Power user clustering method and apparatus, computer device, and storage medium

Abstract

The application relates to a power user clustering method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring time series data of historical power consumption of a power user; adopting a minimum-maximum normalization mode to map the time series data of the historical power consumption to a preset value range; extracting historical power consumption features of the power user below a preset dimension from the normalized time series data of the historical power consumption; clustering by using a Gaussian mixture model according to the historical power consumption features of the power user below the preset dimension, and performing convergence processing on the clustering result; and classifying the power user according to the clustering result after the convergence processing. The method can take into account both high-dimensional data dimension reduction and time series characteristics, and improve the clustering accuracy of power user clustering.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction, in particular to a power user clustering method and device, computer equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] With the increasing proportion of new energy power generation, the power system gradually shifts from "source following load" to "source grid load storage" coordination and interaction, and the demand side resource becomes an important part of ensuring system balance. The power user load data has the characteristics of time series, high dimension, strong periodicity and being easily affected by external environment, etc. How to accurately depict the power consumption behavior of different users and implement fine classification has become a key link in demand response and power market design.

[0003] Traditional technology mainly relies on traditional clustering algorithms (such as K-means, fuzzy C-means, spectral clustering) or probabilistic models (Gaussian mixture model, hidden Markov model) to cluster user load curves. However, the traditional clustering algorithm has insufficient ability to distinguish small samples and complex boundary samples, and potential load patterns are easily missed; high-dimensional load data is prone to "dimension disaster", resulting in a sharp decline in clustering effect; the above algorithms cannot fully exploit the time dependence in the load curve, and the clustering result deviates from the actual power consumption behavior.

[0004] Therefore, there is an urgent need for a power user clustering method, device, computer equipment, computer readable storage medium and computer program product that can take into account the dimensionality reduction of high-dimensional data and the time series characteristics, and improve the clustering accuracy of power user clustering. SUMMARY

[0005] Therefore, there is an urgent need for a power user clustering method, device, computer equipment, computer readable storage medium and computer program product that can take into account the dimensionality reduction of high-dimensional data and the time series characteristics, and improve the clustering accuracy of power user clustering.

[0006] In a first aspect, the present application provides a power user clustering method, comprising:

[0007] obtaining time series data of historical power consumption of power users;

[0008] using the minimum-maximum normalization method to map the time series data of historical power consumption to a preset value range;

[0009] extracting the historical power consumption features of power users below the preset dimension from the normalized historical power consumption time series data;

[0010] According to the historical power consumption features of power users below the preset dimension, using a Gaussian mixture model for clustering, and performing convergence processing on the clustering result;

[0011] According to the clustering result after the convergence processing, the power users are classified.

[0012] In one of the embodiments, the historical power consumption features of the power users below the preset dimension are extracted from the normalized historical power consumption time series data, including:

[0013] A feature extraction model with an encoding-decoding symmetric structure is constructed.

[0014] According to the layer-by-layer dimension reduction method, the normalized historical power consumption time series data is compressed to obtain the historical power consumption features of the power users below the preset dimension.

[0015] Among them, the linear rectifier function is used as the activation function of the output layer of the encoding structure in the feature extraction model; and the Sigmoid activation function is used as the activation function of the output layer of the decoding structure in the feature extraction model.

[0016] In one of the embodiments, the clustering using the Gaussian mixture model includes:

[0017] According to the K-means preset clustering result, the mean vector, the covariance matrix and the mixing weight of the Gaussian mixture model are initialized.

[0018] Based on the Bayesian information criterion, the clustering number that makes the criterion value minimum is selected as the optimal clustering number within the preset candidate clustering number range.

[0019] After setting the Gaussian mixture model with the optimal clustering number, the probability of each data sample belonging to each Gaussian component is iteratively updated based on the expectation maximization algorithm, and the model parameters are synchronously updated until the log-likelihood function converges, and the final clustering result is output.

[0020] In one of the embodiments, after the historical power consumption features of the power users below the preset dimension are extracted, it further includes:

[0021] The feature extraction model is trained using the mean square error loss function and the Adam optimizer until the reconstruction error is below the preset threshold.

[0022] In the case where the real-time monitored reconstruction error exceeds the preset threshold again, the encoding-decoding symmetric structure of the feature extraction model is adaptively adjusted until the reconstruction error falls within the preset range again.

[0023] In one of the embodiments, after the final clustering result is output, it further includes:

[0024] All historical power consumption features below the preset dimension in each category are obtained.

[0025] According to the center vector and the covariance matrix of the historical power consumption characteristics in each category, Mahalanobis distance of each sample to the center of the Gaussian component to which the sample belongs is calculated;

[0026] Samples with Mahalanobis distance greater than a preset abnormal threshold are marked as boundary samples or abnormal users, and are excluded from the category to which the samples belong or are classified separately.

[0027] In one of the embodiments, after the power users are classified, the method further includes:

[0028] Obtaining a load curve template corresponding to each category;

[0029] Generating a typical daily load curve of each category of users according to the load curve template;

[0030] Using the typical daily load curve, calculating a load mean value and a peak-valley difference of each time period to form an adjustable potential index;

[0031] According to the adjustable potential index, formulating a differentiated demand response strategy for each category of users, and issuing the differentiated demand response strategy to a user terminal.

[0032] In a second aspect, the application further provides a power user clustering device, comprising:

[0033] A data acquisition module configured to acquire time series data of historical power consumption of power users;

[0034] A data processing module configured to map the time series data of historical power consumption to a preset value range by using a min-max normalization method;

[0035] The data processing module is further configured to extract historical power consumption characteristics of power users below a preset dimension from the normalized time series data of historical power consumption;

[0036] The data processing module is further configured to perform clustering by using a Gaussian mixture model according to the historical power consumption characteristics of power users below the preset dimension, and perform convergence processing on the clustering result;

[0037] A power user clustering module configured to classify power users according to the clustering result after the convergence processing.

[0038] In a third aspect, the application further provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0039] Acquiring time series data of historical power consumption of power users;

[0040] mapping the time series data of the historical power consumption to a preset value range by using a minimum-maximum normalization method;

[0041] extracting historical power consumption features of the power users below the preset dimension from the normalized historical power consumption time series data;

[0042] performing clustering by using a Gaussian mixture model according to the historical power consumption features of the power users below the preset dimension, and performing convergence processing on the clustering result;

[0043] classifying the power users according to the clustering result after the convergence processing.

[0044] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0045] obtaining time series data of historical power consumption of power users;

[0046] mapping the time series data of the historical power consumption to a preset value range by using a minimum-maximum normalization method;

[0047] extracting historical power consumption features of the power users below the preset dimension from the normalized historical power consumption time series data;

[0048] performing clustering by using a Gaussian mixture model according to the historical power consumption features of the power users below the preset dimension, and performing convergence processing on the clustering result;

[0049] classifying the power users according to the clustering result after the convergence processing.

[0050] In a fifth aspect, the present application further provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the following steps:

[0051] obtaining time series data of historical power consumption of power users;

[0052] mapping the time series data of the historical power consumption to a preset value range by using a minimum-maximum normalization method;

[0053] extracting historical power consumption features of the power users below the preset dimension from the normalized historical power consumption time series data;

[0054] performing clustering by using a Gaussian mixture model according to the historical power consumption features of the power users below the preset dimension, and performing convergence processing on the clustering result;

[0055] classifying the power users according to the clustering result after the convergence processing.

[0056] The power user clustering method, device, computer equipment, computer readable storage medium and computer program product can automatically mine potential power consumption patterns from high-dimensional, strong noise and strong time correlation power consumption sequences without relying on artificial feature engineering through the cascade processing of "normalization, dimension reduction and Gaussian mixture model clustering". The dimension is compressed to below a preset range, the calculation complexity and storage overhead caused by "dimension disaster" are significantly reduced, the Gaussian mixture model is used to depict user behavior differences, the identification accuracy of small samples and boundary samples is improved, the number of categories is automatically determined through the Bayesian information criterion to avoid overfitting or underfitting, and finally the clustering result can be directly mapped to an adjustable potential index to support differentiated demand response strategies and time-of-use pricing schemes, and balance power system supply and demand and improve user side flexibility. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other related drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0058] Figure 1 An application environment diagram of the power user clustering method in one embodiment;

[0059] Figure 2 A flowchart of the power user clustering method in one embodiment;

[0060] Figure 3 A flowchart of the power user clustering method in another embodiment;

[0061] Figure 4 A schematic diagram of the overall technical framework of another embodiment;

[0062] Figure 5 A schematic diagram of the encoding-decoding symmetric structure in another embodiment;

[0063] Figure 6 A block diagram of the power user clustering device in one embodiment;

[0064] Figure 7 An internal structure diagram of the computer equipment in one embodiment. DETAILED DESCRIPTION

[0065] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application.

[0066] It should be noted that the terms "first", "second", and the like used in the present application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "include" and "have" and any variations thereof used in the present application are intended to cover non-exclusive inclusion. The term "multiple" used in the present application refers to two or more. The term "and / or" used in the present application refers to one of the options or any combination of multiple options.

[0067] The power user clustering method provided by the embodiments of the present application can be applied to an application environment as shown in Figure 1 The terminal 102 communicates with the server 104 through a network. The data storage system can store data required to be processed by the server 104. The data storage system can be integrated on the server 104, or placed on a cloud or other network server.

[0068] The server 104 obtains the time series data of the historical power consumption of the power user through the terminal 102; adopts the minimum-maximum normalization method to map the time series data of the historical power consumption to a preset value range; extracts the historical power consumption features of the power user below the preset dimension from the normalized historical power consumption time series data; uses the Gaussian mixture model to cluster according to the historical power consumption features of the power user below the preset dimension, and performs convergence processing on the clustering result; and classifies the power user according to the clustering result after the convergence processing.

[0069] The terminal 102 can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle device, a projection device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The head-mounted device can be a virtual reality (VR) device, an augmented reality (AR) device, smart glasses, etc. The server 104 can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0070] In an exemplary embodiment, as shown in Figure 2 A power user clustering method is provided, and the method is applied toFigure 1 Taking server 104 as an example, the explanation includes the following steps S202 to S210. Wherein:

[0071] Step S202: Obtain time-series data of historical electricity consumption of power users.

[0072] Specifically, the acquired historical electricity consumption time series data of power users comes from the power grid metering system (such as smart meters or AMI platforms). The power or electricity consumption values ​​of users are continuously collected at fixed sampling intervals (every minute, 15 minutes, or every hour) to form load curves at multiple time scales such as daily (e.g., 1440 points), weekly, monthly, and yearly. This data has daily, weekly, and seasonal cycle characteristics, is highly continuous between adjacent moments, and is affected by external factors such as weather, temperature, holidays, and electricity pricing policies. It covers typical load patterns of different types of users such as industrial, commercial, residential, and agricultural users, and is exported through CSV, databases, or standard specifications as the original input for subsequent normalization, dimensionality reduction, and clustering.

[0073] Step S204: Using the minimum-maximum normalization method, the historical electricity consumption time series data is mapped to a preset value range.

[0074] Specifically, for each household, the minimum and maximum values ​​of electricity consumption time series data for each day (or each week) are first identified, and then the minimum-maximum normalization formula is used:

[0075] x norm =(xx min ) / (x max -x min This linearly stretches all original power or electricity values ​​to a pre-defined uniform range (e.g., [0,1] or [-1,1]). This eliminates deviations caused by differences in units and magnitudes between different users and on different dates, while preserving the shape information of the electricity consumption curve, providing comparable and stable input for subsequent feature extraction and clustering.

[0076] Step S206: Extract historical electricity consumption characteristics of electricity users with dimensions lower than the preset dimension from the normalized historical electricity consumption time series data.

[0077] Specifically, a deep neural network with an encoder-decoder symmetric structure is used as the feature extraction model, where:

[0078] Encoder: input layer 96 nodes -> hidden layer 512 -> 128 -> 32 nodes (linear rectified activation); Decoder: 32 -> 128 -> 512 -> 96 nodes (S-shaped activation); the mean square error is taken as the loss function, and the Adam optimizer is trained for 200 epochs, so that the reconstruction error is <0.02. After the training is completed, the 32-dimensional vector at the end of the encoder is intercepted as the "power user historical power consumption feature below the preset dimension".

[0079] In step S208, the Gaussian mixture model is used for clustering according to the historical power consumption feature of the power user below the preset dimension, and a convergence processing is performed on the clustering result.

[0080] Specifically, the 32-dimensional feature vector compressed to "below the preset dimension" is taken as input, and a complete Gaussian mixture model (GMM) clustering is performed according to the following process: set the candidate cluster number K = 2 ~ 10. For each K: initialize the mean μk, covariance Σk and mixing coefficient αk of the GMM with the pre-clustering result of K-means; run the EM algorithm: E step: calculate the posterior probability γik of each 32-dimensional feature vector belonging to each Gaussian component; M step: update μk, Σk and αk according to γik; record the log-likelihood L(K) under the current K.

[0081] Calculate the Bayesian information criterion BIC(K); select the Kopt corresponding to the minimum BIC value as the final cluster number. Fix Kopt, and execute EM again until the log-likelihood function converges (|ΔL| < ε), and output the final clustering result and the category label of each user. At this point, the "clustering result convergence processing" is completed, and can be directly used for subsequent user classification or demand response strategy formulation.

[0082] In step S210, the power users are classified according to the clustering result after the convergence processing.

[0083] Specifically, when the EM iteration of the Gaussian mixture model reaches the convergence condition (the log-likelihood function |ΔL| < ε), the model has calculated the final posterior probability distribution for each 32-dimensional feature vector. For each user, the posterior probability of its feature vector belonging to each Gaussian component is compared, and the component number corresponding to the maximum probability is assigned to the user, that is, a "hard classification" is completed; if needed, the probability vector can also be retained as a "soft classification" weight; finally, a "user-category" mapping table is generated, according to which all power users are divided into several homogeneous groups, which are directly called by subsequent differentiated electricity price, demand response or load prediction and other businesses.

[0084] In the power user clustering method, through the cascade processing of "normalization dimension reduction Gaussian mixture model clustering", the method automatically mines potential power consumption patterns from high-dimensional, strong noise and strong time-dependent power consumption sequences without relying on manual feature engineering; the dimension is compressed to below the preset range, which significantly reduces the computational complexity and storage overhead caused by "dimension disaster"; the Gaussian mixture model is used to describe the differences in user behavior, which not only improves the identification accuracy of small samples and boundary samples, but also automatically determines the number of categories through the Bayesian information criterion, avoiding overfitting or underfitting; and finally the clustering result can be directly mapped to the adjustable potential index, supporting differentiated demand response strategies and time-of-use pricing schemes, achieving power system supply-demand balance and user-side flexibility improvement.

[0085] In an exemplary embodiment, as shown in FIG. 5, the historical power consumption features of the power users below the preset dimension are extracted from the normalized historical power consumption time series data, including: Figure 3

[0086] Step S302, a feature extraction model with an encoding-decoding symmetric structure is constructed;

[0087] Step S304, the normalized historical power consumption time series data is compressed to obtain the historical power consumption features of the power users below the preset dimension according to the layer-by-layer dimension reduction manner;

[0088] Step S306, wherein the linear rectifier function is used as the activation function of the output layer of the encoding structure in the feature extraction model; and the Sigmoid activation function is used as the activation function of the output layer of the decoding structure in the feature extraction model.

[0089] Specifically, the feature extraction model includes an encoding-decoding symmetric structure (i.e., an encoder and a decoder); as shown in FIG. 5, Figure 4 Figure 4 FIG. 6 shows the overall technical framework diagram of another embodiment of the present application; for example, the input layer: 96 nodes (corresponding to 15minx24h daily load curve). The encoding path: 96→512→128→32 nodes, each layer followed by a linear rectifier (ReLU) activation function; the decoding path: 32→128→512→96 nodes, each layer followed by a ReLU, and the output layer uses a Sigmoid activation function to normalize the reconstruction value domain to [0,1].

[0090] As shown in FIG. 5, Figure 5 ​​The structural diagram of the encoding-decoding symmetric structure is shown, which is mirror-symmetric. Among them, the input data is 96-dimensional sample after min-max normalization. After the loss (waste) function: mean square error (MSE). The optimizer selects Adam algorithm, and dynamic adjustment of learning rate is adopted, considering the convergence speed and stability, and the initial learning rate is 1x10 - ³, and it is attenuated by 0.5 times every 50 epochs. The termination condition is that the verification set MSE is less than 0.02 or the training is 200 epochs.

[0091] The training target is to minimize the reconstruction error, and the mean square error MSE is used as the loss function, as follows:

[0092] After the training is completed, the decoding path is truncated, only the encoding part is reserved, and the output of the last layer of the encoder is used as the low-dimensional feature representation of the user. Compared with the traditional method, the DAE can capture complex time sequence patterns (such as power consumption peak and valley, periodic fluctuation), and for any normalized daily load curve, the forward inference obtains 32-dimensional hidden variables, that is, the historical power consumption characteristics of the power user below the preset dimension.

[0093] Among them, the output layer of the encoding structure: ReLU (linear rectifier function) introduces nonlinearity and maintains sparsity; the output layer of the decoding structure: Sigmoid compresses the reconstruction result to [0, 1], which is the same scale as the normalized input.

[0094] ;

[0095] Among them, is the input data, is the output data of the decoder, and the stepwise dimension reduction structure is used to compress the original high-dimensional power consumption data to a low-dimensional feature space below one-tenth of the original dimension, which significantly reduces the computational complexity; this process eliminates the dimensional difference through input data normalization, and cooperates with the encoder to enhance the extraction ability of key time sequence features (such as power consumption peak and valley, periodic fluctuation); at the same time, the decoder uses Sigmoid function to ensure the consistency of the reconstructed data with the original structure, and takes minimizing the reconstruction error as the training target to prevent information from being lost too much in the dimension reduction process.

[0096] In this embodiment, the encoding-decoding symmetric structure is used to compress the high-dimensional daily load sequence into low-dimensional hidden features, while keeping the curve shape, peak and valley and cycle information complete, and greatly reducing the data dimension and computational load; the activation function and symmetric design cooperate to improve the nonlinearity expression ability, so that the low-dimensional features have both discriminability and numerical stability; after the training is completed, only one forward inference is needed to output the features in real time, which supports the rapid clustering of a large number of users, and considers the accuracy, efficiency and scalability.

[0097] In one of the embodiments, clustering is performed using a Gaussian mixture model, comprising:

[0098] According to the K-means preset clustering result, the mean vector, covariance matrix and mixing weight of the Gaussian mixture model are initialized;

[0099] Based on the Bayesian information criterion, the clustering number that makes the criterion value minimum is selected as the optimal clustering number in the preset candidate clustering number range;

[0100] After setting the Gaussian mixture model with the optimal clustering number, the probability of each data sample belonging to each Gaussian component is iteratively updated based on the expectation maximization algorithm, and the model parameters are simultaneously updated until the log-likelihood function converges, and the final clustering result is output.

[0101] Specifically, the 32-dimensional feature vector set output by the deep autoencoder is denoted as X={x1,x2,…,x n}, x i ∈R³². Then, K-means is run for candidate clustering number K∈{2,…,10} to obtain K groups of initial centers {μ k 0}, covariance Σ k 0 and mixing weight α k 0 =N k / N, where N k is the number of samples of the kth class.

[0102] The Bayesian information criterion (BIC) is used to calculate each K:

[0103] ;

[0104] where, is the maximum likelihood number; is the maximum log-likelihood after EM convergence; d=K(32+32²+1)-1, which is the total number of model parameters; N is the total number of samples. The minimum value of BIC(K) is taken as the optimal clustering number.

[0105] The expectation maximization EM iteration is used to reinitialize the GMM parameters with the minimum value of BIC(K), and the E step (expectation step) is repeatedly executed to calculate the posterior probability of sample xi belonging to the kth Gaussian distribution:

[0106] ;

[0107] where, Σk is the probability density function of the multivariate Gaussian distribution; Σk is the N×N covariance matrix of the k-th Gaussian distribution; μk is the mean vector (cluster center) of the k-th Gaussian distribution; θ is the set of parameters of the current model; αk is the mixing coefficient of the k-th Gaussian distribution; xi is the i-th sample in the input data; K is the preset number of components (i.e., the number of clusters) in the Gaussian mixture model.

[0108] The M-step (maximization step) is to update the parameters to maximize the log-likelihood function:

[0109] ;

[0110] ;

[0111] ;

[0112] in, Let αk be the posterior probability of each sample belonging to the k-th Gaussian distribution; n be the total number of samples; the new weight αk is equal to the average probability of all samples belonging to the k-th component; the new mean μk is the weighted average of samples xi (with weight Ψik); the new covariance Σk is the weighted covariance matrix of the samples and the mean. Finally, the class label {c} of each sample is obtained. i} and membership probability, to complete user classification.

[0113] A dual verification mechanism ensures that the dimensionality reduction features are adapted to the GMM clustering requirements: on the one hand, a reconstruction error threshold is set to monitor feature quality; on the other hand, the number of clusters is dynamically optimized using BIC, and the optimal number of clusters is automatically determined based on the feature distribution to avoid deviations caused by subjective presets.

[0114] In this embodiment, the training time is significantly shortened by using a simplified encoder-decoder symmetric structure, while GMM accelerates convergence through K-means pre-initialization of parameters and finely handles the boundary sample attribution problem based on the posterior probability calculation of the probability model. This combination significantly reduces training time in scenarios with minute-level electricity consumption data while ensuring the accurate differentiation of complex electricity consumption patterns by the clustering results.

[0115] In one embodiment, after extracting the historical electricity consumption characteristics of electricity users below a preset dimension, the method further includes:

[0116] The feature extraction model is trained using the mean squared error loss function and the Adam optimizer until the reconstruction error is lower than a preset threshold.

[0117] If the reconstruction error detected in real time exceeds the preset threshold again, the encoding-decoding symmetric structure of the feature extraction model is adaptively adjusted until the reconstruction error falls back into the preset range.

[0118] Specifically, after obtaining the low-dimensional features, the system first trains the feature extraction model through an adaptive optimizer using mean square error as the measurement standard until the reconstruction error between the reconstruction result output by the model and the original data falls below the preset threshold. After the model is put into use, the difference is continuously monitored; once the deviation exceeds the preset threshold again, the system will automatically adjust the number of network layer nodes, learning rate or regularization strength inside the model and quickly retrain and verify until the reconstruction error is controlled within the safe range of the preset threshold. In this way, the model can continuously repair itself in long-term operation, always maintain low-dimensional features that are both compact and accurate, without the need for manual retraining, ensuring that the subsequent clustering and classification results are always reliable.

[0119] In this embodiment, through the closed-loop mechanism of "first training-then monitoring-then fine-tuning", it is ensured that the feature extraction model always maintains controlled error in long-term operation, both maintains high fidelity of low-dimensional features and avoids manual retraining, realizing model self-repair and continuous stable output.

[0120] In one of the embodiments, after outputting the final clustering result, it further includes:

[0121] Obtaining all historical power consumption features below the preset dimension in each category;

[0122] According to the center vector and covariance matrix of the historical power consumption features in each category, calculating the Mahalanobis distance of each sample to the center of the Gaussian component belonging to it;

[0123] Marking the samples with Mahalanobis distance greater than the preset abnormal threshold as boundary samples or abnormal users, and excluding them from the category or classifying them separately.

[0124] Specifically, after clustering, for each category, first find the center position and shape (represented by "center vector" and "covariance matrix") of all low-dimensional features in the category, and then calculate the Mahalanobis distance of each sample to the center of the category. This distance not only considers the distance between the sample and the center, but also takes into account the distribution shape within the category, so it can sensitively identify samples that deviate from the typical pattern of the category. If the distance exceeds the pre-set "abnormal threshold", it is considered that the user belongs to the boundary or is abnormal, and it can be excluded or classified separately.

[0125] In this embodiment, the Mahalanobis distance is used for secondary screening within the category to automatically find and strip boundary or abnormal users, significantly improving the clustering purity; at the same time, it avoids the interference of abnormal samples on the subsequent demand response strategy making, ensuring that the classification result is both accurate and robust.

[0126] In one of the embodiments, after classifying the power users, it further includes:

[0127] Obtaining the load curve template corresponding to each category;

[0128] According to the load curve template, a typical daily load curve of each type of user is generated;

[0129] Using the typical daily load curve, the load average and peak-valley difference of each period are calculated to form an adjustable potential index;

[0130] According to the adjustable potential index, a differentiated demand response strategy is formulated for each type of user, and the differentiated demand response strategy is issued to the user terminal.

[0131] Specifically, after clustering, all daily load curves of the same type of user are averaged at the corresponding time to obtain the "load curve template" of this type. A smooth typical daily load curve is generated based on the template. Then, the load average and peak-valley difference are calculated for each period on the typical curve to obtain the "adjustable potential index" of the user. According to the index, a differentiated demand response strategy is designed for different types of users (such as high-potential users participating in peak cutting first, and low-potential users being given basic compensation), and the strategy is directly issued to the user terminal through SMS, APP or smart meter.

[0132] In this embodiment, accurate load portraits of each type of user can be automatically generated without human experience; the adjustable potential index quantifies the user's ability to participate in demand response, enabling the power grid to accurately allocate incentives according to potential; the differentiated strategy improves the user response rate and overall load adjustment efficiency, reduces the system peak-valley difference and operating cost, and realizes the collaborative optimization of supply and demand.

[0133] It should be understood that although each step in the flowchart involved in the above embodiments is displayed in sequence according to the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in the above embodiments can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0134] Based on the same inventive concept, the embodiments of the present application also provide a power user clustering device for implementing the power user clustering method described above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more power user clustering device embodiments provided below can refer to the limitations of the power user clustering method described above, and will not be repeated here.

[0135] In one example embodiment, as shown in Figure 6 An electric power user clustering apparatus is provided, comprising: a data acquisition module 602 configured to acquire time series data of historical power consumption of electric power users;

[0136] a data processing module 604 configured to map the time series data of historical power consumption to a preset value range by using a min-max normalization method;

[0137] The data processing module 604 is further configured to extract historical power consumption features of electric power users below a preset dimension from the normalized time series data of historical power consumption.

[0138] The data processing module 604 is further configured to perform clustering by using a Gaussian mixture model according to the historical power consumption features of electric power users below the preset dimension, and perform convergence processing on the clustering result.

[0139] An electric power user clustering module 606 is configured to classify electric power users according to the clustering result after the convergence processing.

[0140] In one example embodiment, the data processing module 604 is further configured to construct a feature extraction model with a coding-decoding symmetric structure; and perform compression processing on the normalized time series data of historical power consumption to obtain the historical power consumption features of electric power users below the preset dimension by using a layer-by-layer dimension reduction method; wherein a linear rectifier function is used as an activation function of an output layer of the coding structure in the feature extraction model; and a Sigmoid activation function is used as an activation function of an output layer of the decoding structure in the feature extraction model.

[0141] In one example embodiment, the data processing module 604 is further configured to initialize a mean vector, a covariance matrix and a mixing weight of the Gaussian mixture model according to a K-means preset clustering result; select a clustering number that makes a criterion value minimum as an optimal clustering number within a preset candidate clustering number range based on a Bayesian information criterion; and after setting the Gaussian mixture model with the optimal clustering number, iteratively update probabilities of each data sample belonging to each Gaussian component and synchronously update model parameters based on an expectation maximization algorithm until a log-likelihood function converges, and output a final clustering result.

[0142] In one example embodiment, the data processing module 604 is further configured to train the feature extraction model by using a mean square error loss function and an Adam optimizer until a reconstruction error is below a preset threshold; and in a case where the reconstruction error monitored in real time exceeds the preset threshold again, adaptively adjust the coding-decoding symmetric structure of the feature extraction model until the reconstruction error falls into a preset range again.

[0143] In an example embodiment, the power user clustering module 606 is further configured to obtain all historical power consumption features below the preset dimension in each category; calculate Mahalanobis distance of each sample to the center of the Gaussian component to which the sample belongs according to the center vector and covariance matrix of the historical power consumption features in each category; mark the sample with Mahalanobis distance greater than a preset abnormal threshold as a boundary sample or an abnormal user, and remove it from the category to which it belongs or classify it separately.

[0144] In an example embodiment, the power user clustering module 606 is configured to obtain a load curve template corresponding to each category; generate a typical daily load curve of each category of users according to the load curve template; calculate load mean and peak-valley difference of each period using the typical daily load curve to form an adjustable potential index; formulate a differentiated demand response strategy for each category of users according to the adjustable potential index, and issue the differentiated demand response strategy to the user terminal.

[0145] Each module in the power user clustering device described above can be realized by software, hardware, and combinations thereof, in whole or in part. Each module described above can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a memory in a computer device in software form, so as to be called and executed by a processor to perform the operations corresponding to each module.

[0146] In an example embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 7 The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store time series data of historical power consumption of power users. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement a power user clustering method.

[0147] Those skilled in the art can understand that Figure 7The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0148] In one exemplary embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the following steps when executing the computer program:

[0149] obtaining time series data of historical power consumption of a power user;

[0150] mapping the time series data of historical power consumption to a preset value range by using a min-max normalization method;

[0151] extracting historical power consumption features of the power user below a preset dimension from the normalized time series data of historical power consumption;

[0152] performing clustering by using a Gaussian mixture model according to the historical power consumption features of the power user below the preset dimension, and performing convergence processing on the clustering result;

[0153] classifying the power user according to the clustering result after the convergence processing.

[0154] In one embodiment, the processor further implements the following steps when executing the computer program:

[0155] constructing a feature extraction model with an encoding-decoding symmetric structure;

[0156] performing compression processing on the normalized time series data of historical power consumption according to a layer-by-layer dimension reduction method to obtain the historical power consumption features of the power user below the preset dimension;

[0157] wherein a linear rectifier function is used as an activation function of an output layer of the encoding structure in the feature extraction model, and a Sigmoid activation function is used as an activation function of an output layer of the decoding structure in the feature extraction model.

[0158] In one embodiment, the processor further implements the following steps when executing the computer program:

[0159] initializing a mean vector, a covariance matrix and a mixing weight of the Gaussian mixture model according to a K-means preset clustering result;

[0160] selecting, based on a Bayesian information criterion, a clustering number that makes a criterion value minimum as an optimal clustering number within a preset candidate clustering number range;

[0161] After setting the Gaussian mixture model with the optimal cluster number, the probability of each data sample belonging to each Gaussian component is iteratively updated based on the expectation maximization algorithm, and the model parameters are simultaneously updated until the log-likelihood function converges, and the final clustering result is output.

[0162] In one embodiment, the processor, when executing the computer program, also implements the following steps:

[0163] The feature extraction model is trained using a mean square error loss function and an Adam optimizer until the reconstruction error is below a preset threshold.

[0164] In the case where the real-time monitored reconstruction error again exceeds the preset threshold, the encoding-decoding symmetric structure of the feature extraction model is adaptively adjusted until the reconstruction error falls within the preset range again.

[0165] In one embodiment, the processor, when executing the computer program, also implements the following steps:

[0166] Obtain all historical power consumption features below the preset dimension in each category;

[0167] According to the center vector and covariance matrix of the historical power consumption features in each category, the Mahalanobis distance of each sample to the center of the Gaussian component to which it belongs is calculated.

[0168] Samples with a Mahalanobis distance greater than a preset anomaly threshold are marked as boundary samples or abnormal users, and are excluded from the category to which they belong or classified separately.

[0169] In one embodiment, the processor, when executing the computer program, also implements the following steps:

[0170] Obtain the load curve template corresponding to each category;

[0171] According to the load curve template, a typical daily load curve for each type of user is generated.

[0172] Using the typical daily load curve, the load mean and peak-valley difference of each period are calculated to form an adjustable potential index.

[0173] According to the adjustable potential index, a differentiated demand response strategy is formulated for each category of users, and the differentiated demand response strategy is issued to the user terminal.

[0174] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program, when executed by a processor, implements the following steps:

[0175] Obtain time series data of historical power consumption of power users;

[0176] The time series data of the historical power consumption is mapped to a preset value range by using a minimum-maximum normalization method.

[0177] From the normalized historical power consumption time series data, the historical power consumption features of the power users below the preset dimension are extracted.

[0178] According to the historical power consumption features of the power users below the preset dimension, clustering is performed using a Gaussian mixture model, and the clustering results are processed for convergence.

[0179] According to the clustering results after the convergence processing, the power users are classified.

[0180] In one embodiment, the computer program is executed by the processor to further implement the following steps:

[0181] A feature extraction model with an encoding-decoding symmetric structure is constructed.

[0182] According to a layer-by-layer dimension reduction method, the normalized historical power consumption time series data is compressed to obtain the historical power consumption features of the power users below the preset dimension.

[0183] The linear rectifier function is used as the activation function of the output layer of the encoding structure in the feature extraction model, and the Sigmoid activation function is used as the activation function of the output layer of the decoding structure in the feature extraction model.

[0184] In one embodiment, the computer program is executed by the processor to further implement the following steps:

[0185] According to the K-means preset clustering results, the mean vector, covariance matrix and mixing weight of the Gaussian mixture model are initialized.

[0186] Based on the Bayesian information criterion, the clustering number that minimizes the criterion value is selected as the optimal clustering number within the preset candidate clustering number range.

[0187] After setting the Gaussian mixture model with the optimal clustering number, the probability of each data sample belonging to each Gaussian component is iteratively updated based on the expectation maximization algorithm, and the model parameters are simultaneously updated until the log-likelihood function converges, and the final clustering results are output.

[0188] In one embodiment, the computer program is executed by the processor to further implement the following steps:

[0189] The feature extraction model is trained using the mean square error loss function and the Adam optimizer until the reconstruction error is below the preset threshold.

[0190] In the case that the reconstructed error monitored in real time exceeds the preset threshold again, the encoding-decoding symmetric structure of the feature extraction model is adaptively adjusted until the reconstructed error falls within the preset range again.

[0191] In one embodiment, the computer program, when executed by the processor, further implements the following steps:

[0192] Obtaining all historical power consumption features of each category below the preset dimension;

[0193] According to the center vector and the covariance matrix of the historical power consumption features in each category, the Mahalanobis distance of each sample to the center of the Gaussian component is calculated.

[0194] Samples with a Mahalanobis distance greater than a preset abnormal threshold are marked as boundary samples or abnormal users, and are excluded from the category or classified separately.

[0195] In one embodiment, the computer program, when executed by the processor, further implements the following steps:

[0196] Obtaining the load curve template corresponding to each category;

[0197] According to the load curve template, a typical daily load curve of each type of user is generated;

[0198] Using the typical daily load curve, the load average and the peak-valley difference of each period are calculated to form an adjustable potential index;

[0199] According to the adjustable potential index, a differentiated demand response strategy is formulated for each category of users, and the differentiated demand response strategy is issued to the user terminal.

[0200] In one embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the following steps:

[0201] Obtaining time series data of historical power consumption of power users;

[0202] Using the minimum-maximum normalization method, the time series data of historical power consumption is mapped to a preset value range;

[0203] From the normalized historical power consumption time series data, historical power consumption features of power users below the preset dimension are extracted;

[0204] According to the historical power consumption features of power users below the preset dimension, a Gaussian mixture model is used for clustering, and the clustering result is processed for convergence;

[0205] According to the clustering result after convergence processing, the power users are classified.

[0206] In one embodiment, the computer program, when executed by the processor, further implements the following steps:

[0207] The feature extraction model with the encoding-decoding symmetric structure is constructed.

[0208] According to the layer-by-layer dimension reduction manner, the normalized historical power consumption time series data is compressed to obtain the historical power consumption features of the power user below the preset dimension.

[0209] The linear rectifier function is used as the activation function of the output layer of the encoding structure in the feature extraction model, and the Sigmoid activation function is used as the activation function of the output layer of the decoding structure in the feature extraction model.

[0210] In one embodiment, the computer program, when executed by the processor, further implements the following steps:

[0211] According to the K-means preset clustering result, the mean vector, the covariance matrix and the mixing weight of the Gaussian mixture model are initialized.

[0212] Based on the Bayesian information criterion, the clustering number that makes the criterion value minimum is selected as the optimal clustering number in the preset candidate clustering number range.

[0213] After the Gaussian mixture model is set with the optimal clustering number, the probability that each data sample belongs to each Gaussian component is iteratively updated based on the expectation maximization algorithm, and the model parameters are synchronously updated until the log-likelihood function converges, and the final clustering result is output.

[0214] In one embodiment, the computer program, when executed by the processor, further implements the following steps:

[0215] The feature extraction model is trained using the mean square error loss function and the Adam optimizer until the reconstruction error is lower than the preset threshold.

[0216] In the case that the real-time monitored reconstruction error again exceeds the preset threshold, the encoding-decoding symmetric structure of the feature extraction model is adaptively adjusted until the reconstruction error falls into the preset range again.

[0217] In one embodiment, the computer program, when executed by the processor, further implements the following steps:

[0218] All historical power consumption features below the preset dimension in each category are obtained.

[0219] According to the center vector and the covariance matrix of the historical power consumption features in each category, the Mahalanobis distance of each sample to the center of the Gaussian component to which it belongs is calculated.

[0220] Samples with a Mahalanobis distance greater than a preset abnormal threshold are marked as boundary samples or abnormal users, and are excluded from the category to which they belong or are classified separately.

[0221] In one embodiment, the computer program, when executed by the processor, also implements the following steps:

[0222] Obtain the load curve template corresponding to each category;

[0223] According to the load curve template, generate a typical daily load curve for each type of user;

[0224] Using the typical daily load curve, calculate the load mean and peak-valley difference of each period to form an adjustable potential index;

[0225] According to the adjustable potential index, develop a differentiated demand response strategy for each category of user, and issue the differentiated demand response strategy to the user terminal.

[0226] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0227] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. The non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. The volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., without being limited thereto.

[0228] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0229] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for clustering electricity users, characterized in that, The method includes: Obtain historical time-series data on electricity consumption by power users; The time series data of historical electricity consumption is mapped to a preset value range using a minimum-maximum normalization method. From the normalized historical electricity consumption time series data, historical electricity consumption features of electricity users with a lower than preset dimension are extracted, including: constructing a feature extraction model with an encoding-decoding symmetric structure; compressing the normalized historical electricity consumption time series data according to a layer-by-layer dimensionality reduction method to obtain historical electricity consumption features of electricity users with a lower than preset dimension; wherein, the linear rectifier function is used as the activation function of the output layer of the encoding structure in the feature extraction model; and the sigmoid activation function is used as the activation function of the output layer of the decoding structure in the feature extraction model. The feature extraction model of the encoder-decoder symmetric structure is mirror-symmetric, with the goal of minimizing reconstruction error during training. Mean Squared Error (MSE) is used as the loss function, and the Adam algorithm is selected as the optimizer, with dynamic adjustment of the learning rate. After training, the decoding path is truncated, retaining only the encoding part. The output of the last layer of the encoder serves as the user's low-dimensional feature representation. The output layer of the encoder uses the ReLU activation function to introduce non-linearity while maintaining feature sparsity. The output layer of the decoder uses the Sigmoid function to compress the reconstructed data to a scale consistent with the normalized input data. Based on the historical electricity consumption characteristics of electricity users below the preset dimension, clustering is performed using a Gaussian mixture model, including: pre-setting clustering results based on K-means, initializing the mean vector, covariance matrix, and mixture weights of the Gaussian mixture model; selecting the number of clusters that minimizes the criterion value within a preset range of candidate cluster numbers as the optimal number of clusters based on the Bayesian information criterion; setting the Gaussian mixture model with the optimal number of clusters, iteratively updating the probability of each data sample belonging to each Gaussian component based on the expectation-maximization algorithm, and synchronously updating the model parameters until the log-likelihood function converges, outputting the final clustering result, and performing convergence processing on the clustering result; Electricity users are classified based on the clustering results after convergence processing.

2. The method according to claim 1, characterized in that, After extracting the historical electricity consumption characteristics of electricity users below the preset dimension, the process also includes: The feature extraction model is trained using the mean squared error loss function and the Adam optimizer until the reconstruction error is lower than a preset threshold. If the reconstruction error detected in real time exceeds the preset threshold again, the encoding-decoding symmetric structure of the feature extraction model is adaptively adjusted until the reconstruction error falls back into the preset range.

3. The method according to claim 1, characterized in that, After outputting the final clustering result, the following is also included: Obtain all historical electricity consumption characteristics below the preset dimension within each category; Based on the center vector and covariance matrix of historical electricity consumption characteristics within each category, calculate the Mahalanobis distance from each sample to the center of its Gaussian component. Samples with a Mahalanobis distance greater than a preset anomaly threshold are marked as boundary samples or abnormal users, and are removed from their respective categories or classified separately.

4. The method according to claim 1, characterized in that, After classifying electricity users, the process also includes: Obtain the load curve templates corresponding to each category; Based on the load curve template, generate a typical daily load curve for each type of user; Using the typical daily load curve, the average load and peak-valley difference for each time period are calculated to form an adjustable potential index. Based on the adjustable potential index, differentiated demand response strategies are formulated for each category of users, and the differentiated demand response strategies are distributed to user terminals.

5. A power user clustering device, characterized in that, The device includes: The data acquisition module is used to acquire historical time-series data of electricity consumption by power users; The data processing module is used to map the historical electricity consumption time series data to a preset value range using a minimum-maximum normalization method; The data processing module is also used to extract historical electricity consumption features of electricity users with a lower than preset dimension from the normalized historical electricity consumption time series data, including: constructing a feature extraction model with an encoding-decoding symmetric structure; compressing the normalized historical electricity consumption time series data according to a layer-by-layer dimensionality reduction method to obtain historical electricity consumption features of electricity users with a lower than preset dimension; wherein, the linear rectifier function is used as the activation function of the output layer of the encoding structure in the feature extraction model; and the sigmoid activation function is used as the activation function of the output layer of the decoding structure in the feature extraction model. The feature extraction model of the encoder-decoder symmetric structure is mirror-symmetric, with the goal of minimizing reconstruction error during training. Mean Squared Error (MSE) is used as the loss function, and the Adam algorithm is selected as the optimizer, with dynamic adjustment of the learning rate. After training, the decoding path is truncated, retaining only the encoding part. The output of the last layer of the encoder serves as the user's low-dimensional feature representation. The output layer of the encoder uses the ReLU activation function to introduce non-linearity while maintaining feature sparsity. The output layer of the decoder uses the Sigmoid function to compress the reconstructed data to a scale consistent with the normalized input data. The data processing module is also used to perform clustering using a Gaussian mixture model based on the historical electricity consumption characteristics of the electricity users below the preset dimension. This includes: setting the clustering results based on the K-means, initializing the mean vector, covariance matrix, and mixture weights of the Gaussian mixture model; selecting the number of clusters that minimizes the criterion value within a preset range of candidate cluster numbers based on the Bayesian information criterion as the optimal number of clusters; setting the Gaussian mixture model with the optimal number of clusters, iteratively updating the probability of each data sample belonging to each Gaussian component based on the expectation-maximization algorithm, and synchronously updating the model parameters until the log-likelihood function converges, outputting the final clustering result, and performing convergence processing on the clustering result. The power user clustering module is used to classify power users based on the clustering results after convergence processing.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images