A method and system for predicting the endurance demand of a car based on the multi-dimensional portrait of the car owner

Abstract

The application discloses a kind of based on the automobile endurance demand prediction method and system of car owner multidimensional image. Cluster algorithm is used to classify car owner users based on image label;The endurance demand classification model of multiple features of car owner user group, namely image label data, is trained using an improved self-attention deep learning algorithm, and the trained model outputs the endurance demand prediction value of each user group, including endurance mileage and prediction probability. The problem that the endurance version of pure electric vehicles is planned by car companies only according to statistical analysis, research and other methods of user group travel mileage data, resulting in the problem of endurance version design solidification, reduces the cost of users' car purchase, and improves user satisfaction.

Description

Technical Field

[0001] This invention belongs to the field of vehicle range demand prediction technology, specifically involving a method and system for predicting vehicle range demand based on a multi-dimensional profile of car owners. Background Technology

[0002] The development of pure electric vehicles carries the responsibility of alleviating the energy crisis and solving environmental pollution problems. In recent years, pure electric vehicles have entered a golden age of development. With the widespread adoption of pure electric vehicles, battery range and price have become two key issues that car owners consider when purchasing a vehicle. With the development of battery technology, range has been significantly improved while battery production costs have decreased. However, in recent years, as automakers have deepened their research into car owners' usage scenarios, battery range versions have become increasingly refined. They no longer simply plan for the highest range version; automakers have also launched different battery range versions under their own brands to meet the range needs of different car owner groups as much as possible. Car owners usually choose a version that meets their own usage needs and is reasonably priced, rather than directly pursuing the highest range version. As shown in Table 1, most currently available pure electric vehicles offer range versions concentrated in the 300-600km range range. However, the proportion of 150-200km range versions has increased significantly, while the proportions of 200-300km and 300-400km range versions have gradually decreased. The proportion of versions with a range of over 400km shows a fluctuating trend. This indicates that car owners' demand for different range versions is influenced by diverse factors, and they do not only focus on the highest range versions but also consider factors such as charging convenience and price. Therefore, car owners' demand for pure electric vehicle range versions needs to be predicted and analyzed comprehensively based on multiple factors so that automakers can launch range version classifications that meet the needs of car owners and improve owner satisfaction.

[0003] Table 1. Trends in the Proportion of Pure Electric Vehicles by Range Version

[0004] Driving range (km) 2017 2018 2019 2020 2021 2022 Below 150 8.2％ 0.9％ 0.0％ 7.0％ 7.7％ 6.9％ 150～200 65.4％ 17.1％ 0.7％ 6.4％ 9.9％ 12.6％ 200～300 9.5％ 18.9％ 2.0％ 1.9％ 0.4％ 0.6％ 300～400 13.9％ 38.5％ 36.5％ 22.9％ 17.3％ 16.7％ 400～500 1.3％ 18.2％ 30.2％ 33.9％ 35.6％ 26.3％ 500～600 1.7％ 6.4％ 25.0％ 16.9％ 21.9％ 26.0％ More than 600 0.0％ 0.0％ 5.6％ 10.9％ 7.3％ 10.9％

[0005] Currently, automakers and charging operators generally only conduct predictive analysis on the charging demand of electric vehicles to improve charging support services. However, there is no accurate predictive analysis method for the range demand of different car owner groups for pure electric vehicles. This leads to automakers' rigid design and planning of pure electric vehicle range versions, resulting in a mismatch between range versions and car brand level and price, and low owner satisfaction.

[0006] Current methods for predicting the range demand of pure electric vehicles mostly rely on historical marketing data, owner travel and charging behavior data, and quantitative analysis using mathematical time management techniques, such as multifactor analysis. Multifactor analysis predicts the future by calculating the relationship between demand and its influencing factors. Common causal analysis methods include multiple linear regression, random forest, gradient boosting tree, and SARIMAX.

[0007] Because people's travel patterns vary in time and space, and their driving behaviors are diverse, environmental factors such as climate also affect driving range. Traditional range demand forecasting methods ignore other user needs regarding vehicle type and driving behavior, leading to inaccurate predictions and estimation errors. This results in automakers rigidly categorizing and designing range versions of pure electric vehicles, causing a mismatch between range versions and vehicle brand level and price, leading to low owner satisfaction. Summary of the Invention

[0008] To effectively improve the accuracy of predicting the range demand of pure electric vehicles among car owners, this invention proposes a method and system for predicting vehicle range demand based on a multi-dimensional profile of car owners.

[0009] One of the objectives of this invention is a method for predicting vehicle range demand based on a multi-dimensional profile of car owners, comprising the following steps:

[0010] S1. Use clustering algorithms to classify car owner users based on profile tags;

[0011] S2. An improved self-attention deep learning algorithm is used to train a classification model for range demand on multiple profile label data of classified car owner users. The trained model outputs a range demand prediction value for each user group, and the range demand prediction value includes the range mileage.

[0012] Self-attention deep learning algorithms are models that closely resemble human brain activity, providing a more realistic simulation of the human brain. In user profiling, the quality of results deteriorates in the traditional Encoder-Decoder framework as the number of attributes in the profile increases (for example, in the human brain, with too many attributes, the human cannot remember the entire attribute set for evaluation; instead, it divides them into smaller segments and evaluates them one by one). This is because encoding an excessively long set of attributes into a fixed-length vector easily leads to the loss of information within the attributes. Self-attention deep learning models effectively solve this problem. At each time step, attribute calculation focuses only on the most relevant parts, thus improving the quality of profile attribute calculation.

[0013] Self-attention mechanisms are essentially a network architecture designed to address issues such as handling inputs consisting of numerous vectors of uncertain magnitude. Specifically, a self-attention mechanism incorporates three vectors: Query, Key, and Value. The QKV concept is derived from information retrieval systems; using this invention as an example:

[0014] When creating a profile of a user based on a certain dimension, that dimension is the Query (basic attributes of the car owner). The system matches the Query with the Key (gender, age group, driving experience, and years since purchase). Finally, the system obtains the matched content Value (a weighted matrix of values ​​such as gender, age group, driving experience, and years since purchase) based on the similarity between the Query and the Key.

[0015] In matrix computation, the dot product is one of the methods to calculate the similarity between two matrices. Therefore, QKT is used in the following formula to calculate the similarity. Next, the matching is output based on the similarity. Here, a weighted matching method is used, and the weight is the similarity between the query and the key.

[0016] Att(Q,K,V)=ω(QK T V

[0017] Att(Q,K,V) is the obtained attention value, where Q, K, and V are the query vector, key vector matrix, and value vector, respectively. These three matrices are calculated from the original feature values.

[0018] Self-attention mechanisms reduce reliance on external information and are better at capturing the internal correlations of data or features. However, in traditional self-attention models, because each point needs to capture global contextual information, the self-attention mechanism module has high computational complexity and memory requirements, and generates a large amount of redundant information. Furthermore, in traditional methods, each feature value in the feature vector is calculated with other feature values, resulting in extremely high computational complexity and redundant information. Therefore, in this invention, in step S2, this problem is solved by adding weight masking based on the improved self-attention model.

[0019] Furthermore, the calculation formula for the improved self-attention deep learning algorithm includes:

[0020] Att(Q,K,V)=ω(W1*QK T V*W2

[0021] In the formula:

[0022] Att(Q,K,V): represents the final predicted user battery life demand matrix, where each element in the matrix represents the battery life demand of each group;

[0023] ω: Represents the parameters that need to be trained in deep learning;

[0024] QKV represents the vector parameters of the self-attention model, where Q is the dimension matrix of the car owner user; K is the profile label matrix of the car owner user; and V is the profile label weight matrix of the user, which has the same size as the original data.

[0025] W1 and W2 are setting parameters used to modify the correlation between the dimensions to which the profile tags of the two car owner users belong.

[0026] The characteristic dimensions of car owner users are categories that divide multiple characteristics, such as: basic car owner attributes, car owner travel time and space attributes, car owner driving behavior attributes, car owner charging behavior attributes, and car owner vehicle attributes.

[0027] The basic attributes of car owners are demographic characteristics, such as gender, age group, driving experience, and years since car ownership.

[0028] The spatiotemporal attributes of a car owner's trip refer to the time and space characteristics of the trip, such as: the driving mileage within a set number of days, the driving time within a set number of days, the travel time period and frequency, high-frequency stopping points, high-frequency road condition types, and the highest temperature during the trip.

[0029] The driver's driving behavior attributes, i.e. driving habit characteristics, include average speed per single trip, maximum speed per single trip, average motor output current and / or voltage, maximum motor output current and / or voltage, SOC change per 100 kilometers, and average number of rapid accelerations per single trip.

[0030] The charging behavior attributes of car owners are the characteristics of their charging habits, such as SOC during charging, high-frequency charging duration, charging frequency / week, average charging interval, charging station type (fast charging and slow charging), and high-frequency charging location (POI).

[0031] The vehicle attributes of the owner refer to the characteristics of the vehicles already owned by the owner, such as the range version (mileage), vehicle price range, and the number of configuration differences compared with other range versions of the same model.

[0032] Furthermore, traditional methods for user segmentation include grouping based on a single dimension of the user profile or using traditional clustering algorithms such as k-means. However, the value of K needs to be determined manually beforehand and cannot be changed throughout the algorithm. When dealing with high-dimensional, massive datasets, it is often difficult to accurately estimate the size of K, leading to redundancy in the segmentation dimensions and a high risk of losing group characteristics, which is detrimental to improving the accuracy of the group's range demand prediction model. Therefore, in step S1, a binary k-means clustering algorithm is used to segment car owner users; this solves the problem that when initializing k random centroids, one or more points may disappear during the iteration process due to their extreme positions.

[0033] Furthermore, in step S1, the method for classifying car owner users includes:

[0034] S101. Calculate the sum of squared errors (SSE) with all vehicle owner users as parent nodes;

[0035] S102. On this parent node, the K-means clustering method is used to divide the car owner user group into two groups, and then the SSE is calculated for each group. These are called child node SSE.

[0036] S103. Calculate the difference between the SSE of the parent node and the SSE of each child node, select the group to which the node with the largest difference belongs as the parent node, return to step S102 for partitioning, and do not partition other groups; until the total number of clusters reaches the preset value K.

[0037] The calculation method for SSE is as follows:

[0038]

[0039] In the formula:

[0040] n is the total number of image tags;

[0041] ω i Let y be the weight value of the i-th image label; i Let i be the value of the i-th image tag;

[0042] y * This represents the average value of the profile tag for all car owner users within the group.

[0043] The second objective of this invention is a vehicle range demand prediction system based on a multi-dimensional profile of car owners, which includes a classification module and a range prediction module.

[0044] The classification module is used to classify car owner users based on profile tags using a clustering algorithm;

[0045] The range prediction module is used to train a classification model for grouped car owner users using an improved self-attention deep learning algorithm. The trained model outputs the predicted range demand for each user group.

[0046] Furthermore, the classification module includes an error calculation module, used to calculate the sum of squared errors (SSE) with all vehicle owner users as parent nodes.

[0047] Furthermore, the classification module includes: a first group division module, used to divide the car owner user group into two groups using the K-means clustering method on the parent node.

[0048] Furthermore, the classification module includes a second group division module, which is used to calculate the difference between the SSE of the parent node and the SSE of each child node, and continuously divide the group to which the child node with the largest difference belongs using the K-means clustering method until it is divided into k groups.

[0049] Beneficial effects:

[0050] This invention incorporates charging pile distribution data, comprehensively considers car owners' preferences for other vehicle features and the correlation between multiple features, and employs an improved self-attention deep learning algorithm based on a weighted masking layer to train a range demand prediction model. This solves the problem that car manufacturers plan pure electric vehicle range versions solely based on statistical analysis and surveys of user group travel mileage data, leading to a fixed range version design. This reduces the car purchase cost for users and improves user satisfaction. Attached Figure Description

[0051] Figure 1 The construction process of the group demand prediction method provided in the embodiments of the present invention. Detailed Implementation

[0052] The following detailed embodiments are provided to explain the technical solutions of the claims of this invention, so that those skilled in the art can understand the claims. The scope of protection of this invention is not limited to the following specific embodiments. Any modifications made by those skilled in the art that incorporate the technical solutions of the claims but differ from the following detailed embodiments are also within the scope of protection of this invention.

[0053] The following is in conjunction with the appendix Figure 1 An embodiment of the method described in this invention is presented.

[0054] This invention proposes a method for predicting the range demand of pure electric vehicles based on the profile of car owners. It mainly includes three steps / modules: (1) data preprocessing; (2) car owner profile construction, that is, constructing a car owner profile label system, processing the profile labels, i.e. feature values, and grouping car owners based on the car owner profile; (3) training a self-attention deep learning model based on a weighted masking layer to identify the range demand of different car owner groups for pure electric vehicles and obtain the corresponding range and prediction probability.

[0055] Step 1. Data Preprocessing

[0056] By acquiring data on the spatiotemporal distribution of pure electric vehicle travel, driving behavior, charging behavior, and charging pile distribution through vehicle-to-everything (V2X) big data analytics, Kalman filtering is used to remove noisy and invalid data due to the generation of such data. The required features and data processing procedures are as follows: Figure 1 As shown.

[0057] Step 2. Car Owner Profile Construction

[0058] (1) Construction of a vehicle owner profile tagging system

[0059] To improve the accuracy of range demand prediction when training the range demand prediction model, the following five dimensions are selected to construct a car owner profile tag system based on vehicle network big data.

[0060]

[0061] Table 1 Car Owner Profile Tagging System

[0062] (2) Feature processing and car owner grouping algorithm

[0063] The label values ​​for each dimension are standardized as follows. Then, all car owners are first classified using a clustering algorithm, and finally, the range requirements are predicted for each user group. The specific clustering algorithm process is as follows:

[0064] 1. Standardize the features. Since multiple feature values ​​do not fall within the same value range, it is necessary to standardize the vehicle owner profile label values. For example, range, duration, and frequency need to be normalized using standard deviation. The formula is:

[0065]

[0066] In the formula:

[0067] σ is the standard deviation;

[0068] u is the average value of the characteristic value χ;

[0069] χ` is the normalized feature value, that is, the normalized car owner profile label value.

[0070] 2. Suppose that all car owners need to be divided into K categories of car owner groups. First, treat all car owner users as one category and name it the parent node. Calculate the sum of squared errors (SSE).

[0071]

[0072] In the formula:

[0073] ω i The weight value of the i-th feature (i.e., the owner profile label);

[0074] y i Let i be the value of the profile label for the i-th car owner;

[0075] y * This represents the average value of the profile tag for all car owner users within the group.

[0076] 3. Perform K-means clustering calculation (k=2) on the parent node to divide the car owner group into 2 groups. Name these 2 groups of car owners as child nodes, and then calculate the SSE of the child nodes respectively.

[0077] 4. Calculate the SSE difference, i.e., parent node SSE - child node SSE. Select the child node with the largest SSE difference as the new parent node, return to step 3 above to continue the group partitioning, and do not partition other groups; the algorithm ends when the total number of clusters reaches the preset value K.

[0078] Step 3. Training a self-attention deep learning model based on a weighted masking layer

[0079] After grouping users in step 2, the feature values ​​of the profile tags of each type of car owner user are used as input, and a self-attention deep learning algorithm based on weighted masking layer is used to train the classification model, outputting the range requirements of each user group.

[0080] In this embodiment, the self-attention deep learning self-attention model based on the weighted masking layer is as follows:

[0081] Att(Q,K,V)=ω(W1*QK T V*W2

[0082] In the formula:

[0083] Att(Q,K,V): represents the final predicted user battery life demand matrix, where each element in the matrix represents the battery life demand of each group;

[0084] ω: Represents the parameters that need to be trained in deep learning;

[0085] Q is the dimension matrix of car owner users; K is the profile tag matrix of car owner users; V is the profile tag weight of users, obtained from the deep learning network model;

[0086] W1 and W2 are preset parameters. By setting the values ​​of W1 and W2, the correlation between certain dimensions can be manually modified. For example, if the parameter W1 for the vehicle owner's travel time-space attribute and driving behavior attribute is set to 0, it means that the vehicle owner's travel time-space attribute and driving behavior attribute will have no correlation. This involves classifying the statistical feature vector values, setting the weight matrix corresponding to the correlated feature values ​​that require calculation to 1, and setting the corresponding positions of feature values ​​that are not correlated and do not need calculation to 0, thereby reducing computational load and information redundancy. This correlation is set manually based on experience and the expected behavior of the system.

[0087] 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 this application.

[0088] This application embodiment also provides a vehicle range demand prediction system based on a multi-dimensional profile of the vehicle owner, including a classification module and a range prediction module;

[0089] The classification module is used to classify car owner users based on profile tags using a clustering algorithm;

[0090] The range prediction module is used to train a classification model for grouped car owner users using an improved self-attention deep learning algorithm. The trained model outputs the predicted range demand for each user group.

[0091] Preferably, the classification module further includes an error calculation module, used to calculate the sum of squared errors (SSE) with all vehicle owner users as parent nodes.

[0092] Preferably, the classification module further includes: a first group division module, used to divide the car owner user group into two groups using the K-means clustering method on the parent node.

[0093] Preferably, the classification module further includes: a second group division module, used to calculate the difference between the SSE of the parent node and the SSE of each child node, and continuously divide the group to which the node with the largest difference belongs using the K-means clustering method until it is divided into k groups.

[0094] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A method for predicting vehicle range demand based on multi-dimensional vehicle owner profiles, characterized in that, Includes the following steps: S1. Use clustering algorithms to classify car owner users based on profile tags; S2. An improved self-attention deep learning algorithm is used to train a classification model for range demand on multiple profile tag data of classified car owner users. The trained model outputs the predicted range demand value for each user group. The calculation formula for the improved self-attention deep learning algorithm includes: ； In the formula: : This represents the final predicted user battery life demand matrix, where each element in the matrix represents the battery life demand of each group; : Indicates the parameters that need to be trained in deep learning; Q, K, and V are the vector parameters of the self-attention model, where Q is the matrix of the dimensions to which the profile tags of the car owner user belong; K is the profile tag matrix of the car owner user; and V is the profile tag weight of the user. and The parameters are set to modify the correlation between the dimensions to which the profile tags of the two car owner users belong; In S1, the binary K-means clustering algorithm is used to group car owner users, including: S101. Calculate the sum of squared errors (SSE) with all vehicle owner users as parent nodes; S102. On this parent node, use the K-means clustering method to divide the car owner user group into two groups, and then calculate SSE for each group. S103. Calculate the difference between the SSE of the parent node and the SSE of each child node, select the child node with the largest difference as the new parent node, return to step S102 to perform partitioning, and do not partition other groups; until the total number of clusters reaches the preset value K.

2. The method for predicting vehicle range demand based on multi-dimensional vehicle owner profiles as described in claim 1, characterized in that, The sum of squared errors (SSE) is calculated as follows: ； In the formula: n is the total number of image tags; Let be the weight value of the i-th image tag; Let i be the value of the i-th image tag; This represents the average value of the profile tag for all car owner users within the group.

3. A vehicle range demand prediction system based on a multi-dimensional profile of the vehicle owner, using the method described in claim 1, characterized in that, Includes a classification module and a battery life prediction module; The classification module is used to classify car owner users based on profile tags using a clustering algorithm; The range prediction module is used to train a classification model for grouped car owner users using an improved self-attention deep learning algorithm. The trained model outputs the predicted range demand for each user group.

4. The vehicle range demand prediction system based on multi-dimensional vehicle owner profiles as described in claim 3, characterized in that, The classification module also includes an error calculation module, which is used to calculate the sum of squared errors (SSE) with all vehicle owner users as parent nodes.

5. The vehicle range demand prediction system based on multi-dimensional vehicle owner profiles as described in claim 3, characterized in that, The classification module further includes: a first group division module, which is used to divide the car owner user group into two groups using the K-means clustering method on the parent node.

6. The vehicle range demand prediction system based on multi-dimensional vehicle owner profiles as described in claim 3, characterized in that, The classification module further includes: a second group partitioning module, used to calculate the difference between the SSE of the parent node and the SSE of each child node, select the child node with the largest difference as the new parent node, and use the first group partitioning module to partition the new parent node, while other groups are not partitioned; until the total number of clusters reaches a preset value K.

Images