User type identification method and system based on multi-modal mixed features

Abstract

The invention provides a user type identification method and system based on multi-modal mixed features. The method comprises the following steps: deconstructing time sequence electric quantity data into a plurality of typical power consumption curves according to seasons and holidays; on the basis of the obtained typical power consumption curves, statistical features are extracted and include the electricity price sensitivity coefficient, the power consumption stability and the statistical features combined with the shape distance of the typical power consumption curves, and user type labels are generated through clustering; performing two-dimensional processing on the time sequence electric quantity data based on a Grubrum angle field and a recurrence plot, fusing statistical features, and constructing a user classification recognition model fused with multi-modal mixed features; and inputting the time sequence electric quantity data of the identified user into the constructed model to obtain a user type.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of proxy electricity purchasing user classification and recognition, and particularly relates to a user type recognition method and system based on multi-modal mixed features. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Proxy electricity purchasing users are usually classified according to factors such as electricity nature, voltage level, industry category, etc. The specific policies in different regions may differ slightly. Classification by electricity nature includes residential life electricity, agricultural production electricity, industrial and commercial electricity, etc. Classification by participation in the electricity market includes users directly participating in market transactions and grid enterprise proxy electricity purchasing users, etc. Proxy electricity purchasing companies (or "power selling companies") are market-oriented entities that provide electricity procurement services for users, mainly helping users (especially industrial and commercial enterprises) to obtain more favorable electricity prices or more flexible services through the electricity market.

[0004] Proxy electricity purchasing work puts new requirements on grid enterprise electricity purchasing strategy optimization, and power supply services face new challenges. Proxy electricity purchasing user type classification and recognition is the basis for proxy electricity purchasing companies to analyze the electricity market, manage users, and optimize electricity purchasing strategies. By deeply analyzing the electricity consumption behavior characteristics of proxy electricity purchasing users, accurately identifying the types of electricity users, developing differentiated energy use strategies, and optimizing the allocation of electricity resources, user satisfaction and the market competitiveness of proxy electricity purchasing companies can be greatly improved.

[0005] Proxy electricity purchasing user classification and type recognition is an important part of electricity demand side management. Among them, mining user electricity consumption behavior from multiple time series features and multiple load characteristics, and then realizing user type classification is the current common user classification method.

[0006] However, the current user electricity consumption behavior analysis and clustering method only considers a single typical day curve, usually ignores seasonal and holiday differences, and uses Euclidean distance or dynamic time warping method when calculating intra-cluster distance. However, the calculation method based on Euclidean distance cannot pay attention to the time lag characteristics between electricity consumption curves, and the dynamic time warping method increases the calculation cost when the amount of electricity consumption curve data is large.

[0007] Existing proxy electricity purchasing user type recognition work is mostly combined with peak-valley difference rate, electricity smoothness, daily load rate, etc. to realize user type recognition. However, in the face of massive user electricity consumption curve data, only relying on abstract features is insufficient to represent user electricity consumption behavior.

[0008] Therefore, in the type identification problem, only relying on the abstract features extracted based on the power consumption curve to classify the user types is simple, but cannot fully mine the user power consumption behavior. SUMMARY

[0009] To overcome the above deficiencies of the prior art, the present application provides a user type identification method based on multi-modal hybrid features, which fuses user power consumption behavior abstract features and full-time sequence features to realize user type classification.

[0010] To achieve the above object, one or more embodiments of the present application provide the following technical solutions: In a first aspect, a user type identification method based on multi-modal hybrid features is disclosed, comprising: deconstructing the time series power data into a plurality of typical power consumption curves according to seasons and holidays; extracting statistical features based on the obtained plurality of typical power consumption curves, including price sensitivity coefficient, power consumption stationarity, and clustering to generate user type labels in combination with the shape distance of the typical power consumption curve; using the Gram angle field and recursive graph to process the time series power data in two dimensions, and fusing the statistical features to construct a user classification identification model that fuses multi-modal hybrid features; inputting the time series power data of the identified user into the constructed model to obtain the user type.

[0011] As a further technical solution, it further comprises: first performing mean value processing on the time series power data, and then deconstructing it into a plurality of typical power consumption curves.

[0012] As a further technical solution, when obtaining the shape distance of the typical power consumption curve, the K-shape clustering algorithm is used to calculate the cross-correlation coefficient between two time series power data, and the shape distance is used as the core measurement index.

[0013] As a further technical solution, when calculating the price sensitivity coefficient, it is calculated based on the ratio of the power of each period to the total power of the day and the monthly proxy electricity purchase price.

[0014] As a further technical solution, when calculating the power consumption stationarity, it is calculated based on the power of each period and the average power of each period of the typical day.

[0015] As a further technical solution, the statistical features further include: daily minimum load rate and daily peak-valley difference rate.

[0016] As a further technical solution, the time series power data is processed in two dimensions based on the Gram angle field, comprising: linearly normalizing the time series power data to a set interval; mapping the normalized value to a polar coordinate system; Based on the polar coordinate system, the superposition relationship between time points is reflected by constructing a summation matrix through angle and cosine value calculation; the difference between dynamic changes is captured by constructing a difference matrix through angle difference and sine value calculation; Based on the summation matrix and the difference matrix, a Gramian matrix is generated; The generated Gramian matrix encodes the time correlation of time series into image pixels, and the diagonal line of the matrix represents time evolution, and the non-diagonal line represents cross-period interaction.

[0017] As a further technical solution, the time series power data is processed in two dimensions based on a recurrence plot, including: Embedding time series power data into high-dimensional m-phase space; Calculating the distance of all state vectors in the phase space to form a distance matrix; Convert the distance matrix to a binary recurrence matrix to mark similar states; Convert the recurrence matrix to an image, where darker colors represent more similar period states, and lighter colors represent less similar states; short lines parallel to the main diagonal line represent periodic or quasi-periodic behavior, vertical or horizontal lines represent long-term stagnation of states, and discrete points or patches represent chaotic dynamics or noise interference.

[0018] As a further technical solution, the user classification and recognition model uses a convolutional neural network to process the Gramian angle field and the recurrence plot, extracts and maps features in the image to each layer structure, and finally uses a fully connected layer to obtain one-dimensional image features. Specifically: The convolutional neural network includes three branches; The first branch one-dimensionizes abstract features such as user load characteristics, price sensitivity, and power stability; The second branch uses the GAFs image to capture global trends and provide angle information representing user load changes; The third branch uses the RP image to reveal the nonlinear dynamic characteristics in the time series; Before the output layer, a BiLSTM classifier is introduced to capture the bidirectional time sequence dependency of the fused features and accurately identify the user type.

[0019] In the second aspect, a user type recognition system based on multi-modal hybrid features is disclosed, including: A typical power curve acquisition module configured to deconstruct time series power data into multiple typical power curves according to seasons and holidays; A user type label generation module configured to extract statistical features including price sensitivity coefficient and power stability based on the obtained multiple typical power curves, and generate user type labels by clustering statistical features combined with shape distance of typical power curves; The user classification recognition model construction module is configured to use the gram angle field and the recurrent graph to process the time series power data in two dimensions, and fuse statistical features to construct a user classification recognition model fusing multi-modal mixed features. The recognition module is configured to input the time series power data of the recognized user into the constructed model to obtain the user type.

[0020] The above one or more technical solutions have the following beneficial effects: The technical scheme constructs the time series power data into a plurality of typical power consumption curves according to seasons and holidays, pays attention to the difference of power consumption behaviors in different seasons and holidays, enriches the user power consumption features, and facilitates subsequent extraction of statistical features.

[0021] Based on the obtained plurality of typical power consumption curves, statistical features are extracted, including a price sensitivity coefficient and power consumption stability, the statistical features are combined with a shape distance of the typical power consumption curve, a user type label is generated through clustering, and a K-shape time series clustering method is used to pay attention to sequence translation invariance of any two typical power consumption curves by means of a cross-correlation measurement calculation mode, so as to optimize a calculation method of the shape distance and a clustering centroid, and improve the calculation efficiency. The scheme is sensitive to the shape distance of the time series, can effectively capture the similarity of the power consumption curve, and realizes user type classification.

[0022] The time series power data is processed by using the gram angle field and the recurrent graph, and the statistical features are fused to construct a user classification recognition model fusing multi-modal mixed features. The model fuses abstract features such as load characteristics, price sensitivity, power consumption stability, and full-time sequence features extracted by the gram angle field and the recurrent graph, realizes rapid and accurate classification of the user type, and the like.

[0023] In the feature processing process of the technical scheme, typical power consumption curves of a user in spring, summer, autumn and winter and during holidays are calculated, the seasonal and holiday characteristics of the user power consumption behavior are analyzed based on the typical power consumption curves, a K-shape shape clustering method is used to consider sequence translation invariance of the power consumption curve, a cross-correlation measurement value of two curves is obtained, and then a shape distance of the two curves is obtained, and the time lag characteristics between the power consumption curves are analyzed.

[0024] The advantages of the additional aspects of the application will be partially given in the following description, partially will become obvious from the following description, or will be understood by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0025] The drawings accompanying the specification of the application form a part of the application and serve to provide further understanding of the application, the illustrative embodiments of the application and their descriptions serve to explain the application, and do not constitute an improper limitation on the application.

[0026] Figure 1 Sequence translation invariance diagram for K-shape calculation sequence shape distance in embodiments of the application; Figure 2 Flow chart for K-shape to realize user type clustering in embodiments of the application; Figure 3 Multi-modal mixed feature user type recognition model architecture diagram in embodiments of the application; Figure 4 Profile coefficient result diagram for each clustering case in embodiments of the application; Figure 5 Clustering center based on K-shape shape distance clustering in embodiments of the application; Figure 6 Visualization diagram of each clustering result in embodiments of the application; Figure 7 Four user gram angle field feature maps in embodiments of the application; Figure 8 Four user recursive feature maps in embodiments of the application; Figure 9 Ablation experiment confusion matrix result diagram in embodiments of the application. DETAILED DESCRIPTION

[0027] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application belongs.

[0028] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the application.

[0029] The embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0030] The traditional clustering scheme selects the statistical features based on the power consumption curve and the Euclidean distance between the features to realize user classification when clustering agent electricity purchasing users. However, this scheme has high requirements for the acquisition of typical user power consumption curves, and the statistical features abstracted based on the typical power consumption curves cannot fully represent the user's power consumption behavior, so the clustering effect needs to be considered.

[0031] The overall idea of the application is as follows: Multi-modal features include abstract features, gram angle fields and recursive graphs.

[0032] The classification model of the agent electricity purchasing user based on multi-modal feature fusion first acquires the abstract features of user load characteristics, electricity price sensitivity, and electricity flatness extracted by traditional means, then converts the normalized electricity consumption data to the polar coordinate space by means of the Gram angle field, calculates the Gram angle and field matrix according to the polar coordinates, and converts the matrix into an image to complete the two-dimensionalization; by means of recursion Figure 2 The Euclidean distance between data points is calculated by the recursive method, the 0, 1 binary data of the full data is realized by setting the distance threshold, and the recursive graph matrix is generated, which is converted into an image to complete the two-dimensionalization. Further, the periodicity feature, stability feature, mutation feature, time dependence feature, and trend feature of the user electricity consumption curve are obtained from the two-dimensional image by means of the neural network. The abstract features and two-dimensional features are fused as the electricity consumption behavior characteristics of the user by the full connection layer of the neural network, and the user type label obtained by K-shape clustering is combined as the classification target of the neural network, and the user type classification is realized by means of the bidirectional long short-term memory network classifier.

[0033] The clustering scheme based on the shape distance of time data first determines the typical electricity consumption curve of the user in different time periods, calculates the shape distance between the electricity consumption curves of the user based on this, and iteratively calculates the final clustering result with the goal of relatively small intra-cluster distance and relatively large inter-cluster distance.

[0034] With the improvement of computing power and optimization of algorithms, time series clustering methods are widely used in fields mainly based on time series features. The K-shape time series clustering method used in this scheme improves the calculation efficiency by optimizing the distance calculation and centroid calculation method, and is sensitive to the shape of time series, which can effectively capture the similarity of electricity consumption curves and realize user type classification.

[0035] The experimental results show that the K-shape clustering method based on shape distance can effectively divide the user types, and the user type recognition scheme can effectively recognize the user types based on electricity consumption behavior, and the recognition accuracy of various users is more than 95%, which provides effective support for the demand side management of the agent electricity purchasing company.

[0036] Embodiment one The embodiment discloses a user type recognition method based on multi-modal mixed features, comprising: Step one: according to the time-of-use electricity price division policy, the time series electricity data is decomposed into five typical electricity consumption curves of spring, summer, autumn, winter, and holidays; Step two: extract the statistical features of price sensitivity coefficient and electricity flatness based on the typical electricity consumption curve, and generate user type labels by clustering based on the shape distance of the typical electricity consumption curve; Step three: using the Gram angle field and recursive graph to process one-dimensional data in two dimensions, and fusing the electricity price sensitivity coefficient, electricity consumption level stability and other characteristics, a user classification recognition model is constructed by fusing multi-modal mixed features.

[0037] In step one, the processing of typical electricity consumption curves is based on the time-of-use electricity price policy published by a certain power company to determine the division of spring, summer, autumn and winter months, and to superimpose the independent time period of statutory holidays. Each user forms 5 typical electricity consumption scenes in each time period. Then, the average value of the 24-point electricity consumption data in each period is processed to eliminate abnormal value interference, and then the typical electricity consumption curve of the user in each season divided by the time-of-use electricity price is obtained. The typical electricity consumption curves of the users participating in clustering are spliced according to the spring, summer, autumn and winter holiday scenes in turn to form a user data set.

[0038] User electricity consumption data set participating in clustering (1) Wherein, i is the number of users, j is the electricity consumption period, k is the division of electricity consumption scene, .

[0039] In time series analysis, the distance calculation method is a key influencing factor of clustering effect. The Euclidean distance (ED) has high calculation efficiency, but needs to align the sequence, and the effect of the sequence with time shift data is poor; dynamic time warping (DTW) realizes non-alignment matching through planning method, and is suitable for analysis of sequences with different lengths, but the calculation cost is high. In practical application, ED is often used in real-time monitoring scene, and DTW algorithm is suitable for fine analysis.

[0040] And the K-shape clustering algorithm calculates the cross-correlation coefficient between the time series data of the typical electricity consumption curves of any two users, and uses the shape distance as the core measurement index, which can still effectively calculate the data with different lengths or not effectively aligned.

[0041] Given two typical user electricity consumption curves , Time series, the sequence length is m (in this application, the length of the typical electricity consumption curve m=120), to realize the translational invariance, invariant, Sliding with s as the step, and calculating the difference between each step and The K-shape calculates the translational mode of the two sequences as follows.

[0042] (2) (3) ​To clearly demonstrate the translation invariance calculation, Figure 1 Take m=3 as an example, assuming the sequence constant, Slide with a step size of s. In the example, s is 1. There are 5 possible values ​​for s. For sequence The total number of moves, is the alignment position of the cross-correlation coefficient.

[0043] The calculation method of the cross-correlation measure (CC) is defined as shown in formula (4), and the calculation method of each step of the two sequences is shown in formula (5): (4) (5) The shape-based distance (SBD) formula is defined as shown in formula (6).

[0044] (6) Where, Typical electricity consumption curve for user 1 The autocorrelation coefficient of Typical electricity consumption curve for user 2 The autocorrelation coefficient is defined based on the cross-correlation. The more similar the shapes of the two sequences are, the larger the CC is. 1-max(CC) is used as the shape distance. That is, the more similar the shapes are, the smaller the SBD is. Due to the amplitude differences between the sequences, this embodiment uses the normalized cross-correlation coefficient when calculating sequence correlation. The range of values ​​is [-1, 1]. Therefore, the range of SBD is [0, 2].

[0045] When adjusting the sequence cluster center according to the distance logic, the shape change trend may not be accurately expressed due to sequence misalignment. In this implementation case, the cluster center calculation method is designed with the goal of maximizing the normalized similarity between the cluster center and the typical electricity consumption curve sequence of each user in the data cluster, as shown in formula (7).

[0046] (7) Where, is the cluster center of the k-th cluster user, is the set of typical electricity consumption curve sequences of the k-th cluster of users.

[0047] Combining linear algebra methods, Equation (7) can be expanded as follows: (8) In formula (7) and formula (8), in order to solve the cluster center , set here , , I is a unit matrix, and O is an all-one matrix.

[0048] The clustering centers of each type of user are calculated by iteration, and all the power consumption curve sequences are re-assigned to different clusters according to the distance between the typical power consumption curve sequence of each user and the new clustering center, until the iteration label no longer changes, or the maximum iteration number is reached. The shape clustering process based on the K-shape algorithm is shown in Figure 2 .

[0049] According to the demand of time series clustering, the silhouette coefficient is improved into formula (9) and formula (10).

[0050] Silhouette Score (SS): (9) (10) wherein, is the cohesion degree of the power consumption curve sequence of user i, is the separation degree of the power consumption curve sequence of user i.

[0051] (11) (12) Elbow Method (EM): (13) (14) In the formula, K is the optimal clustering number of the typical power consumption curve of the user determined based on the elbow method.

[0052] On the basis of K-shape shape clustering, the power consumption type label of different users is obtained, The clustering result obtained based on the K-shape method is taken as the user type label, and here the user type is not specifically defined, but only clustering 0, clustering 1, clustering 2, …, clustering k are shown. Further combining the user abstract features obtained from formula (15-18), and combining the Gram angle field and the recursive graph to obtain the global features of the time series data, and sending them into the combined neural network model of CNN and BiLSTM, the accurate identification of the power consumption type of the new user can be realized.

[0053] In the power market transaction, only the clustering of electricity consumption curve cannot clearly describe the user's electricity consumption behavior. In the area where time-of-use electricity price is implemented, considering the user's sensitivity to electricity price, as well as the user's own electricity smoothness and load difference rate and other characteristics, the user's electricity consumption characteristics can be further enriched. The above characteristics are connected from beginning to end and flattened into a one-dimensional vector as one of the input features of the BiLSTM model.

[0054] Electricity price sensitivity coefficient : (15) In the formula, E represents the ratio of 24-hour electricity quantity to total daily electricity quantity, and P represents the monthly proxy electricity time-of-use electricity price. E t is the ratio of electricity quantity in each time period after adjustment of time-of-use electricity price to total daily electricity quantity; E t-1 is the ratio of electricity quantity in each time period before adjustment of time-of-use electricity price to total daily electricity quantity; P t is the monthly proxy electricity time-of-use electricity price after adjustment of time-of-use electricity price, P t-1 is the monthly proxy electricity time-of-use electricity price before adjustment of time-of-use electricity price.

[0055] Electricity smoothness : (16) In the formula, is the 24-hour electricity quantity, is the 24-hour time period average electricity quantity of a typical day, , n = 24.

[0056] Daily minimum load rate : (17) Daily peak-valley difference rate : (18) Electricity time series data usually presents a certain fluctuation trend, which is influenced by multiple factors such as time period, date, season, event, etc., forming complex periodic and trend changes, while the typical day electricity consumption curve often ignores some detailed characteristics in the total data quantity.

[0057] Therefore, according to the characteristics of power consumption data, the global features of power consumption data are extracted by means of Gramian Angular Fields (GAFs) and Recurrence Plot (RP). Gramian Angular Fields are constructed by polar mapping and Gramian matrix, which retain the global dependence of the typical power consumption curve of the user, and facilitate the analysis of the time series data of the power consumption curve combined with the deep learning model, providing a new perspective for user type recognition. Recurrence Plot reveals the internal structure and dynamic pattern of the typical power consumption curve data in an intuitive visual way. By drawing the mutual relationship between each point in the typical power consumption curve, the periodicity and trend of the user's power consumption curve can be better understood.

[0058] Gramian Angular Fields (GAFs): Data normalization: the original time series power consumption data is linearly normalized to the interval [-1, 1].

[0059] (19) Polar coordinate conversion: the normalized value is mapped to the polar coordinate system. The angle of the normalized user power consumption data after mapping to the polar coordinate system is calculated by the inverse cosine function. The radius is associated with the time step, so that the time sequence is implicitly contained in the polar coordinate.

[0060] (20) Gramian matrix generation: by calculating the cosine value of the angle, a summation matrix is constructed to reflect the superposition relationship between time series points; by calculating the sine value of the angle difference, a difference matrix is constructed to capture the difference in dynamic changes.

[0061] (21) The summation Gramian angular field matrix and the difference Gramian angular field matrix of are generated respectively, and are converted into images. The time series correlation of the time series power consumption curve is encoded as image pixels. The diagonal line of the matrix represents the time lapse, and the non-diagonal line represents the cross-period interaction. The periodicity, trend and dynamic change of the Gramian angular field are obtained by the neural network.

[0062] Recurrence Plot (RP): Phase space reconstruction: given the time series power consumption data , the one-dimensional time series data X is mapped to the vector representation in the high-dimensional m phase space : (22) wherein, is the time delay.

[0063] Distance matrix: Calculate the distance of all state vectors in m-dimensional phase space, form a distance matrix : (23) Recurrence matrix: Convert the distance matrix into a binary recurrence matrix, mark similar states.

[0064] (24) where is a threshold value, usually take a fixed proportion of 5% of the radius of the phase space.

[0065] Convert the two-dimensional recurrence matrix According to the set threshold value, the matrix is visualized as an image, and the short lines parallel to the main diagonal line represent periodic or quasi-periodic behavior, the vertical or horizontal lines represent long-term stagnation of states, and the discrete points or patches represent chaotic dynamics or noise interference.

[0066] Convolutional Neural Network (CNN) can obtain global, trend and other features in images according to the human biological visual system, and can be used as a feature extraction tool for extracting GAFs and RP graphs. CNN mainly includes convolutional layer, pooling layer and fully connected layer, which extracts and maps the features in the Gram angle field graph and the recurrence graph to the structure of each layer through the cooperation of the convolutional layer and the pooling layer, and finally uses the fully connected layer to obtain one-dimensional image features. The core calculation method of CNN is as follows: Convolution operation: (25) In the formula, X is the Gram angle field graph or the recurrence graph of the input layer, is the hidden layer feature obtained by the lth convolution kernel; , and represent the weight and bias of the lth convolution kernel respectively, and s is the size of the convolution kernel; represents the activation function, and ReLU function is selected in this application.

[0067] Max-pooling operation: (26) In the formula, m, n are the area covered by the corresponding pool kernel, is the pooling layer feature obtained by the lth pool kernel.

[0068] The top of the CNN is composed of a fully connected layer and a classifier. The fully connected layer realizes the flattening of the feature map into a one-dimensional vector, and then the classifier obtains the recognition result. In this embodiment, the CNN is mainly used to obtain one-dimensional image features.

[0069] BiLSTM (Bidirectional Long Short-Term Memory) is a special kind of recurrent neural network. BiLSTM processes one-dimensional sequence through two LSTM layers, one layer processes data from the beginning to the end of the sequence, and the other layer processes data from the end to the beginning, which is used to capture bidirectional time series dependence and improve the model's ability to capture user types. The BiLSTM network structure layer has been described in detail in the literature, and will not be repeated here.

[0070] The main structure of the BiLSTM classifier is shown in equation (27).

[0071] (27) where X represents Figure 3 the flattened features of the last layer of the three branches, y represents the probability that the user corresponding to the feature X belongs to each category, , and are the representations of different features, is the weight matrix of the classification layer.

[0072] Based on the foregoing feature analysis, this embodiment proposes a feature fusion method as shown in Figure 3 , which combines user electricity abstract features and global trend features extracted based on Gram angle field diagram and recurrence diagram to realize the recognition of power user types. In this embodiment, the global feature extraction capability of the CNN model is used to extract the image features of the Gram angle field diagram and the recurrence diagram. The abstract feature input branch one-dimensionally and expands the abstract features such as user load characteristics, price sensitivity, and electricity smoothness; the Gram angle field diagram branch uses a CNN module to capture global trends and provide angle information representing user load changes; the recurrence diagram branch also uses a CNN module to extract global features in the recurrence diagram and reveal the nonlinear dynamics features in the time series; before the BiLSTM classifier, the one-dimensional features of the three branches are fused and input into the BiLSTM classifier, through capturing the bidirectional time series dependence of the fused features, the probability value of the input user electricity curve belonging to each clustering result is input, and the clustering result with the maximum probability is the type to which the input user belongs.

[0073] Test verification In this embodiment, the experimental data comes from 1884 users in a city in Shandong Province, 2023, a total of 12 months of daily 24-point historical electricity data, which is used for calculation and analysis verification of the method proposed in the foregoing. The data of all users are calculated according to the price sensitivity coefficient and the peak-valley difference rate, and the user type clustering is used to obtain the user label. In the user type recognition, 1500 households are used as the training set, and 384 households are used as the test set.

[0074] In this embodiment, the abstract features such as user seasonal price sensitivity coefficient and electricity smoothness calculated are shown in Table 1.

[0075] Table 1 User abstract features

[0076] This example uses the silhouette coefficient and elbow method to determine the optimal number of clusters, adaptively adjusting the number of clusters from 3 to 8. Based on typical daily electricity consumption curves for five time periods, users with similar electricity usage behaviors are clustered using the K-shape clustering method.

[0077] like Figure 4 The clustering result determination method shown in the figure shows that the error of the intra-cluster cluster sum of squares is the smallest when the number of clusters is 8. However, too many clusters will lead to problems such as complex type descriptions, and the silhouette coefficient is closer to no correlation. Therefore, the final number of clusters is selected as 4 based on the silhouette coefficient. The cluster centers in each cluster are as follows Figure 5 shown.

[0078] from Figure 5 The cluster centers of each cluster show that they are significantly recognizable based on the typical electricity consumption curves of spring, summer, autumn, winter and holidays. The effectiveness of the K-shape clustering results can be further verified from the shape of the cluster centers. Figure 6 shown.

[0079] from Figure 6 The user clustering results show that in the two-dimensional space, in the results of the clustering scheme in this embodiment, the recognition of three types of users can be clearly observed. The fourth type of user overlaps with the other three types of users. Figure 5 The electricity consumption curves of cluster 0 in each time period overlap with the curves of the other three types of users in different time periods. It can be judged that all users are identifiable in three-dimensional space.

[0080] The clustering results in this section serve as labels for subsequent user type identification. The 1884 users are divided into four categories based on the shape clustering results. The number of users in each category is shown in Table 2: Table 2 User type labels and number of user types

[0081] The graphical features used in this section are derived directly from the annual electricity consumption curves for each user type. The 8760-point one-dimensional annual electricity consumption curve is converted into a two-dimensional image, thereby capturing global trend characteristics and nonlinear dynamic features within the user's electricity consumption curve. The recurrence graph and Gram's angle field graph for each user type are shown below.

[0082] like Figure 7The Gram angle field diagram features shown reveal the complex dynamics of the user's annual electricity consumption curve. The main diagonal line represents the instantaneous value of the original time series data over time, and the non-diagonal line represents the period interaction, representing the time lag, similarity, etc. between time points. Figure 7 (b) and Figure 7 (d) show monthly or quarterly regular stripes, i.e. strong periodicity at long time scales; Figure 7 (a) and Figure 7 (c) show stripes evenly distributed, showing daily regularity, i.e. strong periodicity at each daily period.

[0083] The recurrence plot features shown reveal the deterministic and periodic behavior of the user's annual electricity consumption curve. The short lines or grids parallel to the main diagonal line represent the periodic or quasi-periodic behavior of electricity consumption behavior, and the evenly distributed points or plates represent chaotic dynamics or noise interference. Figure 8 (a) and Figure 8 (c) show evenly distributed points and clear grid features, i.e. quasi-periodic and chaotic characteristics at short time scales; Figure 8 (b) and Figure 8 (d) have clear diagonal and grid features, i.e. strong periodicity at long time scales. Figure 8 To verify the reliability of the multi-modal fusion feature user type recognition model, three evaluation indicators, precision (Precision), recall (Recall), and F1 score (F1-score), are selected to evaluate the performance of the user recognition model in avoiding false positives and false negatives.

[0084]

[0085] (28) (29) (30) Where TP is the correctly predicted positive example; FP is the false positive; FN is the false negative; and TN is the correctly predicted negative example.

[0086] To verify the effect of the multi-modal fusion feature user type recognition model proposed in this embodiment, based on the Windows 11 system, the processor is 13th Gen Intel(R) Core(TM) i7-13650HX, and the experiment is carried out in the Python 3.10.3 programming environment using the Tensorflow 2.10.0 framework.

[0087] The convolutional network structure for two-dimensional image feature extraction is shown in Table 3.

[0088] Table 3 Convolutional network structure​

[0089] The BiLSTM structure used for user classification consists of two LSTMs in the forward and reverse directions, with the number of neurons set to 16. After 50 iterations of model training, the network loss and classification accuracy are almost stable, indicating that the network weights are stable.

[0090] The ablation experiment is used to verify the effectiveness of the feature fusion scheme proposed in this example. To ensure the rationality of the ablation experiment, only the feature parameters are changed without changing the network structure. Figure 9 shown.

[0091] Depend on Figure 9 From the confusion matrix results in (a), we can see that the model with the addition of abstract features and global image features has the highest recognition accuracy. The recognition accuracy for type 0 and type 2 both reaches 100%, and the recognition accuracy for type 1 and type 3 also reaches about 95%, indicating that the fusion features can basically accurately describe the user's electricity consumption behavior. Figure 9 (b) It can be seen that the accuracy of using only two-dimensional image features is relatively good, with the recognition accuracy of type 2 reaching 100% and the recognition accuracy of type 0 exceeding 95%. The recognition accuracy of types 1 and 3 is relatively poor, indicating that without adding abstract features and only using image detail features, the representation of the electricity consumption behavior of types 1 and 3 is too refined and insufficient to support the identification of these two types of users; Figure 9 (c) It can be seen that the recognition effect of using only abstract features of electricity usage behavior is the worst. The recognition accuracy of type 2 is 100%, the recognition accuracy of type 0 is 73.2%, and the recognition accuracy of types 1 and 3 is only about 50%, indicating that using only abstract features is not sufficient to identify all types of users.

[0092] comprehensive Figure 9 It can be seen that under various feature conditions, the user classification and recognition model has the best recognition effect on type 0 and type 3, which shows that the user type label acquisition method based on K-shape shape clustering has certain advantages.

[0093] The precision, recall, and F1-score of each case in the ablation experiment are shown in Table 4.

[0094] Table 4 Evaluation indicators

[0095] Example 2 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0096] Embodiment three The purpose of this embodiment is to provide a computer-readable storage medium.

[0097] A computer-readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the above method.

[0098] Embodiment four The purpose of this embodiment is to provide a user type recognition system based on multi-modal hybrid features, comprising: The typical power consumption curve acquisition module is configured to: deconstruct the time-series power data into a plurality of typical power consumption curves according to seasons and holidays; The user type label generation module is configured to: based on the obtained plurality of typical power consumption curves, extract statistical features including price sensitivity coefficients and power consumption smoothness, and combine the statistical features with the shape distance of the typical power consumption curves to generate user type labels through clustering; The user classification recognition model construction module is configured to: use the Gram angle field and the recursive graph to process the time-series power data in two dimensions, and fuse the statistical features to construct a user classification recognition model that fuses multi-modal hybrid features; The recognition module is configured to: input the time-series power data of the identified user into the constructed model to obtain the user type.

[0099] Embodiment five The purpose of this embodiment is to provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods and functions involved in any of the above embodiments. The steps and methods involved in the above embodiments correspond to Embodiment One, and the specific implementation can refer to the relevant description of Embodiment One. The term "computer-readable storage medium" should be understood to include a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry instruction sets for execution by a processor and cause the processor to perform any of the methods in the present application.

[0100] Those skilled in the art should understand that the above modules or steps of the present application can be realized by a general computer device, alternatively, they can be realized by program code executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be made into individual integrated circuit modules, or a plurality of modules or steps among them can be made into a single integrated circuit module. The present application is not limited to any specific combination of hardware and software.

[0101] The above describes the specific embodiments of the present application in combination with the drawings, but is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A user type identification method based on multimodal hybrid features, characterized by: include: Deconstruct time-series electricity consumption data into multiple typical electricity consumption curves according to seasons and holidays; Based on multiple typical electricity consumption curves, statistical features are extracted, including electricity price sensitivity coefficient and electricity consumption stability. These statistical features are combined with the shape distance of the typical electricity consumption curves to cluster and generate user type labels. Using Gram's angle field and recurrence graph to process time-series power data in two dimensions, and integrating statistical features to build a user classification and recognition model that integrates multimodal hybrid features; The time series electricity consumption data of the identified users is input into the constructed model to obtain the user type.

2. The user type identification method based on multimodal hybrid features according to claim 1 is characterized in that: When obtaining the shape distance of a typical electricity consumption curve, the K-shape clustering algorithm is used to calculate the correlation coefficient between two time series electricity data, and the shape distance is used as the core measurement indicator.

3. The user type identification method based on multimodal hybrid features according to claim 1 is characterized in that: The electricity price sensitivity coefficient is calculated based on the ratio of electricity consumption in each period to the total electricity consumption for the day and the monthly agency electricity purchase price; The calculation of the electricity consumption stability is based on the electricity consumption in each time period and the average electricity consumption in each time period of a typical day; The statistical characteristics also include: daily minimum load rate and daily peak-to-valley difference rate.

4. The user type identification method based on multimodal hybrid features according to claim 1 is characterized in that: The time series power data is processed in two dimensions based on the Gram angle field, including: Normalize the time series power data linearly to the set range; Map the normalized values ​​to the polar coordinate system; Based on the polar coordinate system, the cosine value is calculated by the angle sum to construct a sum matrix to reflect the superposition relationship between time series points; the sine value is calculated by the angle difference to construct a difference matrix to capture the dynamic change difference; Gramian matrices generated based on sum and difference matrices; The generated Gramian matrix encodes the temporal correlation of time series into image pixels, with the matrix diagonal representing the passage of time and the off-diagonal reflecting the interaction across time periods.

5. The user type identification method based on multimodal hybrid features according to claim 1 is characterized in that: Use recursive graphs to process time series power data in two dimensions, including: Embed time series electricity data into high-dimensional m-phase space; Calculate the distances of all state vectors in the phase space to form a distance matrix; Convert the distance matrix into a binary recursive matrix and mark similar states; The recursive matrix is ​​converted into an image. The darker the color, the more similar the state of the time period, and the lighter the color, the less similar it is. Short lines parallel to the main diagonal indicate periodic or quasi-periodic behavior, vertical or horizontal lines indicate long-term stagnation of the state, and discrete points or patches reflect chaotic dynamics or noise interference.

6. The user type identification method based on multimodal hybrid features according to claim 1 is characterized in that: The user classification and recognition model uses a convolutional neural network to process the Gram angle field and recursive graph, extracts features from the image and maps them to each layer structure, and finally uses a fully connected layer to obtain one-dimensional image features. Specifically: The convolutional neural network consists of three branches; The first branch quantifies the abstract features such as user load characteristics, electricity price sensitivity, and electricity consumption stability into one dimension; The second branch uses GAFs images to capture global trends and provide angular information that characterizes user load fluctuations. The third branch uses RP images to reveal the nonlinear dynamic characteristics in the time series; Before the output layer, a BiLSTM classifier is introduced to accurately identify user types by capturing the bidirectional temporal dependencies of the fused features.

7. A user type identification system based on multimodal hybrid features, characterized by: include: The typical power consumption curve acquisition module is configured to: deconstruct the time series power consumption data into multiple typical power consumption curves according to seasons and holidays; The user type label generation module is configured to: extract statistical features, including electricity price sensitivity coefficient and electricity consumption smoothness, based on multiple typical electricity consumption curves obtained, and cluster the statistical features and shape distance of the typical electricity consumption curves to generate user type labels; The user classification and recognition model construction module is configured to: use Gram's angle field and recursive graph to process the time series power data in two dimensions, fuse statistical features, and build a user classification and recognition model that integrates multimodal hybrid features; The identification module is configured to: input the time series power data of the identified user into the constructed model to obtain the user type.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method described in any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 6 are performed.

Images