A battery SOH estimation method and device based on wasserstein distance

Abstract

The present invention provides a battery SOH estimation method and device based on the wasserstein distance, and the method includes the following steps: S1, collecting aging data of multiple pouch batteries, and respectively obtaining the IC curve of each aging cycle of each pouch battery; S2, respectively calculating the wasserstein distance between each adjacent IC curve peak window to obtain a feature data set; S3, using a prediction model based on the Gaussian regression process method, for the hyperparameters α, l and p of the prediction model, respectively set their value ranges to (0.1, 100), (0, 1) and (0, 100), first take any value within the value range, and then use the previous data set to perform initial training on the prediction model, and adjust and determine the value of each hyperparameter according to the gap between the predicted value and the true value; S4, using the later data set to train the prediction model with the determined hyperparameters. The present invention does not need to establish a battery model, takes the wasserstein distance as the feature of the prediction model, is suitable for online estimation of SOH, and has high estimation accuracy and strong practicality.

Description

Technical Field

[0001] The invention relates to a battery SOH estimation method and device based on Wasserstein distance. Background Art

[0002] The commonly used SOH (battery health state) is defined as SOH = C now / C0, where C now The maximum available capacity of the battery after it is fully charged at the current moment, and C0 represents the maximum available capacity of the battery when it leaves the factory. Accurate battery health status estimation is a key link in BMS (battery management system). SOH is an important indicator of battery aging. It is generally believed that when this value drops to 70%-80%, it is considered to be the end of life.

[0003] The measurement of SOH is different from direct voltage and current measurement. It can only be estimated and predicted by observing the battery characteristics and combining with the corresponding mathematical model. In the existing technology, prediction can be made through model-driven method, but the model-driven method needs to establish a battery model, which is complex in calculation and not suitable for online application of BMS. Summary of the invention

[0004] The present invention proposes a battery SOH estimation method and device based on the Wasserstein distance, which does not require the establishment of a battery model and uses the Wasserstein distance as a feature of a prediction model. The method is suitable for online estimation of SOH and has high estimation accuracy and strong practicability.

[0005] The present invention is achieved through the following technical solutions:

[0006] A battery SOH estimation method based on wasserstein distance includes the following steps:

[0007] S1. Collecting aging data of multiple pouch batteries, and obtaining IC curves of each aging cycle of each pouch battery, wherein IC represents the battery power increased by a continuous voltage increment;

[0008] S2, respectively calculating the wasserstein distance between each two adjacent IC curve peak windows to obtain a feature data set, and dividing the feature data set into an early data set and a late data set;

[0009] S3, using a prediction model based on the Gaussian regression process method, with the feature data set in S2 as the input of the prediction model, and the battery SOH as the output of the prediction model. The hyperparameters α, l, and p of the prediction model are set to have a value range of (0.1, 100), (0, 1), and (0, 100), respectively. First, any value within the value range is taken, and then the prediction model is initially trained using the previous data set. The value of each hyperparameter is adjusted and determined according to the gap between the predicted value and the true value.

[0010] S4. Use the later dataset to train the prediction model with the determined hyperparameters.

[0011] Furthermore, in S2, each IC curve is sampled using a differential formula to obtain a sampled data set, and Sinkhorn iteration is performed through Pytorch to obtain the wasserstein distance between two adjacent IC curve peak windows.

[0012] Furthermore, in S1, the aging data of the plurality of pouch-type batteries are specifically pre-recorded aging data of eight pouch-type batteries with a nominal capacity of 740 Ma·h.

[0013] Furthermore, in S2, the peak window is an interval with a length of 3 centered at the peak of the IC curve.

[0014] Furthermore, in S3, in the parameter determination module, when the gap between the predicted value and the true value is greater than a first threshold, the value of the hyperparameter α is lowered and the values ​​of the hyperparameters l and p are increased; when the gap between the predicted value and the true value is less than a second threshold, the value of the hyperparameter α is increased and the values ​​of the hyperparameters l and p are lowered. The adjustment amount should be within 10% of the empirical interval, and the empirical interval is selected from the value range of each hyperparameter in step S3.

[0015] Furthermore, in the prediction model, the kernel function of the Gaussian process is a covariance function.

[0016] Furthermore, in S1, the specific calculation formula of IC is: Among them, Q is the battery capacity and U is the battery voltage.

[0017] The present invention is also achieved through the following technical solutions:

[0018] A battery SOH estimation device based on wasserstein distance, comprising:

[0019] Acquisition module: used to collect aging data of multiple pouch batteries and obtain IC curves of each aging cycle of each pouch battery, where IC represents the battery power added in a continuous voltage increment;

[0020] Feature data set acquisition module: used to calculate the wasserstein distance between each two adjacent IC curve peak windows, obtain the feature data set, and divide the feature data set into an early data set and a late data set;

[0021] Parameter determination module: A prediction model based on the Gaussian regression process method is used, with the feature data set as the input of the prediction model and the battery SOH as the output of the prediction model. The hyperparameters α, l and p of the prediction model are set to have a range of (0.1, 100), (0, 1) and (0, 100) respectively. Any value within the range is first taken, and then the prediction model is initially trained using the previous data set. The value of each hyperparameter is adjusted and determined based on the gap between the predicted value and the true value.

[0022] Model training module: Use the later dataset to train the prediction model with determined hyperparameters.

[0023] Furthermore, in the feature data set acquisition module, each IC curve is sampled using a differential formula to obtain a sampled data set, and Sinkhorn iteration is performed through Pytorch to obtain the wasserstein distance between two adjacent IC curve peak windows.

[0024] Furthermore, in the parameter determination module, when the gap between the predicted value and the true value is greater than a first threshold, the value of the hyperparameter α is lowered and the values ​​of the hyperparameters l and p are increased; when the gap between the predicted value and the true value is less than a second threshold, the value of the hyperparameter α is increased and the values ​​of the hyperparameters l and p are lowered. The adjustment amount should be within 10% of the empirical interval, and the empirical interval is selected from the value range of each hyperparameter.

[0025] The present invention has the following beneficial effects:

[0026] 1. The present invention first obtains the IC curve of each aging cycle of each bag battery, and then calculates the wasserstein distance between each adjacent IC curve peak window to obtain a feature data set. The feature data set is divided into an early data set and a late data set. The early data set is used to train the prediction model based on the Gaussian regression process method, and according to the gap between the predicted value and the true value, the value of each hyperparameter is adjusted and determined. Finally, the late data set is used to train the prediction model, and the slowly changing charging voltage is converted into the peak height of the IC curve, and the wasserstein distance between each adjacent IC curve peak window is calculated. More accurate battery state characteristics can be obtained. Combined with the Gaussian regression process method, the prediction results can be more accurate, and there is no need to establish a battery model. The calculation is also simpler and more practical, which is suitable for online estimation of SOH.

[0027] 2. Each IC curve is sampled using the differential formula to obtain a sampled data set, and Sinkhorn iteration is performed through Pytorch to obtain the wasserstein distance between the peak windows of two adjacent IC curves, so that the prediction model establishment process has more flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present invention is further described in detail below with reference to the accompanying drawings.

[0029] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0030] like Figure 1 As shown, the battery SOH estimation method based on the wasserstein distance includes the following steps:

[0031] S1. Collect the aging data of eight pre-recorded pouch batteries with a nominal capacity of 740mAh, and obtain the IC curve of each aging cycle of each pouch battery through Matlab, where IC represents the battery power added in a continuous voltage increment. The specific calculation formula of IC is:

[0032] Among them, Q is the battery capacity, U is the battery voltage;

[0033] Different from standard voltage aging measurement, IC analysis studies the aging mechanism of the battery from the electrode level. As the number of aging increases, the IC curve becomes gradually flatter, the curve peak decreases, and changes with the changes in the internal materials of the battery during the aging process.

[0034] S2, respectively calculating the wasserstein distance between each two adjacent IC curve peak windows to obtain a feature data set, and dividing the feature data set into an early data set and a late data set;

[0035] In this embodiment, each IC curve is sampled using a differential formula to obtain a sampled data set, and then the sampled data in the peak window is subjected to Sinkhorn iteration through Pytorch to obtain the wasserstein distance between two adjacent IC curve peak windows;

[0036] In this embodiment, the peak window is an interval with a length of 3 centered at the peak of the IC curve; in this embodiment, the peak window is set to a voltage interval of 3.26v-3.54v;

[0037] S3, using a prediction model based on the Gaussian regression process method, with the feature data set in S2 as the input of the prediction model, and the battery SOH as the output of the prediction model. The proportional coefficient hyperparameter α, the feature length scale hyperparameter l, and the period hyperparameter p of the prediction model are set to have a value range of (0.1, 100), (0, 1), and (0, 100), respectively. First, any value within the value range is taken, and then the prediction model is initially trained using the previous data set, and the value of each hyperparameter is adjusted and determined according to the gap between the predicted value and the true value;

[0038] More specifically, the data point is first determined to be a sampling point of a Gaussian process, and the kernel function of the Gaussian regression process is the covariance function; then the predicted expression is determined based on the lag probability, and the likelihood estimate is usually required to be maximized. The hyperparameter is solved through the maximum likelihood estimate, and the hyperparameter is the unknown quantity that cannot be determined in the expression;

[0039] The predicted value is the predicted capacity curve, and the true value is the recorded real capacity curve. When the difference between the predicted value and the true value is greater than the first threshold, the value of the hyperparameter α is lowered and the values ​​of the hyperparameters l and p are increased. When the difference between the predicted value and the true value is less than the second threshold, the value of the hyperparameter α is increased and the values ​​of the hyperparameters l and p are lowered. The adjustment amount should be within 10% of the experience interval, and the value range of the above hyperparameters in the experience interval is selected. In this implementation, the difference between the predicted value and the true value can also be calculated using the wasserstein distance. In the prior art of the calculation process, the first threshold is set to 0.89 and the second threshold is set to 0.67.

[0040] S4. Use the later dataset to train the prediction model with the determined hyperparameters.

[0041] The above description is only a preferred embodiment of the present invention, and therefore cannot be used to limit the scope of implementation of the present invention. That is, equivalent changes and modifications made according to the scope of the patent application of the present invention and the contents of the specification should still fall within the scope covered by the patent of the present invention.

Claims

1. A battery SOH estimation method based on wasserstein distance, characterized in that: The steps include: S1. Collecting aging data of multiple pouch batteries, and obtaining IC curves of each aging cycle of each pouch battery, wherein IC represents the battery power increased by a continuous voltage increment; S2, respectively calculating the wasserstein distance between each two adjacent IC curve peak windows to obtain a feature data set, and dividing the feature data set into an early data set and a late data set; S3, using a prediction model based on the Gaussian regression process method, with the feature data set in S2 as the input of the prediction model, and the battery SOH as the output of the prediction model. The hyperparameters α, l, and p of the prediction model are set to have a value range of (0.1-100), (0,1), and (0,100), respectively. First, any value within the value range is taken, and then the prediction model is initially trained using the previous data set. The value of each hyperparameter is adjusted and determined according to the gap between the predicted value and the true value. S4. Use the later data set to train the prediction model of the determined hyperparameters; in S3, when the gap between the predicted value and the true value is greater than the first threshold, the value of the hyperparameter α is lowered and the values ​​of the hyperparameters l and p are increased; when the gap between the predicted value and the true value is less than the second threshold, the value of the hyperparameter α is increased and the values ​​of the hyperparameters l and p are lowered. The adjustment amount should be within 10% of the experience interval, and the experience interval is selected from the value range of each hyperparameter in step S3.

2. The battery SOH estimation method based on wasserstein distance according to claim 1, characterized in that: In S2, each IC curve is sampled using a differential formula to obtain a sampled data set, and Sinkhorn iteration is performed through Pytorch to obtain the wasserstein distance between two adjacent IC curve peak windows.

3. A battery SOH estimation method based on wasserstein distance according to claim 1 or 2, characterized in that: In S2, the peak window is an interval with a length of 3 centered at the peak of the IC curve.

4. A battery SOH estimation method based on wasserstein distance according to claim 1 or 2, characterized in that: In the prediction model, the kernel function of the Gaussian process is a covariance function.

5. A battery SOH estimation method based on wasserstein distance according to claim 1 or 2, characterized in that: In S1, the specific calculation formula of IC is: , where Q is the battery capacity and U is the battery voltage.

6. A battery SOH estimation device based on wasserstein distance, characterized in that: include: Acquisition module: used to collect aging data of multiple pouch batteries, and obtain the IC curve of each aging cycle of each pouch battery, where IC represents the battery power added in a continuous voltage increment; Feature data set acquisition module: used to calculate the wasserstein distance between the peak windows of each adjacent IC curve, obtain the feature data set, and divide the feature data set into an early data set and a late data set; Parameter determination module: adopts a prediction model based on the Gaussian regression process method, takes the feature data set as the input of the prediction model, and the battery SOH as the output of the prediction model. For the hyperparameters α, l and p of the prediction model, their value ranges are set to (0.1,100), ( 0,1) and (0,100), first take any value within the value range, then use the previous data set to perform initial training on the prediction model, and adjust and determine the value of each hyperparameter according to the gap between the predicted value and the true value; model training module: use the later data set to train the prediction model with determined hyperparameters; in the parameter determination module, when the gap between the predicted value and the true value is greater than the first threshold, the value of the hyperparameter α is lowered and the values ​​of the hyperparameters l and p are increased; when the gap between the predicted value and the true value is less than the second threshold, the value of the hyperparameter α is increased and the values ​​of the hyperparameters l and p are lowered, and the adjustment amount should be within 10% of the experience interval, and the experience interval is selected from the value range of each hyperparameter.

7. The battery SOH estimation device based on wasserstein distance according to claim 6, characterized in that: In the feature data set acquisition module, each IC curve is sampled using a differential formula to obtain a sampled data set, and Sinkhorn iteration is performed through Pytorch to obtain the wasserstein distance between two adjacent IC curve peak windows.

Images