A method for lithium-ion battery capacity estimation and remaining useful life prediction

Abstract

The application belongs to the technical field of battery management system, and provides a method for lithium ion battery capacity estimation and residual useful life prediction, wherein firstly, feature voltage and capacity data of a discharge phase under each cycle in a lithium ion battery aging experiment are extracted, then a feature voltage-cycle number model and a capacity-feature voltage model are respectively constructed, parameters of the models are optimized and identified by using a predetermined PF algorithm and online extracted feature voltage and discharge capacity data, the identified parameters are substituted into the two models to complete online estimation of the capacity of the lithium ion battery, and finally, when the capacity decays to a RUL prediction threshold, RUL and probability density distribution of the lithium battery are offline predicted by model extrapolation. The application can not only realize SOH monitoring and RUL prediction of the whole life cycle of the lithium battery, but also greatly reduce storage and calculation amount of online data, and effectively improve prediction accuracy of the model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of battery management system, and particularly relates to a method for estimating battery capacity and predicting residual service life of a lithium ion battery. BACKGROUND

[0002] With the development of society and the progress of science and technology, people's demand and dependence on automobiles increase year by year, and energy shortage and environmental pollution problems are becoming increasingly serious. Petroleum will be exhausted in a few years, and developing new energy vehicles to replace traditional fuel vehicles is the trend of the times. Lithium ion batteries have excellent comprehensive performance such as high specific energy, high specific power, low self-discharge rate, long cycle life and environmental friendliness, and thus become the most important power source for electric vehicles. However, the capacity degradation of lithium ion batteries is inevitable, which shortens the service life of the battery and reduces the safety of the battery. Therefore, the state of health (SOH) monitoring and residual useful life (RUL) prediction of lithium batteries are of great significance to the safety of lithium batteries.

[0003] The SOH and RUL of lithium batteries are generally evaluated by the capacity of the battery. At present, there are three technical routes for battery capacity estimation, namely, model-based method, data-driven method and fusion estimation method. The model-based method is usually based on an empirical model, which directly establishes a model between the cycle number and the remaining capacity of the battery to estimate the capacity. This method has a large estimation error; the data-driven method mainly relies on battery charge and discharge data to establish a mapping relationship between a certain feature and battery capacity degradation. Common features include commonly used features such as differential voltage (DV), charge and discharge curves, incremental capacity (IC) curves and internal features found by deep learning algorithms. This method requires a large amount of data and a lot of calculation to ensure accuracy; the fusion estimation based on fusion technology refers to a hybrid prediction method that integrates multiple methods to give full play to the advantages of different methods to obtain higher accuracy and robustness. However, how to determine the credibility of each fusion quantity to improve the prediction accuracy has always been a difficulty in fusion estimation technology. SUMMARY

[0004] The present application is made to solve the above problems, and aims to provide a method for estimating battery capacity and predicting residual service life of a lithium ion battery, which is based on a predetermined particle filter (PF) algorithm and a feature voltage model to predict the residual service life of the lithium ion battery.

[0005] The application provides a method for estimating the battery capacity and predicting the remaining service life of a lithium ion battery, which has the following characteristics: the method comprises the following steps: step S1, extracting the sample characteristic voltage U and sample capacity Q of the sample lithium ion battery at each cycle in the discharge phase of the aging experiment; step S2, constructing a characteristic voltage-cycle number model and a capacity-characteristic voltage model respectively, and simultaneously using a predetermined PF algorithm and the sample characteristic voltage and discharge capacity data extracted online to optimize the parameters of the characteristic voltage-cycle number model and the capacity-characteristic voltage model to obtain optimized parameters; step S3, substituting the optimized parameters into the characteristic voltage-cycle number model and the capacity-characteristic voltage model respectively, calculating the characteristic voltage online according to the cycle number of the lithium ion battery to be measured through the characteristic voltage-cycle number model, and estimating the battery capacity of the lithium ion battery to be measured according to the characteristic voltage; and step S4, when the battery capacity of the lithium ion battery to be measured decays to a predetermined threshold, stopping online estimation, and using the characteristic voltage-cycle number model, the capacity-characteristic voltage model and the optimized parameters to offline predict the remaining service life of the lithium ion battery to be measured.

[0006] In the method for estimating the battery capacity and predicting the remaining service life of a lithium ion battery provided by the application, the characteristic voltage-cycle number model can also have the following characteristics: the characteristic voltage-cycle number model is:

[0007] U=aC x 2 +bC x +c,

[0008] wherein a, b and c are parameters of the characteristic voltage-cycle number model, C x is the cycle number, and U is the corresponding characteristic voltage at different cycle numbers.

[0009] The capacity-characteristic voltage model is:

[0010] Q=m*exp(n*U)+p*exp(q*U),

[0011] wherein m, n, p and q are parameters of the capacity-characteristic voltage model, U is the characteristic voltage, and Q is the battery capacity at different cycle numbers.

[0012] In the method for estimating the battery capacity and predicting the remaining service life of a lithium ion battery provided by the application, the process of optimizing the parameters in step S2 can also have the following characteristics:

[0013] Step Sa-1, using the capacity data set Q iThe model initial parameter set x0 when the loss function E is minimum is substituted into the formula of the least square algorithm, and a characteristic voltage-cycle number model and a capacity-characteristic voltage model are obtained.

[0014] Step Sa-2, for k=0, the sampling particles are directly generated from the optimized prior distribution p(x0) Wherein i=1, 2, …, N;

[0015] Step Sa-3, for k=1, 2, …, the optimal state estimation value at different time is completed as an optimization parameter through the spatial state equation of the characteristic voltage-cycle number model until the cycle ends.

[0016] In the method for estimating the capacity of a lithium ion battery and predicting the remaining service life provided by the application, the spatial state equation of the characteristic voltage-cycle number model can be:

[0017]

[0018] U k = a k C x 2 +b k C x +c k +n U,k ,

[0019] Wherein v represents the transfer noise, n U,k represents the observation noise, the subscript k represents the parameter value at k time, ν U1,K-1 , ν U2,K-1 and ν U3,K-1 represent different transfer noises,

[0020] The spatial state equation of the capacity-characteristic voltage model is as follows:

[0021]

[0022] Q k = m k *exp(n k *U)+p k *exp(q k *U)+n Q,k ,

[0023] Wherein v represents the transfer noise, n Q,k represents the observation noise, the subscript k represents the parameter value at k time, ν Q1,K-1 , ν Q2,K-1 , ν Q3,K-1 and νQ4,K-1 These represent different types of transmitted noise.

[0024] The method for estimating the capacity and predicting the remaining lifespan of a lithium-ion battery provided by this invention may also have the following feature: wherein step S2-3 specifically includes the following sub-steps:

[0025] Step S2-3-1: The initial particle set obtained through importance sampling... Substituting these values ​​into the characteristic voltage-cycle number model, we obtain the weights of each particle at each time step.

[0026] Step S2-3-2: Calculate the weights of each particle based on the weights calculated in step S2-3-1. Resampling is performed to generate a new set of particles. At this point, the weight of each particle is 1 / N;

[0027] In step S2-3-3, after resampling, the weights of each particle are the same, so the estimated value at different times is the average of the sum of all particles. x k These are the identified values ​​of the model parameters.

[0028] The method for estimating the capacity and predicting the remaining lifespan of a lithium-ion battery provided by this invention may also have the following feature: step S4 specifically includes the following sub-steps:

[0029] Step S4-1, assuming that in the S-th cycle (i.e., k = S), the new particle set matrix generated after resampling is Substitute N sets of parameters into the characteristic voltage-cycle number relationship model to calculate the decayed characteristic voltage vector. Then, by substituting the characteristic voltage set into the capacity-characteristic voltage model, the capacity degradation vector is calculated. Based on the predicted capacity degradation vector? With cyclic number vector The mapping between them yields the cyclic number vector:

[0030]

[0031] That is: the range of the estimated cycle number is

[0032] Step S4-2: After resampling, the weights of the N particles are all 1 / N. The interval is divided into W equal subintervals. The probability of a particle appearing is represented by the frequency of the particle appearing in each subinterval, denoted as:

[0033] (pro1,pro2,...,pro w) = (fre1, fre2,..., fre w ) / N, wherein, pro (1,2,...,w) represents the probability of W intervals, fre (1,2,...,w) represents the frequency of W intervals, and N is the number of particles.

[0034] In the method for estimating the capacity of a lithium ion battery and predicting the remaining useful life provided by the application, the feature voltage U can be the voltage value at a fixed time in the discharge phase in the cycle aging experiment, and the capacity Q can be the discharge capacity when the cell voltage in the discharge phase drops to the discharge cutoff voltage.

[0035] Effects of the application

[0036] According to the method for estimating the capacity of a lithium ion battery and predicting the remaining useful life provided by the application, the feature voltage and capacity data in the discharge phase of each cycle in the lithium ion battery aging experiment are first extracted, then a feature voltage-cycle number model and a capacity-feature voltage model are respectively constructed, the predetermined PF algorithm and the feature voltage and discharge capacity data extracted online are used to optimize and identify the parameters of the model, the identified parameters are substituted into the two models to complete the online estimation of the capacity of the lithium ion battery, and finally when the capacity decays to the RUL prediction threshold, the RUL and the probability density distribution of the lithium battery are predicted offline through model extrapolation. The application not only can realize the SOH monitoring and RUL prediction of the lithium battery in the whole life cycle, but also greatly reduces the storage and calculation amount of online data, and effectively improves the prediction accuracy of the model. Compared with the prior art, the application has the following advantages:

[0037] (1) A new feature for capacity estimation is proposed, that is, only the voltage value at a fixed time in the discharge phase of each aging cycle is collected, which greatly reduces the storage and calculation of online data.

[0038] (2) The predetermined PF algorithm no longer obtains the optimized prior distribution by iteratively calculating each particle, but only needs to obtain the full life cycle decay data of the sample battery through the experiment, and then uses the least square method to fit to obtain the initial parameters of the relationship model, and the initial parameters are directly substituted into the generation of the new prior distribution as the initial value of the particle filter. This greatly simplifies the optimization process and improves the calculation accuracy.

[0039] (3) By applying the method of the application, the capacity estimation error can be maintained within 3%, the prediction error of RUL can be maintained within 5%, the probability density distribution curve estimated by the predetermined PF algorithm is more consistent with the real situation of battery decay, and has better uncertainty expression ability. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1is a method flow chart of lithium ion battery capacity estimation and remaining useful life prediction in embodiments of the present application;

[0041] Figure 2 is a relationship chart between each battery monomer characteristic voltage and cycle number and a relationship chart between capacity and characteristic voltage in embodiments of the present application;

[0042] Figure 3 is a characteristic voltage and capacity estimation result chart of No. 5 battery monomer in embodiments of the present application;

[0043] Figure 4 is a capacity estimation result error chart of No. 5 battery monomer in embodiments of the present application;

[0044] Figure 5 is a comparison chart of RUL prediction and probability density distribution results of No. 5 battery monomer before and after PF algorithm improvement in embodiments of the present application; and

[0045] Figure 6 is a RUL prediction result error chart of No. 5 battery monomer in embodiments of the present application. DETAILED DESCRIPTION

[0046] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application is specifically described below in combination with embodiments and drawings.

[0047] In the present application, the predetermined PF algorithm is an improved PF algorithm.

[0048] <EMBODIMENT>

[0049] Figure 1 is a method flow chart of lithium ion battery capacity estimation and remaining useful life prediction in embodiments of the present application.

[0050] As shown in Figure 1 , the method for lithium ion battery capacity estimation and remaining useful life prediction of the present embodiment is used for estimating the battery capacity and predicting the remaining useful life of the lithium ion battery to be measured, and includes the following steps:

[0051] Step S1, extracting the sample characteristic voltage U and sample capacity Q of the sample lithium ion battery at each cycle in the discharge phase of the aging experiment. The characteristic voltage is the voltage value at a fixed time in the discharge phase of each cycle, and the fixed time is selected as 2000s. The capacity is the battery discharge capacity when the voltage drops to the discharge cutoff voltage.

[0052] In this embodiment, in order to better illustrate the universal generality of the characteristic voltage and capacity proposed by the present application, the present application verifies the lithium ion battery life data provided by the database of the United States National Aeronautics and Space Administration (NASA), extracts the characteristic voltage and capacity data of Cell#5, Cell#6 and Cell#7 battery cells respectively, and the results are shown in Figure 2 Figure 2 (a) is a graph of the characteristic voltage and cycle number of three battery cells, Figure 2 (b) is a graph of the relationship between capacity and characteristic voltage.

[0053] Step S2, respectively constructing a characteristic voltage-cycle number model and a capacity-characteristic voltage model, and simultaneously using a predetermined PF algorithm and the sample characteristic voltage and discharge capacity data extracted online to optimize the parameters of the characteristic voltage-cycle number model and the capacity-characteristic voltage model, to obtain the optimized parameters.

[0054] The characteristic voltage-cycle number model is: U=aC x 2 +bC x +c, wherein a, b and c are parameters of the characteristic voltage-cycle number model, C x is the cycle number, and U is the corresponding characteristic voltage under different cycle numbers.

[0055] The capacity-characteristic voltage model is: Q=m*exp(n*U)+p*exp(q*U), wherein m, n, p and q are parameters of the capacity-characteristic voltage model, U is the characteristic voltage, and Q is the battery capacity under different cycle numbers.

[0056] The space state equation of the characteristic voltage-cycle number model is as follows:

[0057]

[0058] U k =a k C x 2 +b k C x +c k +n U,k ,

[0059] wherein v represents the transmission noise, n U,k represents the observation noise, the subscript k represents the parameter value at time k, ν Q1,K-1 , ν Q2,K-1 , ν Q3,K-1 and ν Q4,K-1 represent different transmission noises. Because of the transmission noise generated in the parameter transmission process, ν Q1,K-1 , ν Q2,K-1 , ν​Q3,K-1 and ν Q4,K-1 Transmission noise with different parameters.

[0060] The space state equations for the capacity-characteristic voltage model are as follows:

[0061]

[0062] Q k =m k *exp(n k *U)+p k *exp(q k *U)+n Q,k ,

[0063] Where v represents transmitted noise, n Q,k The parameter ν represents the observation noise, where the subscript k represents the parameter value at time k. Q1,K-1 ν Q2,K-1 ν Q3,K-1 and ν Q4,K-1 These represent different types of transmitted noise. Q,k and the above n U,k Both represent observation noise, but they represent different types of observation noise and are assigned different values ​​during the calculation process.

[0064] Depend on Figure 2 (a) It can be seen that the characteristic voltage gradually decreases as the number of cycles (i.e., the loop count) increases. The first derivative of the characteristic voltage curve is non-positive and its absolute value increases, while the second derivative is also negative. The curve is a convex function and a decreasing function within a specific range. Simply put, the decreasing trend of the voltage characteristic point is first flat and then accelerates, exhibiting obvious parabolic model characteristics. In addition, although using a cubic or higher-order function can better fit the results, the improvement in accuracy is small, while the computational load is greatly increased. Therefore, considering all factors, this embodiment uses a parabolic equation to describe the correlation model between the characteristic voltage and the number of cycles, i.e., the characteristic voltage-cycle count model U = aC. x 2 +bC x Similarly, the capacity-characteristic voltage relationship clearly conforms to an exponential relationship, but the accuracy of the first-order exponential relationship is much lower than that of the second-order exponential relationship, and higher-order relationships may lead to overfitting. Therefore, in this embodiment, a double exponential model is selected, namely Q=m*exp(n*U)+p*exp(q*U).

[0065] In step US, the parameter optimization process is as follows:

[0066] Step Sa-1, optimize the initial parameter set: using the capacity dataset Q of the entire life cycle of the sample lithium-ion battery cells. iThe model initial parameter set x0 is obtained by substituting the formula of the least square algorithm, and the characteristic voltage-cycle number model and the capacity-characteristic voltage model are obtained when the loss function E is minimum. The model initial parameter set x0 is taken as the initial value x0 of the particle filter proposal density function.

[0067] In the embodiment, the particle filter proposal density function is a normal distribution function.

[0068] The loss function E is:

[0069] In the formula, y i represents the dependent variable corresponding to the independent variable in the corresponding relationship model under different cycles, represents the given historical data under the corresponding cycle. The relationship model refers to the characteristic voltage-cycle number model or the capacity-characteristic voltage model. That is, when the relationship model is the characteristic voltage-cycle number model, y i is the corresponding characteristic voltage under different cycles; when the relationship model is the capacity-characteristic voltage model, y i is the corresponding battery capacity under different cycles. The parameters in the characteristic voltage-cycle number model and the parameters in the capacity-characteristic voltage model corresponding to the minimum loss function E are taken as the model initial parameter set x0.

[0070] Step Sa-2, particle set initialization: for k=0, the sampling particles are directly generated from the model initial parameter set x0 and the particle filter proposal density function.

[0071] Step Sa-3, update iteration: for k=1, 2, …, the optimal state estimation value under different times is obtained by the spatial state equation of the characteristic voltage-cycle number model as the optimization parameter until the cycle ends.

[0072] In the embodiment, the optimization parameter is the optimal state estimation value under different cycles k=0, 1, 2, 3, …, and the optimal state estimation value here is for the spatial state equation, and the optimization parameter is for the model.

[0073] Step S2-3 specifically includes the following sub-steps:

[0074] Step S2-3-1, importance sampling: the initial particle set obtained by importance sampling is brought into the characteristic voltage-cycle number model to obtain the weight of each particle under each time wherein i refers to the number of particles.

[0075] Step S2-3-2, resampling: according to the weight of each particle calculated in step S2-3-1, the particle set is resampled to obtain the resampled particle set Resampling is performed to generate a new set of particles At this time, the weight of each particle is 1 / N;

[0076] Step S2-3-3, output: after resampling, the weight of each particle is the same, and the estimated value (i.e. mathematical expectation) at different times is the average of the sum of each particle x k That is, the identified value of the model parameter.

[0077] Through iterative calculation, a new set of parameters is obtained at each cycle (time), but the parameter value is different at each cycle because it is updated by iteration, which is called the identified value.

[0078] Further, the capacity online estimation in step 3 needs to first substitute the optimized parameters into the characteristic voltage-cycle number model to realize online estimation of the characteristic voltage, and then substitute the estimated characteristic voltage and the optimized parameters in the capacity-characteristic voltage model into the capacity-characteristic voltage model to realize online estimation of the capacity.

[0079] In this embodiment, the parameter optimization is specifically as follows.

[0080] The characteristic voltage and capacity data at a part of the cycle number extracted online are used to perform online optimization of the parameters of the above two models. From Figure 1 It can be found from the above that the judgment standard for whether to continue to extract data online in the embodiment of the present application is whether the battery capacity is degraded to the RUL predetermined threshold. The RUL predetermined threshold refers to the percentage of the current remaining capacity of the battery to the nominal capacity. The RUL predetermined threshold in the embodiment of the present application is 80%. That is, when the capacity of the battery is degraded to 80% of the nominal capacity, the extraction of the characteristic voltage and capacity data is stopped, and the iterative update and optimization of the model parameters are also stopped. The predetermined PF algorithm needs sample battery data. The three batteries in step 1 are all 18650 ternary lithium batteries with the same ampere-hour number and freshness, and the experimental conditions are the same. Under the same production process and experimental conditions, there is no obvious difference in electrical performance before and after the experiment of the three batteries. Any one of the three batteries has generality, and the data of the three batteries can be used as sample battery data. In this embodiment, the experimental data of the full life cycle of Cell#7 is used as sample data, the initial optimized solution of the model parameter set is fitted by the least square method, and this set of parameters is used as the initial parameter value in the capacity estimation and RUL prediction model of the same type battery Cell#5. In this embodiment, the experimental data of the No. 5 monomer is extracted. When the capacity of the battery is degraded to the RUL threshold, the battery is cycled for 100 times. The two model parameters optimized at the 100th time are listed in Table 1 and Table 2.

[0081] Table 1 Parameter identification result of the characteristic voltage-cycle number model of the No. 5 monomer

[0082]

[0083] Table 2 Parameter identification results of the capacity-characteristic voltage model of No. 5 monomer

[0084]

[0085] Step S3, the optimization parameters are respectively substituted into the characteristic voltage-cycle number model and the capacity-characteristic voltage model, the characteristic voltage is calculated on-line according to the cycle number of the lithium ion battery to be measured through the characteristic voltage-cycle number model, and the battery capacity of the lithium ion battery to be measured is estimated on-line according to the characteristic voltage.

[0086] In the process of model parameter updating iteration, a set of optimized parameters can be obtained in each cycle, and the new characteristic voltage can be calculated after the set of parameters is substituted into the model. After the characteristic voltage is estimated on-line, the new characteristic voltage and the optimized parameters of the capacity-characteristic voltage model are substituted into the model, so that the on-line estimation of the capacity can be realized, and the estimation result is as shown in Figure 3 .

[0087] Figure 3 Fig. 5 is a characteristic voltage and capacity estimation result diagram of No. 5 battery monomer in the embodiment of the present application. Among them, Figure 3 (a) is a comparison diagram of the experimental value and the estimated value of the characteristic voltage of #5 battery monomer, Figure 3 (b) is a comparison diagram of the experimental value and the estimated value of the capacity of #5 battery monomer.

[0088] From Figure 3 (a) and Figure 3 (b), it can be seen that the estimated characteristic voltage can stably reflect the attenuation trend of the battery capacity.

[0089] In addition, based on the estimated characteristic voltage value and the characteristic voltage-capacity model, the battery capacity can be estimated in real time, and the estimation error result is as shown in Figure 4 .

[0090] Figure 4 Fig. 5 is a characteristic voltage and capacity estimation result diagram of No. 5 battery monomer in the embodiment of the present application. Among them,

[0091] From Figure 4 , it can be seen that the error of the on-line estimation of the capacity can be kept within 3%, and a high estimation accuracy and stability are presented in the battery aging process, which shows that the predetermined PF algorithm and the voltage characteristic model for on-line estimation of the battery capacity are effective.

[0092] Step S4, when the battery capacity of the lithium ion battery to be measured decays to a predetermined threshold, the on-line estimation is stopped, and the remaining service life and the probability density distribution of the lithium battery to be measured are predicted off-line by using the characteristic voltage-cycle number model, the capacity-characteristic voltage model and the optimization parameters.

[0093] The RUL prediction and probability density distribution estimation method of this step is specifically as follows:

[0094] Step S4-1, assuming that at the Sth cycle (i.e., k=S), the new particle set matrix generated after resampling is The N sets of parameters are brought into the characteristic voltage-cycle number relationship model to calculate the characteristic voltage vector after attenuation The characteristic voltage set is brought into the capacity-characteristic voltage model to calculate the capacity degradation vector According to the predicted capacity degradation vector and the cycle number vector , the cycle number vector is obtained:

[0095]

[0096] That is, the estimated cycle number range is

[0097] Step S4-2, the weight of the N particles after resampling is 1 / N, and The interval is equally divided into W small intervals, and the frequency of particles appearing in each small interval represents the probability of particle appearance, denoted as:

[0098] (pro1,pro2,...,pro w )=(fre1,fre2,...,fre w ) / N,

[0099] Where pro (1,2,...,w) represents the probability of the W intervals, fre (1,2,...,w) represents the frequency of the W intervals, and N is the number of particles, thereby completing the prediction of the remaining useful life of the lithium battery to be tested.

[0100] In this embodiment, the specific operation is as follows:

[0101] When the capacity reaches the RUL threshold, stop the online estimation of the capacity, and record the optimized parameters of the two models at the 100th cycle and the distribution of each particle, respectively. The optimized parameters are the parameter data in Table 1 and Table 2, and the distribution of each particle is too much to be listed here (the same below). Put these parameter data back into the model to enter the offline prediction mode. First, predict the characteristic voltage value and its distribution at the subsequent cycle according to the characteristic voltage-cycle number model, then put the predicted characteristic voltage value into the capacity-characteristic voltage model to extrapolate to predict the remaining capacity of the battery and its distribution. Again, because there is a certain mapping between the predicted degradation capacity and the cycle number , denoted as Where f represents the functional relationship between degradation capacity and number of cycles. Indicates error. While ensuring... To minimize the mean square error, the equation becomes

[0102] In this way, the number of iterations can be calculated.

[0103]

[0104] Therefore, the estimated range of the number of iterations is: The weights of the N particles after resampling are all 1 / N. Divide the interval into W equal subintervals. The probability of a particle's occurrence is represented by the frequency of its occurrence in each subinterval, denoted as (pro1, pro2, ..., pro...). w ) = (fre1,fre2,...,fre w ) / N,

[0105] Among them, pro (1,2,...,w) fre represents the probability of W intervals. (1,2,...,w) Let W represent the frequency of the W intervals, and N be the number of particles. This completes the offline prediction of the RUL (Rate Length Usage) and its probability density distribution estimation for cell No. 5. The results are as follows: Figure 5 and Figure 6 As shown.

[0106] Figure 5 This is a comparison chart of the RUL prediction and probability density distribution results of the No. 5 battery cell before and after the PF algorithm improvement in this embodiment of the invention; Figure 6 This is an error diagram showing the RUL prediction results of cell No. 5 in an embodiment of the present invention.

[0107] In this embodiment, only 70% of the nominal capacity of a single cell was predicted, but this method is not limited to predicting up to 70%. Table 3 lists the mean absolute error (MAE) and root mean square error (RMSE) of the Cell#5 capacity prediction values. It can be seen that compared with the standard PF algorithm, the MAE and RMSE of the proposed predetermined PF algorithm prediction values ​​are smaller. The prediction accuracy and stability of the capacity obtained using the predetermined PF algorithm are higher than those obtained by the standard PF algorithm throughout the entire prediction period. Furthermore, from Figure 5It can be seen that for Cell#5, the RUL prediction cut-off circle (PCC) obtained by the predetermined PF algorithm is 168 and 93 respectively, the RUL prediction PCC obtained by the standard PF is 151 and 114 respectively, and the true cut-off circle (TCC) is 162 and 102 respectively. It can be seen that the RUL predicted by the predetermined PF algorithm is more accurate. At the same time, from the PDF estimation result, it can be seen that the peak cycle number of the PDF curve obtained by the predetermined PF algorithm is close to the true cut-off cycle number, while the peak of the PDF prediction curve obtained by the standard PF algorithm has a large deviation from the true value. Finally, we can clearly find from Table 2 that the prediction error of the improved PF algorithm is less than 5%. Figure 6

[0108] In order to comprehensively evaluate the influence of different life prediction thresholds on the capacity and RUL prediction accuracy, the capacity and RUL prediction results under different prediction thresholds are compared, and the results are shown in Table 3. It can be seen that the RUL predicted by the improved PF algorithm is closer to the true value than the RUL predicted by the standard PF algorithm. In addition, increasing the life prediction threshold range (the SOH threshold of Cell#5 is increased to 90%-70%) reduces the prediction accuracy, but the improved PF algorithm is less affected by the threshold, and can still guarantee a high-precision RUL prediction. The standard PF algorithm is very limited by the threshold setting in the prediction, and the prediction error is large when the threshold range is expanded. At the same time, it can be seen that expanding the threshold range will reduce the capacity online estimation accuracy, but the estimation accuracy based on the improved PF algorithm is still higher than the capacity estimation accuracy obtained by the standard PF algorithm. In summary, the accuracy and stability of the improved PF algorithm in estimating and predicting the capacity and RUL are better than those of the standard PF algorithm.

[0109] Table 3 Error analysis results of No. 5 cell algorithm before and after improvement

[0110]

[0111] Effects of the embodiments

[0112] ​The embodiment provides a method for lithium ion battery capacity estimation and remaining useful life (RUL) prediction. Firstly, feature voltage and capacity data in a discharge phase of each cycle in a lithium ion battery aging experiment are extracted, then a feature voltage-cycle number model and a capacity-feature voltage model are respectively constructed, a predetermined PF algorithm and feature voltage and discharge capacity data extracted on line are used to optimize and identify parameters of the model, the identified parameters are substituted into the two models to complete on-line estimation of the capacity of the lithium ion battery, and finally when the capacity decays to a RUL prediction threshold, the RUL and probability density distribution of the lithium battery are predicted off-line through model extrapolation. The predetermined PF algorithm is an optimized PF algorithm. In the improved PF algorithm, initial values of an optimized suggested probability density are obtained by fitting sample battery aging data, the accuracy and rapidity of model parameter identification are improved, and thus the accuracy of the established correlation model is improved. The method can not only realize SOH monitoring and RUL prediction of the lithium battery in a whole life cycle, but also greatly reduces storage and calculation of on-line data and effectively improves the prediction accuracy of the model. Compared with the prior art, the method has the following advantages:

[0113] (1) A new feature for capacity estimation is proposed. Only voltage values at fixed time points in a discharge phase of each aging cycle are collected, and storage and calculation of on-line data are greatly reduced.

[0114] (2) The improved PF algorithm does not obtain an optimized prior distribution by iterative calculation of each particle, but only needs to obtain full-life cycle attenuation data of a sample battery through an experiment, and then initial parameters of a relationship model are fitted by using a least square method to directly substitute the initial parameters into generation of a new prior distribution as initial values of particle filtering. In this way, the optimization process is greatly simplified, and the calculation accuracy is improved.

[0115] (3) The capacity estimation error can be kept within 3%, the prediction error of the RUL can be kept within 5%, the probability density distribution curve estimated by the improved PF algorithm is more in line with the real situation of battery attenuation, and has better uncertainty expression ability.

[0116] The above embodiment is a preferred case of the present application, and is not used to limit the protection scope of the present application.

Claims

1. A method for lithium-ion battery capacity estimation and remaining useful life prediction, for estimating the battery capacity and predicting the remaining useful life of a lithium-ion battery under test, characterized in that, The method comprises the following steps: Step S1, extracting sample characteristic voltage U and sample capacity Q of a sample lithium ion battery at a discharge stage of each cycle in an aging experiment; Step S2, respectively constructing a characteristic voltage-cycle number model and a capacity-characteristic voltage model, and simultaneously optimizing parameters of the characteristic voltage-cycle number model and the capacity-characteristic voltage model by using a predetermined PF algorithm and sample characteristic voltage and discharge capacity data extracted online, to obtain optimized parameters; The characteristic voltage-cycle number model is as follows: U = aC x 2 + bC x + c, where a, b, and c are parameters of the characteristic voltage-cycle number model, C x is the cycle number, and U is the corresponding characteristic voltage at different cycle numbers. The capacity-characteristic voltage model is as follows: Q = m * exp(n * U) + p * exp(q * U), wherein m, n, p and q are parameters of the capacity-characteristic voltage model, U is the characteristic voltage, and Q is the battery capacity at different cycle numbers; The process of optimizing the parameters in step S2 is as follows: Step Sa-1, using the capacity dataset Q of the full life cycle of the cell of the sample lithium-ion battery i The model initial parameter set x0 of the characteristic voltage-cycle number model and the capacity-characteristic voltage model at the minimum loss function E is obtained by substituting the formula of the least square algorithm, and the model initial parameter set x0 is taken as the expectation of the initial value extracted in the particle filtering proposal density function. Step Sa-2, for k = 0, sample particles are generated directly from the model initial parameter set x0and the particle filter proposal density function where i = 1, 2,..., N; Step Sa-3, for k = 1, 2, …, completing optimal state estimation values at different times as the optimized parameters by using a spatial state equation of the characteristic voltage-cycle number model until the cycle ends; Step S3, substituting the optimized parameters into the characteristic voltage-cycle number model and the capacity-characteristic voltage model respectively, calculating the characteristic voltage according to the cycle number of the lithium ion battery to be tested online by using the characteristic voltage-cycle number model, and estimating the battery capacity of the lithium ion battery to be tested according to the characteristic voltage; and Step S4, when the battery capacity of the lithium ion battery to be tested decays to a predetermined threshold, stopping online estimation, and using the characteristic voltage-cycle number model, the capacity-characteristic voltage model and the optimized parameters to offline predict the remaining useful life of the lithium ion battery to be tested.

2. The method for estimating the battery capacity and predicting the remaining useful life of a lithium ion battery according to claim 1, characterized in that: wherein The spatial state equation of the characteristic voltage-cycle number model is as follows: U k = a k C x 2 + b k C x + c k + n U,k , where v represents the transmission noise, n U,k represents the observation noise, the subscript k represents the parameter value at time k, v U1,K-1 , v U2,K-1 , and v U3,K-1 represent different transmission noises, respectively, The spatial state equation of the capacity-characteristic voltage model is as follows: Q k = m k * exp(n k * U) + p k * exp(q k * U) + n Q,k , where v represents the transmission noise, n Q,k represents the observation noise, the subscript k represents the parameter value at time k, v Q1,K-1 , v Q2,K-1 , v Q3,K-1 , and v Q4,K-1 represent different transmission noises, respectively.

3. The method for estimating the battery capacity and predicting the remaining useful life of a lithium ion battery according to claim 2, characterized in that: wherein Step S2-3 specifically comprises the following sub-steps: Step S2-3-1, obtaining an initial particle set through importance sampling The feature voltage-cycle number model is brought in, and the weight of each particle at each time is obtained Step S2-3-2, calculating the weight of each particle according to step S2-3-1 Resampling is performed to generate a new particle set At this time, the weight of each particle is 1 / N; Step S2-3-3, after resampling, the weight of each particle is the same, and the estimated value at different times is the average of the sum of each particle x k That is, the identified value of the model parameter.

4. The method for estimating the battery capacity and predicting the remaining useful life of a lithium ion battery according to claim 3, characterized in that: wherein Step S4 specifically comprises the following sub-steps: Step S4-1, assuming at the Sth cycle (i.e., k=S), the new particle set matrix generated after resampling is The N sets of parameters are brought into the characteristic voltage-cycle number relationship model to calculate the characteristic voltage vector after attenuation The characteristic voltage sets are then brought into the capacity-characteristic voltage model to calculate the capacity degradation vector According to the predicted capacity degradation vector and the mapping between the cycle number vector the cycle number vector is obtained: That is, the estimated number of cycles is in the range of Step S4-2, the weight of the N particles after resampling is 1 / N, and the weight of the particle is The interval is equally divided into W small intervals, and the frequency of the particles appearing in each small interval represents the probability of the particle appearing, denoted as: (pro1, pro2,..., pro w ) = (fre1, fre2,..., fre w ) / N, wherein pro (1,2,...,w) representing the probability of W intervals, fre (1,2,...,w) representing the frequency of W intervals, N is the number of particles, thereby completing the prediction of the remaining service life of the lithium battery under test.

5. The method for estimating the battery capacity and predicting the remaining useful life of a lithium ion battery according to claim 1, characterized in that: wherein The characteristic voltage U is a voltage value at a fixed time in a discharge stage in a cycle aging experiment, and the capacity Q is a discharge capacity when a cell voltage in the discharge stage drops to a discharge cutoff voltage.

Images