A method, system, device and medium for identifying energy storage users based on multi-dimensional feature analysis and feature dimension reduction technology

Abstract

The application provides a kind of energy storage user identification method, system, equipment and medium based on multi-dimensional feature analysis and feature dimension reduction technology, comprising: selecting the load curve data of continuous days from the multiple feature months of user, and extracting multi-dimensional features from each load curve data respectively;Based on multi-dimensional features, correlation coefficient method and sparsification strategy are used to construct feature correlation graph;The high-dimensional feature vector formed by the feature correlation graph and the multi-dimensional features is input into the pre-trained feature correlation graph variational autoencoder, and the low-dimensional latent feature is obtained;The low-dimensional latent feature is input into the pre-trained user identification model, and the binary classification identification result of energy storage user or non-energy storage user is obtained;The application converts the originally unstructured high-dimensional feature vector into a feature representation form with topological structure by constructing a multi-dimensional feature correlation graph, and performs dimension reduction modeling on this basis, effectively suppresses the negative impact of feature redundancy on model performance, and improves the compactness and discriminability of feature representation.

Description

Technical Field

[0001] This invention relates to the field of power system technology, specifically to a method, system, device, and medium for identifying energy storage users based on multidimensional feature analysis and feature dimensionality reduction technology. Background Technology

[0002] With the development of large-scale data analysis and intelligent recognition technologies, complex systems often require the extraction of numerous multidimensional features describing object behavior from raw data. However, in practical applications, these features typically exhibit significant redundancy and strong correlation. High-dimensional feature spaces not only increase model computational complexity but also easily lead to overfitting and decreased generalization ability, becoming one of the key factors restricting model performance improvement. To address the high-dimensional feature problem, researchers have proposed various feature dimensionality reduction and representation learning methods. Among these, linear dimensionality reduction methods such as Principal Components Analysis (PCA) are widely used due to their simplicity, but they struggle to characterize nonlinear feature relationships in complex data, limiting their applicability in real-world complex scenarios.

[0003] To overcome the limitations of linear methods, deep learning-based autoencoders (AEs) and variational autoencoders (VAEs) have gradually become mainstream dimensionality reduction techniques. VAEs learn core behavioral pattern features through probabilistic generative models, demonstrating good performance in nonlinear feature modeling and noise suppression, and have gained widespread attention in representation learning and feature compression. However, traditional VAEs typically treat input features as independent vector dimensions, ignoring the statistical correlations and structural dependencies that are prevalent among multidimensional features, making it difficult to fully utilize the intrinsic connections between features.

[0004] In recent years, Graph Neural Networks (GNNs) have been introduced into feature selection and representation learning research due to their advantages in relation modeling. Related research attempts to treat features as nodes and construct feature relationship graphs based on correlations or dependencies, thereby achieving structure-aware feature modeling through graph convolution and other methods. Building on this, models such as Graph Variational Autoencoders (GVAEs) further combine graph structure modeling with probabilistic latent variable learning, providing new technical paths for unsupervised representation of complex feature relationships. Some recent studies have also introduced attention mechanisms or improved graph structure modeling strategies to enhance the model's ability to express local and global feature correlations.

[0005] Overall, while existing research has made progress in nonlinear dimensionality reduction and feature relationship modeling, there are still shortcomings in the deep integration of explicit modeling of multidimensional feature correlations and the dimensionality reduction process. There is still a lack of a unified methodological framework that can simultaneously characterize feature correlation structures and achieve dimensionality reduction optimization. Summary of the Invention

[0006] To address the shortcomings of existing technologies in the deep integration of explicit modeling and dimensionality reduction of multidimensional feature correlations, this invention proposes a method for identifying energy storage users based on multidimensional feature analysis and feature dimensionality reduction techniques, comprising: Load curve data for multiple consecutive days is selected from each of the user's multiple characteristic months, and multidimensional features are extracted from each load curve data. Based on the aforementioned multidimensional features, a feature correlation graph is constructed using the correlation coefficient method and a sparsity strategy. The high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the pre-trained feature correlation map variational autoencoder to obtain low-dimensional latent features. By inputting low-dimensional latent features into a pre-trained user identification model, a binary classification identification result of energy storage users or non-energy storage users is obtained. The feature correlation graph variational autoencoder includes an encoder and a decoder. The encoder uses a graph neural network to aggregate information from the feature correlation graph, comprehensively considering the attributes of the feature nodes themselves and their correlation with neighboring nodes, and learns the embedded representation of each feature node. The user identification model adopts a convolutional neural network plus a bidirectional gated recurrent unit structure.

[0007] Preferably, the multidimensional features include one or more of the following: time-domain statistical features, time-series morphological features, periodic features, and energy features; The time-domain statistical features include one or more of the following: daily load peak value, valley value, peak-valley difference, mean, variance, skewness or kurtosis; The time-series morphological features include one or more of the following: extracting charging and discharging duration, frequency, load fluctuation rate, load difference entropy between adjacent time periods, or ramp rate; The periodic characteristics include one or more of the following: Fast Fourier Transform fundamental amplitude, harmonic proportion, or main periodic energy concentration. The energy characteristics include one or more of the following: daily cumulative load, charge / discharge energy difference, or time-of-day energy ratio.

[0008] Preferably, the step of constructing a feature correlation map based on the multidimensional features using a correlation coefficient method and a sparsity strategy includes: Calculate the Pearson correlation coefficients among the features of each dimension; Using each dimension feature as a node, select the K groups of dimension features with the largest absolute value of Pearson correlation coefficient to establish undirected weighted edges between nodes, and use the Pearson correlation coefficient between dimension features as edge weights. Construct a feature-related graph from each node and edge; Where K is an integer greater than or equal to 1.

[0009] Preferably, the training process of the feature correlation map variational autoencoder includes: The high-dimensional feature vector, composed of the feature correlation map and multi-dimensional features, is input into the encoder of the feature correlation map variational autoencoder to obtain the low-dimensional feature representation. The dimensionality-reduced feature representation is input into the decoder of the feature correlation graph variational autoencoder to reconstruct the original high-dimensional feature vector. The feature correlation graph variational autoencoder is then trained by minimizing the weighted sum of the reconstruction error and the dimensionality-reduced feature distribution constraint term, thus obtaining the trained feature correlation graph variational autoencoder.

[0010] Preferably, the loss function of the feature correlation map variational autoencoder is calculated as follows:

[0011] In the formula, L This represents the value of the loss function. Indicates feature reconstruction error, Indicates the weighting coefficient. Indicates the KL divergence. Represents a given original high-dimensional load feature vector At that time, the low-dimensional feature representation learned by the encoder Z The conditional probability distribution; Representation of low-dimensional features Z The distribution of .

[0012] Preferably, the step of inputting the high-dimensional feature vector composed of the feature correlation map and multi-dimensional features into the encoder of the feature correlation map variational autoencoder to obtain a low-dimensional feature representation includes: The high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the encoder of the feature correlation map variational autoencoder. The graph convolutional layer of the encoder is used to aggregate information of nodes corresponding to each feature. The attribute value of each node is weighted and fused with the feature information of its neighboring nodes to obtain a node feature vector containing topological association information. The feature vectors of all nodes are integrated and compressed through a fully connected layer, and the mean vector and variance vector representing the overall feature distribution are output. By employing reparameterization techniques, low-dimensional feature representations are obtained by sampling from the dimensionality-reduced feature distribution based on the mean vector and variance vector.

[0013] Preferably, the training process of the user identification model includes: The training set is composed of low-dimensional feature representations and corresponding labels indicating whether the user is an energy storage user. A convolutional neural network is used to perform one-dimensional convolution on the low-dimensional feature representations in the training set to extract local temporal pattern features, resulting in a convolutional sequence. A bidirectional gated recurrent unit structure is used to perform bidirectional temporal modeling on the convolutional sequence to extract long-term dependency features. Finally, the probability of energy storage users is output through a fully connected layer and either Sigmoid or Softmax, and the binary classification result of energy storage users or non-energy storage users is obtained accordingly. With the goal of minimizing the binary classification recognition result and the corresponding label error, the parameters in the user recognition model are trained to obtain a trained user recognition model.

[0014] Based on the same inventive concept, the present invention also provides an energy storage user identification system based on multidimensional feature analysis and feature dimensionality reduction technology, including: a feature extraction module, a feature correlation graph module, a dimensionality reduction module and an identification module; The feature extraction module is used to select load curve data for multiple consecutive days from each of the user's multiple characteristic months, and extract multidimensional features from each load curve data respectively; The feature correlation graph module is used to construct a feature correlation graph based on the multidimensional features using a correlation coefficient method and a sparsity strategy. The dimensionality reduction module is used to input a high-dimensional feature vector composed of feature correlation maps and multi-dimensional features into a pre-trained feature correlation map variational autoencoder to obtain low-dimensional latent features. The identification module is used to input low-dimensional latent features into a pre-trained user identification model to obtain a binary classification identification result of energy storage users or non-energy storage users. The feature correlation graph variational autoencoder includes an encoder and a decoder. The encoder uses a graph neural network to aggregate information from the feature correlation graph, comprehensively considering the attributes of the feature nodes themselves and their correlation with neighboring nodes, and learns the embedded representation of each feature node. The user identification model adopts a convolutional neural network plus a bidirectional gated recurrent unit structure.

[0015] Preferably, the multidimensional features in the feature extraction module include one or more of the following: time-domain statistical features, time-series morphological features, periodic features, and energy features; The time-domain statistical features include one or more of the following: daily load peak value, valley value, peak-valley difference, mean, variance, skewness or kurtosis; The time-series morphological features include one or more of the following: extracting charging and discharging duration, frequency, load fluctuation rate, load difference entropy between adjacent time periods, or ramp rate; The periodic characteristics include one or more of the following: Fast Fourier Transform fundamental amplitude, harmonic ratio, or main periodic energy concentration. The energy characteristics include one or more of the following: daily cumulative load, charge / discharge energy difference, or time-of-day energy ratio.

[0016] Preferably, the feature correlation graph module is specifically used for: Calculate the Pearson correlation coefficients among the features of each dimension; Using each dimension feature as a node, select the K groups of dimension features with the largest absolute value of Pearson correlation coefficient to establish undirected weighted edges between nodes, and use the Pearson correlation coefficient between dimension features as edge weights. Construct a feature-related graph from each node and edge; Where K is an integer greater than or equal to 1.

[0017] Preferably, the training process of the feature correlation graph variational autoencoder in the dimensionality reduction module includes: The high-dimensional feature vector, composed of the feature correlation map and multi-dimensional features, is input into the encoder of the feature correlation map variational autoencoder to obtain the low-dimensional feature representation. The dimensionality-reduced feature representation is input into the decoder of the feature correlation graph variational autoencoder to reconstruct the original high-dimensional feature vector. The feature correlation graph variational autoencoder is then trained by minimizing the weighted sum of the reconstruction error and the dimensionality-reduced feature distribution constraint term, thus obtaining the trained feature correlation graph variational autoencoder.

[0018] Preferably, the loss function of the feature correlation graph variational autoencoder in the dimensionality reduction module is calculated as follows:

[0019] In the formula, L This represents the value of the loss function. Indicates feature reconstruction error, Indicates the weighting coefficient. Indicates the KL divergence. Represents a given original high-dimensional load feature vector At that time, the low-dimensional feature representation learned by the encoder Z The conditional probability distribution; Representation of low-dimensional features Z The distribution of .

[0020] Preferably, in the dimensionality reduction module, a high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the encoder of the feature correlation map variational autoencoder to obtain a low-dimensional feature representation, including: The high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the encoder of the feature correlation map variational autoencoder. The graph convolutional layer of the encoder is used to aggregate information of nodes corresponding to each feature. The attribute value of each node is weighted and fused with the feature information of its neighboring nodes to obtain a node feature vector containing topological association information. The feature vectors of all nodes are integrated and compressed through a fully connected layer, and the mean vector and variance vector representing the overall feature distribution are output. By employing reparameterization techniques, low-dimensional feature representations are obtained by sampling from the dimensionality-reduced feature distribution based on the mean vector and variance vector.

[0021] Preferably, the training process of the user identification model in the identification module includes: The training set is composed of low-dimensional feature representations and corresponding labels indicating whether the user is an energy storage user. A convolutional neural network is used to perform one-dimensional convolution on the low-dimensional feature representations in the training set to extract local temporal pattern features, resulting in a convolutional sequence. A bidirectional gated recurrent unit structure is used to perform bidirectional temporal modeling on the convolutional sequence to extract long-term dependency features. Finally, the probability of energy storage users is output through a fully connected layer and either Sigmoid or Softmax, and the binary classification result of energy storage users or non-energy storage users is obtained accordingly. With the goal of minimizing the binary classification recognition result and the corresponding label error, the parameters in the user recognition model are trained to obtain a trained user recognition model.

[0022] In another aspect, the present invention also provides an electronic device, comprising: at least one processor and a memory; The memory is used to store one or more programs; When the one or more programs are executed by the one or more processors, the energy storage user identification method based on multidimensional feature analysis and feature dimensionality reduction technology, as described above, is implemented.

[0023] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed, it implements the energy storage user identification method based on multidimensional feature analysis and feature dimensionality reduction technology as described above.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method, system, device, and medium for energy storage user identification based on multidimensional feature analysis and feature dimensionality reduction technology. The method includes: selecting load curve data for multiple consecutive days from various characteristic months of a user, and extracting multidimensional features from each load curve data; constructing a feature correlation graph based on the multidimensional features using a correlation coefficient method and a sparsity strategy; inputting a high-dimensional feature vector composed of the feature correlation graph and the multidimensional features into a pre-trained feature correlation graph variational autoencoder to obtain low-dimensional latent features; and inputting the low-dimensional latent features into a pre-trained user identification model to obtain a binary classification result for energy storage users or non-energy storage users. The feature correlation graph variational autoencoder includes an encoder and a decoder. The encoder uses a graph neural network to aggregate information from the feature correlation graph, comprehensively considering the attributes of the feature nodes themselves and their correlation with adjacent nodes to learn the embedded representation of each feature node. The user identification model uses a convolutional neural network with a bidirectional gated recurrent unit structure. To address the issue of degraded model performance caused by the large number, strong correlation, and high redundancy of multidimensional load features, this invention constructs a multidimensional feature correlation graph. This transforms unstructured high-dimensional feature vectors into feature representations with topological structures, and then performs dimensionality reduction modeling based on this, effectively suppressing the negative impact of feature redundancy on model performance and improving the compactness and discriminativeness of feature representations. To address the problem that traditional dimensionality reduction methods cannot explicitly characterize the correlation structure between multidimensional features, this invention introduces a feature correlation graph modeling approach. This approach expresses the statistical correlations between features in a graph structure and aggregates information through a graph neural network, enabling the dimensionality reduction process to explicitly perceive the correlations between features, thereby achieving structure-aware feature dimensionality reduction. Attached Figure Description

[0025] Figure 1 A flowchart of an energy storage user identification method based on multidimensional feature analysis and feature dimensionality reduction technology provided by the present invention; Figure 2 A flowchart illustrating a specific example of an energy storage user identification method based on multidimensional feature analysis and feature dimensionality reduction technology provided by the present invention; Figure 3 A flowchart for constructing the feature correlation graph provided by this invention; Figure 4 Example diagram of feature relationships provided by the present invention; Figure 5 The FCG-VAE model diagram provided by this invention; Figure 6 A diagram illustrating the energy storage user identification model provided by this invention; Figure 7 A schematic diagram of the structure of an energy storage user identification system based on multidimensional feature analysis and feature dimensionality reduction technology provided by the present invention; Figure 8This is a schematic diagram of an electronic device structure provided by the present invention. Detailed Implementation

[0026] This invention proposes a method, system, device, and medium for energy storage user identification based on multidimensional feature analysis and feature dimensionality reduction technology, aiming to solve the following technical problems: (1) To address the problem of declining model performance caused by the large number, strong correlation, and high redundancy of multidimensional load features; (2) Solve the problem that traditional dimensionality reduction methods cannot explicitly characterize the correlation structure between multidimensional features; (3) Solve the problem that the dimensionality reduction results lack physical and statistical interpretability in the existing energy storage user identification methods.

[0027] This invention constructs a multidimensional feature correlation graph and introduces a dimensionality reduction framework combining graph neural networks and variational autoencoders to learn low-dimensional feature representations that are most discriminative for energy storage users, thereby improving the accuracy, stability, and interpretability of energy storage user identification.

[0028] Example 1: A method for identifying energy storage users based on multidimensional feature analysis and feature dimensionality reduction techniques, such as Figure 1 As shown, it includes: Step 1: Select load curve data for multiple consecutive days from each of the user's multiple characteristic months, and extract multidimensional features from each load curve data; Step 2: Construct a feature correlation map based on multidimensional features using the correlation coefficient method and sparsification strategy; Step 3: Input the high-dimensional feature vector composed of the feature correlation map and multi-dimensional features into the pre-trained feature correlation map variational autoencoder to obtain low-dimensional latent features; Step 4: Input the low-dimensional latent features into the pre-trained user identification model to obtain the binary classification identification result of energy storage users or non-energy storage users; The feature correlation graph variational autoencoder includes an encoder and a decoder. The encoder uses a graph neural network to aggregate information from the feature correlation graph, comprehensively considering the attributes of the feature nodes themselves and their correlation with neighboring nodes, and learns the embedded representation of each feature node. The user recognition model uses a convolutional neural network with a bidirectional gated recurrent unit structure.

[0029] A specific example of an energy storage user identification method based on multidimensional feature analysis and feature dimensionality reduction techniques is as follows: Figure 2 As shown.

[0030] First, daily user load curve data is acquired (sampled every 15 minutes, totaling 96 points / day), and preprocessed including missing data repair, anomaly removal, and normalization. Then, time-domain statistical features, time-series morphological features, periodic (FFT) features, and energy features are extracted from the daily load curves to construct a high-dimensional feature vector F. The Pearson correlation coefficient of any two-dimensional feature across all user samples is calculated to obtain an initial correlation matrix C. Subsequently, for each feature node in C, a Top-K sparsity strategy is adopted (i.e., only the top K edges with the largest absolute values ​​of the correlation coefficients are retained) to construct a topologically structured feature correlation graph G(V,E,W), where V represents a node, E represents an edge, and W represents the edge weight. During this process, the feature correlation graph G is transformed into a weighted adjacency matrix A for model calculation, where the elements of A are... ij The value of is the value of node v in graph G. i With v j The edge weight W between ij (i.e., the real value of the Pearson correlation coefficient), the corresponding A for node pairs that have not established connections. ij = 0. Input (F,A) into the Feature Correlation Graph Variational Autoencoder (FCG-VAE). The graph encoder extracts structural information and outputs low-dimensional latent features Z (new feature vectors obtained after dimensionality reduction). Simultaneously, the graph decoder reconstructs the features to complete model training. For the same user, select 7 consecutive days of data from each of the four representative months to obtain 28 days of dimensionality-reduced new features, which are then concatenated in chronological order to form sequence Z. Input this sequence Z into a CNN+BiGRU (convolutional neural network with bidirectional gated recurrent units) classification model to output the probability of energy storage users and the energy storage / non-energy storage identification results.

[0031] Step 1 specifically includes: To address the technical challenges of "redundancy and interference in multidimensional load features," multidimensional features with energy storage user identification characteristics are extracted from the original load data, providing accurate input for subsequent correlation analysis and dimensionality reduction.

[0032] Input data: The raw data consists of load curve data for 7849 households, covering 28 non-continuous periods in different seasons (including representative periods of winter, summer, spring, and autumn), with 96 sampling points per day. The data is preprocessed (cleaned and standardized) before being input into this module. The multidimensional features are shown in Table 1.

[0033] Table 1 Multidimensional Features

[0034] Feature extraction dimensions: Time-domain statistical features: Extract daily load peak value, valley value, peak-valley difference, mean, variance, etc., to characterize the temporal distribution characteristics of the load; Temporal morphological characteristics: Extract charging and discharging duration, frequency, load fluctuation rate, and load difference entropy between adjacent time periods to capture the temporal variation pattern of energy storage users' charging and discharging behavior; Periodic characteristics: The fundamental amplitude and harmonic ratio are extracted by Fast Fourier Transform (FFT) to reflect the periodic fluctuation characteristics of the load; Energy characteristics: Extract daily cumulative load, charge-discharge energy difference, etc., to quantify the energy transfer characteristics of energy storage users.

[0035] Step 2 specifically includes: Since multidimensional load features are generally correlated and redundant, directly modeling high-dimensional feature vectors can easily overlook the structural information between features. To explicitly characterize the correlations between different features and improve the effectiveness of subsequent dimensionality reduction and modeling, this invention proposes a method for constructing multidimensional feature correlation graphs.

[0036] The flowchart for constructing the feature correlation graph is as follows: Figure 3 As shown.

[0037] Each dimension of the multidimensional load feature vector is treated as a node in a graph, and a set of feature nodes is constructed.

[0038] Where N is the feature dimension, and V is the set of nodes. These are node 1, node 2, and node N, respectively.

[0039] To measure the correlation between different feature nodes, a correlation index for any two-dimensional feature across all user samples is calculated. The correlation index uses the Pearson correlation coefficient method. When the correlation index between two features exceeds a preset threshold... When the time comes, an undirected weighted edge is established between the corresponding nodes.

[0040]

[0041] In the formula, E represents an edge. This represents the correlation coefficient between node vi and node vj.

[0042] Edge weight W ij Used to characterize the correlation strength between features i and j (i.e., nodes i and j).

[0043] In this way, the original unstructured multidimensional feature vectors are transformed into a feature correlation graph G=(V,E) with a topological structure, providing structured input for subsequent feature dimensionality reduction and representation learning based on graph models.

[0044] A specific example of a feature relationship is as follows: Figure 4 As shown.

[0045] The example shows the feature correlation graph topology obtained after Top-K sparsification. In this embodiment, the constructed feature correlation graph contains 6 feature nodes, and undirected weighted edges are established between the nodes based on feature correlation to form a sparse feature association structure. Figure 4 As shown, there is a correlation between nodes v5 and v6, while no connections are established between the remaining node pairs, indicating that the correlation between this feature pair and other features is lower than a preset threshold. Furthermore, the edge weight between nodes v1 and v2 is the largest, represented by the line thickness in the graph; the edge weight between nodes v2 and v4 is the second largest; and the edge weight between nodes v1 and v3 is the smallest, representing the differences in correlation strength between different feature pairs.

[0046] Step 3 specifically includes: To achieve efficient dimensionality reduction representation of multidimensional load features while preserving feature-related structural information, this invention proposes a Feature Correlation Graph Variational Autoencoder (FCG-VAE) model. This model takes a multidimensional feature correlation graph as input and models and compresses high-dimensional features through graph structure constraints.

[0047] The FCG-VAE model comprises an encoder and a decoder. The encoder uses a graph neural network to aggregate information from the feature correlation graph, comprehensively considering the attributes of each feature node and its correlation with neighboring nodes to learn the embedded representation of each feature node. Based on this, the encoder uses graph convolutional layers to aggregate information from each feature node, weighting and fusing the attribute values ​​of each feature node with the feature information of its neighboring nodes to obtain a node feature vector containing topological association information. Subsequently, a fully connected layer integrates and compresses all node feature vectors, outputting a mean vector μ and a variance vector σ representing the overall feature distribution, which serve as distribution parameters for subsequent compressed feature sampling.

[0048] By employing reparameterization techniques, a low-dimensional feature representation Z is obtained by sampling from the dimensionality-reduced feature distribution based on the mean vector μ and variance vector σ. This low-dimensional feature representation Z is used to characterize the core information in the original high-dimensional feature vector, thereby achieving feature space compression and representation optimization.

[0049] The decoder takes the reduced-dimensional feature representation Z as input to reconstruct the original high-dimensional feature vector, and trains the model by minimizing the weighted sum of the reconstruction error and the constraint term of the reduced-dimensional feature distribution. The model's loss function is defined as:

[0050] in, L This represents the value of the loss function. Indicates feature reconstruction error. Indicates the weighting coefficient. The KL divergence is represented by the KL term, which is used to constrain the continuity and stability of the distribution of latent variables. Represents a given original high-dimensional load feature vector At that time, the low-dimensional feature representation learned by the encoder Z The conditional probability distribution; Representation of low-dimensional features Z The distribution of features. Through the above methods, the FCG-VAE model achieves effective dimensionality reduction and representation learning of multidimensional load features while preserving relevant structural information.

[0051] The load distribution of ordinary users is chaotic and irregular, while the key characteristics of energy storage users (such as peak-valley balance rate) will show a highly concentrated specific distribution pattern due to the constraints of battery physical characteristics (charge and discharge cycles) and electricity pricing strategies.

[0052] The "latent distribution" learned through the VAE model is essentially a "standard feature spectrum" of the electricity consumption behavior patterns of energy storage users. By projecting the data of the user to be identified onto this distribution, we can accurately determine whether the user possesses energy storage characteristics based on the degree to which the user deviates from the "standard spectrum".

[0053] FCG-VAE model such as Figure 5 As shown, the main function of the FCG-VAE model is unsupervised learning to transform the original high-dimensional data features into low-dimensional latent features Z.

[0054] Step 4 specifically includes: After obtaining the low-dimensional feature representation Z from the FCG-VAE model output, the daily low-dimensional feature representation is denoted as Zi. m,d Where m represents the month index, and d represents the d-th day of the selected consecutive week (d=1,…,7). For the same user, 7 consecutive days of data are selected from each of the 4 representative months, resulting in 28 days of feature representation, which constitutes the low-dimensional latent feature sequence Z=[Z m,1 Z m,2 ,…,Z m,d The feature sequence is used as input to the energy storage user identification model (i.e., user identification model or identification model).

[0055] The recognition model adopts a CNN (Convolutional Neural Network) + BiGRU (Bidirectional Gated Recurrent Unit) structure. The CNN performs one-dimensional convolution on the feature sequence to extract local temporal pattern features, and the BiGRU performs bidirectional temporal modeling on the convolutional sequence to extract long-term dependency features. Finally, the probability of energy storage users is output through a fully connected layer and either Sigmoid or Softmax, and the binary classification recognition result of energy storage users / non-energy storage users is obtained accordingly.

[0056] The identification model is a supervised learning model trained on data labeled "whether or not an energy storage user" to learn how to distinguish the feature patterns of energy storage users from Z. Specifically: a low-dimensional feature representation and corresponding "whether or not an energy storage user" label are used as the training set; a convolutional neural network is used to perform one-dimensional convolution on the low-dimensional feature representation in the training set to extract local temporal pattern features, resulting in a convolutional sequence; a bidirectional gated recurrent unit structure is used to perform bidirectional temporal modeling on the convolutional sequence to extract long-term dependency features; finally, a fully connected layer and either sigmoid or softmax are used to output the probability of an energy storage user, resulting in a binary classification result of either an energy storage user or a non-energy storage user; the parameters in the user identification model are trained with the goal of minimizing the error between the binary classification result and the corresponding label, resulting in a trained user identification model.

[0057] Energy storage user identification model, such as Figure 6 As shown.

[0058] The technical effects achieved by this invention include: (1) To address the problem of high number, strong correlation, and high redundancy of multidimensional load features leading to a decline in model performance: This invention constructs a multidimensional feature correlation graph to transform the originally unstructured high-dimensional feature vector into a feature representation with a topological structure, and performs dimensionality reduction modeling on this basis, effectively suppressing the negative impact of feature redundancy on model performance and improving the compactness and discriminability of feature representation.

[0059] (2) To address the problem that traditional dimensionality reduction methods cannot explicitly characterize the correlation structure between multidimensional features: This invention introduces a feature correlation graph modeling method, which expresses the statistical correlation between features in the form of a graph structure, and aggregates information through a graph neural network, so that the dimensionality reduction process can explicitly perceive the correlation between features, thereby achieving structure-aware feature dimensionality reduction.

[0060] (3) In response to the problem that the dimensionality reduction results lack physical and statistical interpretability in existing energy storage user identification methods: This invention constructs a correlation graph at the feature level, so that the generation process of compressed features is constrained by the original feature association structure. Engineers can trace the correspondence between compressed features and original load features (such as peak-valley difference, charging and discharging duration, etc.), and give the compressed features a clear physical meaning. At the same time, the variational autoencoder framework makes the compressed features follow a continuous probability distribution through KL divergence constraints, ensuring that users with similar electricity consumption behaviors are close to each other in the compressed feature space, so that the dimensionality reduction results have statistically verifiable distribution characteristics.

[0061] This invention proposes a non-intrusive energy storage user identification method based on feature correlation maps and variational autoencoders. This method provides an innovative solution for the power system and related industries to accurately identify and quantify user-side energy storage behavior, possessing significant application value in the context of the increasing prevalence of distributed resources and the urgent need for grid interaction. By deeply mining multi-dimensional correlation patterns in electricity load data, this method can penetrate massive amounts of data and accurately identify user energy storage behavior characteristics without relying on additional hardware investment, effectively supporting the precise perception of core resources of the "virtual power plant."

[0062] The application of this invention will directly empower key links in the energy service industry chain. For power grid companies, this technology is a key tool for achieving lean management on the distribution side, helping them accurately assess the penetration rate and spatiotemporal distribution of energy storage within their jurisdiction, providing core data input for planning upgrades, operation scheduling, and market transactions. For energy storage system integrators, operators, and electricity retailers, this method can serve as an efficient market insight and customer relationship management tool, used to accurately locate potential customers, assess the operational performance of existing energy storage, and provide users with data-driven performance optimization and value-added service packages. For regulatory authorities and policymakers, the macro-level energy storage deployment map and behavioral analysis report generated by this method can provide reliable quantitative decision-making basis for calibrating energy storage subsidy policies, optimizing peak-valley electricity pricing mechanisms, and designing electricity market rules.

[0063] From a broader socio-economic and energy transition perspective, this invention forms the cornerstone for achieving efficient aggregation and value realization of flexible user-side resources. Through data intelligence, it significantly reduces the visibility threshold and management costs of large-scale, distributed energy storage resources, providing a clear technical path for activating massive user-side resources to participate in grid regulation and promoting the consumption of new energy sources. Therefore, the promotion of this invention will help accelerate the construction of a new power system that integrates generation, grid, load, and storage, improve the overall energy utilization efficiency and grid security level of society, and provide solid technical support for structural changes on the energy consumption side and the achievement of "dual-carbon" goals.

[0064] Example 2: Based on the same inventive concept, this invention also provides an energy storage user identification system based on multidimensional feature analysis and feature dimensionality reduction technology, such as... Figure 7 As shown, it includes: a feature extraction module, a feature correlation graph module, a dimensionality reduction module, and a recognition module; The feature extraction module is used to select load curve data for multiple consecutive days from each of the user's multiple characteristic months, and extract multidimensional features from each load curve data respectively; The feature correlation graph module is used to construct a feature correlation graph based on the multidimensional features using a correlation coefficient method and a sparsity strategy. The dimensionality reduction module is used to input a high-dimensional feature vector composed of feature correlation maps and multi-dimensional features into a pre-trained feature correlation map variational autoencoder to obtain low-dimensional latent features. The identification module is used to input low-dimensional latent features into a pre-trained user identification model to obtain a binary classification identification result of energy storage users or non-energy storage users. The feature correlation graph variational autoencoder includes an encoder and a decoder. The encoder uses a graph neural network to aggregate information from the feature correlation graph, comprehensively considering the attributes of the feature nodes themselves and their correlation with neighboring nodes, and learns the embedded representation of each feature node. The user identification model adopts a convolutional neural network plus a bidirectional gated recurrent unit structure.

[0065] Preferably, the multidimensional features in the feature extraction module include one or more of the following: time-domain statistical features, time-series morphological features, periodic features, and energy features; The time-domain statistical features include one or more of the following: daily load peak value, valley value, peak-valley difference, mean, variance, skewness or kurtosis; The time-series morphological features include one or more of the following: extracting charging and discharging duration, frequency, load fluctuation rate, load difference entropy between adjacent time periods, or ramp rate; The periodic characteristics include one or more of the following: Fast Fourier Transform fundamental amplitude, harmonic ratio, or main periodic energy concentration. The energy characteristics include one or more of the following: daily cumulative load, charge / discharge energy difference, or time-of-day energy ratio.

[0066] Preferably, the feature correlation graph module is specifically used for: Calculate the Pearson correlation coefficients among the features of each dimension; Using each dimension feature as a node, select the K groups of dimension features with the largest absolute value of Pearson correlation coefficient to establish undirected weighted edges between nodes, and use the Pearson correlation coefficient between dimension features as edge weights. Construct a feature-related graph from each node and edge; Where K is an integer greater than or equal to 1.

[0067] Preferably, the training process of the feature correlation graph variational autoencoder in the dimensionality reduction module includes: The high-dimensional feature vector, composed of the feature correlation map and multi-dimensional features, is input into the encoder of the feature correlation map variational autoencoder to obtain the low-dimensional feature representation. The dimensionality-reduced feature representation is input into the decoder of the feature correlation graph variational autoencoder to reconstruct the original high-dimensional feature vector. The feature correlation graph variational autoencoder is then trained by minimizing the weighted sum of the reconstruction error and the dimensionality-reduced feature distribution constraint term, thus obtaining the trained feature correlation graph variational autoencoder.

[0068] Preferably, the loss function of the feature correlation graph variational autoencoder in the dimensionality reduction module is calculated as follows:

[0069] In the formula, L This represents the value of the loss function. Indicates feature reconstruction error, Indicates the weighting coefficient. Indicates the KL divergence. Represents a given original high-dimensional load feature vector At that time, the low-dimensional feature representation learned by the encoder Z The conditional probability distribution; Representation of low-dimensional features Z The distribution of .

[0070] Preferably, in the dimensionality reduction module, a high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the encoder of the feature correlation map variational autoencoder to obtain a low-dimensional feature representation, including: The high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the encoder of the feature correlation map variational autoencoder. The graph convolutional layer of the encoder is used to aggregate information of nodes corresponding to each feature. The attribute value of each node is weighted and fused with the feature information of its neighboring nodes to obtain a node feature vector containing topological association information. The feature vectors of all nodes are integrated and compressed through a fully connected layer, and the mean vector and variance vector representing the overall feature distribution are output. By employing reparameterization techniques, low-dimensional feature representations are obtained by sampling from the dimensionality-reduced feature distribution based on the mean vector and variance vector.

[0071] Preferably, the training process of the user identification model in the identification module includes: The training set is composed of low-dimensional feature representations and corresponding labels indicating whether the user is an energy storage user. A convolutional neural network is used to perform one-dimensional convolution on the low-dimensional feature representations in the training set to extract local temporal pattern features, resulting in a convolutional sequence. A bidirectional gated recurrent unit structure is used to perform bidirectional temporal modeling on the convolutional sequence to extract long-term dependency features. Finally, the probability of energy storage users is output through a fully connected layer and either Sigmoid or Softmax, and the binary classification result of energy storage users or non-energy storage users is obtained accordingly. With the goal of minimizing the binary classification recognition result and the corresponding label error, the parameters in the user recognition model are trained to obtain a trained user recognition model.

[0072] Example 3 like Figure 8 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0073] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the energy storage user identification method based on multi-dimensional feature analysis and feature dimensionality reduction technology in the above embodiments.

[0074] Example 4 Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the energy storage user identification method based on multi-dimensional feature analysis and feature dimensionality reduction technology in the above embodiments.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for identifying energy storage users based on multidimensional feature analysis and feature dimensionality reduction techniques, characterized in that, include: Load curve data for multiple consecutive days is selected from each of the user's multiple characteristic months, and multidimensional features are extracted from each load curve data. Based on the aforementioned multidimensional features, a feature correlation graph is constructed using the correlation coefficient method and a sparsity strategy. The high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the pre-trained feature correlation map variational autoencoder to obtain low-dimensional latent features. By inputting low-dimensional latent features into a pre-trained user identification model, a binary classification identification result of energy storage users or non-energy storage users is obtained. The feature correlation graph variational autoencoder includes an encoder and a decoder. The encoder uses a graph neural network to aggregate information from the feature correlation graph, comprehensively considering the attributes of the feature nodes themselves and their correlation with neighboring nodes, and learns the embedded representation of each feature node. The user identification model adopts a convolutional neural network plus a bidirectional gated recurrent unit structure.

2. The energy storage user identification method as described in claim 1, characterized in that, The multidimensional features include one or more of the following: time-domain statistical features, time-series morphological features, periodic features, and energy features; The time-domain statistical features include one or more of the following: daily load peak value, valley value, peak-valley difference, mean, variance, skewness or kurtosis; The time-series morphological features include one or more of the following: extracting charging and discharging duration, frequency, load fluctuation rate, load difference entropy between adjacent time periods, or ramp rate; The periodic characteristics include one or more of the following: Fast Fourier Transform fundamental amplitude, harmonic proportion, or main periodic energy concentration. The energy characteristics include one or more of the following: daily cumulative load, charge / discharge energy difference, or time-of-day energy ratio.

3. The energy storage user identification method as described in claim 1, characterized in that, The construction of the feature correlation map based on the multidimensional features using the correlation coefficient method and sparsity strategy includes: Calculate the Pearson correlation coefficients among the features of each dimension; Using each dimension feature as a node, select the K groups of dimension features with the largest absolute value of Pearson correlation coefficient to establish undirected weighted edges between nodes, and use the Pearson correlation coefficient between dimension features as edge weights. Construct a feature correlation graph from each node and edge; Where K is an integer greater than or equal to 1.

4. The energy storage user identification method as described in claim 1, characterized in that, The training process of the feature correlation map variational autoencoder includes: The high-dimensional feature vector, composed of the feature correlation map and multi-dimensional features, is input into the encoder of the feature correlation map variational autoencoder to obtain the low-dimensional feature representation. The dimensionality-reduced feature representation is input into the decoder of the feature correlation graph variational autoencoder to reconstruct the original high-dimensional feature vector. The feature correlation graph variational autoencoder is then trained by minimizing the weighted sum of the reconstruction error and the dimensionality-reduced feature distribution constraint term, thus obtaining the trained feature correlation graph variational autoencoder.

5. The energy storage user identification method as described in claim 4, characterized in that, The loss function of the feature correlation map variational autoencoder is calculated as follows: In the formula, L This represents the value of the loss function. Indicates feature reconstruction error. Indicates the weighting coefficient. Indicates the KL divergence. Represents a given original high-dimensional load feature vector At that time, the low-dimensional feature representation learned by the encoder Z The conditional probability distribution; Representation of low-dimensional features Z The distribution of .

6. The energy storage user identification method as described in claim 4, characterized in that, The step of inputting a high-dimensional feature vector composed of feature correlation maps and multi-dimensional features into the encoder of a feature correlation map variational autoencoder to obtain a low-dimensional feature representation includes: The high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the encoder of the feature correlation map variational autoencoder. The graph convolutional layer of the encoder is used to aggregate information of nodes corresponding to each feature. The attribute value of each node is weighted and fused with the feature information of its neighboring nodes to obtain a node feature vector containing topological association information. The feature vectors of all nodes are integrated and compressed through a fully connected layer, and the mean vector and variance vector representing the overall feature distribution are output. By employing reparameterization techniques, low-dimensional feature representations are obtained by sampling from the dimensionality-reduced feature distribution based on the mean vector and variance vector.

7. The energy storage user identification method as described in claim 1, characterized in that, The training process of the user identification model includes: The training set is composed of low-dimensional feature representations and corresponding labels indicating whether the user is an energy storage user. A convolutional neural network is used to perform one-dimensional convolution on the low-dimensional feature representations in the training set to extract local temporal pattern features, resulting in a convolutional sequence. A bidirectional gated recurrent unit structure is used to perform bidirectional temporal modeling on the convolutional sequence to extract long-term dependency features. Finally, the probability of energy storage users is output through a fully connected layer and either Sigmoid or Softmax, and the binary classification result of energy storage users or non-energy storage users is obtained accordingly. With the goal of minimizing the binary classification recognition result and the corresponding label error, the parameters in the user recognition model are trained to obtain a trained user recognition model.

8. An energy storage user identification system based on multidimensional feature analysis and feature dimensionality reduction technology, characterized in that, include: The module includes a feature extraction module, a feature correlation graph module, a dimensionality reduction module, and a recognition module. The feature extraction module is used to select load curve data for multiple consecutive days from each of the user's multiple characteristic months, and extract multidimensional features from each load curve data respectively; The feature correlation graph module is used to construct a feature correlation graph based on the multidimensional features using a correlation coefficient method and a sparsity strategy. The dimensionality reduction module is used to input a high-dimensional feature vector composed of feature correlation maps and multi-dimensional features into a pre-trained feature correlation map variational autoencoder to obtain low-dimensional latent features. The identification module is used to input low-dimensional latent features into a pre-trained user identification model to obtain a binary classification identification result of energy storage users or non-energy storage users. The feature correlation graph variational autoencoder includes an encoder and a decoder. The encoder uses a graph neural network to aggregate information from the feature correlation graph, comprehensively considering the attributes of the feature nodes themselves and their correlation with neighboring nodes, and learns the embedded representation of each feature node. The user identification model adopts a convolutional neural network plus a bidirectional gated recurrent unit structure.

9. The energy storage user identification system as described in claim 8, characterized in that, The multidimensional features in the feature extraction module include one or more of the following: time-domain statistical features, time-series morphological features, periodic features, and energy features; The time-domain statistical features include one or more of the following: daily load peak value, valley value, peak-valley difference, mean, variance, skewness or kurtosis; The time-series morphological features include one or more of the following: extracting charging and discharging duration, frequency, load fluctuation rate, load difference entropy between adjacent time periods, or ramp rate; The periodic characteristics include one or more of the following: Fast Fourier Transform fundamental amplitude, harmonic ratio, or main periodic energy concentration. The energy characteristics include one or more of the following: daily cumulative load, charge / discharge energy difference, or time-of-day energy ratio.

10. The energy storage user identification system as described in claim 8, characterized in that, The feature correlation map module is specifically used for: Calculate the Pearson correlation coefficients among the features of each dimension; Using each dimension feature as a node, select the K groups of dimension features with the largest absolute value of Pearson correlation coefficient to establish undirected weighted edges between nodes, and use the Pearson correlation coefficient between dimension features as edge weights. Construct a feature correlation graph from each node and edge; Where K is an integer greater than or equal to 1.

11. The energy storage user identification system as described in claim 8, characterized in that, The training process of the feature correlation graph variational autoencoder in the dimensionality reduction module includes: The high-dimensional feature vector, composed of the feature correlation map and multi-dimensional features, is input into the encoder of the feature correlation map variational autoencoder to obtain the low-dimensional feature representation. The dimensionality-reduced feature representation is input into the decoder of the feature correlation graph variational autoencoder to reconstruct the original high-dimensional feature vector. The feature correlation graph variational autoencoder is then trained by minimizing the weighted sum of the reconstruction error and the dimensionality-reduced feature distribution constraint term, resulting in a fully trained feature correlation graph variational autoencoder.

12. The energy storage user identification system as described in claim 11, characterized in that, The loss function of the feature correlation graph variational autoencoder in the dimensionality reduction module is calculated as follows: In the formula, L This represents the value of the loss function. Indicates feature reconstruction error. Indicates the weighting coefficient. Indicates the KL divergence. Represents a given original high-dimensional load feature vector At that time, the low-dimensional feature representation learned by the encoder Z The conditional probability distribution; Representation of low-dimensional features Z The distribution of .

13. The energy storage user identification system as described in claim 11, characterized in that, The dimensionality reduction module inputs a high-dimensional feature vector composed of the feature correlation map and multi-dimensional features into the encoder of the feature correlation map variational autoencoder to obtain a low-dimensional feature representation, including: The high-dimensional feature vector composed of the feature correlation map and multi-dimensional features is input into the encoder of the feature correlation map variational autoencoder. The graph convolutional layer of the encoder is used to aggregate information of nodes corresponding to each feature. The attribute value of each node is weighted and fused with the feature information of its neighboring nodes to obtain a node feature vector containing topological association information. The feature vectors of all nodes are integrated and compressed through a fully connected layer, and the mean vector and variance vector representing the overall feature distribution are output. By employing reparameterization techniques, low-dimensional feature representations are obtained by sampling from the dimensionality-reduced feature distribution based on the mean vector and variance vector.

14. The energy storage user identification system as described in claim 8, characterized in that, The training process of the user identification model in the identification module includes: The training set is composed of low-dimensional feature representations and corresponding labels indicating whether the user is an energy storage user. A convolutional neural network is used to perform one-dimensional convolution on the low-dimensional feature representations in the training set to extract local temporal pattern features, resulting in a convolutional sequence. A bidirectional gated recurrent unit structure is used to perform bidirectional temporal modeling on the convolutional sequence to extract long-term dependency features. Finally, the probability of energy storage users is output through a fully connected layer and either Sigmoid or Softmax, and the binary classification result of energy storage users or non-energy storage users is obtained accordingly. With the goal of minimizing the binary classification recognition result and the corresponding label error, the parameters in the user recognition model are trained to obtain a trained user recognition model.

15. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the energy storage user identification method based on multidimensional feature analysis and feature dimensionality reduction technology as described in any one of claims 1 to 7 is implemented.

16. A computer-readable storage medium, characterized in that, It contains an execution program, which, when executed, implements the energy storage user identification method based on multidimensional feature analysis and feature dimensionality reduction technology as described in any one of claims 1 to 7.

Images