A method for predicting the retirement of an electric vehicle based on usage requirements

Abstract

The application provides a kind of electric vehicle retirement prediction method based on use demand, comprehensively consider the influence of various state of vehicle itself static index, actual operation and user's charging and driving behavior and habit and other factors on different types of vehicle retirement possibility, also specially put forward based on Weibull distribution aging score and multidimensional travel demand quantification means;Method in turn through original data collection and preprocessing, feature parameter extraction, feature screening to establish training set, then based on machine learning algorithm to establish retirement prediction model, the training set is established to train and verify the effect of prediction model.In the final training of model is applied online, can more objectively and accurately predict the vehicle retirement state, is beneficial to the early planning of the later use, retirement and recycling of vehicle and battery, compared with prior art has more extensive practicality.

Description

Technical Field

[0001] This invention belongs to the field of performance and lifespan assessment and prediction technology for electric vehicles and power batteries, and specifically relates to a method for predicting the retirement of electric vehicles based on usage requirements. Background Technology

[0002] For new energy vehicles approaching the end of their lifespan, current practices in this field often use fixed thresholds for battery capacity or internal resistance degradation as the lifespan endpoint to assess and predict their remaining service life. For example, in terms of capacity, a decrease in usable battery capacity to 80% of the initial capacity is often chosen as the lifespan endpoint, while an increase in battery internal resistance to 200% of the initial internal resistance is chosen. However, this method of determining the lifespan endpoint using fixed thresholds is unsuitable for the complex and varied operating conditions of real vehicles due to its singular evaluation dimension. It lacks consideration for the differences in user habits and operating environments, and in actual use, the vehicle's purpose, type, economy, and functionality should not be ignored in influencing the final retirement decision. On the other hand, with the continuous increase in the capacity of automotive power batteries, the continued use of fixed threshold retirement judgment methods is gradually failing to meet the requirements for the sustainable use of electric vehicles. It cannot provide accurate remaining lifespan predictions for each vehicle's specific characteristics, potentially leading to insufficient battery utilization and premature vehicle retirement. Therefore, there is an urgent need in this field to develop a method for predicting the retirement status of electric vehicles that can accurately predict the retirement status of a single vehicle based on real operating data and by integrating multiple factors affecting retirement. Summary of the Invention

[0003] In view of this, and to address the technical problems existing in this field, the present invention provides a method for predicting the retirement of electric vehicles based on usage needs, specifically including the following steps:

[0004] Step 1: Collect static data related to the vehicle itself and battery of retired electric vehicles of different uses and types, as well as actual vehicle operation data within a certain period before retirement; perform preprocessing on the collected data, including setting outliers to null, deleting or averaging duplicate values, and filling missing values, and establish corresponding static datasets and actual vehicle operation datasets for retired vehicles.

[0005] Step 2: Process the static data to obtain the service life and rated battery energy of each vehicle, and fit the corresponding Weibull distribution probability density function of the vehicle service life for different vehicle types; use the probability density function to calculate the aging score of different types of vehicles; define the service life and aging score as vehicle dimension feature parameters.

[0006] Step 3: Process the real vehicle operation data to obtain segmented charging data segments and driving data segments; extract charging state characteristic parameters from the charging data segments, including battery capacity, single cell voltage, charging time, cumulative driving mileage, average temperature, and average charging current; extract driving state characteristic parameters from the driving data segments, including full charge range, cumulative driving mileage, average current during driving segment, average speed, and energy consumption rate; define the charging state characteristic parameters and driving state characteristic parameters together as state dimension characteristic parameters;

[0007] Step 4: Set corresponding quantile ratios for single-day trips, single trips, and trips between charging stops in driving behavior, and count the corresponding trip mileage that meets the quantile ratios under different conditions; extract charging habit features including daily charging amount and safe SOC retained after each trip for charging behavior, and extract vehicle shutdown features including the number of times the vehicle is stopped, the total shutdown time, and the time interval between two shutdowns within a specific period; define the habit features and vehicle shutdown features as usage dimension feature parameters;

[0008] Step 5: Match the feature parameters of vehicle dimension, status dimension, and usage dimension according to the cumulative mileage, and add corresponding retired and non-retired status labels; filter the feature parameters of each dimension based on the status labels and through correlation analysis; construct a training set using the retained feature parameters;

[0009] Step 6: Establish a vehicle retirement prediction model based on machine learning algorithms, train it using the constructed training set, and verify the training effect.

[0010] Step 7: Apply the trained model online to predict the retirement status of electric vehicles, so that it can predict and output the corresponding vehicle retirement status based on the real-time data corresponding to the feature parameters of each dimension.

[0011] Furthermore, the static data collected in step one includes at least the vehicle's scrap date, production date, vehicle type, battery type, and battery rated parameters. The vehicle's purpose includes passenger cars, special transport vehicles for freight or special transport, and commercial buses. The actual vehicle operation data includes at least vehicle speed, mileage, total voltage, and total current.

[0012] Data preprocessing includes setting abnormal data items that are outside the reasonable range to null values; deleting all data such as voltage, current and temperature that are repeated over time when the differences are too large, and taking the average value when the differences are small; and filling missing values ​​with linear interpolation or nearest neighbor interpolation according to the data type.

[0013] Furthermore, in step two, the service life of each vehicle is calculated using the scrap date and production date data; the rated energy of the battery is obtained by multiplying the rated voltage and rated capacity in the battery's rated parameters; and the following forms of Weibull distribution probability density functions are obtained by distinguishing between passenger cars, special vehicles, and buses:

[0014]

[0015] Where x is the vehicle's usage time, k is the scaling parameter, and λ is the shape parameter;

[0016] Data with service life of less than 1 year and more than 10 years were removed to reduce the impact of outliers, and least squares fitting was performed on different types of vehicles to determine the corresponding proportional and shape parameters.

[0017] This yields the aging score (SR) for different types of vehicles:

[0018]

[0019] Furthermore, in step three, the original data is segmented into charging data segments and driving data segments based on data such as vehicle status, charging status, current, and vehicle speed. For cases where data is still missing after segmentation, a new data frame is formed by linearly interpolating multiple frames of data. Data segments with a total duration greater than a specific value or a SOC change less than a specific value are removed.

[0020] Furthermore, the specific process of extracting battery capacity from charging data segments in step three includes:

[0021] First, remove the SOC data at the left and right boundaries of the charging data segment, and then use a portion of the SOC range to calculate the battery capacity using the following formula:

[0022]

[0023] Where I is the charging current, t0 is the initial time of the charging segment, t1 is the end time of the charging segment, and SOC0 and SOC1 are the SOC values ​​at the initial and end times of the charging segment, respectively.

[0024] The process of extracting the full-charge driving range from driving data segments includes:

[0025] First, define the distance a fully charged vehicle can continuously travel under the current driving conditions as the equivalent fully charged driving range R. d The cumulative mileage and SOC from the driving data segment are used to calculate:

[0026]

[0027] Where R0 and R1 are the cumulative mileage at the start and end of the driving segment, respectively, and SOC0 and SOC1 are the SOC values ​​at the start and end of the driving segment, respectively.

[0028] The calculation results are selected for driving segments with a SOC change greater than 20.

[0029] Furthermore, in step four, quantile ratios of 98%, 95%, and 85% are set for single-day trips, single trips, and trips to charging stations, respectively, and the corresponding trip mileage that meets the quantile ratios in each case is calculated.

[0030] The charging volume is specifically based on the daily charging volume statistics for every two-month cycle, and the daily charging volume corresponding to the 80th percentile is selected as the corresponding usage dimension feature.

[0031] The vehicle decommissioning feature is specifically identified and extracted based on the cumulative mileage and time difference of adjacent driving segments. When the cumulative mileage difference between adjacent driving segments is small but the time difference is greater than 30 days, it is judged that the vehicle has been decommissioned. The number of decommissioning times for each vehicle is counted, and the total decommissioning time and the time interval between two decommissioning times are used as the corresponding usage dimension features.

[0032] Furthermore, in step five, the operational data of each retired vehicle with an expected cumulative mileage greater than 95% of the final mileage and the last two months of use are marked with a retirement status label 1; other data are marked with a non-retirement status label 0.

[0033] In the correlation analysis of feature parameters, feature selection is specifically accomplished by calculating the following Pearson correlation coefficient:

[0034]

[0035] Where PCC is the feature variable x i and retired status y i The Pearson correlation coefficient between them and They are x i and y i The average value.

[0036] Furthermore, in step six, any one of the following machine learning algorithms is used to build and train a vehicle retirement prediction model: Random Forest, K-Nearest Neighbors, LightGBM, XGBoost, CatBoost, or TabNet; or two or more models are trained separately, while hyperparameter optimization is performed using grid search, and the optimal model is determined through cross-validation. When validating the training effect, precision, recall, F1 score, and area under the curve (AUC) are used as evaluation metrics for the model.

[0037] The electric vehicle retirement prediction method based on usage demand provided by this invention comprehensively considers the impact of factors such as the vehicle's static indicators, various states during actual operation, and users' charging and driving behaviors and habits on the potential retirement of different types of vehicles. It also specifically proposes an aging score based on the Weibull distribution and multi-dimensional quantification of travel demand. The method sequentially establishes a training set through raw data collection and preprocessing, feature parameter extraction, and feature selection. Then, a retirement prediction model is built based on machine learning algorithms. The established training set is used to train the prediction model and verify its effectiveness. After the model is finally trained, it is applied online, enabling more objective and accurate prediction of vehicle retirement status. This facilitates advance planning for the later use, retirement, and recycling of vehicles and batteries, and has broader applicability compared to existing technologies. Attached Figure Description

[0038] Figure 1 This is a general framework diagram of the method provided by the present invention;

[0039] Figure 2 This is an optional flowchart for data segmentation in the method provided by the present invention. Detailed Implementation

[0040] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The electric vehicle retirement prediction method based on usage demand provided by this invention, such as... Figure 1 As shown, the specific steps include:

[0042] Step 1: Collect static data related to the vehicle itself and battery of retired electric vehicles of different uses and types, as well as actual vehicle operation data within a certain period before retirement; perform preprocessing on the collected data, including setting outliers to null, deleting or averaging duplicate values, and filling missing values, and establish corresponding static datasets and actual vehicle operation datasets for retired vehicles.

[0043] Step 2: Process the static data to obtain the service life and battery rated energy of each vehicle, and fit the corresponding Weibull distribution probability density function of the vehicle service life for different vehicle types; use the probability density function to calculate the aging score of different types of vehicles; define the obtained service life and aging score as vehicle dimension feature parameters.

[0044] Step 3: Process the real vehicle operation data to obtain segmented charging data segments and driving data segments; extract charging state characteristic parameters from the charging data segments, including battery capacity, single cell voltage, charging time, cumulative driving mileage, average temperature, and average charging current; extract driving state characteristic parameters from the driving data segments, including full charge range, cumulative driving mileage, average current during driving segment, average speed, and energy consumption rate; define the charging state characteristic parameters and driving state characteristic parameters together as state dimension characteristic parameters;

[0045] Step 4: Set corresponding quantile ratios for single-day trips, single trips, and trips between charging stops in driving behavior, and count the corresponding trip mileage that meets the quantile ratios under different conditions; extract charging habit features including daily charging amount and safe SOC retained after each trip for charging behavior, and extract vehicle shutdown features including the number of times the vehicle is stopped, total shutdown time, and the time interval between two shutdowns within a specific period; define the charging habit features and vehicle shutdown features as usage dimension feature parameters;

[0046] Step 5: Match the feature parameters of vehicle dimension, status dimension, and usage dimension according to the cumulative mileage, and add corresponding retired and non-retired status labels; filter the feature parameters of each dimension based on the status labels and through correlation analysis; construct a training set using the retained feature parameters;

[0047] Step 6: Establish a vehicle retirement prediction model based on machine learning algorithms, train it using the constructed training set, and verify the training effect.

[0048] Step 7: Apply the trained model online to predict the retirement status of electric vehicles, so that it can predict and output the corresponding vehicle retirement status based on the feature parameters collected in real time.

[0049] In a preferred embodiment of the present invention, a static dataset of 23,741 retired vehicles and a real-world operational dataset of 720 retired electric vehicles from the three years prior to their retirement are collected. The collected static data includes at least 11 data items such as vehicle scrapping date, production date, vehicle type, battery type, and battery rated parameters. The vehicle uses include passenger cars, freight or special transport vehicles, and buses. The real-world operational data includes at least 20 data items such as vehicle speed, mileage, total voltage, and total current.

[0050] Due to sensor acquisition errors and network transmission errors, the acquired raw operational data contains outliers, duplicates, and missing values. To avoid interference from outliers in subsequent research, these outliers must be identified and processed. For outliers, data items exceeding the nominal range of each data item are set to null values. For time-repeated data, the differences in voltage, current, and temperature among the repeated data are assessed. If the differences are too large, all data are deleted; if the differences are small, the average value is taken. Finally, missing values ​​are filled using linear interpolation or nearest-neighbor interpolation, depending on the data type. For example, linear interpolation is used for continuous values ​​such as voltage and mileage, while nearest-neighbor interpolation is used for discrete values ​​such as vehicle status.

[0051] Step two specifically uses the scrap date and production date data to calculate the service life of each vehicle; the rated energy of the battery is obtained by multiplying the rated voltage and rated capacity in the battery's rated parameters; the following forms of Weibull distribution probability density functions are obtained by distinguishing between passenger cars, special vehicles, and buses:

[0052]

[0053] Where x is the vehicle's usage time, k is the scaling parameter, and λ is the shape parameter;

[0054] Data with service lives less than 1 year and greater than 10 years were removed to reduce the impact of outliers. Least squares fitting was performed on different vehicle types to determine the corresponding proportional and shape parameters. Based on the parameter estimation results, the probability distribution curves of vehicle service life applicable to different vehicle types can be reconstructed. By integrating the area under the curve to the right of the current usage time, the vehicle's aging score can be obtained, representing the percentage of a certain type of vehicle that has not been retired by a certain time point, used to assess vehicle reliability at different time points.

[0055] The formula for calculating the aging fraction (SR) is as follows:

[0056]

[0057] In actual assessment, the vehicle type and cumulative usage time can be obtained first, and then substituted into the aging score calculation formula of the corresponding vehicle type to obtain the aging score result.

[0058] Step three involves segmenting the raw data into charging data segments and driving data segments based on data such as vehicle status, charging status, current, and vehicle speed. The process is as follows: Figure 2 As shown; to address the issue of missing data even after data segmentation, the time interval between two frames is generally 10 seconds. For data with time intervals between 20 and 60 seconds, linear interpolation is used for data interpolation, resulting in 10-second frames for all interpolated data. Furthermore, the length of continuous data affects the accuracy of battery capacity and driving range calculations; therefore, segments with a total time length greater than 100 seconds are selected and retained, while segments with a SOC change of less than 5 are removed.

[0059] Because the charging process employs a multi-stage constant current charging strategy, the charging data exhibits smaller current fluctuations and more stable conditions compared to driving data. Therefore, using charging segment data to calculate battery capacity in step three is more reliable. For each segmented charging data point, considering that the SOC data corresponding to the segment boundaries may be incomplete and affect the battery capacity calculation results, the SOC data at the left and right boundaries is removed, and the middle portion of the data is used to calculate the battery charging capacity. For example, if the SOC range of the segmented charging points is 29-71, then the data from 30-70 is selected to calculate the battery capacity using the following formula:

[0060]

[0061] Where I is the charging current, t0 is the initial time of the charging segment, t1 is the end time of the charging segment, and SOC0 and SOC1 are the SOC values ​​at the initial and end times of the charging segment, respectively.

[0062] The process of extracting the full-charge driving range from driving data segments includes:

[0063] First, define the distance a fully charged vehicle can continuously travel under the current driving conditions as the equivalent fully charged driving range R. d The cumulative mileage and SOC from the driving data segment are used to calculate:

[0064]

[0065] Where R0 and R1 are the cumulative mileage at the start and end of the driving segment, respectively, and SOC0 and SOC1 are the SOC values ​​at the start and end of the driving segment, respectively.

[0066] The above formula calculates the driving range when the battery is fully charged by using a specific driving segment as a case study. Therefore, the length of the driving segment has a significant impact on the calculated driving range. Driving segments with larger changes in State of Charge (SOC) encompass more driving information, resulting in more accurate calculated equivalent driving ranges. Therefore, driving segments with SOC changes greater than 20% are selected and their calculation results are retained.

[0067] Step four considers extracting the usage dimension characteristics of vehicles, starting from the actual operation data of retired vehicles. It analyzes the usage habits of different users to understand the time-varying characteristics of user travel needs and the differences in travel needs for vehicles with different uses. Finally, it quantifies and extracts travel needs. Vehicle travel needs mainly reflect the mileage requirements of vehicle owners, which are usually related to the real-world environment and work needs. For example, the travel needs of private passenger cars are related to the distance from home to work, and the travel needs of logistics vehicles are related to the distance between two warehouses. This illustrates that the differences in usage habits and needs of different vehicles should be fully considered in order to conduct specific analysis of vehicle retirement. In the actual analysis process, the quantification of mileage needs is divided into three dimensions: single-day trips, single trips, and trips between two charging stations. The difference lies in that single trips and trips between two charging stations do not involve recharging, trips between two charging stations may include multiple single trips, while single-day trips generally involve recharging. In terms of mileage values, single-day trips > trips between two charging stations > single trips. Considering that travel needs are affected by seasons and temperature and have time-varying characteristics, mileage data for each dimension is statistically analyzed every two months. Then, travel demand was quantified. To remove the influence of certain occasional extreme trips, quantile ratios of 98%, 95%, and 85% were set for single-day trips, single trips, and trips between charging stations, respectively. The corresponding travel mileage that met the quantile ratios under each condition was then calculated. Simultaneously, the travel demand for single-day, single-trip, and trips between charging stations under the three satisfaction ratios was extracted without dividing the time period, broadening the perspective on the data. Setting different satisfaction ratios allows for subsequent feature selection to ultimately determine the optimal indicator for predicting retirement status. This enables a more comprehensive examination of data characteristics, helps uncover potentially important information, and is convenient for application to different types of data.

[0068] The usage dimensions include not only travel needs but also user charging habits and vehicle downtime characteristics, thus comprehensively representing vehicle usage. Charging habits include charging amount and safe SOC. Charging amount is based on charging data, with daily charging amount calculated every two months, and the daily charging amount corresponding to the 80th percentile is selected as the feature. Safe SOC refers to the amount of power retained at the end of each trip. Too low a charge will cause voltage drop, making it difficult to control the performance of the lithium battery and exacerbating range anxiety. Therefore, the usable driving range is generally the driving range corresponding to a full charge minus the safe SOC. This can be compared with the needs of a single trip, the travel needs between two charging sessions, and for daily needs, it is compared with the driving range corresponding to a full charge minus the safe SOC plus the daily charging amount.

[0069] Vehicle downtime characteristics are identified by analyzing the mileage and time differences between different driving segments. When the cumulative mileage difference between adjacent driving segments is small but the time difference is greater than 30 days, it is judged that the vehicle has been downtime. The number of downtimes for each vehicle is counted, and the total downtime and the time interval between two downtimes are selected features.

[0070] After extracting the feature parameters across the three dimensions and proceeding to step five, nearest neighbor matching is performed based on the cumulative mileage to synthesize feature sets corresponding to different cumulative mileage data. Then, the vehicle's retirement status is calibrated. The retirement of electric vehicles not due to malfunctions has a certain time lag; there is a time difference between the owner's dissatisfaction with the vehicle's performance and condition and its eventual discontinuation. Therefore, the retirement status can be calibrated using the last small portion of the mileage and time dimensions. For example, the operating data corresponding to vehicles whose expected cumulative mileage is greater than 95% of the final mileage and whose usage time is the last two months can be labeled with a retirement status tag of 1; other data can be labeled with a non-retirement status tag of 0.

[0071] In the correlation analysis of feature parameters, feature selection is specifically accomplished by calculating the following Pearson correlation coefficient:

[0072]

[0073] Where PCC is the feature variable x i and retired status y i The Pearson correlation coefficient between them and They are x i and y i The average value was calculated. Based on the correlation coefficient, a feature set that was highly correlated with the decommissioning status label and had low autocorrelation between different features was retained. Through feature selection, a total of 17 features across three dimensions were retained, as shown in Table 1.

[0074] Table 1. The 17 features retained after screening and their descriptions.

[0075]

[0076]

[0077] Step six can specifically employ any one of the following machine learning algorithms: Random Forest, K-Nearest Neighbors, LightGBM, XGBoost, CatBoost, or TabNet, to build and train a vehicle retirement prediction model. Alternatively, two or more models can be trained separately, while simultaneously using grid search for hyperparameter optimization and cross-validation to determine the optimal model. To verify the model's accuracy and generalization ability, 10% of the vehicles from each type can be randomly selected as test data, not input into the model training, and then the remaining data can be used to train the model using a 10-fold cross-validation method.

[0078] When validating the training effect, the following formulas are used to calculate precision, recall, and F1 score respectively:

[0079]

[0080] Among them, TP indicates that the predicted class is positive and the actual class is also positive, FP indicates that the predicted class is positive but the actual class is negative, TN indicates that the predicted class is negative and the actual class is also negative, and FN indicates that the predicted class is negative but the actual class is positive.

[0081] At the same time, the area under the curve (AUC) is also used as an evaluation metric for the model.

[0082] In this embodiment, TabNet was ultimately confirmed as the optimal machine learning model, with an F1 score of 0.836 and an AUC of 0.977 on the retained test set, demonstrating that its model prediction accuracy and generalization both reached a high level.

[0083] It should be understood that the sequence number of each step in the embodiments of the present invention does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0084] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the retirement of an electric vehicle based on usage needs, characterized in that: Specifically comprising the following steps: Step one, collect the static data of the vehicle itself and the battery related data, as well as the real vehicle running data within a certain period before retirement of different purposes and types of retired electric vehicles; perform preprocessing on the collected data, including setting abnormal values to null, deleting or averaging repeated values, and filling missing values, and establish the corresponding retired vehicle static data set and real vehicle running data set; Step two, by processing the static data, the service life of each vehicle is calculated by using the scrap date and production date data, the battery rated energy is obtained by multiplying the rated voltage and rated capacity in the battery rated parameters, and the corresponding vehicle service life Weibull distribution probability density function is fitted according to different vehicle types: Wherein, x is the service life of the vehicle, k is the proportional parameter, and λ is the shape parameter; Remove the data with service life less than 1 year and more than 10 years to reduce the influence of abnormal values, and perform least square fitting on different types of vehicles to determine the corresponding proportional parameter and shape parameter; The aging score SR of different types of vehicles is calculated using the probability density function: The service life and aging score SR are defined as vehicle dimensional characteristic parameters; Step three, process the real vehicle running data to obtain the segmented charging data segment and driving data segment; extract the charging state characteristic parameters including battery capacity, single cell voltage, charging time, cumulative mileage, average temperature and average charging current from the charging data segment; extract the driving state characteristic parameters including full charge cruising range, cumulative mileage, driving segment average current, average speed and energy consumption rate from the driving data segment; The charging state characteristic parameters and driving state characteristic parameters are jointly defined as state dimensional characteristic parameters; Step four, set the corresponding quantile proportion for the single day trip, single trip and twice charging in the driving behavior, and count the corresponding trip mileage that meets the quantile proportion under different conditions; For charging behavior, extract the charging habit features including daily charging capacity and safe SOC reserved after each trip, and extract the vehicle downtime features including vehicle downtime times, total downtime duration and time interval between two downtimes within a certain period; The habit features and vehicle downtime features are defined as usage dimensional characteristic parameters; Step five, match the vehicle dimensional, state dimensional and usage dimensional characteristic parameters according to the cumulative mileage, and add the corresponding retired and unretired state labels; Based on the state label and through correlation analysis, the dimensional characteristic parameters are screened; Use the retained characteristic parameters to construct the training set; Step six, establish a vehicle retirement prediction model based on machine learning algorithm, train the constructed training set, and verify the training effect; Step seven, apply the trained model to electric vehicle retirement prediction online, so that it can predict the corresponding vehicle retirement state according to the real-time data corresponding to each dimensional characteristic parameter.

2. The method of claim 1, wherein: The static data collected in step one includes at least vehicle retirement date, production date, vehicle type, battery type, battery rated parameters, and vehicle use including passenger private car, special transport vehicle and commercial passenger car; and the real vehicle running data includes at least vehicle speed, mileage, total voltage, total current parameters; The data preprocessing includes setting abnormal data items beyond a reasonable range as null values; deleting all voltage, current and temperature data with large differences when repeatedly in time, or taking average values when the differences are small; and filling missing values by linear interpolation or nearest neighbor interpolation according to the data type.

3. The method of claim 1, wherein: In step three, the original data is divided into charging data segments and driving data segments based on vehicle state, charging state, current and vehicle speed data; for the case that there is still data missing after division, a new data frame is formed by linearly inserting multiple frames of data; and data segments with a total duration greater than a certain value or a SOC change less than a certain value are removed.

4. The method of claim 1, wherein: The specific process of extracting battery capacity from the charging data segment in step three includes: First, the SOC data of the left and right boundaries of the charging data segment is removed, and then the battery capacity is calculated in a partial SOC range by the following formula: Where I is the charging current, t0 is the initial time of the charging segment, t1 is the end time of the charging segment, SOC0 and SOC1 are the SOC values at the initial and end times of the charging segment, respectively; The process of extracting full-charge driving range from the driving data segment includes: First, define the distance that a fully charged vehicle can continuously travel under the current driving section's working condition as the equivalent full charge range R d , calculate using the cumulative driving distance in the driving data segment and the SOC: Where R0 and R1 are the cumulative driving ranges at the beginning and end of the driving segment, respectively, and SOC0 and SOC1 are the SOC values at the beginning and end of the driving segment, respectively; The calculation results corresponding to the driving segments with a SOC change greater than 20 are selected.

5. The method of claim 1, wherein: In step four, 98%, 95% and 85% quantile ratios are set for single-day travel, single-travel and travel between two charges, respectively, and the corresponding travel ranges that meet the quantile ratios are counted under each condition; The charging amount is specifically based on the daily charging amount counted every two months, and the daily charging amount corresponding to the 80% quantile is selected as the corresponding use dimension feature; The vehicle downtime feature is specifically identified and extracted based on the cumulative mileage and time difference of adjacent driving segments. When the cumulative mileage difference of adjacent driving segments is small but the time difference is greater than 30 days, it is judged that the vehicle downtime occurs. The downtime times, total downtime duration and time interval between two downtimes of each vehicle are counted as the corresponding use dimension features.

6. The method of claim 1, wherein: In step five, the running data corresponding to the expected cumulative driving range of each retired vehicle greater than 95% of the final range and the use time of the last two months is added with a retired state label 1; and other data is added with a non-retired state label 0. In the correlation analysis of feature parameters, the feature selection is completed by calculating the following Pearson correlation coefficients: where PCC is the Pearson correlation coefficient between the feature variable x i and the retirement status y i , and are the mean values of x i and y i , respectively.

7. The method of claim 1, wherein: In step six, any one of the following machine learning algorithms is specifically used: random forest, K-nearest neighbor, LightGBM, XGBoost, CatBoost, and TabNet to establish a vehicle retirement prediction model and train it; or two or more models are used for training, and grid search is used for hyperparameter optimization, and the optimal model is determined through cross-validation; and when verifying the training effect, the precision, recall, F1 value and area under the curve AUC are specifically used as the evaluation indexes of the model.

Images