New energy automobile battery recession prediction method based on charging condition identification

Abstract

The invention provides a new energy vehicle battery recession prediction method based on charging condition identification, and the method comprises the steps: firstly, recognizing typical vehicle charging conditions through a clustering mode according to the dynamic and variable operation characteristics of a new energy vehicle; then, different health features are extracted in a targeted manner based on a charging condition identification result, and a multi-level feature screening strategy is introduced into historical behavior features, so that a comprehensive and effective feature library is constructed; a data-driven model is established using these feature long short term memory (LSTM) networks for reconstructing historical capacity trajectories and predicting battery degradation. In order to improve the efficiency and accuracy of the method in different scenes, a transfer learning strategy is introduced to correct the constructed prediction model.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy vehicle battery degradation and aging monitoring, and specifically relates to a new energy vehicle battery degradation prediction method based on charging condition identification. Background Art

[0002] Since the power batteries currently used in new energy vehicles are still grouped with a large number of battery cells, it is difficult to obtain the degradation of each cell in the battery pack by direct measurement in actual vehicle use. The solution to the battery degradation problem in the prior art is mainly to perform cycle aging and other operations on the battery cells in a test environment, and estimate the battery capacity or SOH in combination with the capacity increment method, and finally indirectly reflect the overall situation of the power battery through the prediction results of the cells. However, due to the imperfect simulation of the actual working conditions of the vehicle in the test environment, the diverse battery materials and structures also make it difficult for the test results to refer to each other. At the same time, in order to ensure that the types of working conditions involved in the above tests are as comprehensive as possible, the workload, calculation cost and low efficiency cannot be ignored. Summary of the invention

[0003] In view of this, in order to solve the technical problems existing in this field, the present invention provides a new energy vehicle battery degradation prediction method based on charging condition identification, which specifically includes the following steps:

[0004] Step 1: extract the original operation data of the normal operation of a specific vehicle model, perform data cleaning and segmentation processing on the original operation data in sequence, and obtain a charging segment data set and a discharging segment data set respectively;

[0005] For each charging segment in the charging segment data set, the maximum available capacity of the battery corresponding to the segment is determined by calculating the ratio of the capacity integral to the SOC change, and the moving sliding window method and quartile method are used to eliminate abnormal capacity values;

[0006] Step 2: Calculate the average charging rate, average driving SOC change, and average daily driving mileage corresponding to each charging segment and discharging segment, input the calculation results into the Gaussian mixture model for operating condition identification, and obtain various typical charging conditions;

[0007] Step 3: For typical charging conditions, extract real-time health features including charging capacity, as well as historical behavior features involving multiple dimensions such as temperature, charging rate, discharging rate, charging SOC change, and driving SOC change;

[0008] The correlation between each historical behavior feature and capacity decay is evaluated by performing correlation analysis, and the importance of each historical behavior feature is evaluated by importance analysis. The historical behavior features are screened by comprehensively considering the correlation and importance analysis results.

[0009] Integrate real-time health features with filtered historical behavior features and conduct cross-correlation analysis, eliminate features that are dependent on other features, and use the features that are finally retained to establish an effective feature matrix;

[0010] Step 4: Based on the LSTM layer in the deep learning framework Tensorflow, a current capacity estimation model and a future capacity decline prediction model are constructed respectively. Both models use the feature matrix obtained in step 3 as input, and use the current maximum available capacity and the maximum available capacity sequence in a certain period of time in the future as output respectively. The two models are trained until convergence or the predetermined maximum number of training times is reached.

[0011] Step 5: Use transfer learning to correct the trained current capacity estimation model and future capacity decline prediction model in turn; apply the two corrected models online, and use the characteristic parameters collected in real time to predict the current maximum capacity of the target vehicle model and the maximum available capacity decline in a certain period of time in the future.

[0012] Furthermore, in step 1, the maximum available capacity C of the battery corresponding to each charging segment a The specific calculation is based on the following formula:

[0013]

[0014] Where t1 and t2 are the start and end times of the charging segment, I(t) is the total current, and t is the time;

[0015] The following quartile method is used to eliminate outliers in the maximum available capacity calculation results:

[0016]

[0017] Among them, Q1, Q2, and Q3 represent the 1st, 2nd, and 3rd quartiles of capacity data at the same mileage interval, respectively, and IQR is the interquartile range.

[0018] Furthermore, in step 3, a Gaussian mixture model is used to identify typical charging conditions including at least a multi-stage constant current fast charging condition and a constant current constant voltage charging condition;

[0019] When extracting features, the moving sliding window method is specifically used, the window length is set to 10000km, the sliding step is set to 1000km, and the corresponding feature extraction is performed based on the data in each window; when extracting real-time health features, for multi-stage constant current fast charging, the peak and valley voltages of the first and second charging steps, as well as the charging capacity between the two peaks are extracted based on the voltage curve; for constant current and constant voltage charging conditions, the capacity increment curve is calculated and established, and the value and area of ​​the second peak and the charging capacity between the first and second peaks are extracted;

[0020] Perform Pearson correlation analysis or Spearman correlation analysis on the initially extracted historical behavior features, as well as importance analysis based on XGBoost.

[0021] Furthermore, the following input and output matrices are established for the current capacity estimation model and the future capacity decline prediction model using the features finally screened:

[0022] [F 1,k ,F 2,k ,...,F n,k ],[C k ]

[0023]

[0024] Among them, F i,k (i=1,…,n) is the feature extracted from the kth mileage interval, C k is the maximum available capacity of the kth mileage interval, m is the training step size, and l is the prediction step size.

[0025] Furthermore, the specific process of correcting the trained model by transfer learning in step 5 includes:

[0026] When performing the capacity decay prediction task for the target domain (TD), the current capacity estimation model is first modified based on a small amount of labeled data extracted from the TD, and then the historical capacity curve of the vehicles in the TD is quickly reconstructed based on the modified current capacity estimation model; the reconstructed historical capacity curve is then applied to the targeted modification of the future capacity decay model, so that it can more accurately predict the capacity decay trend of vehicles in the TD.

[0027] The new energy vehicle battery degradation prediction method based on charging condition identification provided by the present invention first identifies typical vehicle charging conditions by clustering based on the dynamic and changeable operating characteristics of new energy vehicles. Subsequently, different health features are extracted based on the charging condition identification results, and a multi-level feature screening strategy is introduced for historical behavior features to build a comprehensive and effective feature library. Using these features, a long short-term memory (LSTM) network is used to establish a data-driven model for reconstructing historical capacity trajectories and predicting battery degradation. In order to improve the efficiency and accuracy of the method in different scenarios, the present invention also introduces a transfer learning strategy to correct the constructed prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The figure is an overall framework diagram of the method provided by the present invention. DETAILED DESCRIPTION

[0029] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] The present invention provides a new energy vehicle battery degradation prediction method based on charging condition identification, such as Figure 1 As shown, the specific steps include:

[0031] Step 1: extracting the original operation data of specific models including new energy passenger vehicles and commercial vehicles in normal operation, and performing data cleaning and segmentation processing on the original operation data in turn to obtain a charging segment data set and a discharging segment data set;

[0032] For each charging segment in the charging segment data set, the maximum available capacity of the battery corresponding to the segment is determined by calculating the ratio of the capacity integral to the SOC change, and the moving sliding window method and quartile method are used to eliminate abnormal capacity values;

[0033] Step 2: Calculate the average charging rate, average driving SOC change, and average daily driving mileage corresponding to each charging segment and discharging segment, input the calculation results into the Gaussian mixture model for operating condition identification, and obtain various typical charging conditions;

[0034] Step 3: For typical charging conditions, extract real-time health features including charging capacity, as well as historical behavior features involving multiple dimensions such as temperature, charging rate, discharging rate, charging SOC change, and driving SOC change;

[0035] The correlation between each historical behavior feature and capacity decay is evaluated by performing correlation analysis, and the importance of each historical behavior feature is evaluated by importance analysis. The historical behavior features are screened by comprehensively considering the correlation and importance analysis results.

[0036] Integrate real-time health features with filtered historical behavior features and conduct cross-correlation analysis, eliminate features that are dependent on other features, and use the features that are finally retained to establish an effective feature matrix;

[0037] Step 4: Based on the LSTM layer in the deep learning framework Tensorflow, a current capacity estimation model and a future capacity decline prediction model are constructed respectively. Both models use the feature matrix obtained in step 3 as input, and use the current maximum available capacity and the maximum available capacity sequence in a certain period of time in the future as output respectively. The two models are trained until convergence or the predetermined maximum number of training times is reached.

[0038] Step 5: Use transfer learning to correct the trained current capacity estimation model and future capacity decline prediction model in turn; apply the two corrected models online, and use the characteristic parameters collected in real time to predict the current maximum capacity of the target vehicle model and the maximum available capacity decline in a certain period of time in the future.

[0039] In a preferred embodiment of the present invention, in step 1, the maximum available capacity C of the battery corresponding to each charging segment is a The specific calculation is based on the following formula:

[0040]

[0041] Where t1 and t2 are the start and end times of the charging segment, I(t) is the total current, and t is the time;

[0042] The following quartile method is used to eliminate outliers in the maximum available capacity calculation results:

[0043]

[0044] Among them, Q1, Q2, and Q3 represent the 1st, 2nd, and 3rd quartiles of capacity data at the same mileage interval, respectively, and IQR is the interquartile range.

[0045] In a preferred embodiment of the present invention, in step 3, a Gaussian mixture model is used to identify typical charging conditions including at least a multi-stage constant current fast charging condition and a constant current constant voltage charging condition;

[0046] When extracting features, the moving sliding window method is specifically used, the window length is set to 10000km, the sliding step is set to 1000km, and the corresponding feature extraction is performed based on the data in each window; when extracting real-time health features, for multi-stage constant current fast charging, the peak and valley voltages of the first and second charging steps, as well as the charging capacity between the two peaks are extracted based on the voltage curve; for constant current and constant voltage charging conditions, the capacity increment curve is calculated and established, and the value and area of ​​the second peak and the charging capacity between the first and second peaks are extracted;

[0047] Perform Pearson correlation analysis or Spearman correlation analysis on the initially extracted historical behavior features, as well as importance analysis based on XGBoost.

[0048] In a preferred embodiment of the present invention, the following input and output matrices are established for the current capacity estimation model and the future capacity decline prediction model using the features finally screened:

[0049] [F 1,k ,F 2,k ,...,F n,k ],[C k ]

[0050]

[0051] Among them, F i,k (i=1,…,n) is the feature extracted from the kth mileage interval, C k is the maximum available capacity of the kth mileage interval, m is the training step size, and l is the prediction step size.

[0052] In a preferred embodiment of the present invention, the specific process of correcting the trained model by transfer learning in step 5 includes:

[0053] When performing the capacity decay prediction task for the target domain (TD), the current capacity estimation model is first modified based on a small amount of labeled data extracted from the TD, and then the historical capacity curve of the vehicles in the TD is quickly reconstructed based on the modified current capacity estimation model; the reconstructed historical capacity curve is then applied to the targeted modification of the future capacity decay model, so that it can more accurately predict the capacity decay trend of vehicles in the TD.

[0054] In the specific implementation, the capacity estimation and recession prediction levels were quantitatively verified using two indicators: mean absolute percentage error (MAPE) and root mean square error. The results show that this method is significantly superior to other similar existing technologies in terms of prediction accuracy and computational overhead.

[0055] It should be understood that the size of the serial numbers of the steps in the embodiment of the present invention does not mean 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 embodiment of the present invention.

[0056] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A new energy vehicle battery degradation prediction method based on charging condition identification, characterized by: The specific steps include: Step 1: extract the original operation data of the normal operation of a specific vehicle model, perform data cleaning and segmentation processing on the original operation data in sequence, and obtain a charging segment data set and a discharging segment data set respectively; For each charging segment in the charging segment data set, the maximum available capacity of the battery corresponding to the segment is determined by calculating the ratio of the capacity integral to the SOC change, and the moving sliding window method and quartile method are used to eliminate abnormal capacity values; Step 2: Calculate the average charging rate, average driving SOC change, and average daily driving mileage corresponding to each charging segment and discharging segment, input the calculation results into the Gaussian mixture model for operating condition identification, and obtain various typical charging conditions; Step 3: For typical charging conditions, extract real-time health features including charging capacity, as well as historical behavior features involving temperature, charging rate, discharging rate, charging SOC change, and driving SOC change; The correlation between each historical behavior feature and capacity decay is evaluated by performing correlation analysis, and the importance of each historical behavior feature is evaluated by importance analysis. The historical behavior features are screened by comprehensively considering the correlation and importance analysis results. Integrate real-time health features with filtered historical behavior features and conduct cross-correlation analysis, eliminate features that are dependent on other features, and use the features that are finally retained to establish an effective feature matrix; Step 4: Based on the LSTM layer in the deep learning framework Tensorflow, a current capacity estimation model and a future capacity decline prediction model are constructed respectively. Both models use the feature matrix obtained in step 3 as input, and use the current maximum available capacity and the maximum available capacity sequence in a certain period of time in the future as output respectively. The two models are trained until convergence or the predetermined maximum number of training times is reached. Step 5: Use transfer learning to correct the trained current capacity estimation model and future capacity decline prediction model in turn; apply the two corrected models online, and use the characteristic parameters collected in real time to predict the current maximum capacity of the target vehicle model and the maximum available capacity decline in a certain period of time in the future.

2. The method according to claim 1, characterized in that: The maximum available capacity C of the battery corresponding to each charging segment in step 1 a The specific calculation is based on the following formula: Where t1 and t2 are the start and end times of the charging segment, I(t) is the total current, and t is the time; The following quartile method is used to eliminate outliers in the maximum available capacity calculation results: Among them, Q1, Q2, and Q3 represent the 1st, 2nd, and 3rd quartiles of capacity data at the same mileage interval, respectively, and IQR is the interquartile range.

3. The method according to claim 1, characterized in that: In step 3, a Gaussian mixture model is used to identify typical charging conditions including at least a multi-stage constant current fast charging condition and a constant current constant voltage charging condition; When extracting features, the moving sliding window method is specifically used, the window length is set to 10000km, the sliding step is set to 1000km, and the corresponding feature extraction is performed based on the data in each window; when extracting real-time health features, for multi-stage constant current fast charging, the peak and valley voltages of the first and second charging steps, as well as the charging capacity between the two peaks are extracted based on the voltage curve; for constant current and constant voltage charging conditions, the capacity increment curve is calculated and established, and the value and area of ​​the second peak and the charging capacity between the first and second peaks are extracted; Perform Pearson correlation analysis or Spearman correlation analysis on the initially extracted historical behavior features, as well as importance analysis based on XGBoost.

4. The method according to claim 1, characterized in that: The final screened features are used to establish the following input and output matrices for the current capacity estimation model and the future capacity decline prediction model: [F 1,k ,F 2,k ,...,F n,k ],[C k ] Among them, F i,k (i=1,…,n) is the feature extracted from the kth mileage interval, C k is the maximum available capacity of the kth mileage interval, m is the training step size, and l is the prediction step size.

5. The method according to claim 1, characterized in that: The specific process of correcting the model by transfer learning in step 5 includes: When performing the capacity decline prediction task for the target vehicle group TD, the current capacity estimation model is first modified based on a small amount of labeled data extracted from TD, and then the historical capacity curve of the vehicles in TD is quickly reconstructed based on the modified current capacity estimation model; the reconstructed historical capacity curve is then applied to the targeted modification of the future capacity decline model, so that the predicted capacity decline trend of the vehicles in TD is consistent with the actual situation.

Images