An index prediction method and related apparatus

Abstract

The application discloses an index prediction method and related device, and relates to the technical field of artificial intelligence. The method comprises the following steps: obtaining target feature data of a target object. The target feature data comprises feature values of each target dimension in a target dimension set, and the target dimension set comprises at least one target dimension. The target feature data is input into a target prediction model to obtain a prediction value of a to-be-predicted index of the target object output by the target prediction model. The target prediction model is obtained by taking sample data in a first time window as a training set and taking sample data in a second time window as a test set, and the prediction accuracy of the target prediction model on the test set is higher than a target first threshold. The second time window is later than the first time window.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an indicator prediction method and related apparatus. Background Technology

[0002] Traditional prediction models rely on posterior behavioral data for inference and correction, resulting in low prediction accuracy. Summary of the Invention

[0003] In view of the above problems, this application provides an indicator prediction method and related apparatus, the specific solution of which is as follows:

[0004] The first aspect of this application provides an indicator prediction method, including:

[0005] Obtain target feature data of the target object; the target feature data includes feature values ​​of each target dimension in the target dimension set, and the target dimension set includes at least one target dimension.

[0006] The target feature data is input into the target prediction model to obtain the predicted value of the target object's indicator output by the target prediction model;

[0007] The target prediction model is trained using sample data within a first time window as the training set and sample data within a second time window as the test set, and the prediction accuracy of the target prediction model on the test set is higher than the first target threshold; the second time window is later than the first time window.

[0008] In one possible implementation, the training method for the target prediction model includes at least one round of training, which includes:

[0009] Obtain a set of training dimensions, wherein the set of training dimensions includes at least one sample dimension;

[0010] A training set is constructed using sample data within a first time window, and a test set is constructed using sample data within a second time window. The sample data includes feature values ​​for each sample dimension.

[0011] Candidate models are trained based on the training set, and the candidate models are tested based on the test set to obtain the prediction accuracy.

[0012] In response to the prediction accuracy being higher than the first threshold, the candidate model is determined to be the target prediction model, and the training dimension set is determined to be the target dimension set corresponding to the target prediction model;

[0013] In response to the prediction accuracy not being higher than the first threshold, the training dimension set is updated at least based on the prediction accuracy.

[0014] In one possible implementation, the set of training dimensions obtained in the first round of training includes:

[0015] The candidate dimensions that satisfy the first correlation condition with the metric to be predicted are selected from the candidate dimension pool to obtain the training dimension set;

[0016] The training dimension set includes at least one sample dimension; the candidate dimension pool includes multiple candidate dimensions, and the first correlation condition includes a correlation higher than a first correlation threshold.

[0017] In one possible implementation, during non-initial training rounds, the set of training dimensions obtained includes:

[0018] Obtain the set of training dimensions obtained from the previous training update.

[0019] In one possible implementation, updating the training dimension set based at least on the prediction accuracy includes:

[0020] Lower the first correlation threshold, and select candidate dimensions from the candidate dimension pool that satisfy the first correlation condition with the indicator to be predicted, to obtain an updated training dimension set.

[0021] In one possible implementation, updating the training dimension set based at least on the prediction accuracy includes:

[0022] Determine the prediction residual vector of the candidate model on the test set after this round of training;

[0023] Calculate the correlation between each candidate dimension in the candidate dimension pool and the predicted residual vector;

[0024] In the candidate dimension pool, at least one candidate dimension that satisfies the second correlation condition is obtained and added to the training dimension set to obtain an updated training dimension set. The second correlation condition includes that the correlation with the predicted residual vector is higher than the second correlation threshold.

[0025] In one possible implementation, updating the training dimension set based at least on the prediction accuracy includes:

[0026] From the training dimension set, a first special sample dimension is determined. The first special sample dimension includes sample dimensions that satisfy a first special condition. The first special condition includes that the prediction accuracy of the corresponding sample data subset is lower than a second threshold. The sample data subset corresponding to a sample dimension includes multiple samples with the same feature value of the sample dimension.

[0027] Obtain the dimensions related to the first special sample dimension, add them to the training dimension set, and obtain the updated training dimension set.

[0028] In one possible implementation, updating the training dimension set based at least on the prediction accuracy includes:

[0029] Obtain the contribution parameters of each sample dimension in the training dimension set;

[0030] Based on the contribution parameter, a second special sample dimension is determined from the training dimension set. The second special sample dimension includes sample dimensions that meet the second special condition, which includes sample dimensions whose contribution parameter is higher than the third threshold.

[0031] Obtain the dimensions related to the first special sample dimension, add them to the training dimension set, and obtain the updated training dimension set.

[0032] Based on the contribution parameter, a third special sample dimension is determined from the training dimension set. The third special sample dimension includes sample dimensions that meet the third special condition, which includes sample dimensions whose contribution parameter is lower than the fourth threshold. The third threshold is higher than the fourth threshold.

[0033] The dimension obtained by transforming the third special sample dimension;

[0034] Add the training dimension set to obtain the updated training dimension set. The transformation includes binning discretization.

[0035] In one possible implementation, the number of new sample dimensions added to the updated training dimension set in each round of training is negatively correlated with the prediction accuracy.

[0036] A second aspect of this application provides an indicator prediction device, comprising:

[0037] A feature data acquisition unit is used to obtain target feature data of a target object; the target feature data includes feature values ​​of each target dimension in a target dimension set, and the target dimension set includes at least one target dimension.

[0038] The model prediction unit is used to input the target feature data into the target prediction model to obtain the predicted value of the target object's indicator output by the target prediction model;

[0039] The target prediction model is trained using sample data within a first time window as the training set and sample data within a second time window as the test set, and the prediction accuracy of the target prediction model on the test set is higher than the first target threshold; the second time window is later than the first time window.

[0040] Using the above technical solution, this application provides an indicator prediction method and related apparatus to obtain target feature data of a target object. The target feature data includes feature values ​​of each target dimension in a target dimension set, and the target dimension set includes at least one target dimension. The target feature data is input into a target prediction model to obtain the predicted value of the target object's indicator. The target prediction model is trained using sample data within a first time window as the training set and sample data within a second time window as the test set, and the prediction accuracy of the target prediction model on the test set is higher than a first target threshold. The second time window is later than the first time window. Attached Figure Description

[0041] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0042] Figure 1 A flowchart illustrating an indicator prediction method provided in this application;

[0043] Figure 2 A flowchart illustrating another indicator prediction method provided in this application;

[0044] Figure 3 A flowchart illustrating another indicator prediction method provided in this application;

[0045] Figure 4 A flowchart illustrating another indicator prediction method provided in this application;

[0046] Figure 5 A flowchart illustrating another indicator prediction method provided in this application;

[0047] Figure 6 A schematic diagram of an index prediction device provided in this application;

[0048] Figure 7 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0049] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0050] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0051] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar elements and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing elements with the same properties in the description of embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0052] Figure 1 This is a flowchart illustrating an indicator prediction method provided in an embodiment of this application, as shown below. Figure 1 As shown in the embodiments of this application, an indicator prediction method specifically includes:

[0053] S101. Obtain the target feature data of the target object.

[0054] In this embodiment, the target feature data includes feature values ​​of each target dimension in the target dimension set, which includes at least one target dimension. A dimension represents different aspects, angles, or attributes of a feature; one dimension describes one aspect of the target object. For example, if the target object is a battery, dimensions may include usage time, cycle count, average temperature, fast charge cycles, charging cutoff voltage, internal resistance, etc. Similarly, if the target object is a human body, dimensions may include height, weight, age, heart rate, blood pressure, etc.

[0055] In this embodiment, the feature value of the target dimension is the numerical value corresponding to the target object under that target dimension. For example, the feature value of the cycle count dimension is 320 times, and the feature value of the average temperature dimension is 25.5 degrees Celsius. As another example, when the target object is a human body, the height dimension is 175 centimeters, and the weight dimension is 70 kilograms.

[0056] It should be noted that the target dimension set refers to the set of dimensions actually used in this prediction. The dimensions contained in the target dimension set may be different for different prediction tasks or different prediction models.

[0057] In this embodiment, the specific method for obtaining the target feature data of the target object includes at least one of the following:

[0058] The feature values ​​of each target dimension of the target object are collected in real time through sensors, measuring equipment or monitoring systems.

[0059] Read the feature values ​​of each target dimension that have been stored in the target object from a database, data warehouse, or historical record.

[0060] Import feature values ​​of each target dimension of the target object through user input or external interface.

[0061] Taking a battery whose health status is to be predicted as an example, the target dimension set includes three target dimensions: cycle count, average temperature, and fast charge count. By reading the real-time data of the battery through the battery management system, the cycle count is found to be 320, the average temperature is 26.3 degrees Celsius, and the fast charge count is 45. Then the target feature data includes the cycle count of 320, the average temperature of 26.3 degrees Celsius, and the fast charge count of 45.

[0062] It should be noted that after obtaining the target feature data, the target feature data can also be preprocessed, such as missing value imputation, outlier detection and correction, normalization or standardization, etc., to ensure data quality.

[0063] S102. Input the target feature data into the target prediction model to obtain the predicted value of the target object's indicator output by the target prediction model.

[0064] In this embodiment, the metric to be predicted includes the target parameter or state quantity that needs to be predicted. It can be understood that there is a correlation between the metric to be predicted and the various target dimensions in the target feature data. For example, if the target object is a battery, the metric to be predicted is a parameter characterizing the battery state, such as battery health, and the predicted value is a percentage value.

[0065] In this embodiment, the target prediction model includes a pre-trained mathematical model or algorithmic model, used to calculate and output predicted values ​​of the indicator to be predicted based on the input target feature data. Optionally, the target prediction model may be at least one of a regression model, a classification model, or a neural network model.

[0066] In this embodiment, the target prediction model is trained using sample data within a first time window as the training set and sample data within a second time window as the test set, and the prediction accuracy of the target prediction model on the test set is higher than the first target threshold. The second time window is later than the first time window.

[0067] In one optional embodiment, sample data of the battery during a first time window from January 2025 to December 2025 is collected as a training set, and sample data of the same battery model during a second time window from January 2026 to April 2026 is collected as a test set. The initial model of the battery health prediction model is trained using the training set, and the prediction accuracy of the trained model is evaluated using the test set. When the prediction accuracy is higher than a target first threshold, the trained model is determined as the target prediction model, i.e., the battery health prediction model.

[0068] Furthermore, in response to receiving the prediction task, the target feature data obtained includes 320 cycles, an average temperature of 26.3 degrees Celsius, and 45 fast charging cycles. This target feature data is organized into a three-dimensional vector and then input into a pre-trained battery health prediction model. After inference, the battery health prediction model outputs a predicted battery health value of 91.2%.

[0069] As can be seen from the above technical solution, the indicator prediction method provided in this application first obtains the target feature data of the target object, and then inputs the target feature data into the target prediction model to obtain the predicted value of the target object's indicator output by the target prediction model. The target feature data includes the feature values ​​of each target dimension in the target dimension set. The target prediction model is trained using sample data within a first time window as the training set and sample data within a second time window as the test set. Since the second time window is trained later than the first time window, and the prediction accuracy of the target prediction model on the test set is higher than the target first threshold, the training set and test set are strictly divided by the time axis during the training process of the target prediction model. This ensures that the target prediction model cannot know future data during the training phase of the training set, but can truly evaluate the model's ability to predict future data during the testing phase of the test set, thereby improving the prediction accuracy of the target prediction model on future real data.

[0070] Specifically, one prediction method uses a random partitioning approach to divide historical data into training and test sets during the training process. For example, historical data within the same time frame is randomly shuffled and divided according to a preset ratio. The training and test sets come from the same time window, and there is no chronological order between the data. This method verifies the model's ability to fit historical data, rather than its ability to predict the real future. It cannot evaluate the model's generalization performance over time, is prone to overfitting, and results in high prediction accuracy on the test set but low prediction accuracy on future real data.

[0071] In this application, the training set is entirely from the first time window and the test set is entirely from the second time window, realizing the modeling logic of predicting the future from the past, which is highly consistent with the real business scenario. The target prediction model aims to achieve a prediction accuracy higher than the first target threshold on the test set. Therefore, compared with other prediction methods, the target prediction model of this application has a reliable prediction accuracy on real future data.

[0072] In one optional embodiment, the training method for the target prediction model includes at least one round of training. Figure 2 This is a flowchart illustrating another indicator prediction method provided in an embodiment of this application. Figure 2 This illustrates a specific implementation of any training round in the process of training the target prediction model, such as... Figure 2 As shown, one round of training includes:

[0073] S201, Obtain the set of training dimensions.

[0074] In this embodiment, the training dimension set includes at least one sample dimension, which is also the feature dimension used to train the candidate model. For example, in the battery health prediction scenario, the sample dimension may include the number of cycles, average temperature, number of fast charging cycles, charging cut-off voltage, and internal resistance value.

[0075] S202. Construct a training set using sample data within the first time window, and construct a test set using sample data within the second time window.

[0076] In this embodiment, the sample data includes feature values ​​of each sample dimension. Specifically, the sample data includes multiple samples, and each sample includes a set of feature values ​​of each sample dimension.

[0077] In this embodiment, the first time window is the time interval for collecting training sample data, and the second time window is the time interval for collecting test sample data. The second time window is later than the first time window. For example, sample data of batteries collected during the first time window from January 2025 to December 2025 is used as the training set, and sample data of batteries collected during the second time window from January 2026 to April 2026 is used as the test set.

[0078] S203. Train candidate models based on the training set, and test the candidate models based on the test set to obtain the prediction accuracy.

[0079] In this embodiment, the sample data in the training set also includes the sample labels, that is, the true values ​​of the metrics to be predicted. Specifically, the sample data in the training set is used to learn and optimize the parameters of the candidate model. After training, the sample data in the test set and their corresponding labels are used to evaluate the prediction performance of the candidate model to obtain the prediction accuracy. The metric for prediction accuracy includes at least one of the following: root mean square error, mean absolute percentage error, coefficient of determination, accuracy, precision, recall, or F1 score.

[0080] S204. In response to the prediction accuracy being higher than the first threshold, the candidate model is determined as the target prediction model, and the training dimension set is determined as the target dimension set corresponding to the target prediction model.

[0081] In this embodiment, when the prediction accuracy of the candidate model on the test set is higher than the first threshold, it means that the candidate model meets the business requirements and is determined as the target prediction model. The training dimension set used in this round of training is determined as the target dimension set corresponding to the target prediction model, which is used to collect feature data of the target object when using the target prediction model in the prediction stage. The first threshold includes a pre-configured expected threshold, which represents the minimum prediction accuracy that the target prediction model needs to achieve.

[0082] S205. In response to a prediction accuracy not exceeding a first threshold, update the training dimension set at least based on the prediction accuracy.

[0083] In this embodiment, when the prediction accuracy is not higher than the first threshold, it means that the amount of information provided by the sample dimensions contained in the current training dimension set is insufficient to support the candidate model to achieve the expected prediction performance on the test set. Therefore, it is necessary to update the training dimension set at least based on the prediction accuracy and introduce new sample dimensions to supplement the missing information of the model.

[0084] It should be noted that the specific methods for updating the training dimension set based at least on the prediction accuracy can include a variety of methods. In one optional embodiment, updating the training dimension set based at least on the prediction accuracy specifically includes: lowering the first correlation threshold, filtering candidate dimensions from the candidate dimension pool that satisfy the first correlation condition with the indicator to be predicted, and obtaining the updated training dimension set.

[0085] In this embodiment, the candidate dimension pool includes multiple candidate dimensions, and the first relevance condition includes a relevance higher than a first relevance threshold. It is understood that a higher first relevance threshold is used in the early stages of training to filter out feature dimensions strongly correlated with the metric to be predicted, ensuring the reliability of the basic model. When the model's prediction accuracy on the test set is not higher than the first threshold, the first relevance threshold is lowered, introducing more feature dimensions related to the metric to be predicted, supplementing the missing information in the model, and improving prediction accuracy.

[0086] As can be seen from the above technical solutions, the indicator prediction method provided in this application constructs a training dimension set including at least one sample dimension, constructs a training set with sample data within a first time window, constructs a test set with sample data within a second time window, trains a candidate model based on the training set, and tests it based on the test set to obtain the prediction accuracy. When the prediction accuracy is higher than a first threshold, the candidate model is determined as the target prediction model. When the prediction accuracy is not higher than the first threshold, the training dimension set is updated based on the prediction accuracy and the next round of training begins. Thus, through iterative optimization, the training dimension set is continuously adjusted so that the final target prediction model achieves the expected prediction accuracy on the test set with strict time axis segmentation, thereby improving the prediction accuracy of the target prediction model for the indicator to be predicted.

[0087] In one optional embodiment, obtaining the training dimension set in the first round of training includes: selecting candidate dimensions from the candidate dimension pool that satisfy the first correlation condition with the indicator to be predicted, thereby obtaining the training dimension set.

[0088] In this embodiment, the training dimension set includes at least one sample dimension; wherein, the candidate dimension pool includes multiple candidate dimensions, and the first correlation condition includes a correlation higher than a first correlation threshold.

[0089] For example, in the battery health prediction scenario, the candidate dimension pool includes candidate dimensions such as usage time, number of cycles, average temperature, number of fast charges, charging cut-off voltage, and internal resistance. The correlation between each candidate dimension and battery health is calculated, and the first correlation threshold is set to 0.6. Candidate dimensions with a correlation higher than 0.6 are selected. For example, the correlation coefficient of the number of cycles is 0.85, the correlation coefficient of the number of fast charges is 0.72, and the correlation coefficient of the internal resistance is 0.68. These three candidate dimensions are added to the training dimension set for the first round of training.

[0090] In one optional embodiment, obtaining the training dimension set in non-first rounds of training includes: obtaining the training dimension set updated in the previous round of training.

[0091] In this embodiment, during non-first rounds of training, if the prediction accuracy of the candidate model on the test set in the previous round of training is not higher than the first threshold, the result of updating the training dimension set based on the prediction accuracy is used as the training dimension set for the current round, without the need to re-select from the candidate dimension pool.

[0092] Taking the second round of training in a battery health prediction scenario as an example, the first round of training used a training dimension set including three sample dimensions: number of cycles, number of fast charges, and internal resistance value. After training, the prediction accuracy of the candidate model on the test set did not exceed a first threshold. Based on the prediction accuracy, the training dimension set was updated by adding an average temperature dimension, resulting in an updated training dimension set including four sample dimensions: number of cycles, number of fast charges, internal resistance value, and average temperature. In the second round of training, this updated training dimension set is used as the training dimension set for the second round of training.

[0093] In summary, the training process of the target prediction model includes at least one round of training. In the first round, a training dimension set needs to be initialized from the candidate dimension pool. If the prediction accuracy after the first round of training is not higher than the first threshold, the next round of training begins. In each subsequent round of training, the training dimension set updated in the previous round is used as the training dimension set for the current round. After one or more rounds of iterative training, until the prediction accuracy of the candidate model on the test set is higher than the first threshold, the candidate model obtained in that round of training is determined as the target prediction model, and the training dimension set used in that round is determined as the target dimension set. It is evident that this scheme, through the training mechanism of initial initialization and subsequent inheritance updates, allows the training dimension set to gradually converge with the prediction accuracy feedback, avoiding the computational overhead of re-selecting from the candidate dimension pool in each round, while ensuring that the target prediction model meets the business requirements for prediction performance on data with strict time-separated segments.

[0094] In one optional embodiment, this scheme constrains the number of newly added sample dimensions. Specifically, in each round of training, the number of newly added sample dimensions in the updated training dimension set is negatively correlated with the prediction accuracy. That is, the lower the prediction accuracy, the more newly added sample dimensions are required; the higher the prediction accuracy, the fewer newly added sample dimensions are required. This achieves a progressive optimization process from coarse-tuning to fine-tuning. When the prediction accuracy is low, the model lacks a lot of information, requiring the introduction of a large number of sample dimensions at once to quickly improve model performance. As the prediction accuracy gradually increases, the missing information in the model gradually decreases, requiring only a small number of sample dimensions to be introduced for fine-tuning, avoiding overfitting or wasting computational resources due to excessive introduction of redundant dimensions. In summary, the number of newly added sample dimensions gradually decreases as the prediction accuracy increases, achieving an effective transition from coarse screening to fine-tuning.

[0095] In one alternative embodiment, Figure 3 This is a flowchart illustrating another indicator prediction method provided in this application. Figure 3 This illustrates another specific implementation of S205, namely, updating the training dimension set at least based on the prediction accuracy in response to a prediction accuracy not exceeding a first threshold. Figure 3 As shown, this method includes:

[0096] S301. Determine the prediction residual vector of the candidate model on the test set after this round of training.

[0097] In this embodiment, the prediction residual vector is composed of the difference between the predicted value output by the candidate model on the test set and the true label corresponding to the sample in the test set. The prediction residual vector reflects the information that the candidate model failed to capture, and each element in the residual vector corresponds to the prediction error of a sample in the test set.

[0098] S302. Calculate the correlation between each candidate dimension in the candidate dimension pool and the predicted residual vector.

[0099] In this embodiment, the correlation between each candidate dimension and the predicted residual vector is calculated. For example, a correlation coefficient is calculated. Optionally, the correlation coefficient includes at least one of Pearson correlation coefficient, Spearman rank correlation coefficient, or mutual information. It is understood that the higher the correlation between a candidate dimension and the predicted residual vector, the more relevant the information contained in that candidate dimension is to the information currently missing in the candidate model.

[0100] S303. In the candidate dimension pool, obtain at least one candidate dimension that satisfies the second correlation condition, add it to the training dimension set, and obtain the updated training dimension set.

[0101] In this embodiment, the second correlation condition includes a correlation with the predicted residual vector that is higher than a second correlation threshold. Candidate dimensions from the candidate dimension pool whose correlation with the predicted residual vector is higher than the second correlation threshold are selected and added to the current training dimension set, forming an updated training dimension set.

[0102] For example, in a battery health prediction scenario, the correlation between each candidate dimension and the predicted residual vector is calculated using Pearson correlation coefficient or mutual information. The correlation coefficients are: fast charge count and residual = 0.72; internal resistance value and residual = 0.61; cycle count and residual = 0.45; and average temperature and residual = 0.38. Setting a second correlation threshold of 0.5, candidate dimensions with correlations higher than 0.5 are selected as new dimensions, namely fast charge count and internal resistance value. These are then added to the training dimension set, resulting in an updated training dimension set for the next round of training.

[0103] It should be noted that only the candidate dimensions outside the training dimension set in the candidate pool and their correlation with the predicted residual vector need to be calculated. In other words, new dimensions are selected from the candidate pool that have not been selected into the training dimension set.

[0104] As can be seen from the above technical solution, this solution directly uses the prediction error to perform feature selection by predicting the residual vector, which can specifically fill the information gaps in the candidate model and quickly locate the feature dimension that needs to be supplemented.

[0105] In one alternative embodiment, Figure 4 This is a flowchart illustrating another indicator prediction method provided in this application. Figure 4 This illustrates another specific implementation of S205, namely, updating the training dimension set at least based on the prediction accuracy in response to a prediction accuracy not exceeding a first threshold. Figure 4 As shown, this method includes:

[0106] S401. Determine the first special sample dimension from the training dimension set.

[0107] In this embodiment, the first special sample dimension includes sample dimensions that satisfy the first special condition. The first special condition includes that the prediction accuracy of the corresponding sample data subset is lower than the second threshold. The sample data subset corresponding to a sample dimension includes multiple samples with the same feature value of the sample dimension.

[0108] It should be noted that if a subset of sample data with the same feature value in a certain sample dimension has a significantly lower prediction accuracy on the test set than other subsets of sample data, it indicates that there is a strong correlation between the value range of that sample dimension and the prediction difficulty, and the candidate model exhibits a systematic bias on samples with the same feature value in that sample dimension.

[0109] For example, in the scenario of battery health prediction, the fast charging count dimension is divided into four value ranges according to feature values: 0 to 100 times, 101 to 300 times, 301 to 500 times, and greater than 500 times. The prediction accuracy of the sample data subset corresponding to each value range is calculated. If the prediction accuracy of the subset greater than 500 times is lower than the second threshold, then the fast charging count is determined as the first special sample dimension.

[0110] S402. Obtain the dimensions related to the first special sample dimension, add them to the training dimension set, and obtain the updated training dimension set.

[0111] In this embodiment, the specific methods for obtaining the dimension related to the first special sample dimension include:

[0112] Candidate dimensions that satisfy the correlation condition with the first special sample dimension are selected from the candidate dimension pool, or derived dimensions related to the first special sample dimension are directly generated. It should be noted that the dimensions obtained that are related to the first special sample dimension can further distinguish subtle differences between samples within the same value range of the first special sample dimension, thereby helping the candidate model reduce the prediction error of that sample subset.

[0113] In an optional embodiment, the method for directly generating derived dimensions related to the first specific sample dimension includes generating derived dimensions that have at least one correlation with the first specific sample dimension, wherein the correlation includes at least one of statistical correlation, business logic association, derivational construction association, or causal association. Specifically, statistical correlation is measured using Pearson correlation coefficient, Spearman rank correlation coefficient, or mutual information. Business logic association represents the physical connection or logical relationship between dimensions based on domain knowledge. Derivational construction association represents the association formed when generating new dimensions through mathematical transformations or combination operations on the original dimensions. Causal association represents the causal relationship between dimensions.

[0114] Preferably, if the prediction accuracy of the subset of sample data corresponding to the first special sample dimension is lower than the second threshold, it indicates that the candidate model has a systematic bias in a specific value range of that dimension, manifested as local overfitting or underfitting. That is, the candidate model has not learned the differential information of the samples within the specific value range of that dimension, resulting in similar predictions from samples with the same value, even though their actual labels may differ. Therefore, for the first special sample dimension, when obtaining relevant dimensions, the focus is on finding fine-grained correlation dimensions that can distinguish samples with the same value, prioritizing business logic correlations and derived correlations.

[0115] For example, in battery health prediction scenarios, if the prediction accuracy of samples with more than 500 fast charging cycles is below the second threshold, then the average fast charging temperature or the generated fast charging frequency is introduced as a relevant dimension. The average fast charging temperature and fast charging frequency still show significant differences within the sample group with more than 500 fast charging cycles, which can help candidate models distinguish the different impacts of high-frequency but low-temperature fast charging and high-frequency and high-temperature fast charging on battery health.

[0116] As can be seen from the above technical solutions, the index prediction method provided in this application can identify the key dimension that causes the deviation by discovering the phenomenon of inaccurate prediction of samples with the same value, and supplement the dimensions related to the key dimension, effectively solving the problem of low prediction accuracy of candidate models in a specific value range, and improving the prediction performance of the model in local areas.

[0117] In one alternative embodiment, Figure 5 This is a flowchart illustrating another indicator prediction method provided in this application. Figure 5 This illustrates another specific implementation of S205, namely, updating the training dimension set at least based on the prediction accuracy in response to a prediction accuracy not exceeding a first threshold. Figure 5 As shown, this method includes:

[0118] S501. Obtain the contribution parameters of each sample dimension in the training dimension set.

[0119] In this embodiment, the contribution parameter is used to measure the degree of contribution of each sample dimension to the prediction result of the candidate model. The contribution parameter can be calculated using at least one of the following methods: SHAP value, random forest feature importance, or permutation importance. By calculating the contribution parameter, high-contribution dimensions that the candidate model over-relies on and low-contribution dimensions that do not provide effective information can be identified.

[0120] S502. Based on the contribution parameter, determine the second special sample dimension from the training dimension set.

[0121] In this embodiment, the second special sample dimension includes sample dimensions that meet the second special condition, which includes sample dimensions whose contribution parameter is higher than the third threshold.

[0122] It should be noted that when the contribution parameter of a sample dimension exceeds the third threshold, it indicates that the candidate model is overly reliant on that sample dimension. Over-reliance on a single dimension can easily lead to overfitting and poor prediction robustness when information from other dimensions is insufficient. In such cases, it is necessary to supplement other dimensions related to that sample dimension to balance the attention of the candidate model.

[0123] S503. Obtain the dimensions related to the first special sample dimension, add them to the training dimension set, and obtain the updated training dimension set.

[0124] In this embodiment, the specific methods for obtaining the dimension related to the second special sample dimension include: screening candidate dimensions from the candidate dimension pool that meet the correlation condition with the second special sample dimension, or directly generating a derived dimension that has at least one correlation relationship with the second special sample dimension.

[0125] Preferably, when obtaining relevant dimensions for the second special sample dimension, the focus is on finding alternative or supplementary dimensions that can balance the model's attention, prioritizing statistical correlation and business logic association. The obtained relevant dimensions should have a certain correlation with the second special sample dimension while providing independent information, thereby sharing the prediction task and reducing the candidate model's dependence on a single dimension. For example, in a battery health prediction scenario, if the contribution parameter of cycle count is found to be higher than the third threshold, then the average depth of discharge or charging cut-off voltage is introduced as a relevant dimension. The average depth of discharge and charging cut-off voltage are correlated with the cycle count, sharing the prediction task, and containing information independent of the cycle count, thereby reducing the candidate model's excessive dependence on the cycle count.

[0126] S504. Based on the contribution parameter, determine the third special sample dimension from the training dimension set.

[0127] In this embodiment, the third special sample dimension includes sample dimensions that meet the third special condition. The third special condition includes sample dimensions whose contribution parameter is lower than the fourth threshold, wherein the third threshold is higher than the fourth threshold.

[0128] It should be noted that when the contribution parameter of a certain sample dimension is lower than the fourth threshold, it indicates that the sample dimension does not provide effective information to the candidate model. Reasons for a low contribution parameter include an unreasonable feature representation for that sample dimension, such as non-linear relationships in continuous numerical features, or inappropriate encoding of categorical features.

[0129] S505. The dimension obtained by transforming the dimension of the third special sample is added to the training dimension set to obtain the updated training dimension set.

[0130] In this embodiment, the transformation includes binning discretization.

[0131] In this embodiment, transforming the third special sample dimension can make its relationship with the indicator to be predicted more closely match the fitting ability of the candidate model, thereby transforming invalid features into valid features. Transformation methods include binning discretization or embedding vectorization.

[0132] It should be noted that binning discretization is suitable for continuous numerical features, dividing continuous values ​​into discrete categories according to a preset interval. In other optional embodiments, the transformation also includes embedding vectorization, which is suitable for high cardinality categorical features, mapping discrete categories to dense vector representations.

[0133] As can be seen from the above technical solutions, the indicator prediction method provided in this application identifies high contribution dimensions and low contribution dimensions by using contribution parameters, and adopts two different processing strategies, namely supplementing relevant dimensions and feature transformation, which can effectively solve the problem of weak generalization ability caused by the model's over-reliance on a single dimension. At the same time, it transforms invalid features into valid features, thereby comprehensively improving the model's prediction performance and generalization ability.

[0134] For example, in the scenario of battery health prediction, the contribution parameters of each sample dimension to the prediction residual are calculated using SHAP values.

[0135] For handling high-contribution dimensions: If the contribution parameter of the number of cycles is higher than the third threshold, such as higher than 0.3, it indicates that the candidate model is overly dependent on the number of cycles dimension. In this case, we can select dimensions that are highly correlated with the number of cycles from the candidate dimension pool, such as average depth of discharge or charging cut-off voltage, or directly generate derived dimensions related to the number of cycles and add the above dimensions to the training dimension set to balance the candidate model's dependence on a single dimension.

[0136] For the low-contribution dimension: the contribution parameter of usage duration is 0.01, which is lower than the fourth threshold of 0.05, indicating that the usage duration dimension does not provide effective information to the candidate model. The original values ​​of usage duration are continuous numerical values, such as 0.5 years, 1.2 years, 2.8 years, 4.5 years, etc. The usage duration is binned and discretized into intervals of less than 1 year, 1 to 2 years, 2 to 3 years, and more than 3 years. The transformed dimension is added to the training dimension set. Battery aging is relatively rapid in the early stage (less than 1 year), linearly decreasing in the middle stage (1 to 3 years), and accelerating in the later stage (more than 3 years). After binning, it is more in line with the physical law of battery aging, thus transforming the originally ineffective usage duration dimension into an effective feature.

[0137] As can be seen from the above embodiments, in this embodiment, updating the training dimension set based on prediction accuracy includes at least the following four optional methods:

[0138] The first approach is to increase the correlation threshold and select candidate dimensions from the candidate dimension pool that are highly correlated with the metric to be predicted, and then add them to the training dimension set.

[0139] The second approach is to select candidate dimensions that are highly correlated with the predicted residual vector from the candidate dimension pool and add them to the training dimension set based on the predicted residual.

[0140] The third approach is to identify the first special sample dimension whose prediction accuracy in a subset of samples is lower than the second threshold, and then add the dimensions related to this first special sample dimension to the training dimension set.

[0141] The fourth method involves supplementing relevant dimensions for high-contribution dimensions (those with contribution values ​​above the third threshold) based on contribution parameters, and transforming and adding low-contribution dimensions (those with contribution values ​​below the fourth threshold) to the training dimension set.

[0142] The first and second methods only filter existing dimensions from the candidate dimension pool, consuming fewer computational resources. The third and fourth methods, in addition to filtering from the candidate dimension pool, can directly generate derived dimensions or transform existing dimensions, and are not limited to the candidate dimension pool. In an optional embodiment, the above methods can be combined. For example, when the prediction accuracy is low, the first or second method is preferentially used to quickly expand effective features; when the prediction accuracy improves, the third or fourth method is switched for fine-tuning.

[0143] In one optional embodiment, in response to a prediction accuracy exceeding a first threshold, the candidate model is determined as the target prediction model, and the training dimension set is determined as the target dimension set corresponding to the target prediction model. Then, the correspondence between the target prediction model and the corresponding target dimension set is recorded. Thus, multiple target prediction models and their corresponding target dimension sets corresponding to different prediction accuracies (first threshold) can be pre-trained. For example, target prediction models and their corresponding target dimension sets with prediction accuracies of 80%, 85%, 90%, and 95% can be obtained through training, with each target dimension set including at least one target dimension.

[0144] Based on this, obtaining the target feature data of the target object includes: obtaining a target prediction model and a corresponding set of target dimensions that match the first target threshold. The first target threshold can be preset according to current business needs. Then, based on this set of target dimensions, the feature values ​​of each target dimension of the target object are obtained, thus acquiring the target feature data. During online prediction, a suitable set of target dimensions corresponding to the appropriate threshold can be selected as the actual set of target dimensions to be used, based on currently available computing resources or accessible data resources, thereby flexibly adapting to different application scenarios.

[0145] In one optional embodiment, after the first time period has elapsed or the model update condition has been met, sample data and their labels within a third time window are used as a new test set, wherein the third time window is later than the second time window. The sample data within the third time window is obtained based on the target training dimension set corresponding to the target prediction model; that is, the sample data includes the feature values ​​of the sample dimensions within each target training dimension set, and is used to test the target prediction model that has been trained and executed on an online prediction task.

[0146] If the prediction accuracy of the target prediction model on the new test set falls below a first threshold, model retraining or feature re-selection is triggered. This mechanism enables the target prediction model to adapt to changes in data distribution over time, ensuring the model's prediction performance in long-term operation.

[0147] The above describes an indicator prediction method provided by the embodiments of this application. The following will describe the related apparatus for performing the above indicator prediction method.

[0148] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of an indicator prediction device provided in an embodiment of this application. Figure 6 As shown, the indicator prediction device 600 includes:

[0149] The feature data acquisition unit 601 is used to acquire target feature data of the target object; the target feature data includes feature values ​​of each target dimension in the target dimension set, and the target dimension set includes at least one target dimension.

[0150] The model prediction unit 602 is used to input the target feature data into the target prediction model to obtain the predicted value of the target object's indicator output by the target prediction model;

[0151] The target prediction model is trained using sample data within a first time window as the training set and sample data within a second time window as the test set, and the prediction accuracy of the target prediction model on the test set is higher than the first target threshold; the second time window is later than the first time window.

[0152] In one possible implementation, a model training unit is also included for performing at least one round of training to train the target prediction model, wherein the model training unit, when performing one round of training, is specifically used for:

[0153] Obtain a set of training dimensions, wherein the set of training dimensions includes at least one sample dimension;

[0154] A training set is constructed using sample data within a first time window, and a test set is constructed using sample data within a second time window. The sample data includes feature values ​​for each sample dimension.

[0155] Candidate models are trained based on the training set, and the candidate models are tested based on the test set to obtain the prediction accuracy.

[0156] In response to the prediction accuracy being higher than the first threshold, the candidate model is determined to be the target prediction model, and the training dimension set is determined to be the target dimension set corresponding to the target prediction model;

[0157] In response to the prediction accuracy not being higher than the first threshold, the training dimension set is updated at least based on the prediction accuracy.

[0158] In one possible implementation, the model training unit is used to obtain the set of training dimensions in the first round of training, specifically for:

[0159] The candidate dimensions that satisfy the first correlation condition with the metric to be predicted are selected from the candidate dimension pool to obtain the training dimension set;

[0160] The training dimension set includes at least one sample dimension; the candidate dimension pool includes multiple candidate dimensions, and the first correlation condition includes a correlation higher than a first correlation threshold.

[0161] In one possible implementation, the model training unit is used to obtain the training dimension set during non-first rounds of training, specifically for:

[0162] Obtain the set of training dimensions obtained from the previous training update.

[0163] In one possible implementation, when the model training unit is used to update the training dimension set at least based on the prediction accuracy, it is specifically used for:

[0164] Lower the first correlation threshold, and select candidate dimensions from the candidate dimension pool that satisfy the first correlation condition with the indicator to be predicted, to obtain an updated training dimension set.

[0165] In one possible implementation, when the model training unit is used to update the training dimension set at least based on the prediction accuracy, it is specifically used for:

[0166] Determine the prediction residual vector of the candidate model on the test set after this round of training;

[0167] Calculate the correlation between each candidate dimension in the candidate dimension pool and the predicted residual vector;

[0168] In the candidate dimension pool, at least one candidate dimension that satisfies the second correlation condition is obtained and added to the training dimension set to obtain an updated training dimension set. The second correlation condition includes that the correlation with the predicted residual vector is higher than the second correlation threshold.

[0169] In one possible implementation, when the model training unit is used to update the training dimension set at least based on the prediction accuracy, it is specifically used for:

[0170] From the training dimension set, a first special sample dimension is determined. The first special sample dimension includes sample dimensions that satisfy a first special condition. The first special condition includes that the prediction accuracy of the corresponding sample data subset is lower than a second threshold. The sample data subset corresponding to a sample dimension includes multiple samples with the same feature value of the sample dimension.

[0171] Obtain the dimensions related to the first special sample dimension, add them to the training dimension set, and obtain the updated training dimension set.

[0172] In one possible implementation, when the model training unit is used to update the training dimension set at least based on the prediction accuracy, it is specifically used for:

[0173] Obtain the contribution parameters of each sample dimension in the training dimension set;

[0174] Based on the contribution parameter, a second special sample dimension is determined from the training dimension set. The second special sample dimension includes sample dimensions that meet the second special condition, which includes sample dimensions whose contribution parameter is higher than the third threshold.

[0175] Obtain the dimensions related to the first special sample dimension, add them to the training dimension set, and obtain the updated training dimension set.

[0176] Based on the contribution parameter, a third special sample dimension is determined from the training dimension set. The third special sample dimension includes sample dimensions that meet the third special condition, which includes sample dimensions whose contribution parameter is lower than the fourth threshold. The third threshold is higher than the fourth threshold.

[0177] The dimension obtained by transforming the third special sample dimension;

[0178] Add the training dimension set to obtain the updated training dimension set. The transformation includes binning discretization.

[0179] In one possible implementation, in each round of training, the number of new sample dimensions added to the updated training dimension set is negatively correlated with the prediction accuracy.

[0180] This application also provides an electronic device, which includes at least one processor and a memory connected to the processor, wherein: the memory is used to store a computer program; and the processor is used to execute the computer program so that the electronic device can implement the various steps of the indicator prediction method provided in this application.

[0181] The electronic devices in the embodiments of this application may include, but are not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 7 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0182] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 7As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703. When the electronic device is powered on, the RAM 703 also stores various programs and data required for the operation of the electronic device. The processing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0183] Typically, the following devices can be connected to I / O interface 705: input devices 706 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 707 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 708 including, for example, memory cards, hard drives, etc.; and communication devices 709. Communication device 709 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have instead.

[0184] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the indicator prediction methods provided in this application.

[0185] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the indicator prediction methods provided in this application.

[0186] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0187] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0188] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0189] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for predicting an indicator, comprising: Obtain the target feature data of the target object; The target feature data includes the feature values ​​of each target dimension in the target dimension set, and the target dimension set includes at least one target dimension. The target feature data is input into the target prediction model to obtain the predicted value of the target object's indicator output by the target prediction model; The target prediction model is trained using sample data within a first time window as the training set and sample data within a second time window as the test set, and the prediction accuracy of the target prediction model on the test set is higher than the first target threshold. The second time window is later than the first time window.

2. The indicator prediction method according to claim 1, wherein the training method of the target prediction model includes at least one round of training, and one round of training includes: Obtain a set of training dimensions, wherein the set of training dimensions includes at least one sample dimension; A training set is constructed using sample data within a first time window, and a test set is constructed using sample data within a second time window. The sample data includes feature values ​​for each sample dimension. Candidate models are trained based on the training set, and the candidate models are tested based on the test set to obtain the prediction accuracy. In response to the prediction accuracy being higher than the first threshold, the candidate model is determined to be the target prediction model, and the training dimension set is determined to be the target dimension set corresponding to the target prediction model; In response to the prediction accuracy not being higher than the first threshold, the training dimension set is updated at least based on the prediction accuracy.

3. According to the indicator prediction method of claim 2, in the first round of training, the set of training dimensions obtained includes: The candidate dimensions that satisfy the first correlation condition with the metric to be predicted are selected from the candidate dimension pool to obtain the training dimension set; The training dimension set includes at least one sample dimension; the candidate dimension pool includes multiple candidate dimensions, and the first correlation condition includes a correlation higher than a first correlation threshold.

4. In the indicator prediction method according to claim 3, the set of training dimensions obtained in non-first rounds of training includes: Obtain the set of training dimensions obtained from the previous training update.

5. The indicator prediction method according to claim 4, wherein updating the training dimension set based at least on the prediction accuracy includes: Lower the first correlation threshold, and select candidate dimensions from the candidate dimension pool that satisfy the first correlation condition with the indicator to be predicted, to obtain an updated training dimension set.

6. The indicator prediction method according to claim 4, wherein updating the training dimension set based at least on the prediction accuracy includes: Determine the prediction residual vector of the candidate model on the test set after this round of training; Calculate the correlation between each candidate dimension in the candidate dimension pool and the predicted residual vector; In the candidate dimension pool, at least one candidate dimension that satisfies the second correlation condition is obtained and added to the training dimension set to obtain an updated training dimension set. The second correlation condition includes that the correlation with the predicted residual vector is higher than the second correlation threshold.

7. The indicator prediction method according to claim 4, wherein updating the training dimension set based at least on the prediction accuracy includes: From the training dimension set, a first special sample dimension is determined. The first special sample dimension includes sample dimensions that satisfy a first special condition. The first special condition includes that the prediction accuracy of the corresponding sample data subset is lower than a second threshold. The sample data subset corresponding to a sample dimension includes multiple samples with the same feature value of the sample dimension. Obtain the dimensions related to the first special sample dimension, add them to the training dimension set, and obtain the updated training dimension set.

8. The indicator prediction method according to claim 4, wherein updating the training dimension set based at least on the prediction accuracy includes: Obtain the contribution parameters of each sample dimension in the training dimension set; Based on the contribution parameter, a second special sample dimension is determined from the training dimension set. The second special sample dimension includes sample dimensions that meet the second special condition, which includes sample dimensions whose contribution parameter is higher than the third threshold. Obtain the dimensions related to the first special sample dimension, add them to the training dimension set, and obtain the updated training dimension set; Based on the contribution parameter, a third special sample dimension is determined from the training dimension set. The third special sample dimension includes sample dimensions that meet the third special condition, which includes sample dimensions whose contribution parameter is lower than the fourth threshold. The third threshold is higher than the fourth threshold. The dimension obtained by transforming the third special sample dimension; Add the training dimension set to obtain the updated training dimension set. The transformation includes binning discretization.

9. In the indicator prediction method according to any one of claims 5 to 8, in each round of training, the number of newly added sample dimensions in the updated training dimension set is negatively correlated with the prediction accuracy.

10. An indicator prediction device, comprising: Feature data acquisition unit, used to obtain target feature data of target object; The target feature data includes the feature values ​​of each target dimension in the target dimension set, and the target dimension set includes at least one target dimension. The model prediction unit is used to input the target feature data into the target prediction model to obtain the predicted value of the target object's indicator output by the target prediction model; The target prediction model is trained using sample data within a first time window as the training set and sample data within a second time window as the test set, and the prediction accuracy of the target prediction model on the test set is higher than the first target threshold. The second time window is later than the first time window.

Images