A method and apparatus for predicting vehicle sales
By extracting and processing historical data from the Internet of Vehicles (IoV), performing oversampling and feature engineering, and using convolutional neural networks to build a vehicle sales prediction model, the problem of inaccurate prediction in existing technologies is solved, and more accurate vehicle sales prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-03
- Publication Date
- 2026-03-10
AI Technical Summary
Existing vehicle sales forecasting methods rely on empirical judgment and single-time-segment information, making it difficult to guarantee forecast accuracy and failing to fully utilize the correlation between the characteristic variables and time dimension of vehicle network data.
By extracting historical vehicle network data samples from the vehicle network, oversampling is performed to generate new samples, feature engineering is carried out, the similarity and correlation between vehicle feature variables are calculated, a prediction model is built using a convolutional neural network, and the model training data is converted into a two-dimensional matrix structure for training, generating accurate sales prediction results.
It improves the accuracy of vehicle sales prediction models, better reflects the correlation between vehicle usage patterns and characteristic variables, reduces sample imbalance, and expands the application scope to multi-purpose passenger vehicles.
Smart Images
Figure CN120822990B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of electronic information technology, in particular to a vehicle sale prediction method and device. BACKGROUND
[0002] The automobile market has turned from an incremental market to a stock market, and the existing customer group is increasingly large. Understanding when they plan to replace vehicles is crucial for automakers to establish a stronger customer relationship. With the continuous development of Internet of Vehicles technology, it has become possible to analyze customer behavior patterns before replacing vehicles and predict whether customers are preparing to replace vehicles in the near future with the help of Internet of Vehicles driving data.
[0003] The existing vehicle sale prediction method is based on Internet of Vehicles data to calculate a vehicle replacement index and use it to help automakers identify potential vehicle purchase demand. The key is to extract information in multiple dimensions from Internet of Vehicles data according to experience, and then assign weights to each dimension for weighted calculation to obtain the vehicle replacement index.
[0004] However, the current method of predicting the possibility of replacing vehicles using Internet of Vehicles data only relies on experience, which is to determine what dimension of information to extract from Internet of Vehicles data and assign a weight to it based on experience. It is difficult to ensure prediction accuracy, and only the current state information (cross-sectional information) of Internet of Vehicles data is used for prediction, without using the correlation information between time and features, which affects prediction accuracy. SUMMARY
[0005] The present application provides a vehicle sale prediction method and device, which can solve the problem of inaccurate vehicle sale prediction in the prior art.
[0006] To solve the above technical problems, the present application provides a vehicle sale prediction method, comprising:
[0007] According to a preset time granularity, a plurality of historical Internet of Vehicles data samples are extracted in the Internet of Vehicles; wherein the historical Internet of Vehicles data samples include labels; the labels include sales and no sales;
[0008] Based on a plurality of historical Internet of Vehicles data samples, oversampling is performed to generate a plurality of new samples, and a plurality of historical Internet of Vehicles data samples and a plurality of new samples are determined as model training data;
[0009] The model training data is subjected to feature engineering processing to obtain a plurality of vehicle feature variables for representing usage patterns; wherein the vehicle feature variables include vehicle age, driving range, total mileage, and battery health;
[0010] The similarity between each vehicle feature variable is calculated, and the vehicle feature variables are classified to obtain vehicle feature variables under each category;
[0011] rank the vehicle characteristic variables under each category according to the correlation, to generate a variable ranking result of all vehicle characteristic variables;
[0012] convert the model training data into input data with a two-dimensional matrix structure based on the variable ranking result;
[0013] train a convolutional neural network based on the input data to obtain a first prediction model;
[0014] obtain the Internet of Vehicles data of a to-be-predicted vehicle within a preset time length, generate to-be-predicted vehicle input data with a two-dimensional matrix structure based on the variable ranking result, input the to-be-predicted vehicle input data into the first prediction model, and obtain a sale prediction result of the to-be-predicted vehicle.
[0015] As a preferred solution, the oversampling based on the plurality of historical Internet of Vehicles data samples to generate a plurality of new samples comprises:
[0016] selecting a historical Internet of Vehicles data sample with a label of sale from the plurality of historical Internet of Vehicles data samples to determine as an initial sample;
[0017] For each initial sample, selecting a historical Internet of Vehicles data sample with a similarity greater than a preset threshold to the initial sample from the plurality of historical Internet of Vehicles data samples to determine as an auxiliary sample;
[0018] sample feature processing of the initial sample and the auxiliary sample is performed using the following formula to generate a new sample:
[0019] X new =X m +ε*(X m -X s )
[0020] In the formula, X new is a new sample feature; X m is an initial sample feature; X s is an auxiliary sample feature; and ε is a preset parameter.
[0021] As a preferred solution, the feature engineering processing of the model training data to obtain a plurality of vehicle characteristic variables for representing usage habits comprises:
[0022] extracting features in the model training data to obtain a plurality of feature variables; wherein the feature variables include average driving mileage each time, total mileage per day, parking duration, daily driving mileage, and average driving speed;
[0023] In the Internet of Vehicles data, the correlation between each of the feature variables is obtained.
[0024] Based on the correlation, data processing is performed on each of the characteristic variables to form vehicle characteristic variables for representing usage rules.
[0025] As a preferred solution, the similarity between each of the vehicle characteristic variables is calculated, and the vehicle characteristic variables are classified to obtain vehicle characteristic variables in each category, including:
[0026] Based on the corresponding numerical values of each of the vehicle characteristic variables, the correlation coefficient between each of the vehicle characteristic variables is calculated;
[0027] Based on the correlation coefficient between each of the vehicle characteristic variables, the distance between each of the vehicle characteristic variables is calculated;
[0028] Based on the distance between each of the vehicle characteristic variables, a clustering method is used to classify each of the vehicle characteristic variables to obtain vehicle characteristic variables in each category.
[0029] As a preferred solution, the vehicle characteristic variables in each category are sorted according to the correlation to generate a variable sorting result of all vehicle characteristic variables, including:
[0030] The correlation of each category is sorted to generate an inter-class sorting result;
[0031] Each of the vehicle characteristic variables in each category is sorted according to the correlation to generate a plurality of intra-class sorting results;
[0032] The inter-class sorting result and the plurality of intra-class sorting results are combined to generate a variable sorting result of all vehicle characteristic variables.
[0033] As a preferred solution, the correlation of each category is sorted to generate an inter-class sorting result, including:
[0034] Based on the vehicle characteristic variables in each category, the sum of the eigenvalues of the covariance matrix of each category is calculated;
[0035] The category with the largest sum of eigenvalues of the covariance matrix is determined as the core class;
[0036] The distance between each category and the core class is calculated to obtain an inter-class distance;
[0037] Based on the inter-class distance, each category is sorted to generate an inter-class sorting result.
[0038] As a preferred solution, each of the vehicle characteristic variables in each category is sorted according to the correlation to generate a plurality of intra-class sorting results, including:
[0039] The class center of each category is obtained through clustering analysis;
[0040] For each category, calculate the distance between the vehicle feature variables and the class center for each category to obtain the distance between the variables and the classes.
[0041] Based on the variables and the inter-class distance, the vehicle feature variables under each category are sorted to generate intra-class sorting results.
[0042] As a preferred embodiment, the step of acquiring vehicle-to-everything (V2X) data of the vehicle to be predicted within a preset time period, and generating input data of the vehicle to be predicted with a two-dimensional matrix structure based on the variable sorting results, includes:
[0043] Acquire vehicle-to-everything (V2X) data of the vehicle to be predicted within a preset time period;
[0044] The vehicle network data is subjected to feature engineering processing to obtain several vehicle feature variables for characterizing usage patterns.
[0045] Based on the variable ranking results, several vehicle feature variables to be predicted, which are used to characterize usage patterns, are ranked to form a vehicle feature ranking dataset.
[0046] The vehicle feature sorting dataset is converted into a data format to generate two-dimensional matrix structured vehicle input data to be predicted.
[0047] As a preferred embodiment, after inputting the vehicle input data to the first prediction model to obtain the sales prediction result of the vehicle, the method further includes:
[0048] The sales prediction results of the vehicles to be predicted are input into a preset probability calibration model so that the probability calibration model calibrates the sales prediction results using Platt scaling or ordinal-preserving regression.
[0049] Accordingly, the present invention provides a vehicle sales prediction device, comprising: a sample acquisition module, a sample generation module, a feature extraction module, a feature classification module, a feature sorting module, a data conversion module, a model training module, and a prediction module;
[0050] The sample acquisition module is used to extract several historical vehicle network data samples from the vehicle network according to a preset time granularity; wherein, the historical vehicle network data samples include tags; the tags include "for sale" and "not for sale";
[0051] The sample generation module is used to oversample several historical vehicle network data samples to generate several new samples, and to determine the several historical vehicle network data samples and the several new samples as model training data.
[0052] The feature extraction module is used to perform feature engineering processing on the model training data to obtain several vehicle feature variables used to characterize usage patterns; wherein, the vehicle feature variables include vehicle age, driving range, total mileage and battery health.
[0053] The feature classification module is used to calculate the similarity between the vehicle feature variables and classify the vehicle feature variables to obtain the vehicle feature variables under each category.
[0054] The feature sorting module is used to sort the vehicle feature variables under each category according to their correlation, and generate the variable sorting results of all vehicle feature variables.
[0055] The data conversion module is used to convert the model training data into input data with a two-dimensional matrix structure based on the variable sorting results;
[0056] The model training module is used to train the convolutional neural network based on the input data to obtain a first prediction model.
[0057] The prediction module is used to acquire vehicle network data of the vehicle to be predicted within a preset time period, generate vehicle input data with a two-dimensional matrix structure based on the variable sorting results, input the vehicle input data to be predicted into the first prediction model, and obtain the sales prediction result of the vehicle to be predicted.
[0058] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:
[0059] This invention provides a method for predicting vehicle sales. When constructing a first prediction model for vehicle sales prediction, several historical vehicle network data samples are extracted from the vehicle network according to a preset time granularity. This not only considers vehicle network data at a single time segment but also fully utilizes the correlation between vehicle network data in both feature variables and time dimensions, improving the richness of the model training data. New samples are generated by oversampling based on the historical vehicle network data samples, and then the model training data is derived by combining the historical and new samples to reduce sample imbalance and improve model accuracy. Feature engineering is performed on the model training data to obtain multiple vehicle feature variables representing usage patterns, thus improving model stability. The vehicle feature variables are classified by calculating the similarity between them. The vehicle feature variables under each category are sorted according to their correlation, resulting in a variable ranking result for all vehicle feature variables. Based on this ranking result, the model training data is converted into input data with a two-dimensional matrix structure, and a convolutional neural network is trained using this input data to obtain the first prediction model. This invention generates model training data based on the correlation between vehicle feature variables, effectively increasing the information content of the model training data and thus improving the accuracy of the first prediction model. After the first prediction model is trained, vehicle network data of the vehicle to be predicted is collected within a preset time period. The input data of the vehicle to be predicted with a two-dimensional matrix structure is generated by combining the variable ranking results. This data is used as the input data of the first prediction model so that the first prediction model can make more accurate sales prediction results for the vehicle to be predicted. Attached Figure Description
[0060] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments 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 from these drawings without creative effort.
[0061] Figure 1 A flowchart illustrating one embodiment of the vehicle sales prediction method provided by the present invention;
[0062] Figure 2 This is a schematic diagram of one embodiment of the vehicle sales prediction device provided by the present invention. Detailed Implementation
[0063] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.
[0065] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "several" means two or more, unless otherwise explicitly defined.
[0066] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0067] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.
[0068] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).
[0069] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.
[0070] Example 1
[0071] like Figure 1 The diagram shown is a flowchart of an embodiment of the vehicle sales prediction method provided by the present invention. The method includes steps 101 to 108, and the specific steps are as follows:
[0072] Step 101: Extract several historical vehicle network data samples from the vehicle network according to a preset time granularity; wherein, the historical vehicle network data samples include tags; the tags include "sold" and "not sold".
[0073] In this embodiment of the invention, the vehicle network records vehicle data for multiple different vehicles, including mileage, battery level, latitude and longitude, and time corresponding to each vehicle start and stop. Historical operating data of the target vehicle can be collected within the vehicle network according to a preset time granularity, thereby obtaining multiple historical vehicle network data samples.
[0074] In this embodiment of the invention, historical vehicle network data samples are obtained from the vehicle network. First, historical vehicle network data is acquired, and various vehicle types are extracted from it. Based on a preset time granularity and multiple vehicle types, multiple historical vehicle network data samples can be extracted from the historical vehicle network data. Each historical vehicle network data sample corresponds to a specific vehicle and a specific time point. The historical vehicle network data sample includes features and labels. Features are vehicle network data aggregated within a specific time window, and labels include "for sale" and "not for sale," used to indicate whether the vehicle will be sold within a future time ΔT (e.g., 3 months) after window t. Specifically, each vehicle i is treated as a sample at each time granularity t. For example, with 1000 vehicles and a time granularity set to weeks, 1000 * 52 = 52,000 training samples will be generated annually. The time granularity can be determined according to business needs. The label for each training sample is "for sale" or "not for sale," and the features of the training sample are vehicle feature variables generated at the corresponding sample times t,...,tn.
[0075] Step 102: Oversample several historical vehicle network data samples to generate several new samples, and determine the several historical vehicle network data samples and the several new samples as model training data.
[0076] As a preferred embodiment, oversampling is performed on several historical vehicle-to-everything (V2X) data samples to generate several new samples, including:
[0077] From a number of historical vehicle-to-everything (V2X) data samples, those labeled "for sale" were selected as the initial samples.
[0078] For each initial sample, select a historical vehicle network data sample with a similarity greater than a preset threshold from several historical vehicle network data samples marked "for sale" and determine it as an auxiliary sample;
[0079] The initial sample and the auxiliary sample are processed using the following formula to generate new samples:
[0080] X new =X m +ε*(X m -X s )
[0081] In the formula, X new Features for new samples; X m Features of the initial samples; X s ε is an auxiliary sample feature; ε is a preset parameter.
[0082] In this embodiment of the invention, if we assume that each vehicle is considered a sample every day from the start of its use, the probability of it being "sold" will be extremely low. For example, if a vehicle is sold after 3 years of use, the probability of selling is 1 / (365*3), resulting in a severe imbalance between the proportion of "sold" and "not sold" samples. Training the model with these samples would cause the model to be biased towards predicting "not sold," meaning the model might perform well in the "not sold" category but have a very low recognition rate for the "sold" category; it would ignore the features of the "sold" category, leading to the failure of predictions for "sold" vehicles; and it would distort the model's evaluation metrics. This would affect the model's prediction accuracy in practical applications. Therefore, the training samples for the first prediction model not only include sample data obtained from historical vehicle network data but also need to generate new samples to balance the number of "sold" and "not sold" samples.
[0083] In this embodiment of the invention, the sample data obtained from historical vehicle network data is divided into a training set and a test set. The test set remains unchanged, and oversampling is performed on the training set to generate new samples. The oversampling can employ the SMOTE algorithm, specifically: a sample labeled "for sale" is randomly selected from the training set (this sample is labeled as the initial sample m). For this sample, k nearest auxiliary samples labeled "for sale" are found based on feature similarity, ensuring that the similarity of these auxiliary samples s all exceeds a preset threshold. All found auxiliary samples s are iterated through, and sample feature processing is performed based on the initial sample m and these auxiliary samples s respectively to obtain new samples. The new sample can be represented as:
[0084] X new =X m +ε*(X m -X s )
[0085] In the formula, X new Features for new samples; X m Features of the initial samples; X s ε is a preset parameter between 0 and 1, used for auxiliary sample features.
[0086] Step 103: Perform feature engineering on the model training data to obtain several vehicle feature variables used to characterize usage patterns; wherein, the vehicle feature variables include vehicle age, driving range, total mileage and battery health.
[0087] As a preferred embodiment, the model training data is subjected to feature engineering processing to obtain several vehicle feature variables used to characterize usage patterns, including:
[0088] Features are extracted from the model training data to obtain several feature variables; wherein, the feature variables include average mileage per trip, total mileage per day, parking time, daily mileage and average driving speed;
[0089] In the vehicle network data, the correlation between each of the aforementioned feature variables is obtained;
[0090] Based on the aforementioned correlation, data processing is performed on each of the aforementioned feature variables to form several vehicle feature variables used to characterize usage patterns.
[0091] In this embodiment of the invention, feature engineering is performed on the acquired raw model training data to generate variables with more explicit physical meaning, such as the vehicle's actual operating age, driving range, total mileage, and battery health. Furthermore, the model training data includes features and labels. Feature engineering is performed on the model training data by first extracting physically meaningful vehicle feature variables for model training from raw vehicle network data such as driving time and driving latitude and longitude. These include features such as the number of times a day the vehicle goes out, the average mileage per trip, the total daily mileage, parking time, and average driving speed. By performing time pattern analysis on the feature data of these features, the correlation between each feature variable is obtained. Finally, by integrating the fine-grained values of each feature variable, vehicle feature variables that can characterize usage patterns are formed.
[0092] As an example of an embodiment of the present invention, assume that the original record of the model training data is the date and time of each engine start, along with the mileage and latitude / longitude at the time of start. By performing time pattern analysis on the original model samples, it is possible to analyze how many times a day the vehicle goes out, the average mileage per trip, the total daily mileage, parking time, daily mileage, average speed, and the differences between each indicator and recent average values. Furthermore, based on the parking time, duration, and weekday / weekend pattern differences, it is possible to infer whether the vehicle is used for commuting, weekend travel, school pick-up / drop-off, or visits to a 4S dealership. This forms data on vehicle characteristic variables that characterize vehicle usage patterns for training the model, such as vehicle age, mileage range, total mileage, and battery health.
[0093] Step 104: Calculate the similarity between the vehicle feature variables and classify the vehicle feature variables to obtain the vehicle feature variables under each category.
[0094] As a preferred embodiment, the similarity between the vehicle feature variables is calculated, and the vehicle feature variables are classified to obtain the vehicle feature variables under each category, including:
[0095] Based on the values corresponding to each of the vehicle feature variables, the correlation coefficients between each of the vehicle feature variables are calculated respectively.
[0096] Based on the correlation coefficients between the vehicle feature variables, the distances between the vehicle feature variables are calculated.
[0097] Based on the distance between the vehicle feature variables, a clustering method is used to classify the vehicle feature variables and obtain the vehicle feature variables under each category.
[0098] In this embodiment of the invention, after obtaining the extracted vehicle feature variables, these vehicle feature variables are classified according to their correlation, with highly correlated vehicle feature variables grouped into one category. The classification of vehicle feature variables is based on the distance between them; however, before calculating the distance between vehicle feature variables, the correlation coefficient between them is calculated. This embodiment uses the following formula to calculate the correlation coefficient between vehicle feature variables:
[0099]
[0100] In the formula, Corr(i,j) is the correlation coefficient between vehicle characteristic variable i and vehicle characteristic variable j; Let V be the value of vehicle characteristic variable i at time t; t j Let j be the value of the vehicle characteristic variable j at time t; Let i be the mean value of the vehicle characteristic variable i. Let j be the mean value of the vehicle characteristic variable j.
[0101] After calculating the correlation coefficients between each vehicle characteristic variable, the distance between the vehicle characteristic variables is calculated using the following formula:
[0102] d i,j =1-|Corr(i,j)|
[0103] In the formula, d i,j denoted as , where is the distance between vehicle feature variable i and vehicle feature variable j; Corr(i,j) is the correlation coefficient between vehicle feature variable i and vehicle feature variable j.
[0104] After calculating the distances between each vehicle feature variable, clustering methods (such as hierarchical clustering, K-means clustering, etc.) are used to classify each vehicle feature variable, and the vehicle feature variables under each category can be obtained.
[0105] Step 105: Sort the vehicle feature variables under each category according to their correlation, and generate the variable ranking results of all vehicle feature variables.
[0106] As a preferred embodiment, the vehicle feature variables under each category are sorted according to their correlation to generate a variable ranking result for all vehicle feature variables, including:
[0107] Sort the relevance of each category and generate inter-category ranking results;
[0108] The vehicle feature variables under each category are sorted according to their correlation, generating several intra-class sorting results;
[0109] By combining the inter-class sorting results and several intra-class sorting results, a variable sorting result for all vehicle feature variables is generated.
[0110] In this embodiment of the invention, vehicle feature variables of different categories are sorted both inter-class and intra-class, resulting in inter-class sorting results and multiple intra-class sorting results. Combining the inter-class sorting results and multiple intra-class sorting results yields the variable sorting result for all vehicle feature variables. An input with a two-dimensional matrix structure encompasses the vehicle state information reflected by multiple vehicle feature variables at a certain point in time and the changes of these feature variables at different times, thereby increasing the information content of the model samples and improving the accuracy of model training.
[0111] As a preferred embodiment, the relevance of each category is sorted to generate inter-category ranking results, including:
[0112] Based on the vehicle characteristic variables under each category, the sum of the eigenvalues of the covariance matrix for each category is calculated separately;
[0113] The class with the largest sum of eigenvalues of the covariance matrix is determined as the core class;
[0114] Calculate the distance between each category and the core class to obtain the inter-class distance;
[0115] The categories are sorted based on the inter-class distance to generate an inter-class ranking result.
[0116] In this embodiment of the invention, the different categories of vehicle feature variables are sorted. First, the sum of the eigenvalues of the covariance matrix of each category is calculated. The category with the largest sum of the eigenvalues of the covariance matrix is determined as the core class. Then, the distance between other categories and the core class is calculated. The categories are sorted from closest to farthest according to the distance to obtain the inter-class sorting result.
[0117] As a preferred embodiment, the vehicle feature variables under each category are sorted according to their correlation to generate several intra-category ranking results, including:
[0118] Cluster analysis was used to determine the cluster centers for each category.
[0119] For each category, calculate the distance between the vehicle feature variables and the class center for each category to obtain the distance between the variables and the classes.
[0120] Based on the variables and the inter-class distance, the vehicle feature variables under each category are sorted to generate intra-class sorting results.
[0121] In this embodiment of the invention, after sorting each category, multiple vehicle feature variables within each category are sorted. First, cluster analysis is used to calculate the cluster centers for each category. The method for obtaining the cluster centers can be the average method, which involves standardizing each vehicle feature variable within a category and then taking the average value of all vehicle feature variables, using this average value as the cluster center. Alternatively, the nearest distance method can be used, which involves calculating the distance from each vehicle feature variable within a category to other vehicle feature variables within the same category, averaging these distances, and using the variable with the smallest average value as the cluster center. After obtaining the cluster centers for each category, the distance between the vehicle feature variables in each category and the cluster center is calculated. The vehicle feature variables within each category are then sorted from closest to furthest according to the distance, resulting in the intra-category sorting result. The intra-category sorting result can be represented as (X... r1 ,X r2 ,...,X rn ), X r1 X represents the first variable after sorting. r2 X represents the second variable after sorting. rn This represents the nth variable after sorting.
[0122] Step 106: Based on the variable sorting results, convert the model training data into input data with a two-dimensional matrix structure.
[0123] In this embodiment of the invention, after obtaining the inter-class ranking results and the intra-class ranking results, the ranking results of all vehicle feature variables can be obtained, and then the vehicle feature variables under each category of the model training data can be converted into a two-dimensional matrix format. Specifically, the row elements of the two-dimensional matrix are determined according to the variable ranking results, and the column elements of the two-dimensional matrix are determined according to the time data corresponding to each vehicle feature variable. After obtaining the row elements and column elements of the two-dimensional matrix, the input data can be converted into a two-dimensional matrix format, represented as follows:
[0124]
[0125] In the formula, t represents the current time and T represents T time intervals.
[0126] Step 107: Train the convolutional neural network model based on the input data to obtain the first prediction model.
[0127] In this embodiment of the invention, a convolutional neural network (CNN) is used as the model framework, and the input data of the obtained two-dimensional matrix structure is used as the model training data to predict the probability of vehicle sale. The trained model is the first prediction model, used to predict the probability of vehicle sale. Historical vehicle network data samples are obtained from the vehicle network based on a preset time granularity. Compared with obtaining historical vehicle network data corresponding to only one point in time, obtaining historical vehicle network data within a time period can reveal the usage patterns of vehicle samples over time, thereby achieving more accurate predictive analysis. The two-dimensional matrix format input data also contains information from multiple dimensions of historical vehicle network data, improving the richness of the sample data and thus improving the accuracy and precision of model training.
[0128] Step 108: Obtain vehicle network data of the vehicle to be predicted within a preset time period, generate input data of the vehicle to be predicted with a two-dimensional matrix structure based on the variable sorting result, input the input data of the vehicle to be predicted into the first prediction model, and obtain the sales prediction result of the vehicle to be predicted.
[0129] As a preferred embodiment, the process involves acquiring vehicle-to-everything (V2X) data of the vehicle to be predicted within a preset time period, and generating input data for the vehicle to be predicted with a two-dimensional matrix structure based on the variable sorting results, including:
[0130] Acquire vehicle-to-everything (V2X) data of the vehicle to be predicted within a preset time period;
[0131] The vehicle network data is subjected to feature engineering processing to obtain several vehicle feature variables for characterizing usage patterns.
[0132] Based on the variable ranking results, several vehicle feature variables to be predicted, which are used to characterize usage patterns, are ranked to form a vehicle feature ranking dataset.
[0133] The vehicle feature sorting dataset is converted into a data format to generate two-dimensional matrix structured vehicle input data to be predicted.
[0134] In this embodiment of the invention, when predicting the sale of a vehicle to be predicted, the vehicle network data of the vehicle to be predicted within a preset time period is obtained. Compared with only obtaining the vehicle network data of the vehicle to be predicted at the current time, more accurate prediction analysis can be achieved by analyzing the usage pattern information of the vehicle to be predicted over time.
[0135] In this embodiment of the invention, after acquiring the vehicle network data of the vehicle to be predicted, feature engineering is performed on the vehicle network data to obtain information in multiple dimensions. Then, by analyzing the usage pattern information of the vehicle to be predicted over time, more accurate prediction analysis can be achieved.
[0136] In this embodiment of the invention, the variable ranking results obtained during model training are acquired, and the correlation of the vehicle feature variables to be predicted is ranked to obtain a vehicle feature ranking dataset that reflects the correlation. Compared with empirical judgment, real-time predictive analysis based on the correlation between the vehicle feature variables of the vehicle to be predicted can effectively improve the accuracy of sales prediction.
[0137] In this embodiment of the invention, the vehicle feature sorting dataset is converted to a different data format to obtain vehicle input data that better conforms to the input format of the first prediction model. This vehicle input data is then input into the first prediction model, which performs a sales prediction analysis on the vehicle to be predicted, yielding a sales prediction result.
[0138] In a preferred embodiment, after inputting the vehicle input data to the first prediction model and obtaining the sales prediction result of the vehicle, the method further includes:
[0139] The sales prediction results of the vehicles to be predicted are input into a preset probability calibration model so that the probability calibration model calibrates the sales prediction results using Platt scaling or ordinal-preserving regression.
[0140] In this embodiment of the invention, because the first prediction model utilizes SMOTE oversampling, the predicted sale probability obtained by the first prediction model will be higher than the actual probability. Therefore, to improve the accuracy of the sale prediction results, the sale prediction results of the vehicle to be predicted can be input into a preset probability calibration model, and the sale prediction results can be processed using Platt scaling or ordinal-preserving regression, so that the sale prediction probability output by the first prediction model is more consistent with reality.
[0141] In practical applications, current vehicle sales forecasting methods can only be applied to single-purpose trucks, making them difficult to apply to multi-purpose passenger vehicles. The vehicle sales forecasting method provided in this invention not only improves the accuracy of vehicle sales forecasting but also expands the types of vehicles that can be applied. This invention can predict not only truck sales but also the sales of different types of passenger vehicles (such as sedans, SUVs, luxury sedans, and luxury SUVs). Implementing the above embodiments has the following effects:
[0142] This invention provides a method for predicting vehicle sales. When constructing a first prediction model for vehicle sales prediction, several historical vehicle network data samples are extracted from the vehicle network according to a preset time granularity. This not only considers vehicle network data at a single time segment but also fully utilizes the correlation between vehicle network data in both feature variables and time dimensions, improving the richness of the model training data. New samples are generated by oversampling based on the historical vehicle network data samples, and then the model training data is derived by combining the historical and new samples to reduce sample imbalance and improve model accuracy. Feature engineering is performed on the model training data to obtain multiple vehicle feature variables representing usage patterns, thus improving model stability. The vehicle feature variables are classified by calculating the similarity between them. The vehicle feature variables under each category are sorted according to their correlation, resulting in a variable ranking result for all vehicle feature variables. Based on this ranking result, the model training data is converted into input data with a two-dimensional matrix structure, and a convolutional neural network is trained using this input data to obtain the first prediction model. This invention generates model training data based on the correlation between vehicle feature variables, effectively increasing the information content of the model training data and thus improving the accuracy of the first prediction model. After the first prediction model is trained, vehicle network data of the vehicle to be predicted is collected within a preset time period. The input data of the vehicle to be predicted with a two-dimensional matrix structure is generated by combining the variable ranking results. This data is used as the input data of the first prediction model so that the first prediction model can make more accurate sales prediction results for the vehicle to be predicted.
[0143] Example 2
[0144] See Figure 2 This is a schematic diagram of the structure of an embodiment of the vehicle sales prediction device provided by the present invention. The device includes a sample acquisition module, a sample generation module, a feature extraction module, a feature classification module, a feature sorting module, a data conversion module, a model training module, and a prediction module.
[0145] The sample acquisition module is used to extract several historical vehicle network data samples from the vehicle network according to a preset time granularity; wherein, the historical vehicle network data samples include tags; the tags include "for sale" and "not for sale";
[0146] The sample generation module is used to oversample several historical vehicle network data samples to generate several new samples, and to determine the several historical vehicle network data samples and the several new samples as model training data.
[0147] The feature extraction module is used to perform feature engineering processing on the model training data to obtain several vehicle feature variables used to characterize usage patterns; wherein, the vehicle feature variables include vehicle age, driving range, total mileage and battery health.
[0148] The feature classification module is used to calculate the similarity between the vehicle feature variables and classify the vehicle feature variables to obtain the vehicle feature variables under each category.
[0149] The feature sorting module is used to sort the vehicle feature variables under each category according to their correlation, and generate the variable sorting results of all vehicle feature variables.
[0150] The data conversion module is used to convert the model training data into input data with a two-dimensional matrix structure based on the variable sorting results;
[0151] The model training module is used to train the convolutional neural network based on the input data to obtain a first prediction model.
[0152] The prediction module is used to acquire vehicle network data of the vehicle to be predicted within a preset time period, generate vehicle input data with a two-dimensional matrix structure based on the variable sorting results, input the vehicle input data to be predicted into the first prediction model, and obtain the sales prediction result of the vehicle to be predicted.
[0153] As a preferred embodiment, oversampling is performed on several historical vehicle-to-everything (V2X) data samples to generate several new samples, including:
[0154] From a number of historical vehicle-to-everything (V2X) data samples, those labeled "for sale" were selected as the initial samples.
[0155] For each initial sample, select a historical vehicle network data sample from several historical vehicle network data samples whose similarity to the initial sample is greater than a preset threshold, and determine it as an auxiliary sample;
[0156] The initial sample and the auxiliary sample are processed using the following formula to generate new samples:
[0157] X new =X m +ε*(X m -X s )
[0158] In the formula, X new Features for new samples; X m Features of the initial samples; X s ε is an auxiliary sample feature; ε is a preset parameter.
[0159] As a preferred embodiment, the model training data is subjected to feature engineering processing to obtain several vehicle feature variables used to characterize usage patterns, including:
[0160] Features are extracted from the model training data to obtain several feature variables; wherein, the feature variables include average mileage per trip, total mileage per day, parking time, daily mileage and average driving speed;
[0161] In the vehicle network data, the correlation between each of the aforementioned feature variables is obtained;
[0162] Based on the aforementioned correlation, data processing is performed on each of the aforementioned feature variables to form several vehicle feature variables used to characterize usage patterns.
[0163] As a preferred embodiment, the similarity between the vehicle feature variables is calculated, and the vehicle feature variables are classified to obtain the vehicle feature variables under each category, including:
[0164] Based on the values corresponding to each of the vehicle feature variables, the correlation coefficients between each of the vehicle feature variables are calculated respectively.
[0165] Based on the correlation coefficients between the vehicle feature variables, the distances between the vehicle feature variables are calculated.
[0166] Based on the distance between the vehicle feature variables, a clustering method is used to classify the vehicle feature variables and obtain the vehicle feature variables under each category.
[0167] As a preferred embodiment, the vehicle feature variables under each category are sorted according to their correlation to generate a variable ranking result for all vehicle feature variables, including:
[0168] Sort the relevance of each category and generate inter-category ranking results;
[0169] The vehicle feature variables under each category are sorted according to their correlation, generating several intra-class sorting results;
[0170] By combining the inter-class sorting results and several intra-class sorting results, a variable sorting result for all vehicle feature variables is generated.
[0171] As a preferred embodiment, the relevance of each category is sorted to generate inter-category ranking results, including:
[0172] Based on the vehicle characteristic variables under each category, the sum of the eigenvalues of the covariance matrix for each category is calculated separately;
[0173] The class with the largest sum of eigenvalues of the covariance matrix is determined as the core class;
[0174] Calculate the distance between each category and the core class to obtain the inter-class distance;
[0175] The categories are sorted based on the inter-class distance to generate an inter-class ranking result.
[0176] As a preferred embodiment, the vehicle feature variables under each category are sorted according to their correlation to generate several intra-category ranking results, including:
[0177] Cluster analysis was used to determine the cluster centers for each category.
[0178] For each category, calculate the distance between the vehicle feature variables and the class center for each category to obtain the distance between the variables and the classes.
[0179] Based on the variables and the inter-class distance, the vehicle feature variables under each category are sorted to generate intra-class sorting results.
[0180] As a preferred embodiment, the process involves acquiring vehicle-to-everything (V2X) data of the vehicle to be predicted within a preset time period, and generating input data for the vehicle to be predicted with a two-dimensional matrix structure based on the variable sorting results, including:
[0181] Acquire vehicle-to-everything (V2X) data of the vehicle to be predicted within a preset time period;
[0182] The vehicle network data is subjected to feature engineering processing to obtain several vehicle feature variables for characterizing usage patterns.
[0183] Based on the variable ranking results, several vehicle feature variables to be predicted, which are used to characterize usage patterns, are ranked to form a vehicle feature ranking dataset.
[0184] The vehicle feature sorting dataset is converted into a data format to generate two-dimensional matrix structured vehicle input data to be predicted.
[0185] In a preferred embodiment, after inputting the vehicle input data to the first prediction model and obtaining the sales prediction result of the vehicle, the method further includes:
[0186] The sales prediction results of the vehicles to be predicted are input into a preset probability calibration model so that the probability calibration model calibrates the sales prediction results using Platt scaling or ordinal-preserving regression.
[0187] Implementing the above embodiments has the following effects:
[0188] This invention provides a vehicle sales prediction device. When constructing a first prediction model for vehicle sales prediction, it extracts several historical vehicle network data samples from the Internet of Vehicles (IoV) according to a preset time granularity. This not only considers IoV data at a single time point but also fully utilizes the correlation between IoV data in both feature variables and time dimensions, improving the richness of the model training data. New samples are generated by oversampling based on the historical IoV data samples, and then the model training data is derived by combining the historical and new samples to reduce sample imbalance and improve model accuracy. Feature engineering is performed on the model training data to obtain multiple vehicle feature variables representing usage patterns, thus improving model stability. The vehicle feature variables are classified by calculating the similarity between them. The vehicle feature variables under each category are sorted according to their correlation, resulting in a variable ranking result for all vehicle feature variables. Based on this ranking result, the model training data is converted into input data with a two-dimensional matrix structure, and a convolutional neural network is trained using this input data to obtain the first prediction model. This invention generates model training data based on the correlation between vehicle feature variables, effectively increasing the information content of the model training data and thus improving the accuracy of the first prediction model. After the first prediction model is trained, vehicle network data of the vehicle to be predicted is collected within a preset time period. The input data of the vehicle to be predicted with a two-dimensional matrix structure is generated by combining the variable ranking results. This data is used as the input data of the first prediction model so that the first prediction model can make more accurate sales prediction results for the vehicle to be predicted.
[0189] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.
Claims
1. A vehicle sale prediction method characterized by comprising: The application relates to a method for predicting the sale of a vehicle. According to a preset time granularity, a plurality of historical vehicle Internet of Things data samples are extracted in the vehicle Internet of Things; wherein the historical vehicle Internet of Things data samples comprise labels; the labels comprise sale and non-sale; Based on the plurality of historical vehicle Internet of Things data samples, oversampling is performed to generate a plurality of new samples, and the plurality of historical vehicle Internet of Things data samples and the plurality of new samples are determined as model training data; Feature engineering processing is performed on the model training data to obtain a plurality of vehicle feature variables for representing use rules; wherein the vehicle feature variables comprise vehicle age, driving range, total mileage and battery health degree; The similarity between each vehicle feature variable is calculated, and the vehicle feature variables are classified to obtain vehicle feature variables in each category; According to the relevance, the vehicle feature variables in each category are sorted to generate a variable sorting result of all vehicle feature variables; Based on the variable sorting result, the model training data is converted into input data with a two-dimensional matrix structure; Based on the input data, a convolutional neural network is trained to obtain a first prediction model; Vehicle Internet of Things data of a to-be-predicted vehicle within a preset time length is obtained, input data with a two-dimensional matrix structure of the to-be-predicted vehicle is generated based on the variable sorting result, the input data of the to-be-predicted vehicle is input into the first prediction model, and a sale prediction result of the to-be-predicted vehicle is obtained; The method comprises the following steps: In the plurality of historical vehicle Internet of Things data samples, historical vehicle Internet of Things data samples with a label of sale are selected to be determined as initial samples; For each initial sample, historical vehicle Internet of Things data samples with a similarity greater than a preset threshold to the initial sample are selected from the plurality of historical vehicle Internet of Things data samples to be determined as auxiliary samples; The initial samples and the auxiliary samples are processed by using the following formula to generate new samples: In the formula, is a new sample feature; is an initial sample feature; is an auxiliary sample feature; is a preset parameter.
2. The vehicle sale prediction method according to claim 1, characterized by, The method comprises the following steps: Features in the model training data are extracted respectively to obtain a plurality of feature variables; wherein the feature variables comprise average driving mileage each time, total mileage each day, parking time, driving mileage each day and average driving speed; In the vehicle Internet of Things data, the correlation between each feature variable is obtained; Based on the correlation, data processing is performed on each feature variable to form a plurality of vehicle feature variables for representing use rules.
3. The vehicle sale prediction method according to claim 2, characterized by, The method comprises the following steps: Based on the numerical values corresponding to each vehicle feature variable, the correlation coefficients between each vehicle feature variable are calculated respectively; Based on the correlation coefficients between each vehicle feature variable, the distances between each vehicle feature variable are calculated; Based on the distances between each vehicle feature variable, a clustering method is used to classify each vehicle feature variable to obtain vehicle feature variables in each category.
4. The vehicle sale prediction method according to claim 3, characterized by, The vehicle feature variables under each category are sorted according to the correlation, and a variable sorting result of all vehicle feature variables is generated, including: The correlation of each category is sorted, and an inter-class sorting result is generated; The vehicle feature variables under each category are sorted according to the correlation, and a plurality of intra-class sorting results are generated; The inter-class sorting result and the plurality of intra-class sorting results are combined to generate a variable sorting result of all vehicle feature variables.
5. The vehicle sale prediction method according to claim 4, characterized by, The correlation of each category is sorted, and an inter-class sorting result is generated, including: Based on the vehicle feature variables under each category, the sum of the eigenvalues of the covariance matrix of each category is calculated respectively; The category with the largest sum of eigenvalues of the covariance matrix is determined as the core class; The distance between each category and the core class is calculated respectively, and an inter-class distance is obtained; Based on the inter-class distance, each category is sorted to generate an inter-class sorting result.
6. The vehicle sale prediction method according to claim 5, characterized by, The vehicle feature variables under each category are sorted according to the correlation, and a plurality of intra-class sorting results are generated, including: The class center of each category is obtained through clustering analysis; For each category, the distance between the vehicle feature variables under each category and the class center is calculated to obtain the variable and inter-class distance; Based on the variable and inter-class distance, the vehicle feature variables under each category are sorted to generate an intra-class sorting result.
7. The vehicle sale prediction method according to claim 6, characterized by, The Internet of Vehicles data of the vehicle to be predicted within a preset time length is obtained, and the variable sorting result is used to generate vehicle to be predicted input data with a two-dimensional matrix structure, including: Obtain the Internet of Vehicles data of the vehicle to be predicted within a preset time length; Perform feature engineering processing on the Internet of Vehicles data to obtain a plurality of vehicle to be predicted feature variables for representing use rules of the vehicle to be predicted; Based on the variable sorting result, the plurality of vehicle to be predicted feature variables for representing use rules are sorted to form a vehicle feature sorting data set; The vehicle feature sorting data set is converted in data format to generate vehicle to be predicted input data with a two-dimensional matrix structure.
8. The vehicle sale prediction method according to claim 7, characterized by, After inputting the vehicle to be predicted input data into the first prediction model to obtain the sale prediction result of the vehicle to be predicted, the method further includes: Input the sale prediction result of the vehicle to be predicted into a preset probability calibration model, so that the probability calibration model uses Platt scaling method or order preserving regression method to calibrate the sale prediction result.
9. A vehicle sale prediction device characterized by comprising: It includes: A sample acquisition module, a sample generation module, a feature extraction module, a feature classification module, a feature sorting module, a data conversion module, a model training module, and a prediction module; The sample acquisition module is used to extract a plurality of historical Internet of Vehicles data samples from the Internet of Vehicles according to a preset time granularity; wherein the historical Internet of Vehicles data samples include labels; the labels include sales and non-sales; The sample generation module is used to perform oversampling based on the plurality of historical Internet of Vehicles data samples to generate a plurality of new samples, and the plurality of historical Internet of Vehicles data samples and the plurality of new samples are determined as model training data; The feature extraction module is configured to perform feature engineering on the model training data to obtain a plurality of vehicle feature variables for representing usage rules; wherein the vehicle feature variables include vehicle age, driving range, total mileage, and battery health; The feature classification module is configured to calculate the similarity between each of the vehicle feature variables, and classify the vehicle feature variables to obtain vehicle feature variables in each category; The feature sorting module is configured to sort the vehicle feature variables in each category according to the correlation to generate a variable sorting result of all vehicle feature variables; The data conversion module is configured to convert the model training data into input data with a two-dimensional matrix structure based on the variable sorting result; The model training module is configured to train a convolutional neural network based on the input data to obtain a first prediction model; The prediction module is configured to obtain Internet of Vehicles data of a to-be-predicted vehicle within a preset time length, generate to-be-predicted vehicle input data with a two-dimensional matrix structure based on the variable sorting result, input the to-be-predicted vehicle input data into the first prediction model, and obtain a sale prediction result of the to-be-predicted vehicle; The over-sampling based on the plurality of historical Internet of Vehicles data samples to generate a plurality of new samples includes: Selecting historical Internet of Vehicles data samples with a label of sale from the plurality of historical Internet of Vehicles data samples to determine initial samples; For each initial sample, selecting historical Internet of Vehicles data samples with a similarity greater than a preset threshold to the initial sample from the plurality of historical Internet of Vehicles data samples to determine auxiliary samples; Performing sample feature processing on the initial samples and the auxiliary samples using the following formula to generate new samples: In the formula, is a new sample feature; is an initial sample feature; is an auxiliary sample feature; is a preset parameter.
Citation Information
Patent Citations
Method, device, medium and program product for replacing spare parts
CN116957530A