Load feature optimization clustering method based on one-dimensional convolution auto-encoder

Abstract

A load feature optimization clustering method based on a one-dimensional convolution auto-encoder comprises the following steps: acquiring load types including daily load curve data as a data set, and performing preprocessing; the method comprises the following steps: constructing a one-dimensional convolution auto-encoder model which comprises an encoder and a decoder, pre-training load data by taking minimization of reconstruction loss as a target, extracting load features, and storing model parameters obtained by pre-training at the same time; the method comprises the following steps: firstly, decoding a network part, removing a decoding network part, retaining an encoding part of feature extraction, constructing a PCA-Kmeans space conversion model, transmitting training data in an encoder, and converting a potential space into a low-dimensional space by using a PCA algorithm to carry out K-means clustering analysis; the encoder is finely adjusted, the encoder is trained by taking minimization of a clustering loss function as a target, and after the encoder is trained for one epoch, K-means clustering is carried out on a newly generated potential space; new clustering distribution is obtained, and power load mode extraction is achieved. Compared with other traditional clustering methods, the method has the advantages that the application complexity is simplified, and the classification efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of power load clustering analysis, and specifically to a load feature optimization clustering method based on a one-dimensional convolutional autoencoder. Background Technology

[0002] Big data storage and analysis technologies have flourished alongside advancements in information science and technology. Big data analytics has gradually become a crucial tool for mining and analyzing data across various industries. One application of big data is using data analysis to uncover its characteristics and derive conclusions that provide guidance for production. Power big data technology, by integrating massive, multi-dimensional, and real-time power-related data, provides intelligent support for power system planning, operation, optimization, and user services, ensuring the safe operation of the power grid and providing users with high-quality and economical electricity.

[0003] With the optimization and adjustment of industrial structure, the fluctuations in electricity demand are gradually intensifying. However, the existing power supply system is struggling to cope with the ever-increasing demand. User load data is characterized by massive volume and low value density, and direct processing and analysis will face limitations in computational efficiency. Cluster analysis of electricity load is the foundation of electricity data mining and has been widely used in the field of electricity user pattern classification. Identifying the load characteristics of typical users can provide a basis for the aggregation and regulation of large-scale distributed resources, effectively improving the level of orderly electricity consumption management. Developing differentiated power supply strategies based on the load characteristics of each type of user provides direction for strategies such as peak-shifting management and orderly electricity consumption, rationally allocating power resources and reducing electricity waste.

[0004] Common clustering methods can perform well when the data dimension or data feature dimension is low, but the accuracy becomes unsatisfactory when the data dimension is high; moreover, a single typical curve does not fully consider the differences in user load characteristics and cannot meet the needs of load uncertainty analysis. Summary of the Invention

[0005] To address the problem that a single typical curve does not fully consider the differences in user load characteristics and cannot meet the needs of load uncertainty analysis, this invention proposes a load feature optimization clustering method based on a one-dimensional convolutional autoencoder. This method enhances the richness of the inherent information in the cluster structure, accurately classifies multiple users based on time-series electricity consumption characteristics, and dynamically expresses load characteristics using an autoencoder, thereby better describing load uncertainty. Compared to other traditional clustering methods, this method simplifies application complexity and improves classification efficiency.

[0006] The technical solution adopted in this invention is as follows: The load feature optimization clustering method based on one-dimensional convolutional autoencoders includes the following steps: Step 1: Collect daily load curve data for industrial, commercial, and residential load types as a dataset, and preprocess the collected data; Step 2: Construct a one-dimensional convolutional autoencoder model, which includes an encoder and a decoder, and aims to minimize the reconstruction loss. To achieve the goal, pre-train the model on the annual industry load data, extract load features, and save the model parameters obtained from the pre-training. Step 3: Remove the decoding network part, retain the encoding part of feature extraction, construct the PCA-Kmeans spatial transformation model, pass the training data in the encoder, and use the PCA algorithm to transform the latent space into a more refined low-dimensional space for K-means clustering analysis. Step 4: Fine-tune the encoder to minimize the clustering loss function. The encoder is trained for the target, and K-means clustering is performed on the newly generated latent space after training the encoder for one epoch. Step 5: Repeat steps 3 and 4 to obtain new cluster assignments and realize power load pattern extraction.

[0007] In step 1, the collected data is preprocessed, and the k-nearest neighbor algorithm is used to complete the missing values. By finding the k nearest neighbor samples that are most similar to the missing value sample, the missing value is estimated using the corresponding feature values ​​of these nearest neighbor samples.

[0008] Calculate the target value The corresponding sample and other samples The Euclidean distance d between them: (1); In equation (1), Indicates the number of sampling points for the load curve; This represents the index for each sampling point; Sort the calculated Euclidean distances d from smallest to largest, and select the first k data points, which are the target values. The k-nearest neighbor data is used to calculate the replacement value for the missing value. : (2); In equation (2), Let be the weights of the k-nearest neighbors, and be the reciprocal of the Euclidean distance d; This indicates the number of nearest neighbor samples used to fill in missing values.

[0009] Finally, the maximum-minimum normalization method was used to narrow the range of values ​​of the collected data to between 0 and 1.

[0010] (3); In equation (3), x These are the original load data values; x min It is the minimum value among all samples in the dataset; x max The maximum value among all samples in the dataset; x norm This is the normalized value.

[0011] In step 2, the encoder includes convolutional layers, pooling layers, flattening layers, and embedding layers. The convolutional layers extract local features from the input load data through convolution operations; the one-dimensional convolutional kernel slides across the time series to capture load variation patterns at different time scales. The pooling layers downsample the feature maps output by the convolutional layers, reducing data dimensionality while preserving important features. The flattening layers flatten the multi-dimensional feature maps obtained after three convolutional layers into a fixed-length one-dimensional feature vector, facilitating subsequent clustering analysis. The embedding layers map the flattened feature vector to a low-dimensional latent space, which contains the main features of the input data and forms the basis for the decoder's data reconstruction. The decoder is a mirror image of the encoder structure; it gradually upsamples through deconvolutional layers to restore data dimensionality and ultimately outputs the reconstructed load sequence.

[0012] In step 2, the input data is mapped to a low-dimensional latent space using a one-dimensional convolutional autoencoder model, reducing the dimensionality of the original dataset and efficiently extracting the time-series characteristics of the data. (4); (5); In the above formula: The encoder's activation function is ReLU. The mapping function for the decoder uses the sigmoid function; This represents a one-dimensional convolution operation; and These are the weight matrices for the encoder and decoder, respectively; and These are the bias terms for the encoder and decoder, respectively; and These are latent spatial features and reconstructed data, respectively.

[0013] The reconstruction loss is represented by mean squared error, and a one-dimensional convolutional autoencoder model is pre-trained with the goal of minimizing the reconstruction loss function. (6); In equation (6), The reconstruction loss represents the data reconstructed by the decoder. Compared with the original input data The differences between them; This represents a sample of the original load curve; Indicates training batch; This represents the load value of the original input load curve in the i-th hour. This represents the load value of the load curve reconstructed by the decoder in the i-th hour; This represents the Euclidean length of the vector.

[0014] In step 3, the pre-trained complete autoencoder is disassembled, the decoding network is removed, and only the encoder used for feature compression is retained. Based on the latent features extracted by the encoder, a PCA-Kmeans spatial transformation model is constructed to further optimize the feature space; In this process, the PCA stage performs secondary dimensionality reduction and decorrelation on the latent features; the K-means stage performs clustering in the optimized feature space. The PCA algorithm yields a new, dimensionality-reduced, and decorrelated feature set, which is then used for subsequent K-means clustering, thereby optimizing the clustering results.

[0015] (7); In equation (7), X is the standardized latent feature matrix; the covariance matrix Σ describes the distribution shape of the data in the feature space.

[0016] The eigenvalues ​​are obtained by calculating the standardized load covariance matrix and performing eigenvalue decomposition. and eigenvector matrix .

[0017] (8); (9); In the above formula: This represents the proportion of the k-th principal component to the total variance; This represents the cumulative contribution rate of the first k principal components, which can help determine the number of principal components to retain. Indicates the first k The variance of each principal component; Indicates the original feature dimension; It represents the first element of the data covariance matrix. i Large eigenvalues; This indicates the number of principal components selected.

[0018] When the cumulative variance contribution rate of the current k principal components is greater than 90%, the corresponding value of k is the number of principal components. This k-dimensional orthogonal eigenvector is a feature reconstructed from the original n-dimensional features as principal components. Here, the k-dimensional orthogonal eigenvector represents the projection matrix formed by the eigenvectors corresponding to the first k largest eigenvalues; Constructing clustering loss using the K-means method: (10); (11); In the above formula: This represents the original clustering loss, which measures the sum of squared distances from all samples to their cluster centers. Represents a single data point in the latent space; Indicates distance from cluster center More recent data points; This represents the set of data points representing the clustering result of the i-th cluster; This represents the current cluster center; k represents the total number of clusters. Clustering discrete matrices are used to calculate principal components, creating a new projection space. Specifically, it includes: The clustering discrete matrix S is calculated based on the initial clustering results. ω After eigenvalue decomposition, the eigenvectors corresponding to the k largest eigenvalues ​​are selected to construct an orthogonal projection matrix V. Then, the original data is projected onto the latent space spanned by V to obtain a new feature representation Y. By creating a feature space that retains important information and is best suited for cluster analysis, the accuracy and robustness of load pattern classification are significantly improved.

[0019] (12); (13); In the above formula: It represents the clustering discrete matrix, which measures the degree of dispersion of samples within a cluster; Indicates the total number of clusters; This represents the matrix transpose operation; This represents the clustering loss after projection onto the new space; The orthogonal projection matrix is ​​represented by the clustering discrete matrix. It consists of the first k eigenvectors.

[0020] Wherein, orthogonal matrix It is a matrix Arrange the eigenvalues ​​in descending order and select the first k eigenvectors to form an eigenma matrix; orthogonal matrix. VThe selected feature vector direction is the direction in which clustering information is most concentrated, which can better capture the electricity consumption pattern characteristics of different user types, thus making the projected data easier to cluster.

[0021] Will Mapping to the orthogonal matrix V, the clustering structure information of the n subsets with relatively low concentration at the tail end is used. By utilizing the clustering loss constructed from the tail feature vectors, the encoder is forced to consider more diverse clustering information during feature extraction, achieving good clustering performance even in directions with relatively scarce information, and optimizing spatial distribution. This optimizes the representation of clustering structure information and expands the extraction range of clustering structure information. (14); In equation (14), This represents the objective function of the fine-tuning encoder; Represents the projection matrix of the tail feature subspace; Represents the feature vectors in the latent space; Indicates training batch; Indicates the current sample; This indicates the number of clusters in the cluster.

[0022] In step 4, the structure and distribution of the feature space are adjusted to better suit the specific data and task requirements, thereby minimizing the clustering loss. The encoder is fine-tuned for the target. In each training epoch, training data is input into the encoder to obtain latent features; K-means clustering is performed on the current latent feature space to obtain cluster assignments and cluster centers. The loss function value is calculated based on the cluster assignments and cluster centers. The optimizer updates the encoder parameters.

[0023] After training the encoder for one epoch, features are extracted using the updated encoder, and K-means clustering is performed again in the new feature space to obtain new cluster assignments.

[0024] In step 4, the clustering loss function is optimized. Iteratively fine-tune the encoder network, letting the cluster assignment label in the initial iteration be U, and the cluster label in the new iteration be P. Compare the proportion η of the number of data points with different assignments in U and P.

[0025] (15); In equation (15), N represents the total number of user load curves; U represents the cluster assignment label vector of the previous iteration; and P represents the cluster assignment label vector of the current iteration. This indicates an indicator function that returns 1 if the condition is true, and 0 otherwise.

[0026] The threshold is set to 0.001. When the proportion is less than the set threshold, the degree of change in cluster allocation is considered to be low, the method converges, the iteration stops and the clustering results are output.

[0027] This invention discloses a load feature optimization clustering method based on a one-dimensional convolutional autoencoder, with the following technical effects: 1) This method first analyzes the load characteristics of a single user and extracts the typical load curve shape of that user to help to more accurately grasp the user's electricity consumption characteristics.

[0028] 2) In order to meet the requirements of load uncertainty analysis, this invention adopts the proposed one-dimensional convolutional autoencoder load feature optimization clustering algorithm. By optimizing the objective functions of autoencoder reconstruction loss and clustering loss one by one, the clustering performance is improved, effectively making up for the gap between clustering quality and reconstruction efficiency.

[0029] 2) This invention utilizes time-series data to mine the load characteristics of different types of users, which helps to enhance the richness of the inherent information in the cluster structure. It accurately classifies multiple users based on time-series electricity consumption characteristics and uses an autoencoder to dynamically express load characteristics, thereby better describing the uncertainty of the load. Compared to other methods using traditional clustering, it simplifies application complexity and improves classification efficiency. Attached Figure Description

[0030] The present invention will be further described below with reference to the accompanying drawings and examples; Figure 1 The flowchart shows the clustering method for optimizing the load features of a one-dimensional convolutional autoencoder.

[0031] Figure 2 This is a schematic diagram of a one-dimensional convolutional autoencoder model.

[0032] Figure 3 Flowchart of user load curve analysis method.

[0033] Figure 4 This is a typical load curve result for a single user.

[0034] Figure 5(a) shows the cluster analysis results of typical load curves for multiple users (first type of users); Figure 5(b) shows the cluster analysis results of typical load curves for multiple users (second category of users); Figure 5(c) shows the cluster analysis results of typical load curves for multiple users (third type of users). Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0036] Figure 1The flowchart for the clustering method for optimizing the load features of a one-dimensional convolutional autoencoder is as follows: Step 1: Determine the research data for typical load curve identification. Daily load curve data from actual substation measurements, including industrial, commercial, and residential loads, are used. Load data from 210 urban industrial and commercial users in a certain regional power grid over a period of 365 days are selected. Each user's daily load curve contains 24 sampling points, i.e., sampling once every 1 hour. This data is used as the dataset, and the collected data is preprocessed. The collected data was preprocessed, and the k-nearest neighbor algorithm was used to complete missing values. The target value was then calculated. The corresponding sample and other samples The Euclidean distance d between them: (1); Sort the calculated Euclidean distances d from smallest to largest, and select the first k data points, which are the target values. The k-nearest neighbor data is used to calculate the replacement value for the missing value. : (2); In equation (2), Let be the weights of the k-nearest neighbors, and be the reciprocal of the Euclidean distance d.

[0037] Finally, the maximum-minimum normalization method was used to narrow the range of values ​​in the collected data to between 0 and 1.

[0038] Step 2: Construct a one-dimensional convolutional autoencoder model, which consists of an encoder and a decoder. The encoder includes convolutional layers, pooling layers, flattening layers, and embedding layers. The decoder is a mirror image of the encoder structure, using the Adam optimizer. The learning rate is set to 0.001 to ensure stable convergence during training. The batch size is set to 64, and the number of iterations is 200. The model minimizes the reconstruction loss. The model is trained on annual industry load data to extract load features and save the pre-trained model parameters.

[0039] By using a one-dimensional convolutional autoencoder model to map input data into a low-dimensional latent space, the dimensionality of the original dataset is reduced, and the time-series characteristics of the data are extracted efficiently. (4); (5); In the formula: The encoder's activation function is ReLU. The mapping function for the decoder uses the sigmoid function; This represents a one-dimensional convolution operation; and These are the weight matrices for the encoder and decoder, respectively; and These are the bias terms for the encoder and decoder, respectively; and These are latent spatial features and reconstructed data, respectively.

[0040] The reconstruction loss is represented by mean squared error, and a one-dimensional convolutional autoencoder model is pre-trained with the objective of minimizing the reconstruction loss function. (6); Step 3: Remove the decoding network portion, retaining only the feature extraction encoding portion. Construct a PCA-Kmeans spatial transformation model, passing the training data through the encoder, and using the PCA algorithm to transform the latent space into a more refined low-dimensional space for K-means clustering analysis.

[0041] The decoding network is removed, retaining only the feature extraction encoding part, and a PCA-Kmeans spatial transformation model is constructed. The standardized load covariance matrix is ​​calculated and eigenvalue decomposition is performed to obtain the eigenvalues. and eigenvector matrix Arrange the eigenvalues ​​in descending order and calculate the variance contribution rate and cumulative variance contribution rate.

[0042] (8); (9); In the formula: This represents the proportion of the k-th principal component to the total variance; The cumulative contribution rate of the first k principal components helps determine the number of principal components to retain. When the cumulative variance contribution rate of the first k principal components is greater than 90%, the corresponding value of k is the number of principal components. These k-dimensional orthogonal eigenvectors are features reconstructed from the original n-dimensional features, and they serve as principal components.

[0043] Constructing clustering loss using the K-means method: (10); (11); in, represents a single data point in the latent space, and k represents the total number of clusters. It refers to the set of data points resulting from the clustering of the i-th cluster. Indicates distance from cluster center More recent data points.

[0044] Clustering discrete matrices are used to calculate principal components, creating a new projection space. .

[0045] (12); (13); Where: orthogonal matrix It is a matrix Sort the eigenvalues ​​in descending order and select the eigenma matrix composed of the first k eigenvectors; Mapped to the n subsets of the orthogonal matrix V with relatively low concentration of clustering information at the tail, the clustering structure is... Optimize the representation of cluster structure information and expand the scope of cluster structure information extraction.

[0046] (14); Step 4: Fine-tune the encoder to minimize the clustering loss function. The encoder is trained for the target. After training the encoder for one epoch, K-means clustering is performed on the newly generated latent space. Steps 3 and 4 are repeated to obtain new cluster assignments, thereby realizing the extraction of power load patterns.

[0047] Perform iterative fine-tuning to optimize the clustering loss function. The network is iteratively fine-tuned, with the initial cluster labels being U and the new cluster labels being P. The learning rate is set to 0.0001, and then further reduced to ensure greater stability during fine-tuning. The Adam optimizer is used to optimize the clustering loss function. The network is fine-tuned iteratively for 200 iterations. The clustering results are updated once every 10 epochs. The number of data points with different assignments in U and P is compared. When the proportion is less than the set threshold of 0.001, the degree of change in cluster assignment is considered low, the method converges, the iteration stops, and the clustering results are output.

[0048] Figure 2 This is a schematic diagram of a one-dimensional convolutional autoencoder model. The input is a one-dimensional payload data sequence of length 24×1. The number of convolutional kernels increases progressively with each layer: 16, 32, and 64. Each convolutional layer performs a convolution operation on the input sequence to extract features at different levels. The kernel size is 3, and the strides are 1, 1, and 2 respectively. After three convolutional layers, the model flattens the multi-dimensional feature map into a one-dimensional sequence in the Flatten layer, facilitating processing by subsequent fully connected layers. A one-dimensional sequence of length 12 is generated in the hidden layer; this is then restored to the same size as the Flatten layer through fully connected layers and Reshape layers. Finally, three deconvolutional layers with kernels of 32, 16, and 1 respectively, with a kernel size of 3 and strides of 2, 1, and 1, gradually restore the dimensionality of the original sequence.

[0049] Figure 3 This is a flowchart of the user load curve analysis method. To fully describe the electricity consumption characteristics of different types of users, a two-stage electricity load curve analysis method is adopted. In the first stage, the K-means algorithm is used to extract the electricity consumption data of each user and perform cluster analysis to obtain the user's electricity consumption regularity and typical load curve. In the second stage, the proposed load feature optimization clustering method based on one-dimensional convolutional autoencoder is used for cluster analysis.

[0050] Figure 4 This is a typical load curve result for a single user. The original load data was cleaned, missing and outlier values ​​were handled, and a typical load curve for each user was extracted. The data was then aggregated and classified using the K-means clustering algorithm. The cluster center of the cluster with the highest sample size after classification was extracted as the most representative typical load curve shape.

[0051] Figure 4 In this study, users exhibit a bimodal distribution, with their load curves showing two distinct peaks. This type of load curve best reflects the behavioral characteristics of the user. By generating a clear and representative typical pattern for each user, the amount of data that needs to be processed is greatly reduced, preventing misinterpretation by noise within individual users.

[0052] Figures 5(a) to 5(c) show the clustering analysis results of typical load curves for multiple users. The load feature optimization clustering algorithm using a one-dimensional convolutional autoencoder aggregates three load interval sequence clusters, clearly showing the characteristics of load curves for each user type, indicating good clustering quality. Compared to a single typical curve describing deterministic load, the autoencoder can dynamically express load characteristics, thus better describing the uncertainty of the load.

[0053] In Figure 5(a), the user is a single-peak user, whose load curve shows a single peak period, indicating relatively concentrated electricity demand. In Figure 5(b), the user is a double-peak user, whose load curve shows two distinct peak periods. In Figure 5(c), the user is a peak-avoiding user, whose load curve is opposite to that of other user types during peak hours. The load characteristics of different user types provide a strong basis for the optimal scheduling of the power system. Differentiated power supply strategies can be formulated based on the load characteristics of each user type to rationally allocate power resources and reduce power waste.

[0054] This invention presents a load feature optimization clustering method based on a one-dimensional convolutional autoencoder. The method first analyzes the load characteristics of individual users, extracting their typical load curve shapes to help more accurately understand their electricity consumption characteristics. To meet the needs of load uncertainty analysis, the proposed one-dimensional convolutional autoencoder load feature optimization clustering algorithm is employed. This improves clustering performance by optimizing the objective functions of the autoencoder reconstruction loss and clustering loss, effectively bridging the gap between clustering quality and reconstruction efficiency. The proposed method enhances the richness of the inherent information in the cluster structure, accurately classifies multiple users based on time-series electricity consumption characteristics, and dynamically expresses load characteristics using an autoencoder, thereby better describing load uncertainty. Compared to other traditional clustering methods, it simplifies application complexity and improves classification efficiency.

Claims

1. A load feature optimization clustering method based on one-dimensional convolutional autoencoder, characterized in that The method comprises the following steps: Step 1: collecting load types including daily load curve data as a data set, and pre-processing the collected data; Step 2: constructing a one-dimensional convolutional autoencoder model including an encoder and a decoder, the one-dimensional convolutional autoencoder model minimizing a reconstruction loss The load data is pre-trained for the target, load features are extracted, and the pre-trained model parameters are saved. Step 3: disassembling the decoding network part, retaining the encoding part for feature extraction, constructing a PCA-Kmeans space transformation model, passing the training data in the encoder, and using the PCA algorithm to transform the latent space into a low-dimensional space for K-means clustering analysis; Step 4: Fine-tune the encoder to minimize the clustering loss function Train the encoder for the target, after one epoch of training the encoder, perform K-means clustering on the newly generated latent space; Step 5: obtaining new cluster assignments to realize power load mode extraction. 2.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 1, wherein: In step 1, the collected data is pre-processed, and the k-nearest neighbor algorithm is used to complete the missing value processing; by finding the k nearest neighbor samples most similar to the missing value sample, the corresponding characteristic values of these neighbor samples are used to estimate the missing value; The target value is calculated The Euclidean distance d between the corresponding samples and other samples is calculated （1）； In formula (1), represents the number of load curve sampling points; represents the index for each sampling point; Sort the calculated Euclidean distances d from small to large, select the first k data, that is, the target value k-nearest neighbor data, and then calculate the replacement value of the missing value : （2）； In formula (2), is the weight of k-nearest neighbor data, and is the inverse of the Euclidean distance d; represents the number of nearest neighbor samples used to fill in the missing values. 3.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 2, characterized in that: In step 1, the maximum and minimum normalization method is used to reduce the value range of the collected data to between 0 and 1; （3）； In formula (3), x is the original load data value; x min is the minimum value of all samples in the data set; x max is the maximum value of all samples in the data set; x norm is the normalized value. 4.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 1, characterized in that: In step 2, the encoder includes a convolution layer, a pooling layer, a flattening layer and an embedding layer; the convolution layer extracts local features in the input load data through convolution operation, and a one-dimensional convolution kernel slides on the time series to capture load change patterns at different time scales; the pooling layer down-samples the feature map output by the convolution layer, while reducing the data dimension and retaining important features; the flattening layer flattens the multi-dimensional feature map obtained after three convolution layers into a fixed-length one-dimensional feature vector, facilitating subsequent clustering analysis; the embedding layer maps the flattened feature vector to a low-dimensional latent space, and the latent space contains the main features of the input data and is the basis for the decoder to reconstruct the data; and the decoder is a mirror image of the encoder structure, which gradually up-samples through the deconvolution layer to restore the data dimension, and finally outputs the reconstructed load sequence. 5.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 4, characterized in that: In step 2, the input data is mapped to a low-dimensional latent space through a one-dimensional convolution autoencoder model, which reduces the dimension of the original data set and efficiently extracts the time series characteristics of the data; （4）； （5）； In the above formula: is the activation function of the encoder, and the activation function Relu is adopted; is the mapping function of the decoder, and the sigmoid function is adopted; represents a one-dimensional convolution operation; and are weight matrices of the encoder and the decoder respectively; and are bias terms of the encoder and the decoder respectively; and are latent space features and reconstructed data respectively. The mean square error is used to represent the reconstruction loss, and the one-dimensional convolution autoencoder model is pre-trained with the objective of minimizing the reconstruction loss function; （6）； In Equation (6), reconstructed by the decoder and the original input data ; denotes the original load curve sample; denotes the training batch; denotes the load value of the original input load curve at the i-th hour; denotes the load value of the load curve reconstructed by the decoder at the i-th hour; denotes the Euclidean length of a vector. 6.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 5, characterized in that: In step 3, the pre-trained complete autoencoder is disassembled, the decoding network part is removed, and only the encoder for feature compression is retained; on the basis of the latent features extracted by the encoder, a PCA-Kmeans space transformation model is constructed to further optimize the feature space. 7.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 6, characterized in that: In the PCA stage, the latent features are subjected to secondary dimension reduction and decorrelation; in the K-means stage, clustering is performed in the optimized feature space; a new feature set that is reduced in dimension and decorrelated is obtained by using the PCA algorithm, which is used for subsequent K-means clustering, thereby optimizing the clustering effect; （7）； In formula (7), X is the normalized latent feature matrix; the covariance matrix Σ describes the distribution shape of the data in the feature space; By calculating the normalized load covariance matrix and performing eigenvalue decomposition, the eigenvalues and the eigenvector matrix are obtained. （8）； （9）； In the above formulae: represents the contribution ratio of the kth principal component to the total variance; represents the cumulative contribution rate of the first k principal components, which can help determine the number of retained principal components; represents the variance size representing the kth principal component; k represents the original feature dimension; represents the kth largest eigenvalue of the data covariance matrix; i represents the number of selected principal components;​​ When the cumulative variance contribution rate of the current k principal components is greater than 90%, the corresponding k value is the number of principal components; the k-dimensional orthogonal feature vector is a projection matrix composed of the feature vectors corresponding to the first k largest eigenvalues, which is reconstructed from the original n-dimensional features as principal components; The Kmeans method is used to construct a clustering loss: （10）； （11）； In the above formula: represents the original clustering loss, which measures the sum of the square of the distance of all samples to their cluster centers; represents a single data point in the latent space; represents the distance between the cluster center closer data points; represents the data point set of the i-th cluster clustering result; represents the current cluster center; k represents the total number of clusters. 8.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 7, characterized in that: The cluster discrete matrix is used to calculate the principal components to create a new projection space , and specifically comprises: calculating a clustering dispersion matrix S based on the initial clustering result ω After eigenvalue decomposition, the eigenvectors corresponding to the first k largest eigenvalues are selected to construct an orthogonal projection matrix V; then, the original data is projected into the latent space formed by V to obtain a new feature representation Y; （12）； （13）； In the above formulae: denotes the cluster dispersion matrix, which measures the dispersion degree of samples within the cluster; denotes the total number of clusters; denotes the matrix transposition operation; denotes the cluster loss after projection into the new space; denotes the orthogonal projection matrix, which is composed of the first k eigenvectors of the cluster dispersion matrix ; Wherein, the orthogonal matrix is by the matrix The eigenvalues are arranged in descending order, and the first k eigenvectors are selected to form the characteristic matrix; Will The mapping is to the tail of the orthogonal matrix V, the clustering structure information is relatively low in the n subsets, the representation of the clustering structure information is optimized, and the extraction range of the clustering structure information is expanded. （14）； In formula (14), denotes an objective function of the fine-tuning encoder; denotes a tail feature subspace projection matrix; denotes a feature vector in the latent space; denotes a training batch; denotes a current sample; denotes a cluster number of the clustering. 9.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 8, characterized in that: In step 4, the clustering loss is minimized Fine-tuning the encoder for the target; in each training cycle (epoch), input the training data into the encoder to obtain the latent features; perform K-means clustering on the current latent feature space to obtain the cluster assignment and cluster centers; calculate the loss function value according to the cluster assignment and cluster centers Update the parameters of the encoder using the optimizer; After one epoch of training the encoder, the updated encoder is used to extract features, and K-means clustering is performed again on the new feature space to obtain a new cluster assignment. 10.The load feature optimization clustering method based on one-dimensional convolution auto-encoder according to claim 9, characterized in that: In step 4, the clustering loss function is optimized Iteratively fine-tune the encoder network, let the cluster assignment label of the initial iteration be U, and the cluster label of the new iteration be P; compare the proportion of data points with different assignments in U and P, η: （15）； In formula (15), N represents the total number of user load curves; U represents the cluster assignment label vector of the last iteration; P represents the cluster assignment label vector of the current iteration; represents an indicator function that returns 1 when the condition is true and 0 otherwise. A threshold value is set to 0.001, and when the proportion is less than the set threshold value, it is considered that the degree of change of the cluster assignment is low, the method converges, the iteration is stopped, and the clustering result is output.

Images