Response potential evaluation method, device and equipment based on electric vehicle user portrait

Abstract

The invention provides a response potential evaluation method, device and equipment based on an electric vehicle user portrait, and relates to the technical field of intelligent power grids. The method comprises the following steps: clustering charging load behavior characteristics of an electric vehicle user to obtain a plurality of clusters; wherein each cluster corresponds to one user portrait; obtaining a demand response potential index of the electric vehicle user corresponding to the target user portrait; wherein the target user portrait is any user portrait; and inputting the demand response potential index into a trained LightGBM model to obtain a demand response potential value of the electric vehicle user corresponding to the target user portrait. According to the invention, accurate and efficient evaluation of the demand response potential can be realized.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and in particular to a method, apparatus, and equipment for assessing response potential based on electric vehicle user profiles. Background Technology

[0002] With the advancement of global energy transition and the "dual-carbon" goal, the number of electric vehicles (EVs) is growing at an unprecedented rate due to their clean and efficient characteristics. However, the disorderly and concentrated charging behavior of large-scale EVs poses a serious challenge to the safe, stable, and economical operation of the power system. The large number of EVs connecting to the grid for charging during peak electricity consumption periods exacerbates the peak-valley difference in the grid, leading to overload of local distribution facilities and increasing the grid's peak-shaving costs and investment pressure. To resolve this contradiction, demand response technology is considered a key solution. Demand response refers to the behavior of electricity users proactively adjusting their electricity consumption patterns based on market price signals or incentive mechanisms. Incorporating EV clusters as a dispatchable flexible load resource into demand response projects can transform them from a "grid burden" into a "high-quality regulating resource," providing services such as peak shaving and valley filling, and frequency regulation to the grid by guiding the spatial and temporal transfer of charging.

[0003] Electric vehicles, with their adjustable charging load, are considered a core flexible resource in the demand response system, and their participation potential directly determines the grid peak-shaving efficiency and the capacity for renewable energy absorption. However, current research on assessing the demand response potential of electric vehicles mainly focuses on vehicle parameter-oriented and grid command-oriented approaches, failing to effectively address the core issue of "mismatch between user behavior and grid demand," resulting in significant discrepancies between assessment results and actual response effects. Summary of the Invention

[0004] This invention provides a method, apparatus, and device for assessing response potential based on electric vehicle user profiles, in order to address the problem of poor accuracy in assessing the demand response potential of electric vehicle users.

[0005] In a first aspect, embodiments of the present invention provide a method for assessing response potential based on electric vehicle user profiles, including: The charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters; each cluster corresponds to a user profile. Obtain the demand response potential indicators of electric vehicle users corresponding to the target user profile; where the target user profile can be any user profile. By inputting the demand response potential index into the trained LightGBM model, the demand response potential value of electric vehicle users corresponding to the target user profile is obtained.

[0006] In one possible implementation, before clustering the charging load behavior characteristics of electric vehicle users to obtain multiple clusters, the following steps are also included: Construct an original behavioral feature set; wherein, the original behavioral feature set includes various charging load characteristics; By combining the charging load characteristics in the original behavioral feature set, multiple candidate behavioral feature sets are obtained; The feature set adaptability evaluation coefficient of each candidate behavior feature set is calculated based on the preset feature set adaptability function, and the optimal feature set is selected based on the feature set adaptability evaluation coefficient. Accordingly, the charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters, including: The charging load behavior characteristics of electric vehicle users are obtained based on the optimal feature set, and the charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters.

[0007] In one possible implementation, the feature set fitness function is:

[0008] Where L is the feature set fitness evaluation coefficient of candidate feature set S, G(S) is the information gain index of all features in candidate feature set S, and R(S) is the redundancy index of all features in candidate feature set S.

[0009] In one possible implementation, before inputting the demand response potential index into the trained LightGBM model to obtain the demand response potential value of the electric vehicle user corresponding to the target user profile, the following steps are also included: A sample set was constructed based on demand response potential indicators of multiple electric vehicle users; The initial LightGBM model is trained based on the sample set to obtain the trained LightGBM model.

[0010] In one possible implementation, a sample set is constructed based on demand response potential indicators of multiple electric vehicle users, including: Based on the trained CWGAN network, the demand response potential indicators of electric vehicle users corresponding to the first user profile are synthesized and completed; where the first user profile is any user profile. A sample set is constructed based on the completed demand response potential indicators.

[0011] In one possible implementation, the discriminator loss function of the CWGAN network is:

[0012] in, For the real sample discrimination term, To generate sample discrimination terms, This is a gradient penalty term. Intermediate samples obtained by interpolating between real and generated samples. for The sampling distribution for The sampling distribution. For the distribution of interpolated samples, These are the weight coefficients of the gradient penalty term; The generator loss function is:

[0013] In the formula, Indicates the joint distribution To calculate the mathematical expectation, For the condition generator model, This is a discriminator model.

[0014] In one possible implementation, the charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters, including: The number of clusters is determined based on the sum of squared errors to identify the charging load behavior characteristics of electric vehicle users; Based on the number of clusters, the GMM clustering algorithm is used to cluster the charging load behavior characteristics of electric vehicle users, resulting in multiple clusters.

[0015] Secondly, embodiments of the present invention provide a response potential assessment device based on electric vehicle user profiles, comprising: The clustering module is used to cluster the charging load behavior characteristics of electric vehicle users to obtain multiple clusters; each cluster corresponds to a user profile. The acquisition module is used to acquire the demand response potential indicators of electric vehicle users corresponding to the target user profile; wherein, the target user profile can be any user profile. The evaluation module is used to input the demand response potential index into the trained LightGBM model to obtain the demand response potential value of electric vehicle users corresponding to the target user profile.

[0016] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0017] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.

[0018] The response potential assessment method, apparatus, and equipment based on electric vehicle user profiles provided in this invention scientifically classify typical user profiles through feature optimization and cluster analysis, fundamentally ensuring the accuracy of the assessment targets. Based on this, the system quantifies potential indicators from multiple dimensions such as transferable power and response speed, and utilizes the LightGBM machine learning model for fusion prediction, significantly improving the accuracy and reliability of the assessment results. Furthermore, to address the problem of imbalanced data, a CWGAN network is innovatively used for data augmentation and sample expansion, ensuring the model's generalization ability. Ultimately, this achieves accurate and efficient assessment of demand response potential, providing a direct basis for power grid companies to accurately select users and formulate differentiated incentive strategies, thereby efficiently transforming electric vehicle clusters from a burden on the power grid into a high-quality and controllable regulatory resource, possessing significant engineering application value. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the implementation of the response potential assessment method based on electric vehicle user profiles provided in this embodiment of the invention. Figure 2 This is a flowchart for creating a user electricity consumption behavior profile provided in an embodiment of the present invention; Figure 3 This is a normalized information gain map of various behavioral features provided in the embodiments of the present invention; Figure 4 This is a training curve diagram of the CWGAN network provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the response potential assessment device based on electric vehicle user profiles provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] See Figure 1 The document illustrates a flowchart of the response potential assessment method based on electric vehicle user profiles provided in an embodiment of the present invention, detailed below: Step 101: Cluster the charging load behavior characteristics of electric vehicle users to obtain multiple clusters; each cluster corresponds to a user profile.

[0022] In one possible implementation, the charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters, including: The number of clusters is determined based on the sum of squared errors to identify the charging load behavior characteristics of electric vehicle users; Based on the number of clusters, the GMM clustering algorithm is used to cluster the charging load behavior characteristics of electric vehicle users, resulting in multiple clusters.

[0023] In this embodiment, Figure 2 A flowchart for creating user electricity consumption behavior profiles. (e.g.) Figure 2 As shown, this embodiment uses three steps—GMM clustering algorithm, feature set optimization, and user profile characterization—to characterize the charging behavior of electric vehicle users. Specifically: To analyze the charging behavior of electric vehicle users, this embodiment uses the Growing Mixture Model (GMM) algorithm for clustering, aiming to group similar charging loads into the same cluster for analysis. The GMM clustering algorithm must first determine the number of clusters, k. Based on the principles of the GMM algorithm, the sum of squared errors (SSE) can be used as an evaluation metric for determining the number of clusters. The formula for SSE is as follows: (1) In the formula, k is the number of clusters. For a cluster in the clustering, for The mean of each data point in the middle. Let be a point in the cluster.

[0024] After determining the number of clusters k using SSE, it is assumed that each cluster follows a Gaussian distribution, where the probability model of the Gaussian distribution is: (2) In the formula, As a sample, As weight, Let be the probability density function of a Gaussian distribution. Let be the parameter of the probability density, and its expression is: (3) In the formula, For the sample standard deviation For the sample The mean.

[0025] The probability density functions of these k classes of Gaussian distributions and the weights of each class were estimated using the training data of the Gaussian model. Next, calculate the probability of each data point appearing in each of the k Gaussian distributions, that is, substitute the data point into each of the k Gaussian distributions to find the probability of it belonging to each class. : (4) In the formula, For the i-th sample data, It is the k-th Gaussian distribution.

[0026] Finally, the sample is assigned to the class with the highest probability value by comparison.

[0027] In one possible implementation, before clustering the charging load behavior characteristics of electric vehicle users to obtain multiple clusters, the following steps are also included: Construct an original behavioral feature set; wherein, the original behavioral feature set includes various charging load characteristics; By combining the charging load characteristics in the original behavioral feature set, multiple candidate behavioral feature sets are obtained; The feature set adaptability evaluation coefficient of each candidate behavior feature set is calculated based on the preset feature set adaptability function, and the optimal feature set is selected based on the feature set adaptability evaluation coefficient. Accordingly, the charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters, including: The charging load behavior characteristics of electric vehicle users are obtained based on the optimal feature set, and the charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters.

[0028] In this embodiment, feature selection is a crucial step in electric vehicle user behavior profiling. Obtaining a reasonably sized feature set that accurately characterizes key user traits is the ideal goal of electric vehicle user behavior profiling. The characteristics inherent in electric vehicle user load data can be characterized by various feature indicators; however, different features may describe user characteristics in a repetitive way, meaning there may be correlations between different features. Therefore, when selecting features, the effectiveness and redundancy between features should be considered. Furthermore, to enhance the understandability of user features, quantitative analysis should be performed on the user features.

[0029] The method for selecting and quantifying the electricity consumption characteristics of electric vehicle users adopted in this paper is as follows: First, a multi-dimensional original feature set of electricity users is constructed. On this basis, the degree of information gain of the addition of each feature to the overall information of the electricity user is measured. Then, the Spearman correlation coefficient is introduced to measure the similarity between different features. Finally, taking into account the information gain after the feature is introduced and the similarity between features, the feature set adaptability evaluation coefficient is obtained, and the feature set with the highest evaluation coefficient is selected as the optimal electricity consumption feature set of the electricity user.

[0030] The charging characteristics of electric vehicle users are reflected in the characteristics of the user charging load curve. The main analytical indicators for charging load characteristics are: (1) Statistical characteristics, namely average charging power. Average charging time Weekly charging frequency Total charging capacity Off-peak electricity utilization rate (2) Timing behavior class, namely, charging start-up time distribution Standard deviation of charging time Charging active window length Concentration of charging start time Weekly charging frequency The above-mentioned characteristic indicators constitute the original characteristic set of electric vehicle users, that is, the original characteristic set is... The meanings of each characteristic indicator are shown in Table 1.

[0031] Table 1

[0032] Information gain of features can effectively measure how well different features characterize the behavior of electric vehicle users. To better understand this metric, the concept of information entropy must be introduced. Information entropy measures the degree of uncertainty of a system; the higher the value, the more chaotic the system and the greater the amount of information. Information quantity measures whether an event will occur; the greater the information quantity, the lower the probability of occurrence and the greater the uncertainty.

[0033] For a random variable X, assume that all its possible values ​​are... The corresponding probability of occurrence is The information entropy of the random variable X is defined as shown in equation (5).

[0034] (5) In the analysis of electric vehicle user charging characteristics, all possible values ​​of the random variable X are the clusters formed by clustering, and the corresponding probability is the proportion of the number of samples contained in each cluster to the total number of samples.

[0035] Introducing feature Z changes the information entropy of the random variable X. Assume that all possible values ​​of feature Z are... The corresponding probability of occurrence is This refers to the proportion of samples falling within each feature value range out of the total sample size. For a given feature value... The probability that a sample belongs to a different cluster is given by the internal structure. The formula for calculating the information entropy of random variable X after adding feature Z is shown in equation (6).

[0036] (6) For a given feature Z, its information gain can measure the effectiveness of the introduction of Z in reducing the uncertainty of the random variable X. The information gain is numerically defined as the interpolation of the information entropy of the random variable X before and after the introduction of feature Z, and its calculation formula is shown in Equation (7).

[0037] (7) However, the traditional information gain expression mentioned above has the problem of favoring high-cardinality features (i.e., features with a large number of values), which may lead to high-value features obtaining higher information gain due to their finer-grained segmentation, thus misleading feature selection. Therefore, this paper introduces normalized information gain (Gain Ratio) based on the feature's own information entropy. The information gain is normalized and corrected; the larger the value, the stronger the ability to distinguish variables.

[0038] (8) To measure redundancy among features and thus avoid potential nonlinear relationships, a distance correlation coefficient is used. This method effectively captures nonlinear relationships between features. Given a set of feature vectors... and ,in , n is a feature and The number of samples. Feature vector sample set. and The distance matrices are A and B, respectively, and each element of the matrix is ​​a normalized and centered Euclidean distance matrix between each pair of samples.

[0039] (9) (10) (11) (12) (13) In the formula, and Let be the Euclidean distance between the i-th and j-th samples. and The elements are the distance matrix after doubly centralization. Let the distance covariance be... and Let be the distance variance. The distance correlation coefficient ranges from [0,1] and is true only when... and When independent, it takes the value of 0, thus enabling the detection of statistical dependencies of arbitrary forms, both linear and nonlinear. It is an important tool for capturing nonlinear correlations when improving feature redundancy measures.

[0040] Therefore, for a given feature set U, a redundancy index is defined. , can be represented as: (14) In the formula, The total number of features contained in the feature set. Let i and j be the i-th and j-th features in the feature set.

[0041] In one possible implementation, the feature set fitness function is:

[0042] Where L is the feature set fitness evaluation coefficient of candidate feature set S, G(S) is the information gain index of all features in candidate feature set S, and R(S) is the redundancy index of all features in candidate feature set S.

[0043] In this embodiment, for a certain feature set Based on the effectiveness of the comprehensive feature information gain and the redundancy measured by the correlation coefficient between features, a feature set fitness evaluation function is constructed to calculate the feature set fitness evaluation coefficient, which is defined as follows: (15) In the formula, The information gain index is the sum of all features in the feature set S. is the redundancy index for all features of feature set S.

[0044] By using a traversal method to search for feature subsets in S, the feature subset with the largest feature evaluation function value is the optimal feature subset sought in this paper, which is the optimal behavioral feature for the category classification problem of user behavior profiling.

[0045] To improve the readability of electric vehicle user feature tags, a scoring system is adopted, with a score of 0-1 as the baseline, to measure the scores of different types of electric vehicle users across all feature dimensions. The optimal charging feature set is semantically represented, and the calculation method is as follows: (16) In the formula, For the j-th label of the i-th user, Let j be the average of all users belonging to the i-th category and their j-th label. and These are the minimum and maximum values ​​of the j-th label, respectively.

[0046] To enable business personnel to more intuitively understand the charging characteristics of various electric vehicle users, the obtained user charging tags are visualized to create user behavior profiles for different user types. The visualization consists of two parts: intra-category user charging behavior profiles and inter-category charging characteristic comparisons. Intra-category user charging behavior profiles use methods such as radar charts to display the charging characteristics of various users, while inter-category charging characteristic comparisons are presented using bar charts to highlight the comparison of the frequency of the same tag across different user groups. Combining these two methods allows business personnel to more accurately and conveniently understand the commonalities and individual characteristics of electric vehicle user charging behavior.

[0047] User profiles guide feature construction: Key features extracted from user profiles (such as off-peak electricity utilization rate, charging activity window, etc.) are directly used as part of the input to the LightGBM model. User profiles explain potential results, primarily through profile categories (such as "rigid users" and "flexible users"), explaining why a certain type of user has high or low potential, thus enhancing the model's interpretability. User profiles support strategy formulation, mainly by combining profile tags and potential values ​​to formulate more targeted demand response strategies, achieving refined "one-category-one-policy" operations.

[0048] Step 102: Obtain the demand response potential indicators of electric vehicle users corresponding to the target user profile; wherein, the target user profile is any user profile.

[0049] In this embodiment, demand response potential can be assessed for any type of user profile. The demand response potential indicator system is as follows: (1) Peak hour electricity consumption ratio Peak electricity consumption ratio This refers to the peak hours in a user's daily load curve. The proportion of peak-hour electricity consumption to total daily electricity consumption. A higher peak-hour proportion indicates a more concentrated period of electricity consumption, resulting in better peak-shaving and valley-filling regulation. The expression is as follows: (17) In the formula, The charging load of user l at time t. The sampling time interval, It is a set of sampling time points.

[0050] (2) Load fluctuation rate Load volatility This is the ratio of the standard deviation to the mean of the user load curve, reflecting the degree of dispersion of charging load over time. A higher volatility indicates greater fluctuations in the user charging load curve and greater user dispatch potential. The expression is as follows: (18) In the formula, and These are the standard deviation and mean load of user l, respectively. This represents the number of user load sampling points.

[0051] (3) Adjustment range Adjustment range This refers to the maximum range within which a user's charging load can fluctuate without affecting their basic usage needs during load regulation. This indicator reflects the user's maximum adjustable load capacity and is one of the key parameters for evaluating their participation in peak shaving and demand response. A larger adjustment range indicates a stronger flexible adjustment capability and a higher potential for participating in system operation optimization. The adjustment range can be defined as the ratio of the difference between the user's maximum and minimum charging power within the adjustment time window to the maximum power, expressed as follows: (19) In the formula, The maximum charging power for user l within the adjustable time period. The minimum charging power for user l within an adjustable time period.

[0052] (4) Adjustment rate Adjustment rate This refers to the rate of change of user charging load per unit time, used to measure the agility of user load response adjustment commands. In actual demand response scheduling, users with a faster adjustment rate can respond to control commands more quickly, improving the real-time performance and safety of system operation. It is suitable for high-frequency control scenarios such as emergency response and ancillary services. The adjustment rate is usually characterized by the derivative of the load change with respect to the time interval, as shown in the following expression: (20) In the formula, This represents the charging power after the sampling time interval.

[0053] (5) Response period Response period It refers to the overlapping time period during which electric vehicle users initiate demand response incentives on the grid side, and where they themselves have the conditions for grid connection. In other words, it is the theoretical time range during which electric vehicle users can participate in demand response regulation.

[0054] (twenty one) In the formula, This refers to the time period for demand response control initiated by the power grid side. This refers to the grid connection time of user l (i.e., the actual time period during which the electric vehicle connects to the charging facility).

[0055] In addition, indicators such as off-peak electricity utilization rate and charging activity window can be extracted from user profiles and used as inputs to the LightGBM model to assess the demand response potential of electric vehicle users in the corresponding user profile categories.

[0056] Step 103: Input the demand response potential index into the trained LightGBM model to obtain the demand response potential value of electric vehicle users corresponding to the target user profile.

[0057] In this embodiment, the input to the LightGBM model is a demand response potential feature index extracted from the charging behavior of electric vehicle users, specifically including: peak-hour electricity consumption ratio. Load fluctuation rate Adjustment range Adjustment rate Response period These characteristics constitute a potential feature vector of users participating in demand response in the current or future period.

[0058] The data sources for obtaining input parameters include real-time or historical charging data of electric vehicle charging piles, user grid connection status data, and grid demand response time information. Based on the original data such as user charging power curves, charging duration, and charging time periods, the potential feature values ​​are calculated using formulas (17) to (21). Some features (such as responsive time periods) need to be matched and calculated by combining the grid demand response incentive time period with the user grid connection time period. Before this, data preprocessing can be performed, including missing value processing, outlier removal, and standardization, to ensure that the input data conforms to the data distribution during model training.

[0059] The LightGBM model outputs a user's demand response potential value, typically a continuous numerical value, representing the user's expected ability to participate in demand response within a specified time period. Output format: (Predicted potential value of the s-th user) Value range: usually [0,1] or a standardized real number. The larger the value, the higher the potential.

[0060] Based on users' demand response potential, grid operators can prioritize high-potential users for demand response projects such as peak shaving and valley filling, and frequency regulation. This enables precise load control and improves grid operational efficiency and stability. Resource allocation and incentive strategy formulation: Differentiated incentive policies (such as electricity price discounts and subsidies) can be developed for different user groups based on their potential to increase user participation. This optimizes the allocation of demand response resources and enhances the overall response effect.

[0061] This invention, through feature optimization and cluster analysis, scientifically identifies typical user profiles, fundamentally ensuring the accuracy of the assessment. Based on this, the system quantifies potential indicators from multiple dimensions, such as transferable power and response speed, and utilizes the LightGBM machine learning model for fusion prediction, significantly improving the accuracy and reliability of the assessment results. Furthermore, addressing the issue of imbalanced data, it innovatively employs a CWGAN network for data augmentation and sample expansion, ensuring the model's generalization ability. Ultimately, it achieves accurate and efficient assessment of demand response potential, providing direct evidence for power grid companies to accurately select users and formulate differentiated incentive strategies. This transforms electric vehicle clusters from a burden on the power grid into a high-quality and controllable regulatory resource, demonstrating significant engineering application value.

[0062] In one possible implementation, before inputting the demand response potential index into the trained LightGBM model to obtain the demand response potential value of the electric vehicle user corresponding to the target user profile, the following steps are also included: A sample set was constructed based on demand response potential indicators of multiple electric vehicle users; The initial LightGBM model is trained based on the sample set to obtain the trained LightGBM model.

[0063] In one possible implementation, a sample set is constructed based on demand response potential indicators of multiple electric vehicle users, including: Based on the trained CWGAN network, the demand response potential indicators of electric vehicle users corresponding to the first user profile are synthesized and completed; where the first user profile is any user profile. A sample set is constructed based on the completed demand response potential indicators.

[0064] In this embodiment, to further enhance the generative model's control over specific user profile features, a conditional variable is introduced on top of WGAN to construct a Conditional Wasserstein Generative Adversarial Network (CWGAN), enabling the generation of DR feature samples for users of specific profile categories. This method can significantly alleviate the problem of scarce potential user samples in the early stages of Demand Response (DR) projects, providing high-quality pseudo-samples with rich structure and reasonable category distribution for the subsequent training of semi-supervised learning models.

[0065] Based on the clustering results of the aforementioned electric vehicle user profiles, it is assumed that users are divided into class I, with each class possessing consistent charging behavior characteristics and adjustment capability distributions. The generator input consists of two parts: a latent vector and condition variables. The latent vector z is sampled from a standard normal distribution, denoted as... , representing latent variables in the hidden feature space, used to capture potential changes in user behavior characteristics. Condition variables represent the user's profile category, defined as... This is used to guide the generative model to synthesize samples with specified behavioral profile features, enhancing the generator's feature discrimination ability across different user profile categories. The final joint input form of the generator is... , indicating that in the specified condition variable Next, we utilize random latent vectors.

[0066] The discriminator's input is also a conditional combination structure, consisting of a sample vector and condition variables. The sample vector represents the demand response characteristic parameters of potential users, defined as: (twenty two) The aforementioned features constitute the real sample distribution for potential user modeling, serving as the basis for cWGAN training data. The joint input received by the discriminator is represented as... .

[0067] Using condition variables cWGAN enables targeted data augmentation for different categories of potential users by controlling the category attributes of generated samples while learning the latent distribution of user demand response features. Compared to the traditional WGAN model, this method significantly improves the model's generalization ability in small sample scenarios and effectively alleviates the problem of insufficient samples in the early stages of electric vehicle demand response.

[0068] Within the basic framework of Conditional Generative Adversarial Networks (CWGAN), the structural design of the generator and discriminator is crucial to the model's generative ability and stability. Considering the actual needs and characteristics of electric vehicle users' charging behavior, this paper designs a generator (G) and discriminator (D) network structure within cWGAN that includes conditional inputs, aiming to effectively learn and regulate the feature distribution of potential users under limited sample conditions.

[0069] (1) Generator network structure Generator networks use latent vectors Conditional variables related to user profile categories As input, a simulated regulation potential feature vector is generated through a series of nonlinear mappings. The generator's structure is modeled using a multi-layer fully connected neural network (MLP). The input layer connects z with... After concatenation, the data is mapped to a high-dimensional space. The intermediate layers use the ReLU activation function, and the output layer uses a linear activation function. The output is consistent with the real sample. Dimensionally consistent moderating potential characteristics.

[0070] The generator structure can be represented as: (twenty three) in, The nonlinear mapping function representing the generator, These are the generator network parameters.

[0071] (2) Discriminator network structure The discriminator network takes input feature vectors (real or generated) and corresponding condition variables. After concatenation, it is determined whether the data originates from the true sample distribution. To enhance the discriminator's consistency constraint on class conditions, the discriminator also employs an MLP structure, with each layer using the LeakyReLU activation function to improve training stability. The output layer uses the Sigmoid function to output the discriminant probability. .

[0072] The discriminator structure can be represented as: (twenty three) in, The nonlinear mapping function representing the generator, These are the generator network parameters.

[0073] In one possible implementation, the discriminator loss function of the CWGAN network is:

[0074] in, For the real sample discrimination term, To generate sample discrimination terms, This is a gradient penalty term. Intermediate samples obtained by interpolating between real and generated samples. for The sampling distribution for The sampling distribution. For the distribution of interpolated samples, These are the weight coefficients of the gradient penalty term; The generator loss function is:

[0075] In the formula, Indicates the joint distribution To calculate the mathematical expectation, For the condition generator model, This is a discriminator model.

[0076] In this embodiment, in order to stabilize the training process of CWGAN and achieve class-conditional generation of potential user samples, this paper introduces conditional variables and gradient penalty terms on the basis of Wasserstein GAN to construct a loss function with stronger robustness.

[0077] (1) Discriminator loss function The discriminator loss function is defined based on the Wasserstein distance and includes a gradient penalty term to satisfy 1-Lipschitz continuity. The overall discriminator objective is to maximize the output difference between the real and generated samples, while constraining the gradient norm of the discriminator function, as expressed below: (twenty four) In the formula, For real sample discrimination, it represents the average judgment score of the discriminator for real samples, and it is expected that the larger the value, the better (the more inclined to judge as "true"). The generated sample discrimination term represents the discriminator's judgment score for the generated sample, and it is expected that the smaller the value, the better (the more inclined to judge it as "false"). This is a gradient penalty term that introduces a Lipschitz continuity constraint by penalizing the degree to which the gradient deviates from 1, thereby improving training stability. Intermediate samples obtained by interpolation between real and generated samples are used to calculate gradient penalty terms. Represents the joint distribution of the real data, i.e. The sampling distribution. The joint distribution of the data generated by the generator, i.e. The sampling distribution. The distribution of the interpolated samples is used to calculate the gradient penalty. The weight coefficient for the gradient penalty term (usually set to 10) is used to control the constraint strength of the Lipschitz condition.

[0078] (2) Generator loss function The generator's goal is to minimize the discriminator's judgment value for the generated samples, that is, to maximize the probability that they are misclassified as real samples. The generator continuously optimizes its parameters. This is to maximize the discrimination probability of the generated samples under various profile conditions, thereby improving their consistency with the distribution of real potential users. The generator loss function is defined as: (25) In the formula, Indicates the joint distribution To calculate the mathematical expectation, This is a condition generator model. This is a discriminator model that uses a scoring function to determine whether an input sample is real under specified category conditions. The larger the output, the more likely it is to be judged as a real sample.

[0079] To quantitatively predict the response potential of electric vehicle users in demand-response scenarios, this paper selects the Light Gradient Boosting Machine (LightGBM), which boasts efficient modeling capabilities and good interpretability, as the core regression model to construct a response potential predictor for individual users, especially when potential value labels are missing. This model uses an indicator system as input features and combines training samples synthesized and completed based on cWGAN to achieve a refined estimate of user response capabilities. The LightGBM model is essentially an ensemble of additive regression trees, and its overall prediction function is as follows: (26) In the formula, Let be the predicted potential value of the s-th sample. Let be the feature vector of the s-th sample. Let K represent the k-th decision tree model, where K is the total number of regression trees.

[0080] The objective function of the LightGBM model consists of a loss function and a regularization term. The expression for the objective function of the k-th tree is as follows: (27) In the formula, For potential prediction error loss term, For the s-th sample at the k-th position The predicted value at the first iteration, The regularization term for the k-th tree is usually expressed in the following form: (28) in, To regress the number of leaf nodes, This is the output value of the j-th leaf node. and These are the penalty coefficients used to control the complexity of the model.

[0081] 4.4.2 Evaluation of Model Fit To comprehensively evaluate the fitting accuracy and generalization ability of the constructed LightGBM-based Demand Response (DR) potential regression model on different user samples, this paper introduces a series of goodness-of-fit evaluation indicators, and conducts multi-dimensional quantitative analysis of model performance from the perspectives of prediction error, explanatory power, and relative bias. This evaluation system not only helps to determine the model's approximation ability on known training samples, but also serves as an important reference for future model optimization and promotion.

[0082] First, the most basic evaluation metric is the Mean Squared Error (MSE). A smaller MSE indicates a lower overall model error and higher fitting accuracy. It is defined as: (29) Secondly, to further analyze the directionality of model bias, the Mean Absolute Error (MAE) is introduced. This value is not affected by extreme values ​​and is suitable for reflecting the average deviation of typical samples. Its definition is: (30) Finally, to measure the model's explanatory power for the trend of changes in the overall response potential, the coefficient of determination (R²) is introduced, which is defined as follows: (31) In one specific embodiment, the method provided by this invention is used to create a user behavior profile, resulting in... Figure 3 The normalized information gain plots of various behavioral features shown, and Figure 4 The training curves for CWGAN are shown below. Analysis reveals the following characteristics for various behavioral features: Highly discriminative features include: 1. Valley Ratio (Peak Electricity Utilization Rate) – The Strongest Discriminating Feature This indicates whether users tend to charge during off-peak hours when electricity prices are low, reflecting their price sensitivity and scheduled charging habits.

[0083] This indicates that different types of users have significantly different time selection strategies.

[0084] 2. std_hour and entropy_val (standard deviation of startup time & entropy of startup time) Both indicators represent the stability of a user's behavior regarding "when to charge." The more concentrated the activation time (lower entropy), the more fixed the user's habits; the more dispersed the activation time, the more flexible or random the behavior. This is of great value in distinguishing between "rigid behavior users" and "flexible behavior users."

[0085] 3. active_window (active window) This indicates the time span during which a user participates in charging. The longer the active period, the more continuous the user's charging behavior; the shorter the active period, the more likely the user is an occasional or one-time user.

[0086] Low-discrimination features include: 1. weekly_freq (weekly charging frequency) Although it reflects users' charging activity, the distribution in this dataset may be concentrated or strongly correlated with other features, resulting in low information gain.

[0087] 2. mean_power & total_energy This redundancy may be due to small differences in battery power among user vehicles, or because it is highly correlated with charging time.

[0088] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0089] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0090] Figure 5 A schematic diagram of the response potential assessment device based on electric vehicle user profiles provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 5 As shown, the response potential assessment device 5 based on electric vehicle user profiles includes: Clustering module 51 is used to cluster the charging load behavior characteristics of electric vehicle users to obtain multiple clusters; each cluster corresponds to a user profile. The acquisition module 52 is used to acquire the demand response potential indicators of electric vehicle users corresponding to the target user profile; wherein, the target user profile is any user profile; Evaluation module 53 is used to input the demand response potential index into the trained LightGBM model to obtain the demand response potential value of electric vehicle users corresponding to the target user profile.

[0091] In one possible implementation, clustering module 51 is also used for: Before clustering the charging load behavior characteristics of electric vehicle users to obtain multiple clusters, an original behavior feature set is constructed; the original behavior feature set includes various charging load characteristics. By combining the charging load characteristics in the original behavioral feature set, multiple candidate behavioral feature sets are obtained; The feature set adaptability evaluation coefficient of each candidate behavior feature set is calculated based on the preset feature set adaptability function, and the optimal feature set is selected based on the feature set adaptability evaluation coefficient. Accordingly, the charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters, including: The charging load behavior characteristics of electric vehicle users are obtained based on the optimal feature set, and the charging load behavior characteristics of electric vehicle users are clustered to obtain multiple clusters.

[0092] In one possible implementation, the feature set fitness function is:

[0093] Where L is the feature set fitness evaluation coefficient of candidate feature set S, G(S) is the information gain index of all features in candidate feature set S, and R(S) is the redundancy index of all features in candidate feature set S.

[0094] In one possible implementation, the evaluation module 53 is also used for: Before inputting the demand response potential index into the trained LightGBM model to obtain the demand response potential value of electric vehicle users corresponding to the target user profile, a sample set is constructed based on the demand response potential index of multiple electric vehicle users. The initial LightGBM model is trained based on the sample set to obtain the trained LightGBM model.

[0095] In one possible implementation, the evaluation module 53 is specifically used for: Based on the trained CWGAN network, the demand response potential indicators of electric vehicle users corresponding to the first user profile are synthesized and completed; where the first user profile is any user profile. A sample set is constructed based on the completed demand response potential indicators.

[0096] In one possible implementation, the discriminator loss function of the CWGAN network is:

[0097] in, For the real sample discrimination term, To generate sample discrimination terms, This is a gradient penalty term. Intermediate samples obtained by interpolating between real and generated samples. for The sampling distribution for The sampling distribution. For the distribution of interpolated samples, These are the weight coefficients of the gradient penalty term; The generator loss function is:

[0098] In the formula, Indicates the joint distribution To calculate the mathematical expectation, For the condition generator model, This is a discriminator model.

[0099] In one possible implementation, clustering module 51 is specifically used for: The number of clusters is determined based on the sum of squared errors to identify the charging load behavior characteristics of electric vehicle users; Based on the number of clusters, the GMM clustering algorithm is used to cluster the charging load behavior characteristics of electric vehicle users, resulting in multiple clusters.

[0100] This invention, through feature optimization and cluster analysis, scientifically identifies typical user profiles, fundamentally ensuring the accuracy of the assessment. Based on this, the system quantifies potential indicators from multiple dimensions, such as transferable power and response speed, and utilizes the LightGBM machine learning model for fusion prediction, significantly improving the accuracy and reliability of the assessment results. Furthermore, addressing the issue of imbalanced data, it innovatively employs a CWGAN network for data augmentation and sample expansion, ensuring the model's generalization ability. Ultimately, it achieves accurate and efficient assessment of demand response potential, providing direct evidence for power grid companies to accurately select users and formulate differentiated incentive strategies. This transforms electric vehicle clusters from a burden on the power grid into a high-quality and controllable regulatory resource, demonstrating significant engineering application value.

[0101] Figure 6 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 6 As shown, the electronic device 6 of this embodiment includes a processor 60 and a memory 61. The memory 61 stores a computer program 62. When the processor 60 executes the computer program 62, it implements the steps in the various method embodiments described above. Alternatively, when the processor 60 executes the computer program 62, it implements the functions of each module / unit in the various device embodiments described above.

[0102] For example, computer program 62 may be divided into one or more modules / units, which are stored in memory 61 and executed by processor 60 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 62 in electronic device 6.

[0103] Electronic device 6 may include, but is not limited to, processor 60 and memory 61. Those skilled in the art will understand that... Figure 6 This is merely an example of electronic device 6 and does not constitute a limitation on electronic device 6. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 6 may also include input / output devices, network access devices, buses, etc.

[0104] The processor 60 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0105] The memory 61 can be an internal storage unit of the electronic device 6, such as a hard disk or RAM. The memory 61 can also be an external storage device of the electronic device 6, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 61 can include both internal and external storage units of the electronic device 6. The memory 61 is used to store the computer program 62 and other programs and data required by the electronic device 6. The memory 61 can also be used to temporarily store data that has been output or will be output.

[0106] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0107] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0108] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0109] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0110] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0111] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for evaluating response potential based on a user portrait of an electric vehicle, characterized by, The method comprises the following steps: Clustering charging load behavior characteristics of electric vehicle users to obtain a plurality of clustering clusters; wherein each clustering cluster corresponds to a user portrait; Obtaining a demand response potential index of an electric vehicle user corresponding to a target user portrait; wherein the target user portrait is any user portrait; Inputting the demand response potential index into a trained LightGBM model to obtain a demand response potential value of the electric vehicle user corresponding to the target user portrait. 2.The response potential assessment method based on an electric vehicle user portrait according to claim 1, wherein, Before the step of clustering the charging load behavior characteristics of the electric vehicle users to obtain a plurality of clustering clusters, the method further comprises the following steps: Constructing an original behavior characteristic set; wherein the original behavior characteristic set comprises a plurality of charging load characteristics; Combining the charging load characteristics in the original behavior characteristic set to obtain a plurality of candidate behavior characteristic sets; Calculating a feature set adaptability evaluation coefficient of each candidate behavior characteristic set based on a pre-set feature set adaptability function, and selecting an optimal feature set based on the feature set adaptability evaluation coefficient; Correspondingly, the step of clustering the charging load behavior characteristics of the electric vehicle users to obtain a plurality of clustering clusters comprises the following steps: Obtaining the charging load behavior characteristics of the electric vehicle users based on the optimal feature set, and clustering the charging load behavior characteristics of the electric vehicle users to obtain a plurality of clustering clusters. 3.The response potential evaluation method based on an electric vehicle user portrait according to claim 2, characterized in that, The feature set adaptability function is as follows: Wherein L is the feature set adaptability evaluation coefficient of the candidate feature set S, G(S) is the information gain index of all features in the candidate feature set S, and R(S) is the redundancy index of all features in the candidate feature set S. 4.The response potential assessment method based on an electric vehicle user portrait according to claim 1, wherein, Before the step of inputting the demand response potential index into the trained LightGBM model to obtain the demand response potential value of the electric vehicle user corresponding to the target user portrait, the method further comprises the following steps: Constructing a sample set based on demand response potential indexes of a plurality of electric vehicle users; Training an initial LightGBM model based on the sample set to obtain a trained LightGBM model. 5.The response potential evaluation method based on an electric vehicle user portrait according to claim 4, wherein, The step of constructing a sample set based on demand response potential indexes of a plurality of electric vehicle users comprises the following steps: Synthesizing and completing the demand response potential index of the electric vehicle user corresponding to the first user portrait based on a trained CWGAN network; wherein the first user portrait is any user portrait; Constructing a sample set based on the completed demand response potential index. 6.The response potential assessment method based on an electric vehicle user portrait according to claim 5, wherein, The discriminator loss function of the CWGAN network is as follows: wherein, is a real sample discriminant term, is a generated sample discriminant term, is a gradient penalty term, is an intermediate sample obtained by interpolating between a real sample and a generated sample, is is a sampling distribution of is is a sampling distribution of is a distribution of the interpolated sample, is a weight coefficient of the gradient penalty term; The generator loss function is as follows: wherein denotes the joint distribution taking the mathematical expectation, is a conditional generator model, is a discriminator model. 7.The response potential assessment method based on an electric vehicle user portrait according to claim 1, wherein, The step of clustering the charging load behavior characteristics of the electric vehicle users to obtain a plurality of clustering clusters comprises the following steps: Determining the number of clusters of the charging load behavior characteristics of the electric vehicle users based on the sum of squared errors; Clustering the charging load behavior characteristics of the electric vehicle users based on the number of clusters to obtain a plurality of clustering clusters by using a GMM clustering algorithm.

8. An electric vehicle user portrait-based response potential assessment device, characterized by, The method comprises the following steps: A clustering module is configured to cluster charging load behavior characteristics of electric vehicle users to obtain a plurality of clustering clusters; wherein each clustering cluster corresponds to a user portrait; An obtaining module is configured to obtain a demand response potential index of an electric vehicle user corresponding to a target user portrait; wherein the target user portrait is any user portrait; An evaluation module is configured to input the demand response potential index into a trained LightGBM model to obtain a demand response potential value of the target user portrait corresponding to the electric vehicle user.

9. An electronic device, comprising: A computer readable storage medium stores a computer program, and a processor executes the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and a processor executes the computer program to implement the method according to any one of claims 1 to 7.

Images