Disclosure of Invention
The invention overcomes the defects of the prior art, and provides the method for evaluating the SOH of the battery of the electric vehicle under the multi-time scale, which combines the data under the multi-time scale between the macroscopic time scale and the microscopic time scale to ensure that the microscopic model is iterated step by step along with the increase of the charging times, and simultaneously combines the data of a battery laboratory and the evaluation of actual test data to ensure the accuracy and the real-time property of the SOH evaluation of the battery.
The technical scheme of the invention is as follows:
an SOH (State of health) evaluation method for an electric vehicle battery under multiple time scales comprises the following steps:
step a, acquiring nth charging segment data d (t, U, I) of the ith battery;
b, budgeting the available capacity for the nth charging based on the improved iEKF-GPR;
step c, calculating the battery to obtain the health state of the nth charging;
d, acquiring nth SOH historical data of the m batteries, and clustering the data;
step e, calculating the uncertainty J and the uncertainty index A of the current budget;
f, judging whether A is less than Amin or not, and if so, determining the budget reliability and the cost to be added into the historical training data; if not, the budget result is unreliable, and the actual full charge test is carried out to obtain full charge data;
step g, judging whether the actual test result is A < Amin, if so, updating the initial iteration curve, and executing the step b; if not, the battery is repaired or replaced when the battery is in a problem.
Further, the method for acquiring nth charging segment data d (t, U, I) of the ith battery comprises the following steps:
determining a battery primary cycle loop0Constant current charging current I, constant voltage charging cut-off voltage U, full charging data d under initial constant current charging0=(t0(k),U0(k)),k=1,2,...,k0,k0For the total number of sampling time points, t, at which the battery reaches a cut-off voltage V under constant current I charging0(k) Is a discrete relative time of equal-spaced sampling, the sampling time interval Ts=t0(k+1)-t0(k) Is constant, U0(k) Represents the voltage at the kth sample point; based on the following formula, the measurement function is h0;
Nth daily charging segment data dn=(tn(k),Un(k),In(k) Since the absolute time of the clip data time starts from 0,
tk=0,Ts,2Ts,3Ts…(k-1)Ts。
further, the method for budgeting the available capacity for the n-th charging time based on the improved iEKF-GPR comprises the following steps:
the iEKF-GPR model budget was improved for n times of fragment data:
the state equation is as follows:
the measurement equation is as follows:
wherein
Is the charging time;
charging voltage value for charging time; GP
V(n) the difference from the conventional assumed white gaussian noise, the measured noise is the white gaussian noise, but the mean V is regressed by the gaussian process, the variance is Rn, and w (k) is the white gaussian noise with the mean 0 and the variance Q predicted according to the charging curve;
and (3) state prediction: u (k | k-1) ═ GPf(U(k-1|k-1))
Measurement prediction: t (k | k-1) ═ h (U (k | k-1)) + V
A state transition matrix, approximating the derivative with the difference quotient: phi (k) ═ U (k | k-1) -U (k-1| k-1)
A measurement matrix, approximating the derivative with the difference quotient:
covariance prediction matrix:
and (3) calculating gain:
and (3) updating the state:
and (3) updating the covariance:
obtaining the absolute time corresponding to the starting voltage U, the required time Tn of the full charge of the nth charge, and the available capacity of the nth constant current charge as follows:
wherein I is the current value of constant current charging of the battery.
Further, the method for calculating the state of health of the battery for the nth charging comprises the following steps:
according to the linear correlation of the charge capacity and the discharge capacity, obtaining the relation between the charge capacity and the discharge capacity of the linear regression research;
Cd(n)=β1Cc(n)+β0
wherein the coefficient beta1And beta0Calculated according to the following formula;
the factor that influences the estimation error and the relative trend of the state of health of the battery greatly based on the charging data and the discharging data is the determination of the initial capacity, and the state of health SOH of the battery at the nth charging is as follows:
wherein C isd(n) is the calculated available discharge capacity, CNIs the nominal discharge capacity.
Further, the data clustering method comprises the following steps:
step d1, determination
Data wherein the state of health of the different batteries at the nth charge
m represents the number of each battery, and m clusters are divided into 5 clusters C1-C5 according to influence characteristics;
step d 2: randomly distributing 5 records to become the center position of the initial cluster;
step d 3: finding a nearest cluster center for each data, each cluster center "owning" a subset of the data, thereby determining a partition of the data set;
step d 4: for each of the 5 clusters, finding a cluster centroid and updating the cluster centroid with the new cluster center position;
step d 5: and repeating the steps d 3-d 5 until convergence or termination.
Further, the method for calculating the uncertainty J of the current budget includes:
clustering is carried out on the data to obtain the statistical value of the battery health state when the battery is charged for the nth time, namely the cluster center
The nth charging budget state of health value of the battery with the number i
Clustering the budget data into a cluster with the nearest distance, carrying out next step statistic on the cluster, and obtaining intra-cluster merging standard deviation according to the following formulaFor poor use
Represents;
the uncertainty of the budget is as follows:
wherein,
is the current budget value;
the cluster center closest to the budget value is taken as the cluster center;
the standard deviation is obtained by k cluster statistics; s
pTo consider a budget analysis uncertainty factor.
Further, the method of uncertainty index a comprises: further obtaining an uncertainty index A, optimizing those that have not met the Gaussian distribution
Obeying a probability density function of
Distribution of (2) for visual evaluation
Performing mapping calculation to obtain a numerical value which is easier to directly compare;
compared with the prior art, the invention has the following beneficial effects:
the invention provides a method for evaluating SOH of an electric vehicle battery under multiple time scales, which combines data under multiple time scales with macroscopic time scales and microscopic time scales, wherein the data comprises battery charging working voltage, battery charging time, battery charging times and battery number, a macroscopic data model is used as an iterative convergence condition of the microscopic data model at a 4-dimensional angle, so that the microscopic model is ensured to be iterated step by step along with the increase of the charging times, and meanwhile, the accuracy and the real-time performance of the SOH evaluation of the battery are ensured by combining battery laboratory data and actual test data evaluation.
The invention utilizes the historical charging data of the battery packs of the same type to establish an iterative model which changes in real time according to the actual battery performance, perfects the estimation model in real time, enables the model to be close to the current working state of the battery pack and improves the estimation precision.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings.
Detailed description of the invention
An SOH evaluation method for an electric vehicle battery under multiple time scales is shown in fig. 1, and comprises the following steps:
step a, acquiring nth charging segment data d (t, U, I) of the ith battery;
b, estimating the available capacity of the nth charging based on the improved iEKF-GPR;
step c, calculating the battery to obtain the health state of the nth charging;
d, acquiring nth SOH historical data of the m batteries, and clustering the data;
step e, calculating the uncertainty J and the uncertainty index A of the estimation;
f, judging whether A is less than Amin or not, if so, determining the recomputation reliability price to be added into the historical training data; if not, the estimation result is unreliable, and the actual full charge test is carried out to obtain full charge data;
step g, judging whether the actual test result is A < Amin, if so, updating the initial iteration curve, and executing the step b; if not, the battery is repaired or replaced as a problem.
Specifically, Amin is a minimum threshold for uncertainty;
specifically, the method for acquiring nth charging segment data d (t, U, I) of the ith battery comprises the following steps:
determining a battery primary cycle loop0The current time t, the constant-current charging current I, the constant-voltage charging cut-off voltage U, and the full-charging data d under the initial constant-current charging0=(t0(k),U0(k)),k=1,2,...,k0,k0For the total number of sampling time points, t, at which the battery reaches a cut-off voltage V under constant current I charging0(k) Is a discrete relative time of equal-spaced sampling, the sampling time interval Ts=t0(k+1)-t0(k) Is constant, U0(k) Represents the voltage at the kth sample point; based on the following formula, the measurement function is h0;
Wherein U1, U2 and U3 respectively represent three inflection point voltage values of a charging curve, a1, a2, a3, c1, c2 and c3 respectively represent model coefficients of a piecewise function, and a4 represents a gradually-increased value of initial voltage in the aging process of the battery
Nth daily charging segment data dn=(tn(k),Un(k),In(k) Since the absolute time of the clip data time starts from 0,
tk=0,Ts,2Ts,3Ts…(k-1)Ts
specifically, the improved iEKF-GPR comprises the following components:
a GPR method is used for identifying a system model, and the system model comprises a state transition model, a measurement model and a corresponding noise covariance matrix to replace or enhance an original system model, so that the state estimation problem under the condition that the system model is unknown or is not accurate enough can be solved. The establishment of the parameterized model requires a great deal of professional knowledge, and some systems can only establish simplified models, and the parameterized models are difficult to complete and represent actual systems. The GPR model can overcome the defects of a parameterized model in use, and the model precision is improved through the training of multidimensional data, because the GPR model can represent system information which cannot be captured by the parameterized model; secondly, the generalization performance is enhanced. Essentially, the GPR model characterizes the residual output portion of the system outside the contribution of the isolated parametric model. Therefore, on the basis of the proposed iEKF-GPR model under the microscopic time scale, the GPR model is built by using the measurement noise in the EKF under the macroscopic time scale, so that the model real-time performance of the model is improved.
According to the following formula,
tk=h(Uk)+V(k)
wherein, h (U)k) Is a function of the U-t measurement equation, and V (k) is the measurement noise;
measurement noise V (k) in the iEKF-GPR model measurement equation can be interpreted as deviation of each battery charging curve according to the physical meaning of battery charging, iteration is carried out in the iEKF-GPR through the characteristic that the adjacent charging curves of the batteries do not have mutation, but the assumption ignores short-time fluctuation of the available charging capacity of the batteries, and simultaneously increases the risk of estimation dispersion, so the measurement noise V (k) is introduced into the noise influence quantity of the battery capacity fluctuation under the macroscopic time scale. Combining the following equations:
obtaining the whole measurement equation:
wherein, U2 and U3 respectively represent inflection point voltage values of the charging curve; Δ U1 represents a first plateau voltage differential; vn represents the change in the charging time in the n-times charging process. As shown in fig. 2, the variation of the battery charging time in a macroscopic time scale is shown,
and (3) in a GPR-based SOH prediction fitting model, selecting a neural network kernel function and a Maternard kernel function to be added to determine a GPR covariance function. The definition of SOH is given by the formula:
wherein, C'MTo measure discharge capacitance, C'NA nominal discharge capacitance for the battery;
the SOH of the battery is directly estimated by adopting the battery capacity, namely the SOH of the battery is in positive correlation with the charging and discharging time of the battery, so that a model for predicting the SOH is adopted to carry out prediction fitting on the charging time of the battery.
The state equation and the measurement equation of the improved Gauss-Kalman filtering model are respectively
xk=GPf(x1,x2....xk-1)+W(k)
zk=h(xk)+GPV(n)
Wherein the parameter expression is consistent with the following formula,
Uk=GPf(U1,U2....Uk-1)+W(k)
tk=h(Uk)+V(k)
GPV(n) represents the amount of noise influence caused by battery charging time fluctuations. In addition, GPfAnd GPVAre all non-linear predictions based on GPR, but the physical meanings are completely different, GPfThe next voltage value in the single charge curve is predicted at the microscopic time scale, and GPVThe fluctuation situation of the whole charging time is predicted under the macroscopic time scale.
The daily charging data cannot achieve full charging and full discharging of the electric quantity of the lithium battery, so that the real available charging capacity of the lithium battery cannot be directly obtained. Under the change condition of SOH under the macroscopic time scale, the iEKF-GPR model is improved, and the influence of the battery capacity fluctuation on the model is further considered.
Specifically, the method for estimating the available capacity of the nth charging based on the improved iEKF-GPR comprises the following steps:
an improved iEKF-GPR model estimation was performed on the n fragment data:
the state equation is as follows:
the measurement equation is as follows:
wherein k represents a discrete time,
is the state noise, h is the measurement function,
is the charging time;
charging voltage value for charging time; GP
V(n) the difference from the conventional assumption that the measured noise is white gaussian noise, but the mean V is regressed by gaussian process and the variance is Rn, and w (k) is white gaussian noise with
mean 0 and variance Q predicted by charging curve.
And (3) state prediction: u (k | k-1) ═ GPf(U(k-1|k-1))
Measurement prediction: t (k | k-1) ═ h (U (k | k-1)) + V
A state transition matrix, approximating the derivative with the difference quotient: phi (k) ═ U (k | k-1) -U (k-1| k-1)
A measurement matrix, approximating the derivative with the difference quotient:
covariance prediction matrix:
and (3) calculating gain:
and (3) updating the state:
and (3) updating the covariance:
in the above formula, H is the measurement matrix, P is the covariance prediction matrix, K is the gain, R is the measurement noise variance, HTTranspose for H;
obtaining the absolute time corresponding to the starting voltage U, the required time Tn of the full charge of the nth charge, and the available capacity of the nth constant current charge as follows:
wherein I is the current value of constant current charging of the battery.
Specifically, the method for calculating the state of health of the battery in the nth charging comprises the following steps:
according to the linear correlation of the charge capacity and the discharge capacity, obtaining the relation between the charge capacity and the discharge capacity of the linear regression research;
Cd(n)=β1Cc(n)+β0
wherein, Cd(n) is the discharge available capacity, coefficient beta1And beta0As calculated according to the following formula,
wherein x is
iAnd y
iEstimation of parameters of a linear regression model for random variables
Is a random variable y
iA function of (a);
the factor that influences the estimation error and the relative trend of the state of health of the battery greatly based on the charging data and the discharging data is the determination of the initial capacity, and the state of health SOH of the battery at the nth charging is as follows:
wherein C isd(n) is the calculated available discharge capacity, CNIs the nominal discharge capacity.
Specifically, SOH big data information of the same type of battery is introduced, clustering analysis is carried out on current battery data, an unsupervised learning k-means clustering algorithm is adopted, the method is a simple and effective algorithm for finding data clusters, the robustness of the system is further improved by establishing different clusters, and the method comprises the following steps:
step d1, determination
Data wherein the state of health of the different batteries at the nth charge
m represents the number of each battery, and m clusters are divided into 5 clusters C1-C5 according to influence characteristics;
step d 2: randomly distributing 5 records to become the center position of the initial cluster;
step d 3: finding a nearest cluster center for each data, each cluster center "owning" a subset of the data, thereby determining a partition of the data set;
step d 4: for each of the 5 clusters, finding a cluster centroid and updating the cluster centroid with the new cluster center position;
step d 5: and repeating the steps d 3-d 5 until convergence or termination.
Specifically, the method for calculating the uncertainty J of the current estimation includes:
clustering is carried out on the data to finally obtain the statistical value of the battery health state when the battery is charged for the nth time, namely the cluster center
The estimated state of health value for the nth charge of the battery numbered i
Clustering the estimated data into a cluster with the nearest distance, carrying out next step statistic on the cluster, and obtaining the standard deviation for intra-cluster combination according to the following formula
Represents;
wherein s is
pDenotes the combined standard deviation, V
iThe degrees of freedom of each type are represented,
representing M similar types of prediction results or training data;
the estimated uncertainty is:
wherein,
is the current estimated value;
cluster center closest to the estimate;
the standard deviation is obtained by k cluster statistics; s
pRepresents the combined standard deviation.
Specifically, the method for the uncertainty index a comprises: further obtaining an uncertainty index A, optimizing those that have not met the Gaussian distribution
Obeying a probability density function of
Distribution of (2) for visual evaluation
Performing mapping calculation to obtain a numerical value which is easier to directly compare;
wherein, y
pDenotes the predicted estimate, J
nThe degree of uncertainty is represented by a number of,
represents the ith process output predicted by the ith GPR model,
and
respectively representing the mean and variance of the ith prediction output,
the standard deviation is obtained by k cluster statistics; s
pWhich represents the combined standard deviation of the signals,
represents the statistically derived standard deviation of the k clusters,
the system estimates the variance;
specifically, the reliability of the estimation result and the model optimization space are judged through the A purpose, and the quality of the battery also needs to be judged, namely whether further test and estimation are needed is judged through the uncertainty index, and the uncertainty index A is determined to judge the threshold value
Because the current battery full charge and full discharge test is still the most stable and reliable battery SOH evaluation method, when the uncertainty index A is too high, the battery needs to be subjected to the full charge test to obtain a relatively accurate battery SOH result for updating an iteration initial state and training data, but if actual test data and an estimation result tend to be consistent, the estimation accuracy and the battery are further explained, and the battery cannot be used any more and further maintenance and replacement work is required.
Detailed description of the invention
On the basis of the first specific implementation mode, 20 electric buses of the same type K9B purchased by a certain bus company at the same time are selected, and all the buses operate urban lines, so that the working conditions are basically the same. The battery health state evaluation is carried out on the power battery of the electric bus, the K9B battery pack is formed by combining 63 battery modules, each module is formed by connecting 8 single batteries in series, the 63 modules are divided into 3 relatively independent battery modules, and the 3 battery modules are connected in parallel to form a finished battery pack. The basic characteristics of the power battery pack are as follows:
TABLE 5-1 Power Battery pack basic characteristics
The online evaluation model of the battery health state algorithm is verified by the detection data and the daily charging data, and as shown in fig. 3, a scene photo of battery detection is taken for the electric bus. The testing equipment adopts a charging and discharging integrated machine produced by Sinexcel, and the technical indexes are that the voltage precision is 0.5 percent and the current precision is 1.0 percent.
The developed online capacity and SOH evaluation software of the electric bus analyzes and mines data of the electric bus, and a software display interface is shown in FIG. 4.
And carrying out a first full charge capacity test on 12/3/2013 after delivery, and carrying out online battery health state evaluation on the charging data of the 2014/10/3/based on full charge data and charging historical data by adopting a battery SOH evaluation model under multiple time scales, wherein the initial SOC display value of the battery before charging is 36%. Fig. 5 shows the full charge curve of the battery 2013 at 12/3 and the estimated battery curve at 2014 at 3/10. Finally, the available charging capacity of the bus battery is estimated on line in 2014 within 3, 10 days as follows:
Ck=tk×I=564Ah
thus, the current battery state of health is calculated as:
the electric bus is subjected to full battery charge tests for 3 times between 2013 and 2017, and the actually measured available capacity is shown in a table 5-2.
TABLE 5-2 actual measured Capacity data
Full charge test date
|
Test results
|
2013-12-06
|
562.3Ah
|
2016-05-09
|
522.1Ah
|
2017-10-29
|
472.8Ah |
Selecting charging historical data of which the initial SOC of the electric bus is lower than 40% and which is charged to 100% to perform online evaluation on the battery, wherein charging capacity estimation data of charging data of a segment of the electric vehicle battery with the full charging capacity test interval not exceeding 5 days are selected for comparison. See tables 5-3.
TABLE 5-3 Multi-time Scale Capacity estimation results
Due to the fluctuation of the battery capacity, the accuracy of the battery charging capacity cannot be reproduced, the battery capacity attenuation degree cannot be suddenly changed under the condition that the battery normally works, the fluctuation of the battery capacity in normal work is generally not higher than 2%, the capacity fluctuation is below 6% by comparing the estimated data with the actual data, and the accuracy of the initial estimated capacity estimation is better than 8%. The linear estimation result is obtained by the three-time data SOC, and the result is shown in the table 5-4, so that the discreteness of the whole linear prediction result caused by the accuracy uncontrollable property of the SOC can be seen.
TABLE 5-4 SOC Linear estimation results
122 groups of segment charging data are selected from 2013 to 2017 to verify the SOH (state of health) evaluation algorithm of the multi-time scale battery, and FIG. 6 shows the evaluation result of the state of health of the battery.
The invention firstly defines the definition of the micro time scale and the macro time scale in the battery charging process, and further provides a battery SOH evaluation method under multiple time scales, wherein the combination of the macro time scale and the micro time scale comprises the battery charging working voltage, the battery charging time, the battery charging times and the battery number, and a 4-dimensional angle uses a macro data model as the iteration convergence condition of the micro data model to ensure that the micro model is iterated step by step along with the increase of the charging times, and simultaneously combines the battery laboratory data and the actual test data evaluation to ensure the accuracy and the real-time performance of the battery SOH evaluation.
Due to the difference between the lithium battery monomer and the lithium battery pack, the feasibility of evaluating the SOH of the battery of the electric vehicle in use based on the big data of the charging facility is demonstrated by analyzing the result that the measured data of the charging facility is superior to the measured result of the battery management system.
Providing an SOH (State of health) online evaluation model of the battery of the electric vehicle under the multi-time scale, and establishing a GPR (general purpose function) model by using measurement noise in EKF under the macro-time scale on the basis of the iEKF-GPR model under the micro-time scale, so that the model real-time property of the model is improved; simultaneously, SOH big data information of the same type of battery is introduced, clustering analysis is carried out on the current battery data, and the robustness of the system is further improved by establishing different clusters; and obtaining an uncertainty index by combining the results to optimize the reliability of the objective judgment estimation result and the model optimization space, wherein the quality of the battery also needs to be judged, and whether further test and estimation are needed is judged by the uncertainty index.
When the charging data analysis and the actual capacity test result of 20 buses are carried out, the analysis on-line SOH estimation error can be controlled within 8%.