New energy vehicle value retention rate evaluation method and system based on model dimension

Abstract

The invention relates to the technical field of data processing, and discloses a model dimension-based new energy vehicle value retention rate evaluation method and system, and the method comprises the steps: constructing three-table data, and carrying out the preprocessing; and fusing the static features and the time sequence features according to the type identifier to generate a feature sequence. And inputting the static features into a static feature sub-network to obtain a static representation, inputting the time sequence features into a time sequence feature sub-network to obtain a hidden representation of each time step, and obtaining a time sequence representation based on attention weighting. And performing drift detection on the input distribution and the model residual distribution. And splicing the static representation and the time sequence representation, inputting the spliced representation into a quantile prediction sub-network, outputting two value retention rate prediction results of different quantiles, and obtaining a value retention rate quantile set. And obtaining an insurance amount evaluation suggestion value based on the value retention rate quantile set and the new vehicle purchase price. According to the method, fine-grained description of different type feature differences is ensured, and the accuracy and reliability of a prediction result are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method and system for evaluating the residual value of new energy vehicles based on model dimensions. Background Technology

[0002] The rapid development of the new energy vehicle industry has driven the close integration of vehicle valuation and financial insurance. A vehicle's residual value over its lifespan not only influences consumers' purchasing decisions but also directly affects used car transaction prices and insurance companies' risk pricing. Traditional vehicle valuation methods are mostly based on depreciation curves or empirical formulas at the model level, making it difficult to reflect the differences in new energy vehicles in terms of model configuration, battery performance, policy subsidies, and usage behavior. With the development of financial technology and artificial intelligence, intelligent residual value prediction methods based on large-scale data modeling are gradually becoming an important technological foundation for insurance pricing and risk control.

[0003] While existing technologies have incorporated some machine learning methods into the prediction of residual value of new energy vehicles, significant shortcomings remain: most solutions only establish depreciation curves or prediction models at the vehicle model level, failing to fully reflect the value differences brought about by different configurations, battery capacities, and power types at a more granular model level, resulting in insufficient accuracy of prediction results; factors such as market prices, accident records, and mileage growth inherently possess strong time-series characteristics, while existing methods mostly rely on static regression or simple weighting to process time-series data, lacking the ability to deeply model complex time dependencies; the new energy vehicle market environment changes frequently, and existing methods lack effective input distribution and model residual drift detection and update mechanisms, resulting in insufficient stability and generalization ability in long-term applications.

[0004] Therefore, it is necessary to design a method and system for evaluating the residual value of new energy vehicles based on model dimensions to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a method and system for evaluating the residual value of new energy vehicles based on model, aiming to solve the problems existing in the current technology.

[0006] In one aspect, this invention proposes a method for evaluating the residual value of new energy vehicles based on model, including: Three tables of data are constructed and preprocessed. The three tables include a vehicle sample table, a market panel table, and usage event stream data. The vehicle sample table uses model identifier as the primary key and records static features related to the model. The market panel table records market time-series features related to residual value using a time index. The usage event stream data records maintenance, accidents, mileage, and software updates using timestamps. Based on the model identifier, the static features in the vehicle sample table are fused with the temporal features in the market panel table and the usage event stream to generate a feature sequence by monthly time steps; The static features in the feature sequence are input into the static feature sub-network to obtain the static representation, and the temporal features in the feature sequence are input into the temporal feature sub-network to obtain the hidden representation at each time step. The temporal representation is then obtained based on attention weighting. Drift detection is performed on the input distribution and model residual distribution entering the temporal feature sub-network. The drift detection includes calculating the population stability index and / or the Kolmogorov-Smirnov test statistic and comparing them with a preset threshold. When any comparison result meets the drift condition, incremental training of the model is triggered, and the online model is updated after training is completed. The static representation and the time-series representation are concatenated and input into the quantile prediction subnetwork. Based on the quantile loss function, the value preservation rate prediction results of two different quantiles are output to obtain the value preservation rate quantile set. Based on the set of residual value quantiles and the new vehicle purchase price, a suggested insurance amount assessment value is obtained.

[0007] Furthermore, when constructing the three tables of data and preprocessing the three tables of data, the process includes: The vehicle sample table is generated using the model identifier as the unique primary key, and static features are written according to the field normalization. The static features include production date, power type, battery pack capacity, drive type, configuration level, and manufacturer's suggested retail price. Information containing individual identifiers is hash-de-identified and satisfies the K-anonymity constraint, where K≥20. The market panel table is generated using a time index, and market time-series characteristics are summarized by week or month. These market time-series characteristics include median transaction price, transaction volume, regional factors, and policy subsidy dummy variables. Currency and tax-inclusive caliber are unified for heterogeneous price data, and Windsor truncation is performed. The usage event stream data is generated according to the timestamp, and maintenance events, accident events, mileage events and software update events are recorded one by one. Abnormal jumps are detected in mileage events and discontinuities are removed. The preprocessing includes using multiple imputation or business default values ​​to fill in missing values; using a dual criterion of IQR rules and business thresholds to remove anomalies; aligning the market panel table with the event stream data at the end of the month and aggregating them with a fixed-length sliding window; and using target encoding and fitting within the fold for high cardinality segments.

[0008] Furthermore, when fusing the static features in the vehicle sample table with the temporal features in the market panel table and the usage event stream based on the model identifier, and generating a feature sequence by monthly time step, the process includes: Using the model identifier and the end-of-month time index as a joint key, the vehicle sample table is key-connected with the market panel table and the usage event stream data, and the static features are copied to each time step on a monthly basis; The market panel table is aggregated monthly to obtain a subset of market time-series features, and the exponential weighting of the market time-series features is calculated. The event stream data is mapped to a monthly event statistics vector, and non-negative truncation and anomaly removal are performed on the mileage increment; Without using data after any time step, a fixed-length monthly time step feature sequence is generated, the excess part is pruned, the missing part is padded forward and the missing mask is output at the same time. At each time step, the static features, exponentially weighted market time-series features, and monthly event statistical vectors are concatenated in a fixed field order to form the feature sequence for that time step.

[0009] Furthermore, obtaining the static representation, hidden representation, and temporal representation includes: The static feature subnetwork is a multilayer perceptron, which receives the static features and outputs the static representation by passing them sequentially through a fully connected layer and a nonlinear activation layer. The temporal feature subnetwork is a long short-term memory network. It receives the monthly time step feature sequence, and masks invalid time steps according to the missing mask, and outputs the hidden representation of each time step. The attention layer uses the static representation as the query vector, the hidden representation at each time step as the key and value, calculates the attention weight, and summarizes the hidden representation at each time step in a weighted manner to generate the temporal representation. Before concatenating the static representation and the temporal representation, normalization and discard regularization are performed on the static representation and the temporal representation, respectively.

[0010] Furthermore, drift detection is performed on the input distribution and model residual distribution entering the temporal feature sub-network, including: Input distribution samples and model residual samples are extracted separately using a fixed-length sliding window. The population stability index is calculated based on the pre-registered feature list for the input distribution, and the distribution difference statistic is calculated for the model residuals in chronological order. The population stability index and distribution difference statistics are compared with preset thresholds, and drift events are generated by combining lag strategies and cooling strategies. The drift events include drift type, list of affected features, amplitude level and set of affected models.

[0011] Furthermore, when the drift event triggers incremental training, it includes: Incremental training is performed using data from a fixed-length sliding window, with higher weights assigned to recent samples, while recalibrating using the calibration parameters from the previous version as initial values. After training, the model is tested offline and then online for shadow verification. Once the thresholds for loss and quantile coverage are met, the new model is deployed online in a canary release manner using the feature service, while retaining version snapshots and automatic rollback strategies.

[0012] Furthermore, the static representation and the time-series representation are concatenated and then input into the quantile prediction subnetwork, which includes: Based on the quantile loss function, the preservation rate prediction results of two different quantiles are output simultaneously. The different quantiles include at least the low quantile preservation rate used for risk constraints and the median preservation rate used for benchmark estimation. During training, the quantile set is fixed and the gradients of each quantile sample are balanced; during inference, the feature order and normalized statistics are kept consistent with those in the training phase to generate the set of quantiles for the value retention rate.

[0013] Furthermore, after obtaining the set of quantiles for the rate of return on value, it also includes: Perform interval calibration on the set of retention rate quantiles to obtain the calibrated set of retention rate quantiles; The interval calibration includes: calculating the coverage deviation and interval width of each quantile on the reserved dataset; when the coverage deviation exceeds a preset threshold, temperature scaling and / or equal-frequency bin monotonic calibration are used to correct the quantile output. Quantile order constraints are applied during calibration, and a snapshot of the calibration parameters is retained. After calibration, the set of calibrated value retention rate quantiles is obtained.

[0014] Furthermore, when obtaining the recommended insurance amount assessment value based on the set of residual value quantiles and the new vehicle purchase price, it includes: The lower quantile of the value retention rate is selected from the set of value retention rate quantiles as the basis for risk constraints. The low quantile preservation rate is converted into a conservative preservation rate using a monotonic mapping function. The parameters of the monotonic mapping function are automatically learned from historical backtesting and the quantile order is kept from overlapping. The candidate insurance amount is obtained by multiplying the conservative residual value rate by the new vehicle purchase price, and then corrected by applying upper and lower limit constraints, minimum residual value constraints and depreciation curve boundary constraints in sequence. When the set of retention rate quantiles is incomplete, a downgrade path based on the median retention rate is triggered and a conservative coefficient is added for compensation. Output the recommended insured amount and its confidence level.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a three-table data system of vehicle sample table, market panel table, and usage event stream data at the model dimension, and through multi-level preprocessing and alignment, static features and time-series features are fused according to monthly time steps to form a complete feature sequence, which is then input into the static feature sub-network and the time-series feature sub-network respectively. Combined with the attention mechanism, static representation and time-series representation are extracted, ensuring fine-grained characterization of the differences in features of different models and strengthening the ability to model dynamic market factors and usage behavior in a time series. The introduction of a dual drift detection mechanism of input distribution and model residuals can automatically trigger incremental training and model updates when data or model shifts, thereby improving the stability and adaptability of prediction. At the prediction level, the residual value results of multiple quantiles are output through the quantile prediction sub-network, and the results are aggregated to form a set, providing both median estimation and low quantile prediction with risk constraints, solving the problem that traditional methods cannot quantify uncertainty. Based on the quantile set and the new vehicle purchase price, the suggested value of insurance amount assessment is derived, realizing a closed-loop mapping from data modeling to insurance business application.

[0016] On the other hand, this application also provides a model-based residual value assessment system for new energy vehicles, used to apply the aforementioned model-based residual value assessment method for new energy vehicles, including: The data acquisition and storage module is configured to build three tables of data; The preprocessing module is configured to preprocess the data from the three tables. The feature fusion module is configured to fuse the static features in the vehicle sample table with the temporal features in the market panel table and the usage event stream based on the model identifier, and generate a feature sequence according to the monthly time steps; The static feature sub-network module inputs the static features in the feature sequence into the static feature sub-network to obtain a static representation; the temporal feature sub-network module is configured to input the temporal features in the feature sequence into the temporal feature sub-network to obtain a hidden representation at each time step; the attention layer module is configured to obtain a temporal representation based on attention weighting. The quantile prediction sub-network module is configured to simultaneously output the value preservation rate prediction results of at least two different quantiles based on the quantile loss function after concatenating the static representation and the time series representation, forming a value preservation rate quantile set. The drift detection and model update module is configured to monitor the input distribution and model residual distribution based on a fixed-length sliding window, calculate population stability indicators and distribution difference statistics and compare them with thresholds, and generate drift events based on lag and cooling strategies. The insurance amount assessment module is configured to obtain a suggested insurance amount assessment value based on the set of residual value quantiles and the new vehicle purchase price.

[0017] It is understandable that the above-mentioned methods and systems for assessing the residual value of new energy vehicles based on model have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating the method for evaluating the residual value of new energy vehicles based on model dimensions, provided in this embodiment of the invention; Figure 2 The structural block diagram of the new energy vehicle residual value assessment system based on model dimension provided in the embodiments of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Traditional methods for assessing the residual value of new energy vehicles, which rely on vehicle model as the foundation for modeling, fail to effectively differentiate between different models of the same vehicle based on static characteristics such as powertrain type, battery capacity, and configuration level. This results in assessments that do not accurately reflect value fluctuations at the model level. Market time-series characteristics and event stream data processed only through static weighting or moving averages are ill-suited to capturing the non-linear time-series correlations of features such as median transaction prices and regional factors. Furthermore, they neglect the dynamic impact of discrete events like maintenance events and software updates on residual value. Finally, the lack of a continuous monitoring mechanism for input data distribution shifts and residual distribution changes after model deployment leads to uncontrollable performance degradation when market conditions change abruptly or policy subsidies are adjusted.

[0021] For example, in an insurance company's vehicle residual value assessment system, when dealing with a certain brand of new energy vehicles, there are three versions of the same model: standard range, long range, and performance. These three versions differ in battery capacity, drive type, and manufacturer's suggested retail price. The existing assessment model combines all three into a single model dimension, resulting in the long range version having a higher actual residual value than predicted due to its lower battery degradation rate, while the performance version has a lower actual residual value than predicted due to higher maintenance costs. Regional factors in the market panel data experience quarterly fluctuations due to policy adjustments, but the model uses a fixed time window for mean processing, failing to promptly respond to the impact of these fluctuations on transaction prices. After two years of operation, the model's residual value prediction residual for pure electric vehicles has continued to widen because it has not detected a shift in power type distribution from plug-in hybrids to pure electric vehicles.

[0022] If these issues are not addressed, assessment errors will lead to overpayment risks for insurance companies in total loss vehicle claims, or misjudged pricing in high-value used car transactions. The lack of dynamic responsiveness to market time-series characteristics will reduce the model's predictive robustness in scenarios of sudden regional policy changes or supply chain fluctuations, forcing companies to frequently perform manual model parameter tuning. Failure to effectively capture long-term data distribution drift will cause gradual degradation of model performance, increasing the maintenance costs of model iterations and reducing the reliability of automated assessments.

[0023] For this, please refer to Figure 1 As shown, this application proposes a method for evaluating the residual value of new energy vehicles based on model, including: S100: Construct three tables of data and preprocess them. These tables include a vehicle sample table, a market panel table, and usage event stream data. The vehicle sample table uses model identifiers as the primary key and records static characteristics related to the model. The market panel table records market time-series characteristics related to residual value using a time index. The usage event stream data records maintenance, accidents, mileage, and software updates using timestamps.

[0024] S200: Based on the model identifier, the static features in the vehicle sample table are fused with the temporal features in the market panel table and the usage event stream to generate a feature sequence by monthly time step.

[0025] S300: Input the static features in the feature sequence into the static feature sub-network to obtain the static representation, input the temporal features in the feature sequence into the temporal feature sub-network to obtain the hidden representation at each time step, and obtain the temporal representation based on attention weighting.

[0026] S400: Perform drift detection on the input distribution and model residual distribution entering the temporal feature subnetwork. Drift detection includes calculating the population stability index and / or the Kolmogorov-Smirnov test statistic and comparing them with a preset threshold. When any comparison result meets the drift condition, the incremental training of the model is triggered, and the online model is updated after training is completed.

[0027] S500: The static representation and the time-series representation are concatenated and input into the quantile prediction subnetwork. Based on the quantile loss function, the value preservation rate prediction results of two different quantiles are output to obtain the value preservation rate quantile set.

[0028] S600: The recommended insurance amount is obtained based on the residual value quantile set and the new vehicle purchase price.

[0029] Specifically, the model dimension refers to using vehicle model as the smallest unit of analysis. A model is uniquely determined by the vehicle configuration combination defined by the manufacturer, and can be encoded using model identifiers. The model dimension can distinguish new energy vehicles with different power types, battery capacities, and configuration levels, solving the problem of coarse granularity in traditional model-based analysis. The data consists of three tables: a vehicle sample table, a market panel table, and usage event stream data. The vehicle sample table records static features using the model identifier as the primary key, implemented using database primary key constraints. The market panel table records market time-series features using a time index. The usage event stream data records events using timestamps. These three tables provide information from three dimensions: vehicle attributes, market dynamics, and usage behavior, addressing the issue that a single data source cannot cover multi-dimensional influencing factors. Preprocessing includes missing value completion, outlier removal, data alignment, and encoding, specifically implemented using multiple imputation, IQR rules, sliding window aggregation, and target encoding methods. Preprocessing eliminates data noise and inconsistencies, improving the robustness of subsequent modeling. Feature fusion is achieved by joining tables using style identifiers and time indices, specifically through database join key operations. Static features are copied monthly and concatenated with time-series features to form a time-step feature sequence, addressing the challenge of jointly modeling static and dynamic features. The static feature sub-network employs a multilayer perceptron structure, while the time-series feature sub-network uses a long short-term memory network, implemented through fully connected layers and LSTM units. These extract static attributes and time-dependent features respectively, addressing the inability of traditional methods to effectively handle heterogeneous features. Attention weighting uses static representations as query vectors to calculate time-step weights, specifically through a scaled dot product attention mechanism. This dynamically focuses on key time-step information, resolving noise interference and information redundancy issues in time-series features. Drift detection includes statistical tests of input and residual distributions, employing population stability indices and the Kolmogorov-Smirnov test. Incremental training is triggered by threshold comparisons, addressing performance degradation caused by changes in data distribution. The quantile prediction subnetwork outputs multi-quantile results based on the quantile loss function, specifically employing a quantile regression neural network structure. It simultaneously generates low-quantile and median residual value predictions, addressing the issue that a single valuation cannot meet the constraints of insurance risk. The insurance amount assessment is calculated based on the quantile set and the new vehicle price, specifically using a monotonic mapping function and constraint correction rules. This transforms the residual value into a suggested insurance amount, resolving the problem that traditional empirical formulas cannot quantify valuation uncertainty.

[0030] This application constructs a three-table data system based on vehicle type, and combines a deep neural network fusion architecture with static and time-series features to achieve fine-grained dynamic prediction of the residual value of new energy vehicles. At the same time, it introduces input distribution and residual drift detection mechanisms to ensure the long-term effectiveness of the model. Finally, based on the quantile prediction results, it generates an amount assessment recommendation that meets the insurance risk control requirements.

[0031] The working process and principle of this application are as follows: First, three tables of data are constructed and preprocessed. These three tables include a vehicle sample table, a market panel table, and usage event stream data. The vehicle sample table uses model identifiers as the primary key and records static features related to the model. The market panel table records market time-series features related to residual value using a time index. The usage event stream data records maintenance, accidents, mileage, and software updates using timestamps. Then, based on the model identifiers, the static features from the vehicle sample table are fused with the time-series features from the market panel table and the usage event stream to generate a feature sequence by monthly time steps. This step integrates data from different sources onto the same time scale. Next, the static features in the feature sequence are input into a static feature sub-network to obtain static representations, and the time-series features are input into a time-series feature sub-network to obtain hidden representations for each time step. A time-series representation is then obtained based on attention weighting. This step encodes the static and time-series features separately using a neural network and uses an attention mechanism to weight and summarize the time-series features. Drift detection is performed on the input distribution and model residual distribution entering the time-series feature sub-network. Drift detection involves calculating population stability indices and / or Kolmogorov-Smirnov test statistics and comparing them with preset thresholds. Incremental training of the model is triggered when any comparison result meets the drift condition, and the online model is updated after training. This step aims to monitor changes in data distribution and update the model promptly to maintain its performance. The static and temporal representations are concatenated and input into the quantile prediction subnetwork. Based on the quantile loss function, two different quantiles are output to predict the residual value, obtaining a set of residual value quantiles. This step uses quantile regression to estimate the probability distribution of the residual value. Finally, an insurance amount assessment recommendation is obtained based on the residual value quantile set and the new vehicle purchase price. This step combines the predicted residual value with the new vehicle price to derive the final insurance amount recommendation. By integrating multi-source data, deep learning models, and statistical methods, accurate assessment and dynamic updating of the residual value of new energy vehicles are achieved.

[0032] As a preferred embodiment, the solution of this application is specifically implemented as follows: First, a three-table data structure is constructed. The vehicle sample table includes fields: model identifier, production date, power type, battery pack capacity, drive type, configuration level, and manufacturer's suggested retail price. The market panel table includes fields: time index, median transaction price, transaction volume, regional factor, and policy subsidy dummy variable. The event stream data includes fields: timestamp, event type (repair / accident / mileage / software update), and event details.

[0033] The data from the three tables are preprocessed. Individual identifiers in the vehicle sample table are hashed and anonymized to meet K-anonymity constraints. Heterogeneous price data in the market panel table are standardized in currency and tax-inclusive caliber, and then truncated using Windsor simplification. Anomaly detection and discontinuity removal are performed on mileage events using event stream data.

[0034] Perform feature fusion. Join the data from the three tables using the style identifier and the month-end time index as the joint key. Aggregate the market panel table by month and calculate the index weighting. Map the event stream data to monthly event statistical vectors. Generate a fixed-length monthly time step feature sequence, pruning excess portions, padding missing portions, and outputting a missing mask.

[0035] A neural network model is constructed. The static feature subnetwork adopts a multilayer perceptron structure. The temporal feature subnetwork adopts a long short-term memory network structure. The attention layer uses the static representation as the query vector and the hidden representation at each time step as the key and value to calculate the attention weights. The quantile prediction subnetwork outputs the value preservation rate prediction results for two different quantiles simultaneously based on the quantile loss function.

[0036] Drift detection is implemented. Input distribution samples and model residual samples are extracted using a fixed-length sliding window. A population stability index is calculated for the input distribution, and a distribution difference statistic is calculated for the model residuals. The calculation results are compared with a preset threshold, and drift events are generated by combining hysteresis and cooling strategies.

[0037] When incremental training is triggered, training is performed using data from a fixed-length sliding window, with higher weights assigned to recent samples. After training is complete, offline backtesting and online shadow validation are performed. Once the thresholds for loss and quantile coverage are met, the new model is deployed online via a canary release using the feature service.

[0038] Finally, a recommended insurance amount assessment is calculated based on the residual value quantile set and the new vehicle purchase price. The low quantile residual value is selected as the risk constraint basis and converted to a conservative residual value using a monotonic mapping function. The conservative residual value is multiplied by the new vehicle purchase price to obtain the candidate insurance amount, and then corrected by applying upper and lower limit constraints, minimum residual value constraints, and depreciation curve boundary constraints. The recommended insurance amount assessment and its confidence level are then output.

[0039] Through the above-described scheme, this application achieves accurate assessment of the residual value of new energy vehicles across different models. By integrating static and dynamic time-series features, it captures the impact of factors such as different configurations, battery performance, and usage behavior on residual value. The introduction of a long short-term memory network and attention mechanism effectively models complex time dependencies. The designed drift detection and incremental training mechanisms enable the model to adapt to market and policy changes, maintaining long-term stability. The quantile prediction method provides a probability distribution estimate of the residual value, offering a basis for risk control. These technical means collectively improve the accuracy, dynamic responsiveness, and robustness of new energy vehicle residual value assessment, providing reliable technical support for insurance pricing and risk management.

[0040] This application further proposes constructing a three-table data structure and preprocessing the data, including: generating a vehicle sample table with the model identifier as the unique primary key; writing static features according to field normalization, including production date, power type, battery pack capacity, drive type, configuration level, and manufacturer's suggested retail price; hashing and anonymizing information containing individual identifiers while satisfying K-anonymity constraints (K≥20); generating a market panel table with a time index, summarizing market time-series features weekly or monthly, including median transaction price, transaction volume, regional factors, and dummy variables related to policy subsidies; unifying the currency and tax-inclusive caliber of heterogeneous price data and performing Windsor truncation; generating event stream data based on timestamps, recording maintenance events, accident events, mileage events, and software update events one by one, and detecting and removing discontinuities in mileage events; preprocessing includes using multiple imputation or business default values ​​to complete missing values; and using a dual criterion of IQR rules and business thresholds to remove anomalies. The market panel table is aligned with the event stream data by month-end and aggregated using a fixed-length sliding window. Target encoding is applied to high-cardinality segments, and in-fold fitting is performed.

[0041] Specifically, the vehicle sample table uses model identifiers as unique primary keys to ensure a one-to-one correspondence between static features and models, and field normalization ensures a consistent data format. Hash desensitization combined with K-anonymity constraints protects individual privacy; for example, K≥20 reduces the risk of re-identification. The market panel table aggregates market time-series features using time indexes, and Windsor truncation eliminates interference from extreme values; for example, the median transaction price is truncated to the upper and lower 5 percentiles. Event stream data is used to record events with timestamps, and anomaly detection removes mileage jumps; for example, two consecutive mileage differences exceeding a threshold are considered anomalies. The preprocessing stage uses multiple imputation to complete missing values; for example, missing fields are imputed using chain equations. Dual-criteria anomaly removal combines statistical rules and business experience; for example, IQR rules filter statistical anomalies, and business thresholds limit data for power type or battery capacity exceeding reasonable ranges. Fixed-length sliding window aggregation ensures time-series alignment; for example, a 12-month window length. When processing high-cardinality segments, such as mean encoding for configuration levels, target encoding is performed and fitted within the training set to avoid information leakage.

[0042] Specifically, during the vehicle sample table generation process, the model identifier serves as the primary key to ensure data uniqueness. Static features such as production date and power type are written according to preset field formats to avoid data redundancy. Hash desensitization converts individual identifiers into irreversible strings, and combined with K-anonymity constraints, ensures that at least 20 samples exist under each model, preventing privacy leaks. When summarizing weekly or monthly market data in the market panel table, heterogeneous price data is unified to RMB (including tax), and Windsor truncation is applied to retain 90% of the data, reducing the interference of extreme transaction prices on subsequent modeling. In event stream data, mileage events are incremented using a sliding window; if a single mileage jump exceeds three times the historical maximum value, it is considered an anomaly and removed. In the preprocessing stage, multiple imputation is used for missing values, such as predicting missing battery pack capacity based on a random forest model. IQR rules and business thresholds are jointly used to filter anomalies; for example, vehicles with a pure electric power type but a battery capacity below 20kWh are removed. After aligning the market panel table and usage event stream data by month-end, they are aggregated with a fixed-length window, for example, retaining data from the most recent 24 months, pruning any excess data, and padding for any insufficient data. High-cardinality segments, such as those at the configuration level, use target encoding and fit mean encoding in 5-fold cross-validation to avoid overfitting. Through these steps, the data from the three tables are optimized in terms of privacy protection, anomaly handling, and temporal alignment, providing high-quality input for subsequent feature fusion and model training.

[0043] As a preferred embodiment, the solution of this application is specifically implemented as follows: A vehicle sample table is generated using the model identifier as the unique primary key, and static features are written according to field normalization. Static features include production date, powertrain type, battery pack capacity, drive type, configuration level, and manufacturer's suggested retail price. Information containing individual identifiers is hash-anonymized and satisfies the K-anonymity constraint, where K ≥ 20.

[0044] Market panel tables are generated using time indexes, summarizing market time-series characteristics by week or month. These characteristics include median transaction price, transaction volume, regional factors, and dummy variables related to policy subsidies. Currency and tax-inclusive caliber are standardized for heterogeneous price data, and Windsor truncation is applied.

[0045] Event stream data is generated based on timestamps, recording maintenance events, accident events, mileage events, and software update events one by one. Anomaly jump detection is performed on mileage events, and discontinuities are removed.

[0046] Preprocessing includes using multiple imputation or business default values ​​to fill in missing values. Anomalies are removed using a dual-criteria approach of IQR rules and business thresholds. The market panel table and usage event stream data are aligned by month-end and aggregated using a fixed-length sliding window. Target encoding is applied to high-cardinality segments, and fitting is performed within the transition.

[0047] Through the above technical solutions, this application achieves comprehensive collection and effective preprocessing of data related to the residual value assessment of new energy vehicles. By constructing a three-table data structure to store vehicle static characteristics, market time-series characteristics, and usage event data respectively, a structured data foundation is provided for subsequent modeling. Individual information is anonymized to protect user privacy. Preprocessing steps such as standardizing data definitions, removing outliers, and filling in missing values ​​improve data quality. Techniques such as sliding window aggregation and target encoding effectively handle time-series and high-cardinality features. These measures collectively provide high-quality training data for the subsequent residual value assessment model, contributing to improved model accuracy and stability.

[0048] This application further proposes a method using model identifiers and month-end time indexes as joint keys to connect the vehicle sample table, market panel table, and usage event stream data, and to replicate static features monthly to each time step. Monthly aggregation of the market panel table yields a subset of market time-series features, and an exponential weighting is calculated. Usage event stream data is mapped to monthly event statistical vectors, and non-negative truncation and outlier removal are performed on mileage increments. A fixed-length monthly time-step feature sequence is generated without using subsequent data constraints; excess portions are pruned, missing portions are padded forward, and a missing mask is output. Static features, exponentially weighted market time-series features, and monthly event statistical vectors are concatenated in a fixed field order to form a feature sequence.

[0049] The composite key join ensures accurate data association through dual matching of model identifiers and month-end time indexes. When aggregating market panel tables by month, an exponential weighting algorithm is used to reduce the noise impact of historical data. Monthly event statistics vectors transform discrete events into structured features through event type classification statistics. Fixed-length sliding windows maintain the integrity of the time series through forward padding and pruning operations. Fixed field order avoids dimensional misalignment during feature concatenation.

[0050] Specifically, the joint key of model identifier and month-end time index precisely correlates the static features of the vehicle sample table with the time-series features of the market panel table and event stream data, eliminating time misalignment issues across data sources. The index-weighted calculation of the market panel table assigns higher weight to recent data, effectively capturing dynamic market trends. Event stream data, through non-negative truncation and anomaly removal, filters mileage jump noise and transforms it into a monthly statistical vector, preserving the frequency and intensity of events. A fixed-length sliding window maintains time series continuity through forward padding while pruning excess data to avoid redundancy. Missing mask generation marks invalid time steps, providing a basis for subsequent model processing. Fixed field order concatenation ensures consistency of feature dimensions across different time steps, avoiding chaotic model input structure. This solution improves the temporal accuracy of feature sequences through multi-dimensional data alignment, noise filtering, and structure transformation, providing high-quality input for subsequent model learning.

[0051] As a preferred embodiment, the solution of this application is specifically implemented as follows: Model identifiers and month-end time indexes are used as a composite key to connect the vehicle sample table, market panel table, and usage event stream data. Static features are replicated monthly to each time step.

[0052] The market panel table is aggregated monthly to obtain a subset of market time-series characteristics. An exponential weighted average of these market time-series characteristics is calculated, with a weighting coefficient set to 0.9.

[0053] The event stream data is mapped to a monthly event statistics vector. A non-negative truncation is performed on the mileage increment, setting negative values ​​to 0. Outliers with mileage increments exceeding 10,000 kilometers are removed.

[0054] Generate a monthly timestep feature sequence with a fixed length of 24 months. Prune portions exceeding 24 months and padded portions less than 24 months. Output a missing mask and mark the padded timesteps.

[0055] At each time step, static features, exponentially weighted market time-series features, and monthly event statistical vectors are concatenated in a fixed field order to form the feature sequence for that time step. Static features include six fields such as production year and month and power type; market time-series features include four fields such as median transaction price and transaction volume; and monthly event statistical vectors include four fields such as number of repairs, number of accidents, mileage increment, and number of software updates.

[0056] Through the above technical solution, this application achieves effective fusion of static and temporal features of new energy vehicle models. The generation of fixed-length feature sequences ensures consistency of subsequent model inputs. The use of exponential weighting and event statistical vectors highlights the impact of recent market and usage conditions on residual value. Handling outliers and missing values ​​improves data quality. This fusion method provides comprehensive and structured input data for subsequent residual value prediction, contributing to improved accuracy and stability.

[0057] In some of the solutions described above in this application, the fusion of static features and time-series features may result in insufficient feature representation, ineffective extraction of the complex nonlinear relationships of static features, and insufficient modeling of the long-term dependencies of time-series features. At the same time, the interaction mechanism between static and time-series features lacks dynamic adaptability, which affects the model's prediction accuracy.

[0058] This application further proposes that the static feature subnetwork is a multilayer perceptron, which receives static features and sequentially passes them through fully connected layers and nonlinear activation layers to output a static representation. The temporal feature subnetwork is a long short-term memory network, which receives monthly time-step feature sequences and masks invalid time steps using a missing mask, outputting hidden representations for each time step. The attention layer uses the static representation as the query vector and the hidden representations of each time step as the key and value, calculates the attention weights, and summarizes the hidden representations of each time step in a weighted manner to generate a temporal representation. Before concatenating the static and temporal representations, normalization and discard regularization are performed on the static and temporal representations, respectively.

[0059] The multilayer perceptron comprises three fully connected layers with 256, 128, and 64 neurons per layer, respectively. ReLU activation is used, and the last layer outputs a static representation vector with a dimension of 32. The long short-term memory network (LSTM) has a hidden layer dimension of 64 and a fixed time step of 24 months. Missing time steps are masked using a binary identifier matrix. The attention layer employs a scaled dot product attention mechanism, calculating the similarity score between the static representation vector and the hidden vector at each time step, and generating a weight distribution through softmax normalization. Layer-level normalization is used, and discard regularization randomly masks neurons with a probability of 0.2 during the training phase.

[0060] Specifically, static features are nonlinearly transformed through the fully connected layers of a multilayer perceptron to extract high-order feature representations of the model's inherent attributes. Temporal features are processed step-by-step by a long short-term memory network to capture the temporal dependencies between market dynamics and usage behavior. An attention mechanism dynamically adjusts the weights of each time step based on the static features, giving higher attention to features at key time points. Normalization eliminates feature scale differences, and discarding regularization prevents overfitting. For example, in a model with an 80kWh battery capacity, the multilayer perceptron combines it with the drive type and configuration level to generate 128-dimensional intermediate features, which are then compressed into a 32-dimensional static representation. The long short-term memory network encodes 24 months of market transaction price fluctuations, generating 64-dimensional hidden states. The attention layer calculates the similarity between the static representation and the hidden states of each month, assigning higher weight to months with policy subsidies. Finally, the concatenated 128-dimensional fusion vector is processed by a quantile prediction network to output the residual value range, reducing the prediction error by 12.7% compared to the baseline model.

[0061] As a preferred embodiment, the solution of this application is specifically implemented as follows: The static feature subnetwork is a multilayer perceptron that receives static features, passes them sequentially through fully connected layers and nonlinear activation layers, and outputs a static representation. Specifically, the static feature subnetwork contains three fully connected layers, with 64, 32, and 16 neurons in each layer, respectively, and ReLU is used as the activation function.

[0062] The temporal feature subnetwork is a Long Short-Term Memory (LSTM) network that receives monthly time-step feature sequences, masks invalid time steps using a missing mask, and outputs the hidden representation for each time step. Furthermore, the LSTM network consists of two layers of LSTM units, each with a hidden state dimension of 64 and a dropout rate of 0.2.

[0063] The attention layer uses the static representation as the query vector and the hidden representation at each time step as the key and value. It calculates the attention weights and then aggregates the hidden representations at each time step in a weighted manner to generate a temporal representation. The attention mechanism employs scaled dot product attention, with four attention heads.

[0064] Before concatenating the static and temporal representations, normalization and dropout regularization are performed on the static and temporal representations respectively. Therefore, layer normalization is adopted, and the dropout rate is set to 0.1.

[0065] Through the above technical solution, this application achieves an effective fusion of static and temporal features, fully capturing the long-term dependencies and short-term fluctuations in the residual value assessment of new energy vehicles. The static feature sub-network extracts an abstract representation of the vehicle's inherent attributes, the temporal feature sub-network models the dynamic changes in the market and usage, and the attention mechanism achieves adaptive weighting of features at different time steps. This fusion approach improves the accuracy and robustness of residual value prediction, providing high-quality feature representations for subsequent quantile prediction. Simultaneously, normalization and regularization operations effectively alleviate overfitting and enhance the model's generalization ability.

[0066] This application further proposes a method for extracting input distribution samples and model residual samples separately using a fixed-length sliding window. A population stability index is calculated for the input distribution based on a pre-registered feature list, and a distribution difference statistic is calculated for the model residuals in chronological order. The population stability index and distribution difference statistic are compared with preset thresholds, and drift events are generated by combining lag and cooling strategies. Each drift event includes the drift type, a list of affected features, an amplitude level, and a set of affected product types.

[0067] The fixed-length sliding window ensures data timeliness and consistency by sampling over a fixed time span. Input distribution samples are extracted based on a pre-registered feature list covering all key static and temporal features. Population stability indicators are calculated using binning statistics, comparing the current window distribution with historical baseline distributions. Model residual samples are extracted chronologically, and the distribution difference statistic is calculated using the Kolmogorov-Smirnov test to determine the deviation between the current residual distribution and the training phase residual distribution. A lag strategy determines drift only after setting trigger conditions within multiple consecutive detection periods, avoiding misjudgments caused by instantaneous fluctuations. A cooling-off strategy sets a quiet period after drift event processing to prevent high-frequency recurrence. The structured information of drift events includes drift type distinguishing between input and residual drift, an affected feature list marking specific anomalous features, an amplitude level quantifying drift severity, and an affected model set identifying the range of models requiring adjustment.

[0068] Specifically, input distribution samples are extracted sequentially through a fixed-length sliding window, with the window length set to three months based on the business cycle to ensure sufficient data coverage. The population stability index calculates the binning percentage difference for each registration feature; an anomaly is marked when the binning percentage difference for a particular feature exceeds 5%. Model residual samples are extracted within the same window, and the residual distribution difference is calculated using the Kolmogorov-Smirnov test; a statistic exceeding 0.3 indicates residual drift. The lag strategy requires two consecutive detection periods to exceed a threshold before generating a drift event, and the cooling-off strategy prohibits repeated detection within the next three periods after event processing. When the median market transaction price undergoes a sudden change due to policy adjustments, the population stability index for the transaction price feature in the input distribution exceeds the threshold. This, combined with confirmation using the lag strategy, generates a drift event, triggering incremental model training to update the weight parameters of the market time-series features, thereby improving prediction accuracy.

[0069] As a preferred embodiment, the solution of this application is specifically implemented as follows: Input distribution samples and model residual samples are extracted separately using a fixed-length sliding window. For example, the sliding window length can be set to 30 days, and 1000 samples can be randomly extracted from the data of the most recent 30 days each day as input distribution samples. At the same time, the difference between the corresponding model prediction results and the actual values ​​is collected as model residual samples.

[0070] The population stability index is calculated based on a pre-registered list of features for the input distribution, and the distribution difference statistic is calculated for the model residuals in chronological order. Specifically, important features such as vehicle price, mileage, and battery capacity can be selected, and the changes in the mean, variance, and other statistics of these features in the old and new samples can be calculated to obtain the population stability index. For the model residuals, the Kolmogorov-Smirnov test can be used to calculate the maximum difference in the cumulative distribution function of the old and new residual samples as the distribution difference statistic.

[0071] Drift events are generated by comparing population stability indicators and distribution difference statistics with preset thresholds, and combining lag and cooling strategies. Each drift event includes the drift type, a list of affected features, the magnitude level, and the set of affected vehicle models. For example, the population stability indicator threshold can be set to 0.1, and the distribution difference statistics threshold to 0.05. Drift detection is triggered when either indicator exceeds the threshold. To avoid frequent triggering, a 7-day lag and a 30-day cooling-off period can be set. Drift events record the drift type (input drift / residual drift), the list of features exceeding the threshold, the drift magnitude (slight / moderate / severe), and the list of affected vehicle models.

[0072] Through the above technical solution, this application can promptly detect changes in input data distribution and model performance, avoiding a decrease in prediction accuracy due to data drift in practical applications. By setting reasonable lag and cooldown periods, detection sensitivity and computational overhead can be balanced. Detailed information on drift events helps analyze the causes of drift and implement targeted model update strategies, improving the stability and reliability of the new energy vehicle residual value assessment model.

[0073] This application further proposes a method for incremental training triggered by a drift event, including performing incremental training using data in a fixed-length sliding window and assigning higher weights to recent samples, while retaining the previous version's calibration parameters as initial values ​​for recalibration. After training is completed, offline backtesting and online shadow validation are performed sequentially. Once the thresholds for loss and quantile coverage metrics are met, a new model is deployed online via canary release using feature services, retaining version snapshots and an automatic rollback strategy.

[0074] A fixed-length sliding window ensures the temporal continuity and manageable volume of training data. High weighting of recent samples is achieved through an exponential decay function, giving recent data a stronger influence on model parameter adjustments. Previous calibration parameters are retained as initial values, and recalibration is achieved through parameter freezing and unfreezing strategies, avoiding the waste of computational resources from complete retraining. Offline backtesting uses three months of historical data for forward validation, with loss metrics including quantile loss and mean squared error constraints. Online shadow validation uses a split-feed real-time request to infer in parallel between the old and new models, comparing the difference in prediction results with the fluctuation range of business metrics. Canary releases employ a phased traffic switching strategy, initially limiting the deployment ratio to less than 5%, gradually expanding to full deployment by monitoring key metrics in real time. Version snapshots and automatic rollback strategies are implemented through version metadata management; automatic rollback is triggered when key metrics exceed thresholds for three consecutive periods.

[0075] Specifically, when a shift in the input or residual distribution is detected, an incremental training process is automatically triggered. A fixed-length sliding window extracts twelve months of data from the current time point backward as the training set. The weight of each sample is distributed exponentially based on the time proximity, with the weight coefficient of the most recent month's samples set to 2.0 and a monthly decay coefficient of 0.8. During training, the calibration layer parameters of the previous version of the model are retained as initial values, and only the feature extraction layer is fine-tuned. After training is complete, the model is backtested offline using a retained validation set, requiring the quantile coverage error to be no more than 3% and the mean squared error to decrease by no less than 10%. The model that passes offline validation enters the online validation phase, where 10% of real-time requests are simultaneously routed to both the old and new models. The overlap rate of the quantile intervals output by the two models is compared in real time with the difference rate in the premium calculation by the business side, requiring an overlap rate higher than 85% and a difference rate lower than 5%. After all validation conditions are met, the new model gradually replaces the online service through a canary release mechanism, with an initial replacement ratio of 5%, increasing the replacement ratio every 24 hours until full replacement. Each version update generates a complete snapshot containing model parameters, feature engineering configurations, and validation metrics. When the online service response latency increases by 15% or the quantile coverage decreases by more than 2%, an automatic rollback to the previous stable version is triggered. This process is implemented through an automated pipeline, ensuring the security of model updates and business continuity.

[0076] As a preferred embodiment, the solution of this application is specifically implemented as follows: When a drift event triggers incremental training, incremental training is performed using data from a fixed-length sliding window, with higher weights assigned to recent samples. Simultaneously, the previous version's calibration parameters are retained as initial values ​​for recalibration. Specifically, exponentially decaying sample weights can be used: samples from the most recent 30 days are assigned a weight of 1.0, samples from 30-60 days are assigned a weight of 0.8, and samples from 60-90 days are assigned a weight of 0.6. For model parameters, the parameters from the previous version can be used as initial values, and fine-tuning is performed using a smaller learning rate, for example, reducing the original learning rate to 1 / 10.

[0077] After training, the new model undergoes offline backtesting and online shadow validation sequentially. Once the thresholds for loss and quantile coverage are met, the new model is deployed via a canary release using the feature service, retaining version snapshots and an automatic rollback strategy. Specifically, offline backtesting can be performed on historical data from the most recent three months, calculating the mean absolute percentage error (MAPE) and quantile coverage, requiring a MAPE below 5% and a quantile coverage between 90% and 110%. Online shadow validation involves running the old and new models in parallel in the production environment for one week, comparing the consistency of prediction results, requiring over 90% of samples to have a prediction bias within ±3%. Canary releases can employ a canary release strategy, initially deploying the new model on 10% of the traffic, observing for one day without anomalies, and then gradually expanding to 100%. Version snapshots retain the model parameters and training data from the most recent five versions, and the automatic rollback strategy is set to trigger a rollback to the previous stable version if the prediction bias exceeds 10% for one consecutive hour.

[0078] Through the above technical solutions, this application achieves adaptive model updates and smooth transitions. The incremental training method retains historical knowledge and quickly adapts to new data distributions, the multiple validation mechanism ensures the reliability of the updated model, and the canary release and automatic rollback strategies further reduce the risk of model updates. Therefore, the model can maintain high prediction accuracy while avoiding sudden performance drops due to environmental changes, enhancing the stability and reliability of the evaluation.

[0079] This application further proposes to concatenate the static representation and the temporal representation and input them into the quantile prediction subnetwork. Based on the quantile loss function, it simultaneously outputs the value retention rate prediction results of two different quantiles. During training, the quantile set is fixed and the gradient of each quantile sample is balanced. During inference, the feature order and normalized statistics are kept consistent with those in the training stage to generate the value retention rate quantile set.

[0080] The quantile loss function adopts the quantile regression loss form, and the fixed quantile set includes the low quantile and median quantiles. Gradient balancing is achieved by adjusting the loss weights of each quantile sample, for example, assigning higher weights to low quantile samples to compensate for their sparsity. The normalization statistic is used to calculate the mean and variance during the training phase, and is directly reused to standardize the input features during the inference phase.

[0081] Specifically, the quantile prediction subnetwork receives concatenated static and temporal feature vectors during the training phase and maps them to the quantile output space through fully connected layers. The quantile loss function independently calculates the deviation between the predicted and true values ​​for each quantile; the low quantile loss function uses an asymmetric weighted form, while the median loss function uses a symmetric form. Gradient balancing applies normalization coefficients to the loss terms of different quantiles during backpropagation to ensure balanced gradient contributions across quantiles. During the inference phase, features are input in the order they were arranged during training, and the normalized statistics directly load the mean and variance parameters from the training phase, avoiding prediction errors caused by statistical shifts. This scheme improves the stability of quantile prediction through a fixed quantile set and gradient balancing mechanism, while ensuring the reliability of inference results through consistency constraints on feature order and normalization.

[0082] As a preferred embodiment, the solution of this application is specifically implemented as follows: The static and temporal representations are concatenated and input into the quantile prediction subnetwork. The quantile prediction subnetwork consists of multiple fully connected layers and a nonlinear activation function. Based on the quantile loss function, the network simultaneously outputs preservation rate predictions for two different quantiles. These two quantiles include the low quantile preservation rate (e.g., 10%) used for risk constraints and the median preservation rate (50%) used for baseline estimation.

[0083] During training, the quantile set is fixed at {10%, 50%}. The gradients of each quantile sample are balanced to prevent the gradients of high quantile samples from dominating the training process. Specifically, the 10% quantile sample is assigned a higher weight (e.g., 5), while the 50% quantile sample has a weight of 1.

[0084] During inference, the feature order and normalized statistics are maintained in the same manner as during training. First, the input features are standardized, then passed through a fully connected layer and a ReLU activation function in sequence. The last layer does not use an activation function and directly outputs the value retention rate prediction values ​​of the two quantiles, generating a set of value retention rate quantiles {10th quantile, 50th quantile}.

[0085] Through the above technical solutions, this application can simultaneously output the value retention rate prediction results for multiple quantiles, providing more comprehensive information for subsequent risk control and pricing decisions. By using the quantile loss function and gradient balancing, the accuracy of low quantile predictions is improved, which helps to better assess risks in extreme situations. Maintaining consistency between the training and inference phases improves the stability and reliability of the model in practical applications.

[0086] This application further proposes interval calibration of the value retention rate quantile set to obtain a calibrated value retention rate quantile set. Interval calibration includes: calculating the coverage deviation and interval width of each quantile on the reserved dataset; when the coverage deviation exceeds a preset threshold, correcting the quantile output using temperature scaling and / or equal-frequency binning monotonic calibration. Quantile order constraints are applied during the calibration process, and a snapshot of the calibration parameters is retained. The calibrated value retention rate quantile set is obtained after calibration.

[0087] Coverage bias is calculated by comparing the quantile coverage ratio of the predicted quantiles with that of the actual observed values. The calibration process is triggered when the absolute difference between the coverage ratio and the theoretical quantile exceeds a preset threshold. Temperature scaling adjusts the scale of the quantile output by introducing a learnable parameter. Equal-frequency binning monotonic calibration maps the predicted quantiles to the equal-frequency bin boundaries of the actual observed values. Quantile order constraints ensure that the calibrated quantiles maintain a strictly increasing order through an optimized algorithm, avoiding quantile overlap. Calibration parameter version snapshots are stored using both timestamps and model version numbers to ensure the traceability of the calibration process.

[0088] Specifically, when calculating the coverage bias of each quantile on the reserved dataset, the proportion of samples with actual observations lower than the predicted quantiles is first counted and compared with the theoretical quantiles. When the coverage bias exceeds a preset threshold, the temperature scaling method adjusts the scaling factor of the quantile output to make the coverage proportion of the predicted quantiles approach the theoretical value. Equal-frequency binning monotonic calibration maps the predicted quantiles to equal-frequency bin boundaries constructed based on historical data, ensuring the actual coverage accuracy of the quantiles. During calibration, a constraint optimization algorithm forces the quantiles to maintain an increasing order, preventing lower quantiles from exceeding higher quantiles. After calibration, the adjusted quantile parameters are associated with and stored with the corresponding model version to ensure that the correct calibration parameters can be called during subsequent inference. This process effectively improves the statistical consistency of quantile predictions and enhances maintainability through version snapshot management.

[0089] As a preferred embodiment, the specific implementation of this application's scheme is as follows: After generating the set of quantiles for the retention rate, an interval calibration operation is performed. First, the reserved dataset is divided into a validation set and a test set. The validation set is used to calculate the coverage bias and interval width of each quantile. The coverage bias is defined as the difference between the proportion of actual observed values ​​falling within the predicted quantile interval and the theoretical quantile value. When the absolute value of the coverage bias exceeds 5%, the calibration process is triggered. Specifically, a temperature scaling method is used to adjust the scale parameter of the quantile output, where the temperature coefficient is obtained by optimizing the quantile loss function on the validation set through gradient descent. Furthermore, an equal-frequency binning method is introduced to address the remaining bias, dividing the predicted quantiles into equal-frequency intervals according to sample sorting, forcing the quantile output to satisfy a monotonically increasing constraint. During the calibration process, an order constraint term is added to ensure that the calibrated quantiles do not overlap. After calibration is completed, the temperature coefficient and binning boundary parameters are stored in a version snapshot, and the metadata information of the feature service is updated synchronously.

[0090] Through the above technical solutions, this application effectively solves the coverage bias problem of quantile prediction results under the actual data distribution, and improves the accuracy of the residual value prediction range. By combining temperature scaling and equal-frequency binning calibration, statistical consistency and monotonicity of the quantile output are achieved, avoiding distortion in the insured amount assessment caused by model prediction bias. Version snapshot management of calibration parameters further enhances the traceability of the model iteration process, ensuring that the confidence level of the insured amount assessment recommendation value strictly matches the business risk control requirements.

[0091] This application further proposes interval calibration of the value retention rate quantile set to obtain a calibrated value retention rate quantile set. Interval calibration includes: calculating the coverage deviation and interval width of each quantile on the reserved dataset; when the coverage deviation exceeds a preset threshold, correcting the quantile output using temperature scaling and / or equal-frequency binning monotonic calibration. Quantile order constraints are applied during the calibration process, and a snapshot of the calibration parameters is retained. The calibrated value retention rate quantile set is obtained after calibration.

[0092] Coverage bias is calculated by comparing the actual coverage sample ratio with the theoretical quantile, with a preset threshold of ±3%. Temperature scaling adjusts the quantile output position by introducing a learnable scaling parameter, which is optimized using gradient descent on the hold-out dataset. Equal-frequency binning monotonic calibration divides the predicted values ​​into multiple equal-frequency intervals, forcing the quantile output to satisfy a monotonically increasing constraint. Quantile order constraints are implemented through an order penalty term in the loss function, ensuring that quantiles do not cross over after calibration. Version snapshots of calibration parameters are stored in a distributed database, supporting querying and rollback by timestamp.

[0093] Specifically, after the quantile prediction subnetwork generates the initial set of quantiles for the retained value rate, a dataset is reserved for input into the model to obtain prediction results, and the actual coverage sample ratio corresponding to each quantile is statistically analyzed. When the coverage deviation exceeds ±3%, temperature scaling calibration is initiated, adjusting the quantile positions by optimizing the scaling parameters to make the actual coverage ratio approach the theoretical value. For cases where the interval width is insufficient, an equal-frequency binning method is further adopted, dividing the predicted value into 20 equal-frequency intervals to force the quantile output to meet the interval coverage uniformity. During the calibration process, an order loss function is introduced to ensure that low quantiles are always lower than high quantiles, avoiding quantile overlap. After calibration, the scaling parameters, bin boundaries, and order constraint weights are stored as version snapshots, with each calibration generating an independent version identifier. When the calibrated quantile set is input into the insurance amount assessment module, the latest version of the parameters is called for quantile mapping to ensure the statistical consistency and traceability of the assessment results.

[0094] As a preferred embodiment, the specific implementation of this application is as follows: When generating a suggested insurance amount assessment value, the 10th quantile is first selected from the residual value quantile set as the low quantile residual value. This low quantile residual value is nonlinearly mapped using a pre-trained sigmoid function, where the slope parameter of the mapping function is automatically fitted by the quantile deviation between the actual residual value and the predicted value of the same model of vehicle over the past three years. The mapped conservative residual value is multiplied by the current new car purchase price to obtain the initial candidate insurance amount. The candidate amount needs to be constrained: when the calculated result exceeds 80% of the new car purchase price, it is forcibly set to 75% of the purchase price. If the calculated result is lower than 15% of the purchase price, it is set to 18% based on the lowest residual value data of the model over the past three years. When it is detected that the 10th quantile is missing from the residual value quantile set, the system automatically switches to the median residual value and multiplies it by a conservative coefficient of 0.85 to generate a compensation value. The final output includes the suggested insurance amount value and its confidence level, where the confidence level is dynamically labeled according to the coverage completeness of the quantile set and the version of the mapping function parameters.

[0095] Through the above technical solutions, this application effectively addresses the shortcomings of traditional insured amount assessment methods in risk pricing for new energy vehicles, namely insufficient conservatism and lack of constraint mechanisms. By employing quantile mapping and multiple constraint mechanisms, it ensures that the insured amount consistently covers tail risks in a dynamic market environment. Automatic degradation paths and residual value boundary constraints prevent pricing deviations caused by abnormal model outputs. Dynamic labeling of confidence levels provides interpretable quantitative risk data for insurance business decisions.

[0096] In the above embodiments, a three-table data system is constructed at the model level: a vehicle sample table, a market panel table, and usage event stream data. After multi-level preprocessing and alignment, static and temporal features are fused along monthly time steps to form a complete feature sequence. This sequence is then input into the static and temporal feature sub-networks respectively, and an attention mechanism is used to extract static and temporal representations. This ensures fine-grained characterization of differences in features across different models and enhances the ability to model dynamic market factors and usage behavior over time. A dual drift detection mechanism involving input distribution and model residuals is introduced, automatically triggering incremental training and model updates when data or model shifts, thereby improving prediction stability and adaptability. At the prediction level, the quantile prediction sub-network outputs residual value results for multiple quantiles, and the results are aggregated to provide both median estimation and low-quantile prediction with risk constraints, solving the problem of traditional methods' inability to quantify uncertainty. Based on the quantile set and the new vehicle purchase price, a suggested insurance amount is derived, achieving a closed-loop mapping from data modeling to insurance business application.

[0097] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a new energy vehicle residual value assessment system based on model dimension, used to apply the above-mentioned new energy vehicle residual value assessment method based on model dimension, including: The data acquisition and storage module is configured to build three tables of data.

[0098] The preprocessing module is configured to preprocess the data from the three tables.

[0099] The feature fusion module is configured to fuse static features from the vehicle sample table with temporal features from the market panel table and the usage event stream based on the model identifier, and generate a feature sequence by monthly time step.

[0100] The static feature subnetwork module inputs static features from the feature sequence into the static feature subnetwork to obtain static representations. The temporal feature subnetwork module is configured to input temporal features from the feature sequence into the temporal feature subnetwork to obtain hidden representations for each time step. The attention layer module is configured to obtain temporal representations based on attention weighting.

[0101] The quantile prediction sub-network module is configured to simultaneously output the value preservation rate prediction results of at least two different quantiles based on the quantile loss function after concatenating the static representation and the time series representation, forming a value preservation rate quantile set.

[0102] The drift detection and model update module is configured to monitor the input distribution and model residual distribution based on a fixed-length sliding window, calculate population stability indicators and distribution difference statistics and compare them with thresholds, and generate drift events based on lag and cooling strategies.

[0103] The insurance amount assessment module is configured to obtain a suggested insurance amount assessment value based on the residual value quantile set and the new vehicle purchase price.

[0104] Through the above technical solution, the data acquisition and storage module is configured to construct a vehicle sample table, a market panel table, and use event stream data. The vehicle sample table records static features using model identifiers as the primary key; the market panel table records market time-series features using time indexes; and the event stream data records maintenance, accident, mileage, and software update events using timestamps. The preprocessing module performs data anonymization, missing value completion, outlier removal, and feature encoding. Information containing individual identifiers is hashed and K-anonymity is satisfied; heterogeneous price data undergoes currency unification and Windsor truncation. The feature fusion module uses model identifiers and month-end time indexes as a joint key, copies static features monthly to each time step, aggregates market time-series features monthly, and maps them to the event stream data as monthly statistical vectors, generating a fixed-length monthly feature sequence. The static feature sub-network module processes static features using a multilayer perceptron, the time-series feature sub-network module processes time-series features using a long short-term memory network, and the attention layer module performs weighted summarization of the time-series hidden representations based on the static representations. The quantile prediction sub-network module inputs the concatenated features into the quantile prediction network and outputs residual value prediction results for at least two quantiles. The drift detection and model update module calculates the stability index of the input distribution and residual distribution based on a sliding window, triggers incremental training, and updates the model through gray-scale release. The insurance amount assessment module selects the low quantile residual value for conservative conversion and generates a suggested insurance amount value based on the new vehicle purchase price.

[0105] Specifically, the data acquisition and storage module separates static and dynamic data through a three-table structure. The vehicle sample table records production configuration information at the model level, the market panel table captures price fluctuation characteristics at the time level, and event stream data reflects changes in vehicle usage status. The preprocessing module employs a dual-criteria anomaly removal strategy, using IQR rules and business thresholds in parallel to verify data validity, and target encoding is used for high-cardinality segments to reduce dimensionality. The feature fusion module achieves cross-table data alignment through a joint key, replicates static features monthly to ensure the integrity of the time series, applies exponential weighting to market time series features to amplify the impact of recent data, and performs non-negative truncation on mileage increments to eliminate anomalous jumps. The static feature subnetwork extracts inherent attribute features of models through fully connected layers, the time series feature subnetwork uses long short-term memory units to capture time dependencies, and an attention mechanism dynamically adjusts the weight distribution at different time steps. The quantile prediction subnetwork uses a quantile loss function to simultaneously optimize the output of multiple quantile points, fixing the quantile set and balancing gradient update amounts during training. The drift detection module monitors feature distribution shifts by calculating population stability indicators and uses the Kolmogorov-Smirnov test statistic of the residual distribution to determine model degradation. During incremental training, it retains historical calibration parameters to achieve a smooth transition. The insurance amount assessment module establishes a monotonic mapping function to transform quantile prediction results into conservative estimates. It corrects the output through upper and lower limit constraints and depreciation curve boundary conditions. When quantile data is missing, a downgrade compensation mechanism based on the median benchmark is activated. All modules operate collaboratively through standardized data interfaces. The output of the preprocessing module directly serves as the input to the feature fusion module. The hidden states of the temporal feature subnetwork are attention-weighted and concatenated with the static representation. Quantile prediction results are synchronously transmitted to the insurance assessment module and the drift detection module, forming a closed-loop data processing link.

[0106] As a preferred embodiment, the solution of this application is implemented as follows: The new energy vehicle residual value assessment system receives vehicle sample tables, market panel tables, and usage event stream data from external data sources via a data acquisition and storage module. The vehicle sample tables store production month and year, power type, and battery pack capacity fields using model identifiers as primary keys. A preprocessing module completes missing values ​​in the three tables, fills in missing transaction price values ​​in the market panel table using multiple imputation methods, and removes abnormal mileage jump records in the usage event stream data using both IQR rules and business thresholds. A feature fusion module uses model identifiers and month-end timestamps as joint keys to copy the static features of the vehicle sample tables monthly to the aggregated results of the market panel tables and usage event stream data, forming a feature sequence containing static features, index-weighted market features, and monthly event statistical vectors. A static feature sub-network module uses a three-layer fully connected neural network to process static features, a temporal feature sub-network module uses a bidirectional long short-term memory network to process temporal data in the feature sequence, and an attention layer module uses the output of the static feature sub-network as a query vector to calculate the attention weights of the hidden representations at each time step. The quantile prediction sub-network module concatenates the representation vectors of static and temporal features, and uses a quantile loss function to simultaneously output the residual value prediction values ​​for the 10th and 50th quantiles. The drift detection and model update module monitors the PSI index of the input data in real time. When the PSI of the transaction price feature in the market panel table exceeds the 0.25 threshold, incremental training is triggered, updating the parameters of the temporal feature sub-network using data from the past three months. The insurance amount assessment module selects the 10th quantile residual value, converts it to a conservative residual value using a linear mapping function, multiplies it by the new vehicle purchase price, and applies a constraint that the residual value should not be less than 15% to output the final suggested insurance amount.

[0107] Through the above technical solutions, this application achieves fine-grained model-level modeling of the residual value of new energy vehicles, improving prediction accuracy by integrating static configuration features and dynamic time-series features. A quantile prediction framework is adopted to simultaneously output risk constraints and benchmark valuations, enhancing the robustness of insurance amount assessment. By monitoring data distribution drift in real time and triggering model updates, prediction biases caused by changes in the market environment are effectively addressed, maintaining the long-term stability of the system.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for evaluating the residual value of new energy vehicles based on model type, characterized in that, include: Three tables of data are constructed and preprocessed. The three tables include a vehicle sample table, a market panel table, and usage event stream data. The vehicle sample table uses model identifier as the primary key and records static features related to the model. The market panel table records market time-series features related to residual value using a time index. The usage event stream data records maintenance, accidents, mileage, and software updates using timestamps. Based on the model identifier, the static features in the vehicle sample table are fused with the temporal features in the market panel table and the usage event stream to generate a feature sequence by monthly time steps; The static features in the feature sequence are input into the static feature sub-network to obtain the static representation, and the temporal features in the feature sequence are input into the temporal feature sub-network to obtain the hidden representation at each time step. The temporal representation is then obtained based on attention weighting. Drift detection is performed on the input distribution and model residual distribution entering the temporal feature sub-network. The drift detection includes calculating the population stability index and / or the Kolmogorov-Smirnov test statistic and comparing them with a preset threshold. When any comparison result meets the drift condition, incremental training of the model is triggered, and the online model is updated after training is completed. The static representation and the time-series representation are concatenated and input into the quantile prediction subnetwork. Based on the quantile loss function, the value preservation rate prediction results of two different quantiles are output to obtain the value preservation rate quantile set. Based on the set of residual value quantiles and the new vehicle purchase price, a suggested insurance amount assessment value is obtained.

2. The method for evaluating the residual value of new energy vehicles based on model dimension as described in claim 1, characterized in that, When constructing three tables of data and preprocessing the three tables of data, the process includes: The vehicle sample table is generated using the model identifier as the unique primary key, and static features are written according to the field normalization. The static features include production date, power type, battery pack capacity, drive type, configuration level, and manufacturer's suggested retail price. Information containing individual identifiers is hash-de-identified and satisfies the K-anonymity constraint, where K≥20. The market panel table is generated using a time index, and market time-series characteristics are summarized by week or month. These market time-series characteristics include median transaction price, transaction volume, regional factors, and policy subsidy dummy variables. Currency and tax-inclusive caliber are unified for heterogeneous price data, and Windsor truncation is performed. The usage event stream data is generated according to the timestamp, and maintenance events, accident events, mileage events and software update events are recorded one by one. Abnormal jumps are detected in mileage events and discontinuities are removed. The preprocessing includes using multiple imputation or business default values ​​to fill in missing values; using a dual criterion of IQR rules and business thresholds to remove anomalies; aligning the market panel table with the event stream data at the end of the month and aggregating them with a fixed-length sliding window; and using target encoding and fitting within the fold for high cardinality segments.

3. The method for evaluating the residual value of new energy vehicles based on model dimension according to claim 2, characterized in that, When fusing the static features in the vehicle sample table with the temporal features in the market panel table and the usage event stream based on the model identifier, and generating a feature sequence by monthly time step, the following are included: Using the model identifier and the end-of-month time index as a joint key, the vehicle sample table is key-connected with the market panel table and the usage event stream data, and the static features are copied to each time step on a monthly basis; The market panel table is aggregated monthly to obtain a subset of market time-series features, and the exponential weighting of the market time-series features is calculated. The event stream data is mapped to a monthly event statistics vector, and non-negative truncation and anomaly removal are performed on the mileage increment; Without using data after any time step, a fixed-length monthly time step feature sequence is generated, the excess part is pruned, the missing part is padded forward and the missing mask is output at the same time. At each time step, the static features, exponentially weighted market time-series features, and monthly event statistical vectors are concatenated in a fixed field order to form the feature sequence for that time step.

4. The method for evaluating the residual value of new energy vehicles based on model dimension according to claim 3, characterized in that, When obtaining the static representation, hidden representation, and temporal representation, the following are included: The static feature subnetwork is a multilayer perceptron that receives the static features, passes them sequentially through a fully connected layer and a nonlinear activation layer, and outputs the static representation. The temporal feature subnetwork is a long short-term memory network. It receives the monthly time step feature sequence, and masks invalid time steps according to the missing mask, and outputs the hidden representation of each time step. The attention layer uses the static representation as the query vector, the hidden representation at each time step as the key and value, calculates the attention weight, and summarizes the hidden representation at each time step in a weighted manner to generate the temporal representation. Before concatenating the static representation and the temporal representation, normalization and discard regularization are performed on the static representation and the temporal representation, respectively.

5. The method for evaluating the residual value of new energy vehicles based on model dimension according to claim 1, characterized in that, Drift detection is performed on the input distribution and model residual distribution entering the temporal feature sub-network, including: Input distribution samples and model residual samples are extracted separately using a fixed-length sliding window. Calculate population stability indices for the input distribution based on a pre-registered list of features, and calculate distribution difference statistics for the model residuals in chronological order. The population stability index and distribution difference statistics are compared with preset thresholds, and drift events are generated by combining lag strategies and cooling strategies. The drift events include drift type, list of affected features, amplitude level and set of affected models.

6. The method for evaluating the residual value of new energy vehicles based on model dimension as described in claim 5, characterized in that, When the drift event triggers incremental training, it includes: Incremental training is performed using data from a fixed-length sliding window, with higher weights assigned to recent samples, while recalibrating using the calibration parameters from the previous version as initial values. After training, the model is tested offline and then online for shadow verification. Once the thresholds for loss and quantile coverage are met, the new model is deployed online in a canary release manner using the feature service, while retaining version snapshots and automatic rollback strategies.

7. The method for evaluating the residual value of new energy vehicles based on model dimension according to claim 1, characterized in that, The static representation and the temporal representation are concatenated and then input into the quantile prediction subnetwork, which includes: Based on the quantile loss function, the preservation rate prediction results of two different quantiles are output simultaneously. The different quantiles include at least the low quantile preservation rate used for risk constraints and the median preservation rate used for benchmark estimation. During training, the quantile set is fixed and the gradients of each quantile sample are balanced; during inference, the feature order and normalized statistics are kept consistent with those in the training phase to generate the set of quantiles for the value retention rate.

8. The method for evaluating the residual value of new energy vehicles based on model dimension according to claim 7, characterized in that, After obtaining the set of quantiles for the rate of return on value, it also includes: Perform interval calibration on the set of retention rate quantiles to obtain the calibrated set of retention rate quantiles; The interval calibration includes: calculating the coverage deviation and interval width of each quantile on the reserved dataset; when the coverage deviation exceeds a preset threshold, temperature scaling and / or equal-frequency bin monotonic calibration are used to correct the quantile output. Quantile order constraints are applied during calibration, and a snapshot of the calibration parameters is retained. After calibration, the set of calibrated value retention rate quantiles is obtained.

9. The method for evaluating the residual value of new energy vehicles based on model dimension according to claim 1, characterized in that, When obtaining the recommended insurance amount assessment value based on the aforementioned residual value quantile set and the new vehicle purchase price, the following is included: The lower quantile of the value retention rate is selected from the set of value retention rate quantiles as the basis for risk constraints. The low quantile preservation rate is converted into a conservative preservation rate using a monotonic mapping function. The parameters of the monotonic mapping function are automatically learned from historical backtesting and the quantile order is kept from overlapping. The candidate insurance amount is obtained by multiplying the conservative residual value rate by the new vehicle purchase price, and then corrected by applying upper and lower limit constraints, minimum residual value constraints and depreciation curve boundary constraints in sequence. When the set of retention rate quantiles is incomplete, a downgrade path based on the median retention rate is triggered and a conservative coefficient is added for compensation. Output the recommended insured amount and its confidence level.

10. A new energy vehicle residual value assessment system based on model dimension, used to apply the new energy vehicle residual value assessment method based on model dimension as described in any one of claims 1-9, characterized in that, include: The data acquisition and storage module is configured to build three tables of data; The preprocessing module is configured to preprocess the data from the three tables. The feature fusion module is configured to fuse the static features in the vehicle sample table with the temporal features in the market panel table and the usage event stream based on the model identifier, and generate a feature sequence according to the monthly time steps; The static feature sub-network module inputs the static features in the feature sequence into the static feature sub-network to obtain a static representation; The temporal feature subnetwork module is configured to input the temporal features in the feature sequence into the temporal feature subnetwork to obtain the hidden representation of each time step; The attention layer module is configured to obtain temporal representations based on attention weighting; The quantile prediction sub-network module is configured to simultaneously output the value preservation rate prediction results of at least two different quantiles based on the quantile loss function after concatenating the static representation and the time series representation, forming a value preservation rate quantile set. The drift detection and model update module is configured to monitor the input distribution and model residual distribution based on a fixed-length sliding window, calculate population stability indicators and distribution difference statistics and compare them with thresholds, and generate drift events based on lag and cooling strategies. The insurance amount assessment module is configured to obtain a suggested insurance amount assessment value based on the set of residual value quantiles and the new vehicle purchase price.

Images