A health status assessment method for lithium batteries used in V2G electric vehicles

Abstract

This invention provides a health status assessment method for lithium batteries used in V2G electric vehicles. Based on big data from real new energy vehicles, it sets characterization parameters to represent the state of health (SOH) of lithium batteries and calculates their reference values. At the same time, it extracts multiple operational characteristic parameters that are highly correlated with the characterization parameters to jointly construct a training sample set. Furthermore, it innovatively introduces the NGBoost model to establish an SOH estimation framework, making the trained model significantly superior to existing machine learning methods in terms of both accuracy and robustness in SOH estimation.

Description

Technical Field

[0001] This invention belongs to the technical field of health status estimation of lithium-ion batteries for new energy vehicles, and specifically relates to a health status assessment method for lithium batteries used in V2G electric vehicles. Background Technology

[0002] Currently, there is a lack of effective direct measurement methods for State of Harmony (SOH), a crucial indicator reflecting the aging degree of lithium-ion batteries. Characterization relies on parameters such as usable capacity, internal resistance, and voltage. However, V2G (Vehicle-to-Grid) electric vehicles differ from ordinary electric vehicles in that they can input surplus energy into the grid, enabling bidirectional energy flow. Therefore, the batteries of these vehicles are charged and discharged more frequently, potentially accelerating the aging process and making precise monitoring of SOH even more urgent. Existing SOH estimation methods mainly include: ① direct measurement methods that calculate the maximum usable capacity by directly measuring the battery and using ampere-hour integration; ② model-based methods that establish an electrochemical model or equivalent circuit model of the battery and identify parameters; and ③ data-driven methods that collect real-vehicle big data for machine learning. However, the existing SOH estimation methods mentioned above still have many drawbacks. For example, method ① requires a long measurement cycle and high equipment requirements, and is only suitable for laboratory or other ideal stable environments, but not for actual vehicle use conditions; method ② has extremely high requirements for the realism of the model and parameter tuning, and is not conducive to application in real vehicles; the accuracy of method ③ depends heavily on massive amounts of data, and existing machine learning algorithms still have considerable room for improvement. Summary of the Invention

[0003] In view of this, and to address the technical problems existing in this field, the present invention provides a health status assessment method for lithium batteries used in V2G electric vehicles, specifically including the following steps:

[0004] Step 1: Collect raw data on the operation of each V2G electric vehicle within a certain period, including mileage, voltage, current, SOC, etc., and upload them to the big data platform;

[0005] Step 2: The big data platform performs the following steps on the raw data uploaded by each vehicle in sequence: data cleaning, removing abnormal data and interpolating missing data; data slicing, dividing the raw data into different types according to the vehicle's operating status: charging status, power supply to the grid status, and driving status; and data reconstruction, reconstructing data segments with the same operating status to form an operating segment dataset.

[0006] Step 3: Select characterization parameters for SOH, such as maximum available capacity, internal resistance, and voltage, and calculate reference values ​​for the characterization parameters corresponding to each segment using the data from the charging and power supply segments to the grid; based on the correlation with the reference values ​​of the characterization parameters, select some operating parameters from the operating segment dataset as feature parameters, and construct training, validation, and test sets for training the NGBoost model together with the reference values ​​of the characterization parameters.

[0007] Step 4: Using the reference values ​​of the characterization parameters as output and the feature parameters as input, establish an SOH evaluation model based on the NGBoost model. Train the NGBoost model using the training set, and verify and test the training results using the validation set and test set.

[0008] Step 5: Apply the trained SOH estimation model online, collect the characteristic parameters of the actual vehicle operation and input them into the NGBoost model to estimate the corresponding characterization parameter values, and then convert them to obtain the current SOH evaluation result of the vehicle.

[0009] Furthermore, in step three, the maximum available capacity C is specifically selected. t The characterization parameter is specifically calculated based on the following continuous-time inverse ampere-hour integral:

[0010]

[0011] In the formula, SOC t Represents the SOC value at time t. It is the initial SOC value of the battery, i L (τ) is the charging or discharging current at time τ. Let t0 and t represent the depth of charge and depth of discharge, respectively, and ΔQ represent the change in charge during a single charge or power supply to the grid. The length of the SOC interval;

[0012] Alternatively, the following discrete-time inverse ampere-hour integral method can be used:

[0013]

[0014] In the formula, L w Given the length of a time period, Δt is the current sampling interval, and k is the discrete time.

[0015] To avoid characterization parameter C t To maintain stability amidst drastic fluctuations, the following constraints are set for the sampled runtime data:

[0016] ΔSOC≥20%

[0017] 95% ≥ SOC i≥15%, i=1,2,…,n

[0018] Where n is the sample data length of the operation segment that is charging or discharging into the grid;

[0019] For C t The calculation results are obtained using the following convergence coefficient δ L To evaluate its reliability:

[0020]

[0021] Where, ε L This represents an acceptable range of convergence.

[0022] Furthermore, the feature parameters used to train the NGBoost model are selected based on their correlation with the reference values ​​of the characterization parameters, specifically obtained using Pearson correlation analysis.

[0023] Furthermore, based on correlation, specific parameters were selected, including total mileage, number of charging and discharging to the grid (NCGFC), SOC interval length ΔSOC, mean SOC (SOCmean), and average current I. mean The feature parameters are used as the basis for establishing the training set, validation set, and test set after normalization.

[0024] Furthermore, the NGBoost model described in step four consists of a base learner f (m) Probability distribution P ω (y|x) and scoring rules It consists of three parts, where x is the feature vector of the observation and its corresponding feature parameters, y is the target variable and its corresponding characterization parameters, and ω is the parameter vector of the distribution. For example, if a normal distribution is used, then ω is a parameter vector that includes the mean and standard deviation.

[0025] In the model, the base learner f (m) Specifically, a shallow decision tree is chosen, with probability distribution P. ω (y|x) is selected from a Gaussian distribution, and the scoring rule is based on the negative log-likelihood to establish the scoring function: The specific training process of the NGBoost model is as follows:

[0026] For the training dataset First, estimate a common ω. (0) To minimize the sum of the scoring rules corresponding to all training samples, i.e.:

[0027]

[0028] Where n is the total number of training samples;

[0029] Then, in each decision tree m, i.e. the m-th stage, for each sample data i, the natural gradient from i to the tree, related to the estimated parameters, is calculated.

[0030]

[0031] in, For scoring rules The gradient with respect to ω It is the probability distribution P carried by y. ω The Fisher information is specifically represented as follows:

[0032]

[0033] in, This is the expected value. Once determined, it is used in the m-th stage with the input vector x i Together fit the set of base learners f (m) ;

[0034] Then the ω corresponding to sample data i i (m) The update method is as follows:

[0035]

[0036] Where η is the learning rate, ρ (m) The scaling factor is calculated as follows:

[0037]

[0038] Where ρ is the scaling factor for the initial stage, and ρ = 1 is taken;

[0039] The NGBoost model is based on a gradient boosting framework, which includes the following steps:

[0040] Step 1: Initialize a base model with a constant value (such as the mean of the training set);

[0041] Step 2: Perform each iteration:

[0042] ① Calculate the target variable minus the prediction of the previous model as the residual for the current stage;

[0043] ② Introduce a base model with residuals as the target variable;

[0044] ③ Minimize the negative gradient of the loss function using gradient descent to obtain the optimal fit of the new model;

[0045] ④ Add the new model to the integration;

[0046] Step 3: Weighted fusion of all models to obtain the final model, where the weights of each model are determined based on cross-validation.

[0047] The training performance of NGBoost models can be evaluated using MAPE and RMSE as standards:

[0048]

[0049]

[0050] Among them, y j These are the actual reference values ​​for the characterization parameters. These are the estimation results from the NGBoost model.

[0051] The health status assessment method for lithium batteries used in V2G electric vehicles provided by the present invention sets up characterization parameters for lithium battery SOH based on big data from real new energy vehicles and calculates their reference values. At the same time, it extracts a variety of operating characteristic parameters that are highly correlated with the characterization parameters to jointly construct a training sample set. It also innovatively introduces the NGBoost model to establish a characterization parameter estimation framework, so that the trained model has significantly better accuracy and robustness in SOH estimation than existing machine learning methods. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the method provided by the present invention;

[0053] Figure 2 A schematic diagram illustrating the use of Pearson correlation analysis to screen operational characteristic parameters;

[0054] Figure 3 A flowchart illustrating the gradient boosting algorithm during NGBoost model training.

[0055] Figure 4 This is a comparison chart of the capacity estimation results and the actual values ​​in an example of the present invention;

[0056] Figure 5 This is a comparison chart showing the accuracy of the method of the present invention with other existing technologies based on machine learning. Detailed Implementation

[0057] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] The present invention provides a health status assessment method for lithium batteries used in V2G electric vehicles, such as... Figure 1 As shown, the specific steps include:

[0059] Step 1: Collect raw data on the operation of each V2G electric vehicle within a certain period, including mileage, voltage, current, SOC, etc., and upload them to the big data platform;

[0060] Step 2: The big data platform performs the following steps on the raw data uploaded by each vehicle in sequence: data cleaning, removing abnormal data and interpolating missing data; data slicing, dividing the raw data into different types according to the vehicle's operating status: charging status, power supply to the grid status, and driving status; and data reconstruction, reconstructing data segments with the same operating status to form an operating segment dataset.

[0061] Step 3: Select characterization parameters for SOH, such as maximum available capacity, internal resistance, and voltage, and calculate reference values ​​for the characterization parameters corresponding to each segment using the data related to the charging and power supply segments to the grid; based on the correlation with the reference values ​​of the characterization parameters, select some operating parameters from the operating segment dataset as feature parameters, and construct training, validation, and test sets for training the NGBoost model together with the reference values ​​of the characterization parameters.

[0062] Step 4: Using the reference values ​​of the characterization parameters as output and the feature parameters as input, establish an SOH evaluation model based on the NGBoost model. Train the NGBoost model using the training set, and verify and test the training results using the validation set and test set.

[0063] Step 5: Apply the trained SOH estimation model online, collect the characteristic parameters of the actual vehicle operation and input them into the NGBoost model to estimate the corresponding characterization parameter values, and then convert them to obtain the current SOH evaluation result of the vehicle.

[0064] In step three, specifically select the maximum available capacity C. t The characterization parameter is specifically calculated based on the following continuous-time inverse ampere-hour integral:

[0065]

[0066] In the formula, SOC t Represents the SOC value at time t. It is the initial SOC value of the battery, i L (τ) is the charging or discharging current at time τ. Let t0 and t represent the depth of charge and depth of discharge, respectively, and ΔQ represent the change in charge during a single charge or power supply to the grid. The length of the SOC interval;

[0067] Alternatively, the following discrete-time inverse ampere-hour integral method can be used:

[0068]

[0069] In the formula, L w Given the length of a time period, Δt is the current sampling interval, and k is the discrete time.

[0070] To avoid characterization parameter C t To maintain stability amidst drastic fluctuations, the following constraints are set for the sampled runtime data:

[0071] ΔSOC≥20%

[0072] 95% ≥ SOC i ≥15%, i=1,2,…,n

[0073] Where n is the sample data length of the operation segment that is charging or feeding power to the grid;

[0074] For C t The calculation results are obtained using the following convergence coefficient δ L To evaluate its reliability:

[0075]

[0076] Where, ε L This represents an acceptable range of convergence.

[0077] Feature parameters used for training the NGBoost model are selected based on their correlation with reference values ​​of characterization parameters. Specifically, these parameters are obtained by performing correlation analysis using experimental data, empirical data, or the Pearson correlation coefficients of each parameter. Figure 2 The data shows the total mileage, number of charging and discharging to the grid (NCGFC), SOC interval length ΔSOC, mean SOC (SOCmean), and average current I. mean The results of calculating the Pearson correlation coefficients for each feature parameter are then used to establish the training, validation, and test sets. These feature parameters are first normalized before being used to build the training, validation, and test sets.

[0078] The NGBoost model described in step four consists of a base learner f (m) Probability distribution P ω (y|x) and scoring rules It consists of three parts: x is the feature vector of the observation and its corresponding feature parameters, y is the target variable and its corresponding characterization parameters, and ω is the parameter vector of the distribution. For example, if a normal distribution is used, then ω is a parameter vector that includes the mean and standard deviation.

[0079] In the model, the base learner f (m) Specifically, a shallow decision tree is chosen, with probability distribution P. ω (y|x) is selected from a Gaussian distribution, and the scoring rule is based on the negative log-likelihood to establish the scoring function: The specific training process of the NGBoost model is as follows:

[0080] For the training dataset First, estimate a common ω. (0) To minimize the sum of the scoring rules corresponding to all training samples, i.e.:

[0081]

[0082] Where n is the total number of training samples;

[0083] Then, in each decision tree m, i.e. the m-th stage, for each sample data i, the natural gradient from i to the tree, related to the estimated parameters, is calculated.

[0084]

[0085] in, For scoring rules The gradient with respect to ω It is the probability distribution P carried by y. ω The Fisher information is specifically represented as follows:

[0086]

[0087] in, This is the expected value. Once determined, it is used in the m-th stage with the input vector x i Together fit the set of base learners f (m) ;

[0088] Then the ω corresponding to sample data i i (m) The update method is as follows:

[0089]

[0090] Where η is the learning rate, ρ (m) The scaling factor is calculated as follows:

[0091]

[0092] Where ρ is the scaling factor for the initial stage, and ρ = 1 is taken;

[0093] The NGBoost model is based on the gradient boosting framework, which is as follows: Figure 3 As shown, it includes:

[0094] Step 1: Initialize a base model with a constant value (such as the mean of the training set);

[0095] Step 2: Perform each iteration:

[0096] ① Calculate the target variable minus the prediction of the previous model as the residual for the current stage;

[0097] ② Introduce a base model with residuals as the target variable;

[0098] ③ Minimize the negative gradient of the loss function using gradient descent to obtain the optimal fit of the new model;

[0099] ④ Add the new model to the integration;

[0100] Step 3: Weighted fusion of all models to obtain the final model, where the weights of each model are determined based on cross-validation.

[0101] The training performance of NGBoost models can be evaluated using MAPE and RMSE as standards:

[0102]

[0103]

[0104] Among them, y j This is a true reference value for the maximum available capacity. These are the estimation results from the NGBoost model.

[0105] In a specific example of the present invention, the above method is implemented for V2G vehicles as shown in Table 1:

[0106] Table 1 V2G Vehicle Specifications

[0107]

[0108] The raw data was collected over a period of 19 months, with a collection interval of 30 seconds, covering all four seasons and a cumulative mileage of 6193.5 kilometers. During this period, the selected V2G vehicles performed 168 charging operations and fed power to the grid 113 times, reaching a stable degradation stage. After cleaning, segmenting, and reconstructing the raw data, the maximum usable capacity reference values ​​for each charging and feeding segment were calculated. Pearson correlation analysis was used to extract the following parameters: total mileage, number of charging and grid feeding operations (NCGFC), SOC interval length (ΔSOC), and mean SOC (SOC value). mean ) and average current (I meanThe feature parameters are used to construct the training sample set, with the training and validation sets accounting for 80% and the test set accounting for 20%. After model training, MAPE and RMSE are used as indicators to evaluate model accuracy. Figure 4 The comparison between the estimated capacity of each segment and the true value shown demonstrates the high accuracy achievable by this invention. To further verify the beneficial effects of this invention, the applicant also simultaneously used linear regression (LR), k-nearest neighbors (KNN), support vector machine regression (SVR), random forest regression (RFR), adaptive boosting (AdaBoost), gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), classification boosting (CatBoost), and light gradient boosting machine (LGBM) algorithms, trained and tested on the same dataset. The final trained models achieved the following estimation accuracy for SOH: Figure 5 As shown in the figure, it can also be seen that the SOH estimation method based on NGBoost in this application has a significant accuracy advantage.

[0109] It should be understood that the sequence number of each step in the embodiments of the present invention does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0110] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the health status of lithium batteries used in V2G electric vehicles, characterized in that: Specifically, the following steps are included: Step 1: Collect raw data on the operation of each V2G electric vehicle within a certain period and upload it to the big data platform; Step 2: The big data platform performs the following steps on the raw data uploaded by each vehicle in sequence: data cleaning, removing abnormal data and interpolating missing data; data slicing, dividing the raw data into different types according to the vehicle's operating status: charging status, power supply to the grid status, and driving status. Data reconstruction involves reconstructing data fragments with the same running state and forming a running fragment dataset. Step 3: Select the characterization parameters for SOH and calculate the reference values ​​of the characterization parameters for each segment using the data from the charging segment and the power supply segment to the grid; Based on the correlation with the reference values ​​of the characterization parameters, some running parameters are selected from the running fragment dataset as feature parameters, and together with the reference values ​​of the characterization parameters, they are used to construct the training set, validation set and test set for training the NGBoost model; Step 4: Using the reference values ​​of the characterization parameters as output and the feature parameters as input, establish an SOH evaluation model based on the NGBoost model. Train the NGBoost model using the training set, and verify and test the training results using the validation set and test set. Step 5: Apply the trained SOH evaluation model online, collect the characteristic parameters of the actual vehicle operation and input them into the NGBoost model to estimate the corresponding characterization parameter values, and then convert them to obtain the current SOH evaluation result of the vehicle. The NGBoost model described in step four consists of base learners. probability distribution and scoring rules It consists of three parts, among which x The observed feature vectors correspond to each feature parameter. y The target variable corresponds to the representation parameter. ω Let be the parameter vector of the distribution, and using the normal distribution, then ω It is a parameter vector that includes the mean and standard deviation; Base learners in the model Specifically, a shallow decision tree is selected, with a probability distribution. A Gaussian distribution is chosen, and the scoring rule is based on the negative log-likelihood to establish the scoring function: The specific training process of the NGBoost model is as follows: For the training dataset First, estimate a common one. ω (0) To minimize the sum of the scoring rules corresponding to all training samples, i.e.: in, n The total number of training samples; Then, in each decision tree m That is, the first m At each stage, for each sample data Calculate its natural gradient to the tree, which is related to the estimated parameters. : in, ▽ ω For scoring rules about ω gradient, yes y Carrying about the probability distribution The Fisher information is specifically represented as follows: in, For the expected value, in Once determined, it was used in the first m Each stage and input vector Together fit the set of base learners f (m) ; Then sample data i corresponding The update method is as follows: in, η For learning rate, The scaling factor is calculated as follows: in, As the scaling factor for the initial stage, take =1; The NGBoost model is based on a gradient boosting framework, which includes the following steps: Step 1: Initialize a base model with constant values; Step 2: Perform each iteration: ① Calculate the target variable minus the prediction of the previous model as the residual for the current stage; ② Introduce a base model with residuals as the target variable; ③ Minimize the negative gradient of the loss function using gradient descent to obtain the optimal fit of the new model; ④ Add the new model to the integration; Step 3: Weighted fusion of all models to obtain the final model, where the weights of each model are determined based on cross-validation; The training performance of the NGBoost model is evaluated using MAPE and RMSE as standards: in, These are the actual reference values ​​for the characterization parameters. These are the estimation results from the NGBoost model.

2. The method as described in claim 1, characterized in that: In step three, specifically select the current maximum available capacity. C t As a characterization parameter, it is specifically calculated based on the following continuous-time inverse ampere-hour integral: In the formula, represent SOC value at time t, This is the initial SOC value of the battery. The charging or discharging current at any given moment. , They are respectively Depth of charge and depth of discharge at any given time. The change in charge during a single charging or power supply to the grid. for SOC Interval length; Alternatively, the following discrete-time inverse ampere-hour integral method can be used: In the formula, Given the length of a time period, For current sampling interval, For discrete time intervals; To avoid characterization parameters C t To maintain stability amidst drastic fluctuations, the following constraints are set for the sampled runtime data: D SOC ≥20% 95%≥ SOC i ≥15%， i =1, 2, … , n in, n The length of sample data for an operational segment that is charging or feeding power to the grid; against The calculation results are obtained using the following convergence coefficients. To evaluate its reliability: in, This represents an acceptable range of convergence.

3. The method as described in claim 1, characterized in that: The feature parameters used to train the NGBoost model are selected based on their correlation with the reference values ​​of the characterization parameters, specifically obtained using Pearson correlation analysis.

4. The method as described in claim 3, characterized in that: Based on correlation, specific criteria were selected for total mileage, number of charging sessions, and number of times power was fed back to the grid (NCGFC). SOC Interval length Δ SOC , SOC mean SOC mean and average current I mean These features are used as feature parameters, and after normalization, they are used to establish the training set, validation set, and test set.

Images