if there is only one class, the class center must be the user with the highest class center probability. Therefore, the user with the highest class center probability is used as the initial class center, that is, as the first class center.

[0086] Then, based on the distances of each user from the identified classification center and the probability of each user's classification center, new classification centers are continuously selected until a preset cutoff condition is reached. The identified classification centers include the first classification center. Therefore, the process of continuously selecting new classification centers is an iterative process. In each iteration, a new classification center is obtained until the preset cutoff condition is reached, resulting in all the classification centers.

[0087] It should be understood that when adding a new classification center based on the already determined classification center, the new classification center needs to be as far away from the already determined classification center as possible. For example, the second classification center needs to be as far away from the initial classification center as possible. Ensuring that the distance between classification centers is as large as possible helps to improve the classification discrimination. If the new classification center is too close to the confirmed classification center, then the two classification centers may cover similar power load data, resulting in overlap and ambiguity between categories, thereby reducing the classification effect. Therefore, the newly added classification centers need not only to select a larger classification center probability as much as possible, but also to ensure that they are far away from the existing classification centers to increase the accuracy and discrimination of the classification. In this way, each classification center can be made as independent as possible to avoid overlap, thereby more effectively dividing the data.

[0088] In an exemplary embodiment, Figure 4 As shown, the following cyclic process is continuously iterated to continuously screen out new classification centers:

[0089] Step S3-1: Based on the central constructivity of each dimension and the distance between the power load data of each dimension of the remaining other users and the power load data of each dimension of the confirmed classification centers, the target distance of the remaining other users to the confirmed classification centers is obtained.

[0090] By filtering through classification centers, all users are divided into two categories: the first category is confirmed classification centers, and the second category is the remaining users, who are not confirmed as classification centers. At the end of each loop iteration, the remaining users and confirmed classification centers are updated, with the remaining users decreasing and the confirmed classification centers increasing.

[0091] Based on the central constructivity of each dimension and the distance between the power load data of each dimension of the remaining other users and the power load data of each dimension of the confirmed classification centers, the target distance of the remaining other users with respect to the confirmed classification centers is obtained. Among them, the central constructivity of each dimension is used to weight the distance between the power load data of the remaining other users under the corresponding dimension and the power load data of the confirmed classification centers, thereby obtaining the target distance of the remaining other users with respect to the confirmed classification centers. In an exemplary embodiment, a specific process for obtaining the target distance is given as follows:

[0092] First, the centrality weight of each dimension is obtained according to the central constructivity of each dimension. This part has been described in detail above and will not be explained again.

[0093] For ease of explanation, the reference user is set to any one of the remaining users. For any dimension, the distance between the reference user's power load data for that dimension and the power load data for that dimension of any confirmed classification center is obtained. The distance between two power load data can be the absolute value of the difference between the two power load data. The larger the absolute value of the difference, the greater the distance between the two power load data.

[0094] Based on the centrality weights of each dimension, the distances between the reference user's electricity load data in each dimension and the electricity load data in each dimension of any confirmed classification center are weighted and summed to obtain the reference user's target distance for any confirmed classification center. It should be understood that if multiple classification centers have been confirmed, the above process will be used to obtain the reference user's target distance for each confirmed classification center, and then the average value will be calculated. The result is the reference user's target distance for all confirmed classification centers, which serves as the reference user's target distance for subsequent data processing.

[0095] Using the above process, the target distances corresponding to the remaining users are obtained.

[0096] Step S3-2: Fusing the target distances and category center probabilities corresponding to the remaining users, obtain the target probabilities that the remaining users belong to the new category center.

[0097] While obtaining the target distance corresponding to the reference user, the category center probability of the reference user is obtained, and the target distance corresponding to the reference user and the category center probability of the reference user are fused to obtain the target probability that the reference user belongs to the new category center. From the above analysis, it can be seen that the larger the target distance corresponding to the reference user, the higher the target probability that the reference user belongs to the new category center, and the higher the category center probability of the reference user, the higher the target probability that the reference user belongs to the new category center. Therefore, the target probability that the reference user belongs to the new category center is proportional to the target distance corresponding to the reference user and proportional to the category center probability of the reference user. In an exemplary embodiment, the target distance corresponding to the reference user can be multiplied by the category center probability of the reference user to obtain the target probability that the reference user belongs to the new category center. In this way, the target probabilities of the remaining other users belonging to the new category center are obtained.

[0098] Step S3-3: The user corresponding to the maximum target probability is used as another classification center.

[0099] The higher the target probability, the higher the possibility that the corresponding user belongs to the classification center. Therefore, the user corresponding to the largest target probability among the remaining target probabilities of other users obtained in step S3-2 is used as another classification center.

[0100] Step S3-1 to step S3-3 is a loop iteration process. Each time step S3-1 to step S3-3 is executed, a loop iteration process is executed, thereby continuously looping and iterating steps S3-1 to step S3-3. Each loop iteration process obtains a classification center until the preset cutoff condition is reached and multiple classification centers are obtained.

[0101] For the above iterative process, a specific process example is given below:

[0102] The initial classification center is denoted as .

[0103] For the first loop iteration process, obtain the jth user and the initial classification center among the remaining users The target distance is calculated as follows:

[0104] ;

[0105] in, Represents the jth user among the remaining users and the initial classification center target distance, Indicates the jth user among the remaining users and the initial classification center The distance between the power load data in the Ath dimension.

[0106] The above calculation formula is essentially to obtain the jth user among the remaining users and the initial classification center The Euclidean distance is simply the distance between the power load data of each dimension is weighted by the centrality weight of each dimension.

[0107] Then, the target probability that the jth user among the remaining users belongs to the second classification center is expressed by the following formula:

[0108] ;

[0109] in, represents the target probability that the jth user among the remaining users belongs to the second classification center, represents the category center probability of the jth user among the remaining users.

[0110] Express Normalization, the normalization method here is the maximum and minimum normalization method, that is: get the remaining other users and the initial classification center The maximum and minimum values ​​of the target distance are then normalized using the maximum and minimum values. Perform normalization.

[0111] The user with the maximum target probability among the remaining users belonging to the second classification center is determined as the second classification center, denoted as , so far the first loop iteration process ends, and the following enters the second loop iteration process:

[0112] After the first cycle iteration process, one user is reduced from the remaining users, and one classification center is added.

[0113] For the remaining users, we still use the jth user as the representation. We obtain the jth user and the initial classification center of the remaining users respectively. The target distance and the second classification center The target distance between the jth user and the second classification center The calculation process of the target distance, and the distance between the jth user and the initial classification center Target distance The calculation process is the same as the second classification center The target distance is calculated as follows:

[0114] ;

[0115] in, Indicates the jth user among the remaining users and the second classification center target distance, Indicates the jth user among the remaining users and the second classification center The distance between the power load data in the Ath dimension.

[0116] Calculate target distance Distance to target The average value of the remaining users is used as the target distance corresponding to the jth user, which is recorded as .

[0117] Then, the target probability calculation formula is used to obtain the target probability that the j-th user among the remaining users belongs to the third classification center.

[0118] The user with the highest target probability among the remaining users belonging to the third classification center is determined as the third classification center. This concludes the second iteration and enters the third iteration. The above iteration process is repeated continuously. Therefore, each iteration differs in that the target distance calculation method for each of the remaining users needs to be updated to the average of the target distances to each classification center. For example, for the wth iteration, the wth iteration represents any iteration. After the w-1th iteration, a total of w classification centers are obtained. The target distance between the jth user and each of the w classification centers is then calculated. The average of the target distances to the w classification centers is then calculated as the target distance for the jth user among the remaining users.

[0119] The preset end condition of the loop iteration process can be set according to actual needs. For example, the preset end condition is that the number of loop iterations is a preset number. When the number of loop iterations reaches the preset number, the iteration is stopped. In order to improve the accuracy and reliability of the iteration stop judgment, in an exemplary embodiment, if Figure 5 As shown, a specific implementation process of the preset cutoff condition is given as follows:

[0120] Step S3-4: classify the electricity load data of all users according to the classification centers obtained in any two adjacent iterative processes, obtain two classification results, and obtain the classification effects of the two classification results.

[0121] For any two adjacent loop iteration processes, they are set as the w-th loop iteration process and the w+1-th loop iteration process. At the end of the w-th loop iteration process, w+1 classification centers are obtained, and at the end of the w+1-th loop iteration process, w+2 classification centers are obtained.

[0122] The electricity load data of all users are classified according to the w+1 classification centers obtained in the w-th loop iteration process. In an exemplary embodiment, the electricity load data of all users are clustered according to the target distance between the electricity load data of each user outside the classification center and each classification center, using the K-means clustering algorithm with a K value of w+1, to obtain w+1 clusters as the classification result of the w-th loop iteration process, and the classification effect of the classification result is obtained.

[0123] Similarly, the electricity load data of all users are classified according to the w+2 classification centers obtained in the w+1th loop iteration process. In an exemplary embodiment, the electricity load data of all users are clustered according to the target distance between the electricity load data of each user in the non-classification center and each classification center, using the K-means clustering algorithm with a K value of w+2, to obtain w+2 clusters as the classification result of the w+1th loop iteration process, and the classification effect of the classification result is obtained.

[0124] The classification effect can be obtained by an existing method, such as the silhouette coefficient, the Davis-Bourdin index, etc. In an exemplary embodiment, a specific method for obtaining the classification effect is provided as follows: for any cluster, the average value of the target distance between the power load data of each user in the cluster and the classification center of the cluster is obtained, thereby obtaining the target distance average value of the cluster, and then calculating the average value of the target distance average values ​​of all clusters in the corresponding loop iteration process as the classification effect of the corresponding loop iteration process. The smaller the average value of the target distance average values ​​of all clusters, the better the classification effect. The average value of the target distance average values ​​is inversely proportional to the classification effect.

[0125] Set the average value of the target distance corresponding to the w-th loop iteration process to be , set the average value of the target distance corresponding to the w+1th loop iteration process to be .

[0126] A quantitative method of classification effect is given below. The classification effect of the w-th cycle iteration process is , the classification effect of the w+1th cycle iteration process is .

[0127] It should be understood that the above-mentioned K-means clustering algorithm may involve multiple iterations. In addition to using the K-means clustering algorithm, clustering operations can also be performed using only the classification method provided in step S4 below.

[0128] Step S3-5: If the classification effects of the two classification results meet the preset classification effect degradation condition, the iteration is stopped, and the previous iteration process between any two adjacent iteration processes is regarded as the last iteration process.

[0129] If the classification effect of the w-th iteration process meets the preset classification effect decline condition compared to the classification effect of the w+1-th iteration process, the iteration is stopped and the w-th iteration process is used as the last iteration process, thereby obtaining all the classification centers obtained up to the w-th iteration process. If the classification effect of the w-th iteration process meets the preset classification effect decline condition compared to the classification effect of the w+1-th iteration process, the classification effect of the w+1-th iteration process is worse than that of the w-th iteration process, and the classification effect is no longer improving and has declined to a certain extent, then no further iteration is required.

[0130] In an exemplary embodiment, the classification effect of the w-th loop iteration process and the classification effect of the w+1-th loop iteration process meet a preset classification effect decline condition, specifically: the classification effect of the w+1-th loop iteration process is calculated minus the classification effect of the w-th loop iteration process, that is, the difference between the classification effect of the w+1-th loop iteration process and the classification effect of the w-th loop iteration process is obtained, and the obtained difference is less than a preset threshold. It should be understood that the preset threshold is a negative value, which is used to indicate that the classification effect of the w+1-th loop iteration process is less than the classification effect of the w-th loop iteration process. The specific value of the preset threshold is set according to actual judgment needs, such as -0.2.

[0131] Step S4: Classify the electricity load data of all users according to the finally obtained classification centers.

[0132] Through step S3, several classification centers are finally obtained. Then, the power load data of all users are classified according to the classification centers finally obtained. In an exemplary embodiment, Figure 6 As shown, a specific classification process is given as follows:

[0133] Step S4-1: Calculate the distance between each user's power load data and each classification center;

[0134] Step S4-2: Classify the power load data of each user into the category corresponding to the classification center with the smallest distance.

[0135] In an exemplary embodiment, the target distance between the power load data of each user (i.e., each user not in the classification center) and each classification center is calculated, wherein, for any user, the minimum target distance is determined from the target distances between the user's power load data and each classification center, the classification center corresponding to the minimum target distance is determined, and the power load data of the user is classified into the category corresponding to the classification center corresponding to the minimum target distance, thereby completing the classification of the power load data of each user, which can improve the classification accuracy of the user's power load data.

[0136] In addition, in an exemplary embodiment, after the classification of the power load data of each user is completed, a decision tree may be constructed according to the classification results to facilitate subsequent specific applications.

[0137] Each user's category is quantified, with different quantization values ​​assigned to different categories to distinguish them. Specifically, the categories are first sorted (the sorting rules are set based on actual needs), and then quantized. For example, the first category has a quantization value of 1, the second category has a quantization value of 2, and so on. Based on the multidimensional electricity load data of each user, a feature matrix is ​​constructed for each user. In addition to the multidimensional electricity load data, the feature matrix also includes the quantization value of the category.

[0138] After the feature matrix is ​​constructed, the ID3 algorithm of information gain can be used to construct a classification decision tree. This decision tree can be defined as a classification model to determine the power load categories of different users, such as Figure 7 Figure 1 shows a schematic diagram of a decision tree structure. The constructed decision tree can be used to process new user load data. Simply inputting the new user load data set into the decision tree automatically completes the classification of user loads, thus achieving accurate and automated load classification management.

[0139] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0140] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A user power load classification method based on data feature analysis, characterized in that: include: Acquire multi-dimensional electricity load data of multiple users, wherein the multi-dimensional electricity load data includes electricity load data of the current day; Determine the category center probability of each user, where the category center probability is obtained by the correlation between the power load data of each dimension and the power load data of the day, and the isolation of the power load data of each dimension of each user; The user corresponding to the largest classification center probability is used as the initial classification center. New classification centers are continuously screened out based on the distances of other users from the determined classification center and the classification center probabilities of other users until a preset cutoff condition is reached; the determined classification centers include the initial classification center. Classify the electricity load data of all users according to the final classification centers; The category center probability of each user is obtained by the central constructivity of the power load data of each dimension and the category centrality of the power load data of each dimension of each user. The central constructivity represents the association between the power load data of each dimension and the power load data of the day, and the category centrality represents the isolation of the power load data of each dimension of each user. The method of continuously selecting new classification centers based on the distances of other users from the determined classification center and the classification center probabilities of other users includes: continuously iterating according to the following cyclic process to continuously select new classification centers: Based on the central constructivity of each dimension and the distance between the power load data of each dimension of the remaining users and the power load data of each dimension of the confirmed classification centers, the target distance of each remaining user with respect to each confirmed classification center is obtained; The target distances and category center probabilities corresponding to the remaining users are integrated to obtain the target probabilities of the remaining users belonging to the new category center. The user corresponding to the largest target probability is used as another classification center.

2. The user power load classification method based on data feature analysis according to claim 1 is characterized in that: The process of obtaining the central constructivity includes: The correlation coefficient between the power load data of each dimension and the power load data of the current day is obtained, and the central constructivity of each dimension is obtained from the correlation coefficient.

3. The user power load classification method based on data feature analysis according to claim 1 is characterized in that: The process of obtaining the category centrality includes: Based on the electricity load data of the same dimension of each user, the local outlier factor LOF value of the electricity load data of each dimension of each user is obtained; The local outlier factor (LOF) value of the power load data of each dimension of each user is negatively correlated to obtain the category centrality of the power load data of each dimension of each user.

4. The user power load classification method based on data feature analysis according to claim 1 is characterized in that: The class center probabilities are calculated as follows: The centrality weight of each dimension is obtained according to the centrality constructivity of each dimension; the centrality weight is proportional to the centrality constructivity; According to the centrality weight of each dimension, the category centrality of the power load data of each dimension of the target user is weighted and summed to obtain the category centrality probability of the target user; the target user is any user.

5. The method for classifying user power loads based on data feature analysis according to claim 1, wherein: The target distance acquisition process includes: According to the centrality weight of each dimension, the distances between the power load data of each dimension of the reference user and the power load data of each dimension of each confirmed classification center are weighted and summed to obtain the target distance of the reference user; the reference user is any one of the remaining other users; the centrality weight of each dimension is obtained by the central construct of each dimension.

6. The method for classifying user power loads based on data feature analysis according to claim 1, wherein: The preset cutoff conditions include: Classify the electricity load data of all users according to the classification centers obtained in any two adjacent iterative processes, obtain two classification results, and obtain the classification effects of the two classification results; If the classification effects of the two classification results satisfy a preset classification effect degradation condition, the iteration is stopped, and the previous iteration process of any two adjacent iteration processes is used as the last iteration process.

7. The method for classifying user power loads based on data feature analysis according to claim 6, wherein: The classification effects of the two classification results satisfy a preset classification effect degradation condition, including: The difference between the classification effect corresponding to the latter iterative process and the classification effect corresponding to the previous iterative process in any two adjacent iterative processes is calculated, and the difference is less than a preset threshold.

8. The method for classifying user power loads based on data feature analysis according to claim 1, wherein: The classification of the electricity load data of all users according to the finally obtained classification centers includes: Calculate the distance between each user's electricity load data and each classification center; The electricity load data of each user is classified into the category corresponding to the classification center with the smallest distance.

Images