Echechelon utilization battery SOH estimation method based on random features and neural network

Abstract

The invention discloses an echelon utilization battery SOH estimation method based on random features and a neural network, and the method comprises the steps: extracting aging features from the real use data of a decommissioned battery in an echelon utilization period, building an attention TCN network, taking the aging features of a plurality of charge-discharge cycles as the feature input of the attention TCN network, and carrying out the estimation of the SOH of the decommissioned battery. And taking the corresponding battery capacity as a label, and carrying out adaptive fusion to obtain a predicted battery health state. According to the method, state-of-health estimation for the decommissioned battery in the echelon utilization stage is designed, the randomness and complexity of charging and discharging of the decommissioned battery are fully considered, the aging characteristics extracted from the random charging stage are used as network input, important information in the battery charging process is captured, key time sequence characteristics are focused, and the time sequence characteristics of the decommissioned battery are estimated. The method improves the ability of the network model to research the long-term dependency relationship of the time sequence, is simple in measurement process, and can timely estimate the health state of the retired battery in the echelon utilization period.

Description

Technical Field

[0001] The present invention belongs to the field of new energy technology and relates to power battery management, and in particular to a method for estimating SOH of a second-life battery based on random features and a neural network. Background Art

[0002] With the rapid development of the new energy vehicle industry, the investment in power batteries has also increased significantly. However, when the battery health status drops below 80% of the rated capacity, it not only cannot meet the vehicle power demand, but also increases the possibility of traffic accidents. Therefore, it must be removed from electric vehicles. However, these retired batteries can still be used in low-capacity demand areas. However, the performance of retired batteries is not as good as that of new batteries, and safety failure problems are more likely to occur. Therefore, it is crucial to accurately evaluate the performance of batteries for cascade utilization.

[0003] The battery health status estimation methods in the prior art mainly include formula calculation, model-based methods and data-driven methods. Although the formula calculation method is simple, it can only provide a rough estimate, lacks systematic theoretical support and method system, and it is difficult to accurately evaluate the actual performance of the battery. Although the model-based method can improve the evaluation accuracy to a certain extent, it also faces the challenges of modeling complexity and limited scope of application. For the evaluation of automotive power batteries after retirement, their safety status and electrical performance status are often more complicated. At the same time, the scenarios of cascade utilization are diverse. Simply relying on formula calculation or modeling methods is difficult to meet the needs of cascade utilization, and safety cannot be reliably ensured. Therefore, a data-driven method that uses historical big data and test data as neural network input has become a more effective solution. This method can perform a comprehensive, dynamic and adaptive evaluation of cascade utilization batteries. However, the existing data-driven method usually requires modeling of data from the entire charging and discharging process, which not only increases unnecessary training time, but also limits its efficiency and practicality. In addition, the accuracy and robustness of this type of method still need to be further improved to better meet the needs of actual applications. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes a SOH estimation method for cascade utilization batteries based on random features and neural networks. The aging features are extracted during the random charging stage of the battery, and an attention TCN network is built to avoid the inherent defects of traditional recurrent neural networks and convolutional neural networks, further improve the network efficiency and accuracy, and realize real-time and accurate health status estimation of cascade utilization batteries.

[0005] The SOH estimation method of the second-life battery based on random features and neural network includes the following steps:

[0006] S1. Collect data of retired batteries during cascade utilization and perform preprocessing, including outlier processing, missing value processing, normalization and data smoothing.

[0007] S2. Collect the actual usage data of retired batteries during the period of cascade utilization, extract voltage-time and current-time curves from the actual data, and extract aging characteristics from them.

[0008] The actual usage data of the battery during the cascade utilization period is divided into multiple charge and discharge cycles. Within a charge and discharge cycle, the random charging stage of the battery is divided into several segments according to the time of rising the same voltage. The battery capacity difference is calculated from the voltage-time curve and current-time curve of several segments to obtain a battery capacity difference vector of a charge and discharge cycle. For multiple battery capacity difference vectors obtained from multiple charge and discharge cycles, the average value and standard deviation are calculated to form a two-dimensional aging feature as the input data of the neural network.

[0009] S3. Build an attention TCN network, assign dynamic weights to input data of different time steps through expansion factors with different expansion rates, and adaptively focus on important time points and key features. Then introduce the attention mechanism based on the TCN network to achieve weighted combination of data features to enhance feature selection capabilities.

[0010] S4, using the two-dimensional aging characteristics of the second-life battery as data input, estimates the battery health status, and uses the corresponding true value of the battery capacity as the one-dimensional model label to train the network constructed in S3, and optimize the model according to the SOH prediction error. For the second-life battery that needs to estimate the battery health status, the two-dimensional aging characteristics are extracted by the method of S2, and input into the optimized model to obtain the health status estimate.

[0011] The present invention has the following beneficial effects:

[0012] This method aims at estimating the health of batteries during the cascade utilization period, taking into full account the complexity and diversity of retired batteries, and using the two-dimensional aging characteristics of retired batteries during the cascade utilization period as the input of the model, and the corresponding real capacity of the battery as the model label, which effectively reduces the complexity of calculation and implementation cost, and improves the accuracy of the model. This method uses the attention TCN network, the measurement process is simple, and it can timely and accurately estimate the health status of retired batteries during the cascade utilization period. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The framework diagram of the SOH estimation method for cascade utilization batteries based on random features and neural networks. DETAILED DESCRIPTION

[0014] The present invention will be further explained below with reference to the accompanying drawings;

[0015] like Figure 1 As shown, the SOH estimation method of the second-life battery based on random features and neural network includes the following steps:

[0016] S1. Collect the actual usage data of retired batteries during the cascade utilization period and perform data preprocessing, including outlier processing, missing value processing, normalization and data smoothing. In addition, the data is segmented into charge and discharge cycles and its battery capacity is calculated.

[0017] S1-1. Collect the actual usage data of retired batteries during the cascade utilization period, including the voltage-time curve and current-time curve during the cascade utilization stage. It should be noted that since it is the actual usage data of retired batteries during the cascade utilization period, the charging and discharging of the battery is random and uncertain, which inevitably leads to the collection of useless data during the battery data collection process, or the failure to collect appropriate data due to the insensitivity of the sensor.

[0018] S1-2. Analyze the data of retired batteries during the cascade utilization period through visualization and statistical methods. Since there are many extreme outliers in the data, define the outlier threshold to set the identified outliers to zero; use interpolation to fill the zeroed outliers and missing values ​​in the original data. Finally, perform normalization and data smoothing:

[0019]

[0020] Where x and y are the interpolated and normalized data, μ(x) and σ(x) are the mean and standard deviation of x, respectively. β and γ are learnable translation and scaling parameters. ε 1 is a real number greater than 0, used to prevent the denominator from being 0.

[0021] S1-3, perform charge and discharge cycle segmentation on the pre-processed battery data y to calculate the battery capacity. Since the battery charging is relatively stable and is constant current charging, but the initial and final charging capacities are relatively random, the ratio of the battery capacity change during the entire charging phase to the difference between the final SOC and the initial SOC is taken as the battery capacity SOH%, to achieve normalization of random charging:

[0022]

[0023] Where I(t) represents the actual current value of the retired battery during the charging stage, t 0 Indicates the time when each charge-discharge cycle starts, t 1 Indicates the time when each charge and discharge cycle ends, SOC start Indicates the SOC value measured at the beginning of each charge and discharge cycle.end Indicates the SOC value measured at the end of each charge and discharge cycle.

[0024] S1-4. Improper use of some retired batteries will lead to abnormal battery capacity values. Therefore, it is necessary to identify the abnormal battery capacity values, delete the abnormal battery capacity and the corresponding charge and discharge cycles, and avoid affecting network training and testing. The method for removing outliers in this embodiment is to use the absolute difference filtering and interquartile range (IQR) method. First, use the absolute difference filtering method to preliminarily filter the battery capacity whose absolute difference does not meet the conditions:

[0025] |SOC start -SOC end |≥M

[0026] Where M represents the preset difference, when each charge and discharge cycle starts, the SOC value SOC start The SOC value at the end of charging SOC end When the difference between them is less than M, the data of this charge and discharge cycle is discarded.

[0027] Then use the IQR method to remove outliers, set the lower limit Lower and upper limit Upper of the threshold range, and filter out the charge and discharge cycle data outside the threshold range:

[0028] IQR=Q 3 -Q 1

[0029] Lower=Q 1 -1.5 IQR

[0030] Upper=Q 3 +1.5 IQR

[0031] Among them, Q 3 Represents the third quartile, which is the value at 75% of the data after absolute difference filtering. 1 It represents the first quartile, which is the value at 25% of the data after absolute difference filtering. IQR represents the interquartile range.

[0032] S2. Extract the aging characteristics of retired batteries during the cascade utilization period to describe the aging process of retired batteries. Through the visualization method to analyze the battery data, it is found that due to the internal characteristics of the battery and the different usage behaviors among different users, the discharge of retired batteries is relatively complex and difficult to handle. Therefore, the characteristics are extracted during the battery charging stage, and the random charging stage of the battery is divided into several segments according to the same rising voltage, and then the battery capacity difference is calculated from the segmented current-time curve.

[0033] First, extract the voltage-time curve V and current-time curve I of a single charge and discharge cycle of the battery:

[0034]

[0035] v t 、i t They respectively represent the voltage and current values ​​at time t in a single charge and discharge cycle.

[0036] Then the voltage-time and current-time curves are plotted with the same rising voltage v eq If the voltage difference between the end of charging and the last split point does not reach v eq , then ignore this data and divide the voltage-time curve V and current-time curve I into N segments:

[0037] V={V 1 ,...,V a ,...,V N},V a = {v a1 ,v a2 ,...,v am},v am -v a1 =v eq ,a∈[1,N]

[0038] I={I 1 ,...,I a ,...,I N},I a ={i a1 ,i a2 ,...,i am},a∈[1,N]

[0039] Among them, v a1 、v am Indicates the voltage-time curve V of the ath segment a The first and last voltage values ​​in i a1 、i am Indicates the current-time curve I of the ath segment a The first and last voltage values ​​in .

[0040] Then the battery capacity differences are calculated for the segmented curves in turn to obtain a vector △Q containing N battery capacity differences:

[0041]

[0042] Among them, △Q a represents the battery capacity difference of the segmented a, t a0The current-time curve I of segment a a The starting time point, t am The current-time curve I of segment a a The deadline time.

[0043] Calculate the average value F of each element in the vector △Q 1 and standard deviation F 2 :

[0044]

[0045] The average value F of the vector △Q of the battery capacity difference in a charge and discharge cycle 1 and standard deviation F 2 As two aging characteristics of the charge-discharge cycle. The aging characteristics of M charge-discharge cycles are spliced ​​into X n , as the input features of the network:

[0046] X n =[X 1 ,...,X q ,...,X M ],X q ={F 1 ,F 2},q∈[1,M]

[0047] S3. Build an attention TCN network to predict the battery aging status of retired batteries during the cascade utilization period. The attention TCN network can predict multivariate time series, use expansion factors with different expansion rates and attention modules to achieve adaptive fusion, use the two-dimensional aging features of retired batteries during the cascade utilization period obtained in S2 as network input, and use the corresponding one-dimensional battery capacity as a label to achieve network training and prediction. The specific steps are as follows:

[0048] S3-1. During the training phase, the input data D of the attention TCN network n Includes aging features X n And the corresponding battery real capacity Y n , D n ={X n , Y n}, X n ∈R M×2 , Y n ∈R M×1 The shape of the input features is adjusted to (B, T, F), where B = 64, which indicates the batch size, T = 128, which indicates the sequence length, and F = 2, which indicates the input feature dimension.

[0049] S3-2. Build an attention TCN network, including building a multi-layer TCN network layer and an attention mechanism. The multi-layer TCN network is used to extract high-order features of time series data, using dilated convolution to capture long-term dependencies, while improving the stability of training through residual connections and regularization. The attention mechanism dynamically adjusts the model's attention to each time step or feature by weighted combination of feature representations output by the multi-layer TCN network, forming a more focused expression of key features:

[0050] S3-2-1. Multi-layer TCN network layer for input aging characteristics X n First, perform the convolution operation:

[0051] H i,j,k =Conv1D(X n ; W conv ,b conv )

[0052] Among them, H i,j,k represents the convolution result of the i-th sample at the j-th channel at the k-th time step, W conv is the convolution kernel weight matrix, b conv is the bias term. The convolution operation introduces causal constraints, and the output of a time step only depends on the previous time step, that is, the current battery capacity only depends on the current situation and the past situation, without considering the future situation.

[0053] Then convolve the output H by channel i,j,k Perform batch normalization to improve the stability of training:

[0054]

[0055] Among them, μ j and represents the mean and variance on channel j, ε 2 Represents a small positive number, preventing division by zero.

[0056] Then the batch normalization result Apply the activation function and compare it with the input aging feature X n Perform residual connection to obtain the output feature Y of the multi-layer TCN network.

[0057]

[0058] Y=H′+Residual(X n )

[0059] Among them, ReLU() represents a nonlinear activation function.

[0060] After convolution operations with different expansion rates and residual connections, the long-term dependency of battery capacity can be captured and the gradient vanishing problem of multiple TCN network layers can be alleviated. The feature expression capability is further enhanced through batch normalization and nonlinear activation functions.

[0061] S3-2-2. Dynamically adjust the weight of time steps through the attention mechanism to improve the model's attention to key timing information.

[0062] First, the output feature Y of the multi-layer TCN network is mapped into the query matrix Q, the key matrix K and the value matrix V:

[0063] Q=YW Q , K=YW K 、V=YW V

[0064] Among them, W Q , W K , W V Represents a linear transformation matrix.

[0065] Then the similarity between Q and K is calculated by dot product and normalized to get the attention weight matrix A:

[0066]

[0067] in, It is a scaling factor to prevent the dot product value from being too large and causing the gradient to disappear.

[0068] Finally, the attention weight matrix A is used to weight the value matrix V and output the weighted feature Z:

[0069] Z=AV

[0070] By introducing the attention mechanism, important patterns and relationships in time series can be effectively captured, which can improve the accuracy of prediction.

[0071] S3-2-3. Convert the weighted feature Z output by the attention mechanism back to the original feature dimension through global evaluation pooling Thereby, the information of the time dimension is compressed into a feature vector of fixed size, capturing the overall time series characteristics.

[0072] Then, through the fully connected layer Linearly map the high-dimensional feature representation to the target output dimension to generate the final prediction result

[0073]

[0074] Among them, W is the output weight matrix and b is the output bias term. The task layer compresses the data features and maps the high-dimensional features through global average pooling and linear transformation of the fully connected layer, and finally outputs the predicted SOH.

[0075] S4. Use the verified two-dimensional aging characteristics of the used battery as data input and the one-dimensional real battery capacity as the network label, and divide the data into training set, validation set and test set in a ratio of 7:2:1.

[0076] Determine the initial hyperparameters of the attention TCN network model, including sliding window size, learning rate, number of iterations, and batch size. Use the training set data to train the network built with S3, then input the validation data for battery health state estimation. Optimize the model based on the SOH prediction error, and continuously adjust the learning rate, expansion factor size, and discard rate.

[0077] After training 1000 times, the model was evaluated using the root mean square error RMSE and the mean absolute error MAE. The results are shown in the following table:

[0078] Iteration rate (it / s) RMSE(%) MAE(%) 3.03 5.068 3.655

[0079] From the data in the table, we can see that the model can complete about 3.03 batches of training per second. The root mean square error of the model prediction accounts for 5.068% of the target value, and the mean absolute error accounts for 3.655% of the target value. The model prediction performance is good, and the error accounts for a small proportion of the target value.

Claims

1. A second-life battery SOH estimation method based on random features and neural networks, characterized by: The specific steps include: S1. Collect data of retired batteries during cascade utilization and perform preprocessing; S2, extract the voltage-time curve and current-time curve of each charge and discharge cycle from the preprocessed data, and increase the voltage v in the charging stage. eq The curve is segmented according to the time; the battery capacity difference in each time period is calculated using the segmented curve to obtain a battery capacity difference vector of a charge and discharge cycle, and the average value and standard deviation of each element in the vector are calculated as the two-dimensional aging characteristics of the charge and discharge cycle; S3, construct a time series feature extraction network, extract the two-dimensional aging features obtained in S2, and predict the corresponding battery capacity; use the corresponding true value of the battery capacity as the one-dimensional model label, and train and optimize the time series feature extraction network according to the prediction error; S4. For second-life batteries that require battery health status estimation, two-dimensional aging features are extracted using the method of S2 and input into the optimized model to obtain a health status estimation value.

2. The method for estimating SOH of a second-life battery based on random features and neural network as claimed in claim 1, characterized in that: The data of retired batteries during the period of cascade utilization are processed with outliers, missing values, normalization and data smoothing.

3. The method for estimating SOH of a second-life battery based on random features and neural network as claimed in claim 1 or 2, characterized in that: Define the outlier threshold, set the outliers in the data of retired batteries during the cascade utilization period to zero, and use the interpolation method to fill the outliers that are set to zero and the missing values ​​in the original data, and finally perform normalization and data smoothing; Then the charge and discharge cycle is divided, and the SOC value SOC measured at the end of each charge and discharge cycle is end Compared with the SOC value SOC measured at the beginning of charging start The ratio of the difference is taken as the battery capacity SOH%: Where I(t) represents the actual current value of the retired battery during the charging stage, t0 represents the start time of each charge-discharge cycle, and t1 represents the end time of each charge-discharge cycle; Finally, the calculated abnormal battery capacity values ​​are judged, and the abnormal battery capacity and the corresponding charge and discharge cycle data are deleted.

4. The method for estimating SOH of a second-life battery based on random features and neural network as claimed in claim 3, characterized in that: For battery capacity, first use the absolute difference filtering method to preliminarily filter the battery capacity data whose absolute difference does not meet the conditions: |SOC start -SOC end |≥M Where M represents the preset difference, when each charge and discharge cycle starts, the SOC value SOC start The SOC value at the end of charging SOC end When the difference between them is less than M, the data of this charge-discharge cycle is discarded; Then use the IQR method to remove outliers, set the lower limit Lower and upper limit Upper of the threshold range, and filter out the charge and discharge cycle data outside the threshold range: IQR=Q3-Q1 Lower=Q1-1.5·IQR Upper=Q3+1.5·IQR Among them, Q3 is the third quartile, which means the value at the 75% position in the data after the absolute difference filtering; Q1 is the first quartile, which means the value at the 25% position in the data after the absolute difference filtering; IQR means the interquartile range.

5. The method for estimating SOH of a second-life battery based on random features and neural network as claimed in claim 1, characterized in that: For the voltage-time curve V and current-time curve I of a single charge-discharge cycle, the same voltage v is increased. eq If the voltage difference between the end of charging and the last split point does not reach v eq , then ignore this data and split the remaining data into N segments: V={V1,...,V a ,...,V N },V a ={v a1 ,v a2 ,...,v am },v am -v a1 =v eq ,a∈[1,N] I={I1,...,I a ,...,I N },I a ={i a1 ,i a2 ,...,i am },a∈[1,N] Among them, v a1 、v am Indicates the voltage-time curve V of the ath segment a The first and last voltage values ​​in i a1 、i am Indicates the current-time curve I of the ath segment a The first and last voltage values ​​in ; The battery capacity differences are calculated for the segmented curves one by one, and a vector △Q containing N battery capacity differences is obtained: △Q=[△Q1,...,△Q a ,...,△Q N ], Among them, △Q a represents the battery capacity difference of the segmented a, t a0 The current-time curve I of segment a a The starting time point, t am The current-time curve I of segment a a The deadline for Calculate the mean value F1 and standard deviation F2 of each element in the vector △Q: The average value F1 and standard deviation F2 of the vector △Q of the battery capacity difference in a charge and discharge cycle are used as two aging characteristics of the charge and discharge cycle; the aging characteristics of M charge and discharge cycles are spliced ​​into X n , as the input features of the network: X n =[X1,...,X q ,...,X M ],X q ={F1,F2},q∈[1,M]。 6. The method for estimating SOH of a second-life battery based on random features and neural network as claimed in claim 1, characterized in that: The temporal feature extraction network is an attention TCN network, which first extracts high-order features of input features through multiple TCN network layers, then uses the attention mechanism to dynamically adjust the degree of attention to each time step, and finally uses global average pooling and fully connected layers to map to the final battery capacity prediction result.

7. The method for estimating SOH of a second-life battery based on random features and neural network as claimed in claim 6, characterized in that: The attention mechanism maps the output feature Y of the multi-layer TCN network into the query matrix Q, the key matrix K and the value matrix V, and then calculates the similarity of Q and K through the dot product and normalizes them to obtain the attention weight matrix A: in, It is a scaling factor to prevent the gradient from disappearing due to excessive dot product value; Finally, the attention weight matrix A is used to weight the value matrix V and output the weighted feature Z: Z = AV.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1, 2, 5 to 7.

Images