A method for predicting lithium-ion battery life based on early cycle data

Abstract

The application discloses a kind of based on early cycle data prediction lithium ion battery life method, belong to power battery management technical field.The method is by collecting the discharge voltage and capacity data of 100th charge-discharge cycle before battery, generates Q-U curve, is aligned Q-U curve acquisition point and average processing after extracting life factor vector (LFV) by Gaussian process regression method, as the input feature of gated recurrent unit (GRU), with the cycle number when battery capacity attenuation is 80% as output target, trains life prediction model.Five-fold cross validation is used to optimize hyperparameters, to ensure model accuracy.For the battery to be tested, only the first 100 cycles are required to predict its cycle life.Experiments show that the method can predict the life of batteries with the same capacity but different charge-discharge cycles, with a relative error as low as 3.68%, high accuracy and engineering practicability, and improved battery system safety.

Description

Technical Field

[0001] This invention belongs to the field of power battery management technology and is used to predict the cycle life of vehicle lithium-ion batteries under rapid charge and discharge conditions. Background Technology

[0002] Lithium-ion batteries, with their high energy density, long lifespan, low self-discharge power, and rapid charge / discharge capabilities, are widely used in portable electronic devices, electric vehicles, and energy storage systems, providing a clean and efficient energy storage solution. The State of Health (SOH) of a lithium-ion battery reflects the ratio of its current performance to its initial performance and is a crucial parameter characterizing the battery's state. Depending on the application, when the SOH drops to 60% to 80% of its initial value, the battery is considered to be nearing or reaching the end of its lifespan. Battery performance degradation not only affects device performance but can also pose safety hazards. Therefore, accurately predicting the cycle life of a battery in its early stages allows for advance planning of battery use and maintenance, the development of reasonable replacement schedules, and the reduction of potential safety risks, which is of great significance for improving system reliability and safety. Summary of the Invention

[0003] To predict the cycle life of lithium-ion batteries, this invention proposes a prediction method based on early cycle data. Specifically, the invention involves selecting the cycle life of each battery from existing batch charge-discharge cycle data as the target value (output). From the early cycle data of each battery (taking the first 100 charge-discharge cycles as an example), a life feature vector (LFV) is extracted as input. A gated recurrent unit (GRU) is used to build a data model. The LFV of the aforementioned batch of batteries is used as input, and the cycle life is used as the output to train the model. Cross-validation is then used to optimize the hyperparameters, resulting in a GRU-based data model.

[0004] For the target battery to be predicted, early cycle data, such as data from 100 cycles, is required. A lifetime factor vector is extracted from this data and input into the data model to predict the battery's cycle life. In practice, lifetime prediction was performed on 10 batteries with different charge-discharge cycle modes, achieving a relative error as low as 3.68%. This invention provides a powerful tool for health management and lifetime assessment of batteries of the same capacity but with different charge-discharge cycles.

[0005] The complete technical route of this invention is as follows: Figure 1 As shown.

[0006] The specific implementation steps are explained below:

[0007] Step 1: Organize the existing battery charge-discharge cycle data. The batteries should have the same positive and negative electrode materials and the same capacity. The discharge cycle data should be completely recorded, showing the changes in discharge voltage, current and capacity over time.

[0008] Step 2: Obtain the QU curve.

[0009] The discharge voltage and corresponding cumulative discharge capacity data for cycles 6-10 and 96-100 are extracted from the first 100 cycles of data for each battery to obtain 10 QU curves for each battery. The horizontal axis is the battery discharge voltage U (V) and the vertical axis is the cumulative discharge capacity Q (Ah) from the fully charged state to the corresponding voltage.

[0010] Step 3: Resampling, alignment, and averaging of QU curve data.

[0011] To effectively reduce errors caused by capacity recovery, it is necessary to calculate the average QU curve. Simultaneously, to calculate LFV, the voltage sampling points of the 10 QU curves need to be standardized, ensuring consistency in the voltage sampling points across all QU curves and that each voltage sampling point corresponds to a Q value. Gaussian process regression is used to standardize and align the voltage sampling points of each QU curve. The specific steps are as follows:

[0012] 1) Establish a Gaussian process regression model (GPR) for each QU curve (cycles 6-10 and 96-100), using the discharge voltage U as input and the cumulative discharge capacity Q as output. Use the squared exponential kernel (SE kernel) to model the nonlinear relationship and train the Gaussian regression model:

[0013]

[0014] in Let l be the signal variance and l be the length scale. Let δ be the noise variance. ij This refers to the Kronecker delta function.

[0015] Hyperparameters are automatically optimized using maximum edge likelihood.

[0016]

[0017] The optimization process involves initializing the hyperparameter θ and calculating the covariance matrix K (based on the current θ), then iteratively updating θ using the gradient ascent method until the marginal likelihood converges.

[0018] 2) Specify the voltage range based on the minimum and maximum voltage changes during the discharge process, for example: [2.3V-3.5V], and divide the voltage range into 1000 equally spaced points. Predict each U using the trained GPR model * Click the corresponding Q * For each QU curve (cycles 6-10 and 96-100), a resampled and aligned curve is obtained, as shown in equation (3):

[0019]

[0020] For each battery, after resampling and aligning the 10 QU curves, the average value of Q corresponding to each U sampling point on the QU curves for cycles 6-10 and 96-100 is calculated to obtain two averaged QU curves Q(U). e And Q(U) l .

[0021] Step 4: Calculate the lifetime factor vector. The lifetime factor vector LFV is calculated using equation (4):

[0022] LFV = Q li -Q ei ,i=1,2,3,…,1000 (4)

[0023] In the formula, Q e Q l They are Q(U) respectively. e And Q(U) l The Q value corresponding to each voltage point is Q(U). l The capacity data at each sampling point of the curve and Q(U) e The LFV is obtained by subtracting the capacity data of each corresponding collection point on the curve.

[0024] Step 5: Establish the model's input and output, and train the model.

[0025] The lifespan factor (LFV) of each training battery is used as the model input. When the battery capacity drops to 80% of its initial value, the number of cycles reached at this point is recorded as the battery cycle life, and this is used as the model output. A GRU neural network is used to build the model, and the above input and output data are used for training the GRU model.

[0026] Step 6: Optimize model parameters.

[0027] Cross-validation was used to optimize hyperparameters, including hidden layer units, learning rate, batch size, learning rate dynamic change rate, and number of iterations. The training set was divided into five subsets. Four subsets were selected as training data each time, and the remaining subset was used as validation data. Each set of hyperparameters was repeated five times to ensure that each subset was used for validation once. The average error of the five validations was used as the evaluation index of the impact of the set of hyperparameters on the model performance. The root mean square percentage error (RMSPE) was used to evaluate the result of hyperparameter optimization, as shown in Equation (5):

[0028]

[0029] In the formula, n is the number of samples for the prediction results, and P i M is the predicted lifespan of the i-th battery. i This represents the actual measured lifespan of the i-th battery. When the RMSPE error of a set of hyperparameter combinations is less than 3% after cross-validation, that set of hyperparameters is selected and model training is completed.

[0030] Step 7: Predict the target battery life.

[0031] To predict the lifespan of a target battery, it is only necessary to perform 100 charge-discharge cycles on the battery and extract the lifespan factor LFV according to steps 2 to 4, then input it into the pre-trained model. The model will then output the predicted lifespan of the target battery. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the technical process of the method of the present invention.

[0033] Figure 2 This is the average QU curve of the battery.

[0034] Figure 3 Let LFV be the battery's lifespan factor vector. Detailed Implementation

[0035] To more clearly illustrate the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation steps.

[0036] Step 1: Select 70 laboratory fast-charging lithium iron phosphate batteries as training batteries. The batteries have a rated capacity of 1.1 Ah and a rated voltage of 3.3 V. When the battery capacity reaches 80% of its initial value, the battery's cycle life is considered to have ended. Select another 10 new batteries with the same capacity as the training batteries as target batteries for predicting lifetime.

[0037] Step 2: Extract the cumulative discharge capacity curve and corresponding discharge voltage curve for each of the 70 training batteries during cycles 6-10 and 96-100. A total of 10 QU curves are extracted for each battery.

[0038] Step 3: Using the discharge voltage of each extracted QU curve as input and the corresponding cumulative discharge capacity as output, train a separate GPR model. Distribute the voltage within the range of [2.3V–3.3V] to generate 1000 standardized voltage sampling points, which are then used as new inputs to all GPR models. Utilize the GPR model for the corresponding loop to predict the new capacity value corresponding to the standardized voltage sampling points in that loop, thus obtaining 10 resampled QU curves for each battery.

[0039] The capacity mean of the QU curve for each battery was calculated after the 6th-10th and 96th-100th cycles of resampling, respectively, resulting in two representative average QU curves for each battery, as shown below. Figure 2 As shown.

[0040] Step 4: Calculate the lifespan factor (LFV) of each battery using equation (4) to obtain the LFV vector of the battery. A representative LFV vector is shown below. Figure 3 As shown.

[0041] Step 5: Record the number of cycles performed by each battery when its capacity decays to 80% of its initial value; this number will be used as the cycle life. Using the LFV of the 70 batteries as input and the cycle life as output, build a GRU neural network model and train the model using the above input and output.

[0042] Step 6: Update and optimize the hyperparameters using five-fold cross-validation. In this example, the lowest RMSPE result of five-fold cross-validation reached 2.61%, at which point the relatively optimal hyperparameter combination was considered obtained, and this hyperparameter combination was used to train the final model. Table 2 shows the optimization results of the hyperparameter combination.

[0043] Table 2 Results of Hyperparameter Optimization

[0044]

[0045] Step 7: Perform 100 fast charge-discharge cycles on 10 target batteries and record the data. Following steps 2-4, extract the lifespan factor (LFV) of the 10 batteries from these 100 cycles and input it into the pre-trained GRU neural network model. The model will output the cycle life prediction value for each target battery.

[0046] To verify the effectiveness of the present invention, the accuracy of the cycle life prediction is evaluated using relative error (RE), as shown in equation (6):

[0047]

[0048] In the formula, P is the predicted value of the target battery life d, and M is the actual measured cycle life value of the battery.

[0049] Table 3 shows the predicted cycle life and relative errors for the 10 target batteries.

[0050] Table 3 Prediction Results and Errors

[0051]

[0052] In Table 3, when the relative error is negative, the predicted lifespan is smaller than the actual lifespan, indicating a conservative prediction; when the relative error is positive, the predicted lifespan is larger than the actual lifespan, indicating an optimistic prediction. The prediction results in Table 3 show that the lithium-ion battery lifespan prediction method based on early cycle data and a GRU neural network proposed in this invention has high accuracy. The average absolute value of the relative error between the predicted and actual lifespans of the 10 target batteries is 3.68%, with 7 batteries having errors below 5%, indicating that the model can provide reliable predictions in most cases.

[0053] Based on early battery cycle data, this method uses a GRU neural network model to predict battery cycle life with low prediction error, which is of great significance for the maintenance and safety of power battery systems.

[0054] In practical applications, this method only requires that the positive and negative electrode materials and rated capacity of the battery be the same, and does not require that the battery cycle mode be the same. It can be used for batteries with different charge and discharge cycle modes, thus increasing the versatility of the method.

Claims

1. A method for predicting the lifespan of lithium-ion batteries based on early cycle data, characterized in that: This method extracts the discharge interval QU curves of the first 100 cycles of a lithium-ion battery, unifies and aligns the collection points of the QU curves using Gaussian process regression, calculates the average QU curve and lifetime factor vector LFV, establishes a prediction model using a GRU neural network, optimizes the hyperparameters using five-fold cross-validation, and obtains the final trained prediction model. For the predicted battery, the lifetime factor vector is extracted from its early cycle data and input into the prediction model to predict the cycle life of the battery. The specific implementation steps are as follows: Step 1: Organize the existing battery charge-discharge cycle data. The batteries should have the same positive and negative electrode materials and the same capacity. The discharge cycle data should be completely recorded, showing the changes in discharge voltage, current and capacity over time. Step 2: Extract the discharge voltage and corresponding cumulative discharge capacity data for cycles 6-10 and 96-100 from the first 100 cycles of data for each battery to obtain the QU curve. The horizontal axis is the battery discharge voltage U, and the vertical axis is the cumulative discharge capacity Q from the fully charged state to the corresponding voltage. Step 3: Ensure that the number and value of voltage sampling points are the same for each extracted QU curve; use Gaussian process regression to unify and align the voltage sampling points of each QU curve; Step (4): Calculate and obtain the lifetime factor vector LFV using equation (4): (4)； In the formula, , They are Q(U) respectively. e And Q(U) l The Q value corresponding to each voltage point is Q(U). l The capacity data at each sampling point of the curve and Q(U) e The LFV is obtained by subtracting the capacity data at each corresponding collection point of the curve. Step 5: Use the lifespan factor (LFV) of each training battery as input to the prediction model; when the battery capacity drops to 80% of the initial value, record the number of cycles reached at this point as the battery cycle life, and use this as the output of the prediction model; use the GRU neural network to build the prediction model, and use the above input and output data for training the prediction model. Step 6: The hyperparameters of the prediction model are optimized using the five-fold cross-validation method. The training set data is divided into five subsets. Four subsets are selected as training data each time, and the remaining subset is used as validation data. Each set of hyperparameters is repeated five times to ensure that each subset is used for validation once. The average error of the five validations is used as the evaluation index of the impact of the set of hyperparameters on the performance of the prediction model. The root mean square percentage error (RMSPE) is used to evaluate the result of hyperparameter optimization, as shown in Equation (5): (5)； In the formula The number of samples for the prediction results. Let be the predicted lifespan of the i-th battery. The actual measured lifespan of the i-th battery; when the RMSPE error of a set of hyperparameter combinations is less than 3% after cross-validation, the set of hyperparameters is selected and the prediction model training is completed. Step 7: When predicting the lifespan of the target battery, it is only necessary to perform 100 charge-discharge cycles on the battery, and extract the lifespan factor LFV according to steps 2 to 4 and input it into the pre-trained prediction model. The prediction model will output the predicted lifespan of the target battery.

2. The method for predicting lithium-ion battery life based on early cycle data according to claim 1, characterized in that: The specific steps of step (3) are as follows: 1) Establish a Gaussian process regression model (GPR) for each QU curve, using the discharge voltage U as input, and the cumulative discharge capacity Q as output for cycles 6-10 and 96-100. Use a squared exponential kernel function to model the nonlinear relationship and train the Gaussian regression model. (1)； in For signal variance, For length scale, For noise variance, For the Kronecker delta function; Hyperparameters are automatically optimized using maximum edge likelihood. : (2)； The optimization process specifically involves initializing hyperparameters. The covariance matrix K is calculated and iteratively updated using the gradient ascent method. Until the edges seem to converge; 2) Specify the voltage change range based on the minimum and maximum voltage changes during the discharge process, and divide the voltage range into 1000 equally spaced points. Use the trained prediction model to predict each Point corresponding The value is obtained by resampling and aligning each QU curve, as shown in equation (3): (3) ； For each battery, after resampling and aligning the 10 QU curves, the average value of Q corresponding to each U sampling point on the QU curves for cycles 6-10 and 96-100 is calculated to obtain two averaged QU curves Q(U). e And Q(U) l .

Images