A method for constructing electric vehicle travel features based on a double-layer clustering architecture

Abstract

The application discloses a kind of electric vehicle travel feature construction methods based on double-layer clustering architecture, it is related to electric vehicle travel feature mining technical field, the method includes: constructing city benchmark feature library;The historical operation data of target electric vehicle is preprocessed, and travel segment is divided based on vehicle state change, and the basic information of each travel segment is extracted;Based on city benchmark feature library, construct unified multidimensional feature vector;Application double-layer clustering architecture, first layer clustering uses variation bayes Gaussian mixture model to cluster normalized single travel feature vector, identify micro travel mode, second layer clustering is based on user behavior portrait vector, using the division clustering algorithm of improved initial center point selection to the user group is grouped, identify macroscopic travel habit;When the amount of travel data newly accumulated reaches preset condition, update city benchmark feature library.The feature construction method of the application can realize cross-city, hierarchical fine description.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle travel feature mining technology, and in particular to a method for constructing electric vehicle travel features based on a two-layer clustering architecture. Background Technology

[0002] With the widespread application of electric vehicles, their travel behavior exhibits highly random and spatiotemporally dynamic characteristics, with particularly significant differences in migration behavior between cities. Traditional methods typically rely on questionnaires or limited datasets, making it difficult to comprehensively reflect the diverse travel characteristics of large-scale electric vehicle users. Furthermore, the construction of travel features often fails to adequately consider the essential differences between intra-city and inter-city travel, such as key dimensions like travel distance, time span, and location changes. In addition, existing technologies mostly employ static clustering or periodic analysis methods, which are ill-suited to the dynamic evolution of travel behavior, resulting in lagging feature construction and insufficient accuracy, thereby affecting the overall effectiveness of traffic planning and energy dispatch. Summary of the Invention

[0003] In view of this, the present invention proposes a method for constructing electric vehicle travel features based on a two-layer clustering architecture, in order to solve the problems of poor cross-city generalization and imprecise characterization of travel behavior in existing electric vehicle travel feature mining methods.

[0004] The specific technical solution of this invention is as follows: A method for constructing electric vehicle travel features based on a two-layer clustering architecture includes: A city benchmark feature library is constructed, which contains benchmark parameters in spatial, temporal, and behavioral dimensions to eliminate scale differences between different cities. The historical operating data of the target electric vehicle is preprocessed, including data cleaning, outlier removal, and missing value imputation. Based on changes in vehicle status, travel segments are divided, and basic information of each travel segment is extracted. Based on the city benchmark feature library, the travel segments are subjected to city adaptive normalization processing to construct a unified multi-dimensional feature vector. The normalization processing includes spatial feature normalization, temporal feature alignment and behavioral feature standardization. A two-layer clustering architecture is applied. The first layer uses a variational Bayesian-Gaussian mixture model to cluster the normalized single-trip feature vectors to identify micro-travel patterns. The second layer uses a partitioning clustering algorithm with improved initial centroid selection based on user behavior profile vectors to group user groups and identify macro-travel habits. When the amount of newly accumulated travel data reaches the preset conditions, a smoothing strategy is adopted to dynamically update the city's benchmark feature database to ensure long-term accuracy.

[0005] Specifically, constructing the urban benchmark feature database includes: in the spatial dimension, calculating the average travel distance of all travel segments in the target city as the benchmark value of the city's average travel distance, and calculating the high quantile value as the benchmark value of the maximum travel radius; in the temporal dimension, analyzing the departure time distribution of all travel segments, fitting the probability density curve through kernel density estimation, and determining the time period corresponding to the peak as the benchmark value of the typical peak travel period; in the behavioral dimension, calculating the average daily number of trips of all sample vehicles as the benchmark value of the average daily travel frequency, and the average driving speed of all sample vehicles as the benchmark value of the average driving speed.

[0006] Specifically, the preprocessing of historical operation data includes: collecting operation data of electric vehicles, including vehicle identifier, latitude and longitude coordinates, timestamp, instantaneous speed and cumulative mileage; performing data cleaning to identify and remove outlier data points, and using interpolation methods to fill in missing values; dividing the trip into segments based on the vehicle state sequence, marking the trip start point when the vehicle changes from stationary to moving, and marking the trip end point when it changes from moving to stationary; and extracting basic information for each trip segment, including start time, end time, origin latitude and longitude, destination latitude and longitude, travel distance and duration.

[0007] Specifically, the city adaptive normalization process includes: using a hexagonal spatial grid index system to divide the city's geographical areas, and determining whether the travel type is intra-city or inter-city travel. Intra-city travel is defined as travel where both the origin and destination are located in the city's core area and the normalized relative travel distance is less than a set threshold; for spatial features, the original travel distance is converted into a relative distance relative to the city's maximum travel radius benchmark; for temporal features, the departure time is aligned with the city's peak hour benchmark, and the time offset is calculated; for behavioral features, statistical standardization methods are used to process the daily average travel frequency and average driving speed, and a standard score is calculated based on the city's benchmark.

[0008] Specifically, the first layer of clustering uses a variational Bayesian-Gaussian mixture model, which automatically infers the optimal number of clusters and calculates the probability distribution of each single trip feature vector belonging to different micro-travel modes. The feature vector contains normalized spatial, temporal, and behavioral features. Micro-travel modes include short-distance commuting, long-distance intercity travel, and nighttime travel types, and the identification results are comparable across cities.

[0009] Specifically, the second-level clustering includes: constructing a behavioral profile vector for each user, which consists of the frequency or number of times a user belongs to different micro-travel patterns in all historical travel segments; applying an improved initial centroid selection partitioning clustering algorithm, which randomly selects a user vector as the initial centroid in the first round, and then selects the user vector that is farthest from the selected centroid in subsequent rounds, iteratively performing data allocation and centroid updates until convergence or the maximum number of iterations is reached, thus completing the user group segmentation.

[0010] Specifically, the city benchmark feature database is updated using an exponentially weighted moving average strategy. When the accumulation of new data triggers an update, a new benchmark value is calculated for each benchmark parameter. The formula is that the new benchmark value equals the old benchmark value multiplied by the smoothing factor plus the new data statistical value multiplied by one minus the smoothing factor, ensuring a smooth and stable update process.

[0011] Specifically, the construction of the multidimensional feature vector includes: the spatial feature part includes the relative travel distance value and travel type identifier; the temporal feature part includes the departure time offset, duration and date type identifier; the behavioral feature part includes the standardized travel frequency and average driving speed; all features are normalized to achieve cross-city comparability.

[0012] Specifically, trip segments are divided based on changes in vehicle state. This involves scanning a sequence of vehicle states, marking the start of a trip when the state changes from stationary to moving, and marking the end of a trip when the state changes from moving to stationary. Basic information extraction ensures that each segment contains spatiotemporal and behavioral attributes, providing input for subsequent feature construction.

[0013] Specifically, in the two-layer clustering architecture, the output of the first-layer clustering serves as the input of the second-layer clustering; the user behavior profile vector is formed by the frequency distribution of micro-travel patterns; the improved initial centroid selection partitioning clustering algorithm reduces the number of iterations by optimizing the initial centroid, thereby improving clustering efficiency and accuracy.

[0014] The beneficial effects of this invention are as follows: 1. Propose and construct a city benchmark feature library containing multi-dimensional parameters such as space, time, and behavior. Use it to normalize the original features, eliminate the influence of city differences, and enable travel data from different cities to be effectively analyzed in the same model.

[0015] 2. A two-layer clustering architecture combining VB-GMM and K-Means++ algorithms is applied to identify single-trip patterns at the micro level and group users based on their behavior patterns at the macro level, thereby achieving hierarchical and refined feature construction.

[0016] 3. By introducing a city benchmark feature library and an adaptive normalization mechanism, the problem of poor generalization ability of traditional methods across different cities is effectively solved, and the constructed travel features have cross-city comparability and consistency.

[0017] 4. A two-layer clustering architecture is adopted to achieve a hierarchical and detailed characterization from micro-level single-trip patterns to macro-level user travel habits, effectively solving the problem that single clustering is insufficient for characterizing complex travel behaviors.

[0018] 5. An online update mechanism is introduced. When the newly accumulated travel data reaches the trigger condition, an exponentially weighted moving average strategy is used to smoothly update the benchmark feature library, ensuring the accuracy and reliability of the method in long-term application. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the electric vehicle travel feature construction method based on a two-layer clustering architecture according to the present invention. Detailed Implementation

[0021] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0022] This invention proposes a method for constructing electric vehicle travel characteristics based on a two-layer clustering architecture. This method aims to dynamically and meticulously construct the travel characteristics of electric vehicle users from two levels: micro-level travel patterns and macro-level user habits. This provides a reliable data foundation for improving the accuracy of charging load prediction, thereby supporting the optimized operation of smart distribution networks. The core of this method lies in establishing a cross-city comparable travel characteristic system and revealing the patterns of travel behavior through hierarchical clustering analysis.

[0023] The overall technical architecture of this invention is as follows: Figure 1 As shown, this method first establishes a city benchmark feature library to provide a scale standard for subsequent processing; then, in the feature construction stage, it performs city-adaptive normalization processing to ensure that the features are comparable across cities; finally, through two-layer clustering, it realizes the mining from micro-level travel patterns to macro-level user profiles.

[0024] Specifically, the process begins with the collection and preprocessing of raw data, followed by segmentation of travel segments. Historical operational data of the target electric vehicle is collected, including but not limited to the vehicle's unique identifier, the latitude and longitude coordinates of the recorded points, timestamps accurate to the second, instantaneous speed, and cumulative mileage. The collected raw data undergoes a cleaning process, primarily identifying and removing anomalous data points that significantly deviate from the normal range, and appropriately filling data gaps caused by missing records, such as using nearest-neighbor interpolation or statistical filling methods. After cleaning, travel segments are segmented based on vehicle state changes. The specific segmentation rule is as follows: the vehicle state sequence is scanned chronologically. When the vehicle transitions from a stationary state to a moving state, that moment is marked as the start point of a trip; correspondingly, when the vehicle returns from a moving state to a stationary state, that moment is marked as the end point of the trip. For each identified travel segment, extract its key basic information attributes, including but not limited to the precise start time of the trip, the precise end time of the trip, the latitude and longitude coordinates of the departure point, the latitude and longitude coordinates of the destination point, the absolute distance traveled in the trip, and the duration from start to finish.

[0025] Secondly, an independent urban benchmark feature database is constructed for each target city to overcome the analytical obstacles caused by differences in spatial scale and residents' travel habits between different cities, ensuring that the method has good generalization ability. This benchmark database is built by aggregating the statistical features of all sample vehicles within the city on travel segments.

[0026] The calculation method for key benchmark parameters is as follows: (1) In the spatial scale dimension, statistically analyze the actual driving distance data of all travel segments generated by all vehicles in the city, calculate the arithmetic mean of these distance data, and obtain the benchmark value of the city's average travel distance representing the typical travel distance of the city. Simultaneously, a predetermined high quantile value for these distance data is calculated (e.g., the distance value corresponding to the 95th percentile), and this value is used as the benchmark for the maximum travel radius of the city. (2) In terms of time distribution, the departure time data of all travel segments in the city are collected, and the kernel density estimation algorithm is applied to perform statistical analysis on these time points. The probability density distribution curve of the departure time in the whole day is fitted, and the time period corresponding to the highest peak in the distribution curve is identified. This time period is officially determined as the benchmark value of the typical peak travel period in the city. (3) In terms of behavioral characteristics, for each sample vehicle in the city, the average number of trips per day and the average driving speed of all trip segments are calculated. Then, the average number of trips per day calculated for all sample vehicles is averaged again to obtain the benchmark value of the average daily trip frequency of the city. Similarly, the average driving speeds calculated from all sample vehicles are averaged to obtain the benchmark average driving speed for the city. The city baseline feature library provides a standardized reference at the city's own scale for subsequent feature processing.

[0027] Next, based on the travel segment data obtained from the aforementioned preprocessing and the benchmark feature library of the corresponding cities, we will carry out city-adaptive feature construction and spatial grid division, with the goal of generating a unified multidimensional feature vector that is comparable across cities.

[0028] First, spatial features are constructed and travel type is identified. A hexagonal spatial grid indexing system (H3 Geographic Information System) is used to perform fine-grained grid division of the urban geographic area. Utilizing the spatial indexing capabilities provided by this grid system, it is determined whether both the origin and destination grids of a trip are located within a predefined urban core area. The standard for defining intra-city travel is: both the origin and destination grids of a trip are within the same urban core grid area, and the relative travel distance is calculated after subsequent urban-scale normalization. The distance must be less than a pre-set threshold. Trips that do not meet the above conditions are classified as intercity trips.

[0029] Next, adaptive city-based normalization is performed on key features. For spatial distance features, the original absolute travel distance values ​​are converted into relative values ​​relative to the current city spatial scale: in, This is the original mileage for this trip. This represents the current maximum travel radius of the city.

[0030] For time features, especially departure time features, time alignment processing is performed: in, and These are the departure time and the baseline peak time for the current city, respectively.

[0031] For behavioral characteristics, such as the average daily frequency of vehicle trips. or average driving speed For other parameters, the standard statistical standardization method Z-score is used. Specifically, the corresponding benchmark values ​​from the city's benchmark database are used as the basis for calculating the mean and standard deviation. After processing the above features, a unified feature vector that can be compared across cities is constructed for each travel segment. This feature vector mainly contains three categories of features: the first category is normalized spatial features, including relative travel distance values. The first category is the travel type identifier (intra-city / inter-city); the second category is aligned time characteristics, including departure time offset relative to peak city hours. The first category includes the duration of the trip and the type of the trip (weekday / weekend); the second category includes standardized behavioral characteristics, such as standardized trip frequency and average driving speed.

[0032] Then, the core step is to apply a two-layer clustering architecture to perform hierarchical pattern recognition and user segmentation of travel behavior.

[0033] The first layer of clustering focuses on the micro-level, namely, pattern discovery for individual travel segments. Given the unknown number of travel patterns, their potentially complex forms, and the potential overlap between different patterns, this invention preferably employs a Variational Bayesian Gaussian Mixture Model (VB-GMM) to perform cluster analysis on the feature vectors of all travel segments after the aforementioned normalization steps. This model is a soft clustering algorithm, whose advantage lies in its ability to automatically infer the optimal number of cluster components in the data and calculate the probability distribution of each travel segment belonging to different potential patterns, rather than making rigid divisions. This characteristic makes it particularly suitable for identifying complex travel patterns with mixed features and blurred boundaries. Through this step, several statistically significant typical single-trip pattern types can be identified (such as "short-distance commuting," "long-distance intercity travel," and "nighttime travel"). Since the travel features input to this model have been normalized to city benchmarks, the identified micro-travel patterns have consistent definitions and comparability across different cities.

[0034] The basic principle of variational Bayesian Gaussian mixture models is to utilize multiple Gaussian distribution (sub-distribution) components, combined with different mixing weights, to form a more flexible mixture probability distribution model, thereby better fitting the irregular, multimodal distribution patterns that may appear in real-world data. Each Gaussian component is defined by its position and shape parameters, which together determine the feature center point of the pattern represented by that component and its distribution range in the feature space. Its probability density function is expressed as follows: in, N The number of independent sub-distributions in the model determines the total number of cluster categories; It is the first The weights of each independent sub-distribution should be equal to one when all the weights of the sub-distributions are added together. It is the first The probability density functions of *n* independent sub-distributions, and their corresponding parameter vectors are: The probability density function of the sub-distribution is as follows: Wherein, parameter vector , including average Covariance Matrix Two parameters, The characteristic dimension represents the sample data. The mean determines the location of the sub-distribution, while the covariance matrix determines the shape and direction of expansion of the sub-distribution, reflecting the variance and correlation of the data.

[0035] The second layer of clustering elevates the process to a macro level, aiming to segment user groups based on long-term behavioral habits. Building upon micro-level travel pattern recognition, a behavioral profile vector is constructed for each user. This vector is formed by statistically analyzing the frequency or percentage of each micro-level travel pattern identified in all of the user's historical travel segments. Essentially, it constructs a frequency distribution matrix between users and patterns, where each row represents a user's pattern frequency distribution vector, and each column represents a travel pattern identified by VB-GMM. The values ​​of the matrix elements are the frequency or count of the corresponding pattern in all of the user's travel segments. At this point, each user is represented by a deterministic, numerical, high-dimensional feature vector. The goal of clustering is to divide these user vectors into several non-overlapping categories, ensuring that users within the same category share similar long-term travel habit combinations.

[0036] Considering that the input data is a numerical vector and the objective is hard partitioning, this invention preferably employs an improved initial centroid selection method combined with the traditional partitioning strategy of the K-Means++ clustering algorithm to group users. K-Means is an unsupervised clustering algorithm based on prototype partitioning; its core idea is to iteratively optimize... n Data points are divided into kIn a set of mutually exclusive clusters, the algorithm minimizes the sum of squared Euclidean distances from each data point to the centroid of its cluster. The core optimization lies in its initial cluster centroid selection strategy: unlike completely random selection, this strategy first randomly selects a data point as the first initial centroid. Subsequent newly added initial centroids are selected from the remaining data points, choosing the data point furthest from all currently selected centroids. This process is repeated to select a specified number of initial centroids. After selecting the initial centroids, the algorithm enters a standard iterative process: each user vector is assigned to the cluster represented by the nearest centroid based on its spatial distance to the current cluster centroids; then, the average value of all user vectors in each cluster is recalculated, and this average value is updated as the new centroid for that cluster; this iterative process of data point allocation and centroid update is repeated until a preset maximum number of iterations is reached or the sum of the movement distances of all centroids in two consecutive iterations is less than a set convergence threshold, ultimately completing the user group division. The specific steps are as follows: A. Selecting initial cluster centers. This differs from the traditional K-Means algorithm, which directly selects cluster centers randomly. k Using 10 data points as initial cluster centers, the K-Means++ algorithm randomly selects only the first cluster center, and then selects the data point farthest from the selected cluster center according to Euclidean distance to add as a new cluster center. This process is repeated until the first cluster center is selected. k Initial cluster centers ; B. Divide the data into clusters. Use Euclidean distance to measure the distance between the sample data and the cluster centers, and divide the sample data into the clusters with the closest distance. For those that have undergone... After the second clustering, the... Cluster center of each cluster Sample data relatively The formula for calculating Euclidean distance is: Where m represents the feature dimension of the data.

[0037] C. Recalculate the cluster centers. Divide the sum of samples in each cluster by the number of samples to recalculate the cluster mean, and use this mean as the new cluster center. in, Indicates the first In the next iteration, belonging to the cluster The number of data points.

[0038] D. Iterative calculation. Repeat steps B and C until the maximum number of iterations is reached or the change in cluster centers is less than the set threshold.

[0039] Finally, to ensure that the city benchmark feature database can adapt to the gradual evolution of urban travel habits over time, an online dynamic update mechanism is established. When the system detects that the amount of newly accumulated travel data for a particular city reaches a preset trigger condition, the benchmark feature database update process for that city is automatically initiated. To avoid drastic fluctuations in benchmark values ​​due to short-term fluctuations in new batches of data, this invention employs an Exponentially Weighted Moving Average (EWMA) strategy for smooth updates. For each key parameter in the benchmark database that needs updating... The formula for calculating its new benchmark value is: in, This represents the baseline value that will be written to the database after the update. This refers to the old baseline value stored in the database; This is the current statistical value calculated based on the newly accumulated batch data; This is a smoothing factor. Through this mechanism, the model can gradually incorporate new travel trend information while maintaining stability, ensuring the timeliness of the baseline feature library and the long-term applicability of the model.

[0040] The beneficial effects of this invention are as follows: (1) By constructing a city benchmark feature library containing multi-dimensional parameters such as spatial dimension, time dimension and behavioral dimension, and using it to normalize the original features, the difference between different cities can be eliminated, so that cross-city travel data can be effectively analyzed in a unified model, thereby improving the accuracy of charging load prediction.

[0041] (2) A two-layer clustering architecture combining VB-GMM and K-Means++ is adopted. The first layer identifies the micro-pattern of a single trip through soft clustering, and the second layer performs macro-user grouping based on user behavior pattern combination through hard clustering, so as to realize the hierarchical construction of travel features from specific behaviors to overall habits and enhance the refinement of feature expression.

[0042] (3) By introducing a city benchmark feature library and an adaptive normalization mechanism, the shortcomings of traditional methods in generalizing across different cities are effectively overcome, so that the constructed travel features have cross-city comparability and consistency, providing reliable support for the optimized operation of smart distribution networks.

[0043] (4) By using a two-layer clustering architecture, a hierarchical and detailed characterization of user travel habits from micro-level single travel patterns to macro-level user travel habits can be achieved, effectively solving the problem that single clustering methods are insufficient in expressing complex travel behaviors and improving the model's adaptability to diverse travel scenarios.

[0044] (5) Establish an online update mechanism. When the amount of newly accumulated travel data reaches the preset conditions, the weighted moving average strategy is used to smoothly update the baseline feature library to ensure that the method maintains accuracy and reliability in long-term application and adapts to the dynamic evolution of urban travel habits.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing electric vehicle travel features based on a two-layer clustering architecture, characterized in that, include: A city benchmark feature library is constructed, which includes benchmark parameters in spatial, temporal, and behavioral dimensions to eliminate scale differences between different cities. The historical operating data of the target electric vehicle is preprocessed, including data cleaning, outlier removal, and missing value imputation. Based on changes in vehicle status, travel segments are divided, and basic information of each travel segment is extracted. Based on the aforementioned urban baseline feature library, urban adaptive normalization processing is performed on travel segments to construct a unified multidimensional feature vector. The normalization processing includes spatial feature normalization, temporal feature alignment, and behavioral feature standardization. A two-layer clustering architecture is applied. The first layer uses a variational Bayesian-Gaussian mixture model to cluster the normalized single-trip feature vectors to identify micro-travel patterns. The second layer uses a partitioning clustering algorithm with improved initial centroid selection based on user behavior profile vectors to group user groups and identify macro-travel habits. When the amount of newly accumulated travel data reaches a preset condition, a smoothing strategy is used to dynamically update the city benchmark feature library to ensure long-term application accuracy.

2. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, The construction of the urban benchmark feature database includes: in the spatial dimension, calculating the average travel distance of all travel segments in the target city as the benchmark value of the city's average travel distance, and calculating the high quantile value as the benchmark value of the maximum travel radius; in the temporal dimension, analyzing the departure time distribution of all travel segments, fitting the probability density curve through kernel density estimation, and determining the time period corresponding to the peak as the benchmark value of the typical peak travel period; in the behavioral dimension, calculating the average daily number of trips of all sample vehicles as the benchmark value of the average daily travel frequency, and the average driving speed of all sample vehicles as the benchmark value of the average driving speed.

3. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, The preprocessing of historical operation data includes: collecting operation data of electric vehicles, including vehicle identifier, latitude and longitude coordinates, timestamp, instantaneous speed and cumulative mileage; performing data cleaning to identify and remove abnormal data points, and using interpolation methods to fill in missing values; based on the vehicle state sequence, marking the start point of the trip when the vehicle changes from stationary to moving, and marking the end point of the trip when it changes from moving to stationary, and dividing the trip into segments; extracting basic information for each trip segment, including start time, end time, origin latitude and longitude, destination latitude and longitude, travel distance and duration.

4. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, The city adaptive normalization process includes: dividing the city's geographical areas using a hexagonal spatial grid index system, determining whether the travel type is intra-city or inter-city travel, where intra-city travel is defined as both the origin and destination being located in the city's core area and the normalized relative travel distance being less than a set threshold; for spatial features, converting the original travel distance into a relative distance relative to the city's maximum travel radius benchmark; for temporal features, aligning the departure time with the city's peak hour benchmark and calculating the time offset; and for behavioral features, using statistical standardization methods to process the daily average travel frequency and average driving speed, and calculating a standard score based on the city's benchmark value.

5. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, The first layer of clustering uses a variational Bayesian-Gaussian mixture model, which automatically infers the optimal number of clusters and calculates the probability distribution of each single trip feature vector belonging to different micro-travel modes. The feature vector includes normalized spatial features, temporal features, and behavioral features. The micro-travel modes include short-distance commuting, long-distance intercity travel, and nighttime travel types, and the identification results are comparable across cities.

6. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, The second-level clustering includes: constructing a behavioral profile vector for each user, the vector being composed of the frequency or number of times belonging to different micro-travel patterns in all historical travel segments of the user; applying an improved initial centroid selection partitioning clustering algorithm, which randomly selects a user vector as the initial centroid in the first round, and then selects the user vector farthest from the selected centroid in subsequent rounds, iteratively performing data allocation and centroid updates until convergence or the maximum number of iterations is reached, thus completing the user group segmentation.

7. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, The updated city benchmark feature database adopts an exponential weighted moving average strategy. When the new data accumulation triggers an update, a new benchmark value is calculated for each benchmark parameter. The formula is that the new benchmark value equals the old benchmark value multiplied by the smoothing factor plus the new data statistical value multiplied by one minus the smoothing factor, ensuring a smooth and stable update process.

8. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, The construction of the multidimensional feature vector includes: the spatial feature part includes the relative travel distance value and travel type identifier; the temporal feature part includes the departure time offset, duration and date type identifier; the behavioral feature part includes the standardized travel frequency and average driving speed; all features are normalized to achieve cross-city comparability.

9. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, The segmentation of travel is based on changes in vehicle state. Specifically, it involves scanning a sequence of vehicle states, marking the start of the trip when the state changes from stationary to moving, and marking the end of the trip when the state changes from moving to stationary. The extraction of basic information ensures that each segment contains spatiotemporal and behavioral attributes, providing input for subsequent feature construction.

10. The method for constructing electric vehicle travel features based on a two-layer clustering architecture as described in claim 1, characterized in that, In the two-layer clustering architecture, the output of the first-layer clustering is used as the input of the second-layer clustering; the user behavior profile vector is formed by the frequency distribution of micro-travel patterns; the improved initial centroid selection partitioning clustering algorithm reduces the number of iterations by optimizing the initial centroid, thereby improving clustering efficiency and accuracy.

Images