Similar day clustering method and system based on EV charging station operating status parameters

Abstract

The present invention relates to the technical field of capacity regulation of electric vehicle charging stations, and specifically to a similar day clustering division method and system based on the operating status parameters of EV charging stations. The state parameters of the operating status of the charging station on each typical day are used as the data basis for clustering, and a state parameter matrix describing the operating status of the charging station is defined; the state parameters are classified into the first layer of scenarios at the macro level, and then the scenarios are clustered using a deep learning algorithm to obtain a scenario set consisting of multiple different operating scenarios; a deep convolutional embedding clustering model is constructed and trained, and a convolutional autoencoder is used to extract features and reduce the dimension of the original data of the operating status parameters of the charging station in the scenario set, and then the flattened and reduced-dimensional data is clustered using the K-means algorithm, and data visualization is performed based on the t-random neighbor embedding algorithm. The present invention improves clustering accuracy and efficiency, and solves the dimensionality disaster problem of traditional clustering algorithms when facing high-dimensional complex nonlinear data.

Description

Technical Field

[0001] The present invention relates to the technical field of capacity regulation of electric vehicle charging stations, and in particular to a similar day clustering division method and system based on operating state parameters of EV (Electric Vehicle) charging stations. Background Art

[0002] Clustering similar days based on key factors affecting the operation status of the distribution network, such as distribution network load, distributed renewable energy output, and meteorological conditions, is a key prerequisite and effective means to improve the accuracy of load or renewable energy output forecasts and improve the efficiency of distribution network operation and management.

[0003] First, the existing clustering of the operating status of distribution networks is mainly based on the clustering of similar days based on factors such as the output characteristics of new energy sources such as photovoltaics and wind power, load fluctuation characteristics, and climatic conditions. There is still a lack of clustering for similar days in the operating status of electric vehicle charging station groups.

[0004] Secondly, traditional distance clustering, hierarchical clustering and other algorithms have limited computing power and poor results when facing high-dimensional complex nonlinear data, and it is difficult to capture the global and local nonlinear characteristics of the data. The clustering of the operating status of the charging station group is based on massive historical measurement data of charging stations. The data is high-dimensional, large-volume, and has obvious nonlinear characteristics, so traditional clustering algorithms are no longer used.

[0005] At present, the above research on similar day clustering has made some progress. As a high-power power electronic type flexible and controllable load, electric vehicle charging stations have certain voltage regulation capabilities. The days with similar operating status of charging stations are clustered and divided, and then the adjustable capacity of charging stations is predicted based on similar scene sets. By predicting the adjustable capacity of charging stations, prior information can be provided for power quality regulation and optimal scheduling of distribution networks, helping to make accurate decision-making plans. An effective way to improve the accuracy of prediction is to cluster different typical days with similar adjustable capacity time distribution characteristics, and then predict based on label data similar to the day to be predicted. However, research in this area is still blank.

[0006] The main factor affecting the adjustable capacity of electric vehicle charging stations is the charging demand of electric vehicle users. When the user charging demand is high, the adjustable capacity of the station is less. Charging demand is closely related to climate conditions and day factors (weekdays and holidays). Simply clustering similar days of charging station operation based on these factors does not reflect the differences within the season or within the same type of day; in addition, the division of similar days based on K-means, hierarchical clustering, density clustering and other methods, although the algorithm structure is simple and easy to implement, but for high-dimensional nonlinear data, the clustering effect is poor and the calculation time is long, which is obviously not applicable.

[0007] In summary, the prior art has the following deficiencies:

[0008] 1) Traditional clustering methods (such as K-means, hierarchical clustering, etc.) are usually based on Euclidean distance or simple similarity metrics, and they have weak clustering capabilities for high-dimensional, nonlinear data. These methods cannot effectively capture the deep features of the data, especially when the data has a complex structure or nested distribution, which may lead to inaccurate classification boundaries. DCEC automatically extracts low-dimensional embedded representations of data through deep convolutional networks, thereby effectively identifying and clustering complex high-dimensional nonlinear data.

[0009] 2) Traditional clustering methods usually rely on pre-extracted features and assume that these features can already well describe the distribution of the data. However, the separation of feature extraction and clustering processes may lead to information loss or insufficient features, especially when the original data quality is not high, which will limit the clustering performance. DCEC embeds feature extraction and clustering into the same model, and optimizes feature representation and clustering objectives through end-to-end training, thereby improving clustering effect and efficiency.

[0010] 3) Many existing clustering methods cannot effectively utilize the local structure of data, resulting in the inability to discover the fine-grained distribution of data. DCEC captures the local structural information of data through convolutional neural networks while retaining global embedding features, significantly improving the modeling ability of complex data distribution, and is suitable for scenarios with strong local correlations such as time series. Summary of the invention

[0011] In view of the above problems, the purpose of the present invention is to provide a similar day clustering method and system based on the operating status parameters of EV charging stations, which improves the clustering accuracy and computational efficiency and solves the dimensionality disaster problem of traditional clustering algorithms when facing high-dimensional complex nonlinear data. The technical solution is as follows:

[0012] The similar day clustering method based on the operating status parameters of EV charging stations includes the following steps:

[0013] Step 1: Take the total charging power of the charging station on each typical day, the number of idle fast charging piles, the number of idle slow charging piles, the number of electric vehicles charging at the same time, and the ambient temperature as the state parameters describing the operating status of the charging station, as the data basis for clustering, and define the state parameter matrix describing the operating status of the charging station;

[0014] Step 2: Based on the different charging demands of electric vehicles on weekdays and holidays, rainy and snowy days, and non-rainy and snowy days, the state parameters are first classified into the first-level scenarios at the macro level, and then the deep learning algorithm is used to cluster the scenarios based on the first-level scenario classification results to obtain a scenario set consisting of multiple different operating scenarios;

[0015] Step 3: Construct and train a deep convolutional embedding clustering model, which includes an autoencoder part and a clustering part; the autoencoder part uses a convolutional autoencoder to extract features and reduce the dimension of the original data of the operating status parameters of the charging stations in the scene set; the clustering part uses the K-means algorithm to perform similar day clustering on the flattened and reduced-dimensional data.

[0016] Furthermore, in step 1, d A typical day, at time t n The operating status of a charging station is expressed as:

[0017] (1);

[0018] in, n Number the charging station. is a collection of charging stations; , , , as well as Respectively represent d In a typical day, t Moment n The total charging power of each charging station, the number of idle fast-charging charging piles, the number of idle slow-charging charging piles, the number of electric vehicles charging at the same time, and the ambient temperature.

[0019] Furthermore, step 2 is specifically as follows:

[0020] Step 2.1: First level classification:

[0021] All typical days are divided into working days and holidays; working days and holidays are further divided into rainy and snowy days and non-rainy and snowy days, and finally four label classes are obtained: rainy and snowy days on working days are class I, rainy and snowy days on holidays are class II, non-rainy and snowy days on working days are class III, and non-rainy and snowy days on holidays are class IV; each label class contains multiple typical days;

[0022] Step 2.2: Second layer clustering:

[0023] In each label class, the label class is clustered based on the charging station status parameter matrix of each typical day. Several similar days constitute an operation scenario, and the operation scenario set contains multiple different operation scenarios.

[0024] Furthermore, in step 2.2, for rainy and snowy days on weekdays, i.e., typical days with label class I, the clustering method is as follows:

[0025] The operating status of each typical day of the charging station in category I is characterized by the status parameters at each time:

[0026] (2);

[0027] In the formula, Indicates the first d Charging station operation status for one working day; Indicates the first d working days m Charging station operation status parameters at sampling time; T is the transposition symbol; It is expressed as follows:

[0028] (3);

[0029] In the formula, Indicates the first d working days m The sampling time n Charging station operating status parameters;

[0030] By clustering each typical day in class I, a scenario set consisting of multiple different operating scenarios is formed:

[0031] (4);

[0032] In the formula, To run the scenario set; For the l An operation scenario consists of several operation status parameters of similar days, and is expressed as:

[0033] (5);

[0034] In the formula, i Indicates the number of similar days among rainy and snowy working days.

[0035] The present invention uses convolutional autoencoders to extract features and reduce the dimension of raw data, and then clusters the reduced-dimensional data through the K-means algorithm. There are four steps from data sample input to clustering result output: the first is data preprocessing, the second is model construction and training, the third is raw data feature extraction and dimension reduction, and the fourth is clustering and visualization.

[0036] Furthermore, in step 3, data preprocessing is first performed:

[0037] The robust normalization method is used to normalize the charging station operation status parameter samples in the scene set, as shown in the following formula:

[0038] (6);

[0039] In the formula, x m Indicates that at the sampling timem Charging station data; median( x m ) is the median of the measured data, IQR( x m ) represents the difference between the 75th percentile and the 25th percentile, represents the normalized data; It is the charging station dataset at all sampling moments;

[0040] The deep convolutional embedding clustering model is then trained:

[0041] In the model training phase, the weights, biases, and cluster centers of each layer of the model are iteratively updated to jointly optimize the joint loss function consisting of the reconstruction loss of the autoencoder and the clustering loss; the joint loss function L Defined as:

[0042] (7);

[0043] In the formula, L r is the reconstruction loss, L c is the clustering loss, α , β are the weights of the corresponding loss functions respectively;

[0044] The mean square error and KL Divergence is used to characterize the reconstruction loss L r and clustering loss L c :

[0045] (8);

[0046] In the formula, represents the normalized original data before reconstruction, Represents the reconstructed data; D is the total number of sample days;

[0047] (9);

[0048] In the formula, Y dj Representation sample d Belong to j The target probability distribution of the class, H dj Represents the model prediction sample d Belong to j Probability distribution of classes; k is the number of classes; P and QFor the two probability distributions set; express KL Divergence;

[0049] Probability distribution of model predictions H dj for:

[0050] (10);

[0051] In the formula, z d The sample data is obtained by feature extraction and dimension reduction of the original data through the encoding layer; u j For the j The cluster center of the class; τ is the degree of freedom of the student-t distribution; is the shape parameter;

[0052] Target probability distribution Y dj The expression is:

[0053] (11);

[0054] Then, a mini-batch gradient descent method is used to optimize the joint loss function L , when the training reaches the maximum number of iterations, the model training is completed.

[0055] Finally, based on the trained model, the original data is subjected to feature extraction and dimensionality reduction, and finally clustering is performed in the low-dimensional space.

[0056] A similar day clustering partitioning system based on EV charging station operation status parameters, including a charging station operation status parameter definition unit, a two-layer clustering framework, and a charging station group operation status partitioning unit based on deep convolution embedding clustering;

[0057] The charging station operation status parameter definition unit uses the total charging power of the charging station on each typical day, the number of idle fast charging piles, the number of idle slow charging piles, the number of electric vehicles charging at the same time, and the ambient temperature as the state parameters describing the operation status of the charging station, as the data basis for clustering, and defines the state parameter matrix describing the operation status of the charging station;

[0058] The two-layer clustering framework first classifies the state parameters into the first-layer scenarios at the macro level according to the different charging demands of electric vehicles on weekdays and holidays, rainy and snowy days, and non-rainy and snowy days. Then, the deep learning algorithm is used to cluster the scenarios based on the first-layer scenario classification results to obtain a scenario set consisting of multiple different operating scenarios.

[0059] The charging station group operation status division unit based on deep convolution embedding clustering constructs and trains a deep convolution embedding clustering model, wherein the deep convolution embedding clustering model includes an autoencoder part and a clustering part; the autoencoder part uses a convolution autoencoder to extract features and reduce the dimension of the original data of the charging station operation status parameters in the scene set; the clustering part uses the K-means algorithm to perform similarity day clustering on the flattened and reduced-dimensional data.

[0060] The beneficial effects of the present invention are:

[0061] 1) This paper first proposes an operating status parameter matrix that describes the operating status of a charging station cluster. Secondly, a two-layer clustering framework is proposed. The first layer performs an initial division of seasonality and time from a macro level, and the second layer uses a deep learning algorithm to perform similar day clustering based on the classification results of the first layer.

[0062] 2) For the second-layer clustering, a deep convolutional embedding clustering algorithm is used to cluster the charging station operating status parameter matrix on each typical day, and a set of different operating scenarios consisting of multiple similar days is obtained to complete the clustering division; this improves the clustering accuracy and efficiency, and solves the dimensionality curse problem of traditional clustering algorithms when faced with high-dimensional, complex, nonlinear data.

[0063] 3) The charging station operation status clustering method proposed in this invention fully considers the seasonal and temporal characteristics of the charging station operation, and can fully exploit the local and global nonlinear fluctuation characteristics of massive historical data, thereby improving clustering accuracy and computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is the charging station state parameter matrix diagram.

[0065] Figure 2 It is a basic flow chart of the similar day clustering method based on the operating status parameters of the electric vehicle charging station group of the present invention. DETAILED DESCRIPTION

[0066] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0067] The operating status of a charging station can be characterized by features such as charging power, the number of idle charging piles of different types, and the number of electric vehicles being charged. The main factors affecting the operating status of a charging station include meteorological conditions and time factors, which indirectly determine the charging power and adjustable capacity of the charging station. The adjustable capacity is the product of the number of idle charging piles in the station and the capacity. Among meteorological conditions, ambient temperature and sunny and rainy days are the dominant factors; among daily factors, the charging power on weekends and holidays is quite different. In order to improve the prediction accuracy of the adjustable capacity of a charging station and better cope with the impact of seasonal and cyclical changes on the prediction results, the present invention clusters and divides the days with similar operating status of charging stations under each typical day based on the two dimensions of meteorological conditions and operating status parameters of the charging station, which can provide a scenario and data basis for prediction.

[0068] 1. Charging station operating status parameters:

[0069] The ambient temperature essentially reflects the seasonal variation characteristics and greatly affects the operating status of the charging station. Therefore, the present invention uses the total charging power of the charging station on each typical day, the number of idle fast charging piles, the number of idle slow charging piles, the number of electric vehicles charging at the same time and the ambient temperature as the state parameters describing the operating status of the charging station and as the data basis for clustering. The state parameter matrix (State Variable Matrix, SVM) describing the operating status of the charging station is defined as follows: Figure 1 shown.

[0070] The operating status of the charging station in each typical day is represented by the above state parameter matrix. t is the sampling time of the state variable, Indicates n Status parameters of the charging station. d A typical day, t Moment n The operating status of a charging station can be expressed as follows:

[0071] (1);

[0072] In the formula, n Number the charging station. is the charging station set. Each element in the matrix represents the d In a typical day, t Moment n The total charging power of each charging station, the number of idle fast-charging charging piles, the number of idle slow-charging charging piles, the number of electric vehicles charging at the same time, and the ambient temperature.

[0073] 2. Two-layer clustering framework:

[0074] At present, most of the relevant research on the clustering of charging station operating status scenarios in distribution networks still focuses on the clustering of the operating status of new energy sources such as distributed photovoltaics and wind power and the load operating characteristics of distribution networks. The present invention proposes a clustering division method for similar days of charging station operating status. Unlike photovoltaics, wind power and other new energy sources, which are mainly affected by factors such as weather conditions, the operating characteristics of charging stations are also strongly related to the charging demand of electric vehicles. There are obvious differences in charging demand for weekdays and holidays, rainy and snowy days and non-rainy and snowy days. Therefore, the present invention first performs the first-level scene classification at the macro level, and then uses a deep learning algorithm to perform scene clustering based on the first-level classification results.

[0075] First-level classification: All typical days are divided into working days and holidays (including weekends); working days and holidays are further divided into rainy and snowy days and non-rainy and snowy days, and finally four label classes are obtained, each of which contains multiple typical days, as shown in Table 1.

[0076] Table 1 Classification at the first level

[0077] .

[0078] Second-level clustering. In each label class, the label class is clustered based on the charging station status parameter matrix of each typical day. Several similar days constitute an operation scenario, and the operation scenario set contains multiple different operation scenarios. These operation scenarios with different characteristics are the data basis for the subsequent prediction of charging power and adjustable capacity of charging stations. Taking the clustering of rainy and snowy days on weekdays as an example, the operation status of charging stations on each typical day in class I can be characterized by the state parameters at each time:

[0079] (2);

[0080] In the formula, Indicates the first d Charging station operation status for one working day; Indicates the first d working days m Charging station operation status parameters at each sampling moment; It is expressed as follows:

[0081] (3);

[0082] In the formula, Indicates the first d working days m The sampling time n Charging station operating status parameters;

[0083] By clustering each typical day in class I, a scenario set consisting of multiple different operating scenarios is formed:

[0084] (4);

[0085] In the formula, To run the scenario set; For the l An operation scenario consists of several operation status parameters of similar days:

[0086] (5);

[0087] In the formula, i Indicates the number of similar days among rainy and snowy working days.

[0088] 3. Charging station group operation status division method based on deep convolution embedding clustering:

[0089] The operating status of a charging station on a typical day is characterized by time series state parameters such as idle charging piles, the number of electric vehicles charged, and the charging power at each sampling time. Therefore, when clustering the operating status of charging stations, it is essentially clustering multiple state parameter matrices. Since the operating status parameters contained in the state variable matrix are time-series, the longer the time scale, the larger the data base, and the higher the dimension. Taking 2024 as an example, assuming that the data sampling frequency of the charging station is 15 minutes / time, there are 251 weekday operation scenarios and 115 holiday operation scenarios in a year, and the sample set data size for each scenario is If traditional algorithms such as K-mean and hierarchical clustering are directly used to cluster high-dimensional running data, it will inevitably cause the curse of dimensionality, resulting in inaccurate clustering results, and even the algorithm cannot be executed due to high computational complexity. Deep convolutional embedding clustering (DCEC) is a clustering algorithm based on deep learning theory. The encoder maps complex high-dimensional data to low-dimensional space through nonlinear transformation; in addition, the algorithm simultaneously reduces the dimensionality and clusters the original data to achieve end-to-end learning, avoiding problems such as error transmission caused by staged data processing. The present invention uses a convolutional autoencoder to extract features and reduce the dimensionality of the original data, and then clusters the reduced dimensionality data through the K-means algorithm.

[0090] There are four steps from data sample input to clustering result output. The first is data preprocessing, the second is model building and training, the third is raw data feature extraction and dimensionality reduction, and the fourth is clustering and visualization. The following is an introduction to each step.

[0091] 1) Data preprocessing:

[0092] In order to eliminate the influence of different feature scales and improve the accuracy of model training, the present invention first normalizes the charging station operation status parameter samples. In the real-time measurement data of the charging station, it is inevitable that there is a small amount of data with abnormal values. In order to reduce the adverse effect of abnormal data on the accuracy of the model, the present invention uses the Robust normalization method to process the original sample data, as shown below:

[0093] (6);

[0094] In the formula, x m Indicates that at the sampling time m Charging station data; median( x m ) is the median of the measured data, IQR( x m ) represents the difference between the 75th percentile and the 25th percentile, represents the normalized data; It is the charging station dataset at all sampling moments.

[0095] For the missing parts of the measured data, data supplementation is required. Since the number of idle piles and charging power at each sampling time of the charging station usually changes smoothly and the data volume is huge, and the spline interpolation method is suitable for time series data with smooth data changes and many data points, the present invention uses the commonly used cubic spline interpolation method to supplement the missing data.

[0096] 2) Model building and training:

[0097] like Figure 2 As shown, the DCEC model constructed by the present invention includes two parts, namely, the autoencoder part and the clustering part. The autoencoder part includes an encoding layer and a decoding layer. The encoding layer includes 4 convolutional layers, 4 pooling layers, a flattening layer and a fully connected layer. The decoding layer includes a fully connected layer, a reshaping layer, 4 deconvolution layers, and 4 upsampling layers. In the clustering layer, the K-means algorithm is used to cluster the flattened and reduced-dimensional data, and the data is visualized based on the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm.

[0098] In the model training phase, the weights, biases, and cluster centers of each layer of the model are iteratively updated to jointly optimize the joint loss function consisting of the reconstruction loss and clustering loss of the autoencoder. The joint loss function L is defined as:

[0099] (7);

[0100] In the formula,L r is the reconstruction loss, L c is the clustering loss, α、β are the weights of the corresponding loss functions respectively.

[0101] The present invention adopts mean square error and KL Divergence is used to characterize the reconstruction loss L r and clustering loss L c :

[0102] (8);

[0103] In the formula, represents the normalized original data before reconstruction, Represents the reconstructed data.

[0104] (9);

[0105] In the formula, Y dj Representation sample d Belong to j The target probability distribution of the class, H dj Represents the model prediction sample d Belong to j Probability distribution of classes; k The number of classes. P and Q For the two probability distributions set; express KL Divergence.

[0106] Probability distribution of model predictions H dj for:

[0107] (10);

[0108] In the formula, z d The sample data is obtained by feature extraction and dimension reduction of the original data through the encoding layer; u j For the j The cluster center of the class; τ is the degree of freedom of the student-t distribution; is the shape parameter.

[0109] Target probability distribution Y dj The expression is:

[0110] (11);

[0111] Finally, this paper uses the Mini-Batch Gradient Descent (MBGD) method to optimize the joint loss function L , when the training reaches the maximum number of iterations, the model training is completed.

[0112] 3) Raw data feature extraction and dimensionality reduction:

[0113] Based on the trained model, the original data is subjected to feature extraction and dimensionality reduction, and finally clustering is performed in a low-dimensional space. Taking the sample data of 2024 as an example, for weekdays and holidays, 251 and 115 charging station state parameter matrices with dimensions of 96*30 are input respectively, and each element of the matrix is ​​composed of the four state parameters of formula (1). Then, the local features of the data are extracted through 4 layers of convolution, and the feature dimension is increased from 4 to 256. At the same time, the data dimension is reduced to 6*2 through 4 layers of pooling. Finally, the original data is flattened and reduced to 12 through the flattening layer and the fully connected layer, realizing the low-dimensional embedding representation of the original data.

[0114] In summary, the present invention proposes a two-layer clustering framework for the operating status of a charging station group, and proposes a clustering method for similar days of the operating status of a charging station group based on a deep embedded clustering algorithm, which improves clustering accuracy and efficiency, solves the dimensionality curse problem of traditional clustering algorithms when facing high-dimensional complex nonlinear data, and improves clustering accuracy and efficiency. The method of the present invention can be used in the software development of a charging station data acquisition system, and clusters similar days of the operating status of a charging station group through a data acquisition system.

Claims

1. A similar day clustering method based on EV charging station operating status parameters, characterized in that: The following steps are involved: Step 1: Take the total charging power of the charging station on each typical day, the number of idle fast charging piles, the number of idle slow charging piles, the number of electric vehicles charging at the same time, and the ambient temperature as the state parameters describing the operating status of the charging station, as the data basis for clustering, and define the state parameter matrix describing the operating status of the charging station; Step 2: Based on the different charging demands of electric vehicles on weekdays and holidays, rainy and snowy days, and non-rainy and snowy days, the state parameters are first classified into the first-level scenarios at the macro level, and then the deep learning algorithm is used to cluster the scenarios based on the first-level scenario classification results to obtain a scenario set consisting of multiple different operating scenarios; Step 3: Construct and train a deep convolutional embedding clustering model, which includes an autoencoder part and a clustering part; the autoencoder part uses a convolutional autoencoder to extract features and reduce the dimension of the original data of the operating status parameters of the charging stations in the scene set; the clustering part uses the K-means algorithm to perform similar day clustering on the flattened and reduced-dimensional data.

2. The similar day clustering method based on EV charging station operating status parameters according to claim 1 is characterized in that: In step 1, d A typical day, t Moment n The operating status of a charging station is expressed as: （1）； In the formula, n Number the charging station. is a collection of charging stations; , , , as well as Respectively represent d In a typical day, t Moment n The total charging power of each charging station, the number of idle fast-charging charging piles, the number of idle slow-charging charging piles, the number of electric vehicles charging at the same time, and the ambient temperature.

3. The similar day clustering method based on EV charging station operating status parameters according to claim 1 is characterized in that: Step 2 is as follows: Step 2.1: First level classification: All typical days are divided into working days and holidays; working days and holidays are further divided into rainy and snowy days and non-rainy and snowy days, and finally four label classes are obtained: rainy and snowy days on working days are class I, rainy and snowy days on holidays are class II, non-rainy and snowy days on working days are class III, and non-rainy and snowy days on holidays are class IV; each label class contains multiple typical days; Step 2.2: Second layer clustering: In each label class, the label class is clustered based on the charging station status parameter matrix of each typical day. Several similar days constitute an operation scenario, and the operation scenario set contains multiple different operation scenarios.

4. The similar day clustering method based on EV charging station operating status parameters according to claim 3 is characterized in that: In step 2.2, for rainy and snowy days on weekdays, i.e., typical days with label class I, the clustering method is as follows: The operating status of each typical day of the charging station in category I is characterized by the status parameters at each time: （2）； In the formula, Indicates the first d Charging station operation status for one working day; Indicates the first d working days m Charging station operation status parameters at each sampling moment; T is the transposition symbol; It is expressed as follows: （3）； In the formula, Indicates the first d working days m The sampling time n Charging station operating status parameters; By clustering each typical day in class I, a scenario set consisting of multiple different operating scenarios is formed: （4）； In the formula, To run the scenario set; For the l An operation scenario consists of several operation status parameters of similar days, and is expressed as: （5）； In the formula, i Indicates the number of similar days among rainy and snowy working days.

5. The similar day clustering method based on EV charging station operating status parameters according to claim 3 is characterized in that: In step 3, data preprocessing is first performed: The robust normalization method is used to normalize the charging station operation status parameter samples in the scene set, as shown in the following formula: （6）； In the formula, x m represents the charging station data at sampling time m; median( x m ) is the median of the measured data, IQR( x m ) represents the difference between the 75th percentile and the 25th percentile, represents the normalized data; It is the charging station dataset at all sampling moments; The deep convolutional embedding clustering model is then trained: In the model training phase, the weights, biases, and cluster centers of each layer of the model are iteratively updated, and the reconstruction loss and clustering loss of the autoencoder are jointly optimized to form a joint loss function; the joint loss function L Defined as: （7）； In the formula, L r is the reconstruction loss, L c is the clustering loss, α、β are the weights of the corresponding loss functions respectively; The mean square error and KL Divergence is used to characterize the reconstruction loss L r and clustering loss L c : （8）； In the formula, represents the normalized original data before reconstruction, Represents the reconstructed data; D is the total number of sample days; （9）； In the formula, Y dj Representation sample d Belong to j The target probability distribution of the class, H dj Represents the model prediction sample d Belong to j Probability distribution of classes; k is the number of classes; P and Q For the two probability distributions set; express KL Divergence; Probability distribution of model predictions H dj for: （10）； In the formula, z d The sample data is obtained by feature extraction and dimension reduction of the original data through the encoding layer; u j For the j The cluster center of the class; τ is the degree of freedom of the student-t distribution; is the shape parameter; Target probability distribution Y dj The expression is: （11）； Finally, a mini-batch gradient descent method is used to optimize the joint loss function L , when the training reaches the maximum number of iterations, the model training is completed.

6. A similar day clustering system based on EV charging station operating status parameters, characterized by: It includes a charging station operation status parameter definition unit, a two-layer clustering framework, and a charging station operation status division unit based on deep convolution embedding clustering; The charging station operation status parameter definition unit uses the total charging power of the charging station on each typical day, the number of idle fast charging piles, the number of idle slow charging piles, the number of electric vehicles charging at the same time, and the ambient temperature as the state parameters describing the operation status of the charging station, as the data basis for clustering, and defines the state parameter matrix describing the operation status of the charging station; The two-layer clustering framework first classifies the state parameters into the first-layer scenarios at the macro level according to the different charging demands of electric vehicles on weekdays and holidays, rainy and snowy days, and non-rainy and snowy days. Then, the deep learning algorithm is used to cluster the scenarios based on the first-layer scenario classification results to obtain a scenario set consisting of multiple different operating scenarios. The charging station operation status division unit based on deep convolution embedding clustering constructs and trains a deep convolution embedding clustering model, wherein the deep convolution embedding clustering model includes an autoencoder part and a clustering part; the autoencoder part uses a convolution autoencoder to extract features and reduce the dimension of the original data of the charging station operation status parameters in the scene set; the clustering part uses the K-means algorithm to perform similarity day clustering on the flattened and reduced-dimensional data.

7. The similar day clustering system based on EV charging station operating status parameters according to claim 6 is characterized in that: In the charging station operation status parameter definition unit, d A typical day, t Moment n The operating status of a charging station is expressed as: （1）； In the formula, n Number the charging station. is a collection of charging stations; , , , as well as Respectively represent d In a typical day, t Moment n The total charging power of each charging station, the number of idle fast-charging charging piles, the number of idle slow-charging charging piles, the number of electric vehicles charging at the same time, and the ambient temperature.

8. The similar day clustering system based on EV charging station operating status parameters according to claim 6 is characterized in that: The two-layer clustering framework includes a first-layer classification unit and a second-layer clustering unit: The first-level classification unit: all typical days are divided into working days and holidays; further divided into rainy and snowy days and non-rainy and snowy days, and finally four label classes are obtained: rainy and snowy days on working days are class I, rainy and snowy days on holidays are class II, non-rainy and snowy days on working days are class III, and non-rainy and snowy days on holidays are class IV; each label class contains multiple typical days; Second-layer clustering unit: In each label class, the label class is clustered based on the charging station status parameter matrix of each typical day. Several similar days constitute an operation scenario, and the operation scenario set contains multiple different operation scenarios.

9. The similar day clustering system based on EV charging station operating status parameters according to claim 8, characterized in that: In the second-layer clustering unit, for rainy and snowy days on weekdays, that is, typical days with label class I, the clustering method is as follows: The operating status of each typical day of the charging station in category I is characterized by the status parameters at each time: （2）； In the formula, Indicates the first d Charging station operation status for one working day; Indicates the first d working days m Charging station operation status parameters at each sampling moment; T is the transposition symbol; It is expressed as follows: （3）； In the formula, Indicates the first d working days m The sampling time n Charging station operating status parameters; By clustering each typical day in class I, a scenario set consisting of multiple different operating scenarios is formed: （4）； In the formula, To run the scenario set; For the l An operation scenario consists of several operation status parameters of similar days, and is expressed as: （5）； In the formula, i Indicates the number of similar days among rainy and snowy working days.

10. The similar day clustering system based on EV charging station operating status parameters according to claim 6, characterized in that: The autoencoder part of the deep convolutional embedding clustering model includes an encoding layer and a decoding layer; the encoding layer includes 4 convolutional layers, 4 pooling layers, a flattening layer and a fully connected layer; the decoding layer includes a fully connected layer, a reshaping layer, 4 deconvolution layers and 4 upsampling layers; the model training process is as follows: The robust normalization method is used to normalize the charging station operation status parameter samples in the scene set, as shown in the following formula: （6）； In the formula, x m Indicates that at the sampling time m Charging station data; median( x m ) is the median of the measured data, IQR( x m ) represents the difference between the 75th percentile and the 25th percentile, represents the normalized data; It is the charging station dataset at all sampling moments; In the model training phase, by iteratively updating the weights, biases, and cluster centers of each layer of the model, the reconstruction loss and clustering loss of the autoencoder are jointly optimized to form a joint loss function; the joint loss function L is defined as: （7）； In the formula, L r is the reconstruction loss, L c is the clustering loss, α , β are the weights of the corresponding loss functions respectively; The mean square error and KL Divergence is used to characterize the reconstruction loss L r and clustering loss L c : （8）； In the formula, represents the normalized original data before reconstruction, Represents the reconstructed data; D is the total number of sample days; （9）； In the formula, Y dj Representation sample d Belong to j The target probability distribution of the class, H dj Represents the model prediction sample d Belong to j Probability distribution of classes; k is the number of classes; P and Q For the two probability distributions set; express KL Divergence; Probability distribution of model predictions H dj for: （10）； In the formula, z d The sample data is obtained by feature extraction and dimension reduction of the original data through the encoding layer; u j For the j The cluster center of the class; τ is the degree of freedom of the student-t distribution; is the shape parameter; Target probability distribution Y dj The expression is: （11）； Finally, the mini-batch sample gradient descent method is used to optimize the joint loss function L. When the training reaches the maximum number of iterations, the model training is completed.

Images