1. A method for estimating the residual electric quantity of a power lithium battery is characterized by comprising the following steps:

and (3) an off-line model training stage:

step 1: in the complete charging and discharging process of the power lithium battery, a plurality of groups of working data are selected as training data, and the data information contained in each group of training data comprises: the operating voltage, the operating current, the average temperature of the battery and the remaining capacity of the battery;

step 2: normalizing the training sample data in the step 1 to enable the training sample data to be in an interval of [ -1,1 ];

and step 3: set dynamic weighted least squares vector machine modelNuclear parameter of₀，

And 4, step 4: completing dynamic weighting least square vector machine model training by using training data so as to obtain the trained

An on-line estimation stage:

and 5: setting an initial sliding window length to P₀

Step 6: collecting the working voltage U, the working current I and the average temperature T of the battery on line, and inputting the voltage U, the working current I and the average temperature T into the trained batteryIn the model, outputting a corresponding SOC value; simultaneously turningU，I，T，SOC]Placing the window into an initial window and occupying one window width;

and 7: repeating step 6 until the width in the initial sliding window is occupied;

and 8: calculating variance of SOC variations in an initial windowΔ SOC represents the variation of SOC at two adjacent time points, where var represents the variance of Δ SOC

And step 9: suppose that the current k-th sliding window has a width P_kMeanwhile, normalizing the data in the k-1 th window; while assuming a variance of the SOC variation in the k-1 th window of

Step 10: assuming that the core parameter of the current latest dynamic weighted least square vector machine model is sigma_k-1；

Step 11: on the basis of step 10, the data in the k-1 window after normalization in step 9 is used to complete the training of the current dynamic weighted least squares vector machine model, and the model is recorded as

Step 12: collecting the working voltage, current and average temperature of the battery at the current moment t, and inputting the working voltage, current and average temperature into the modelAnd outputting a corresponding SOC value; at the same time, the [ U, I, T, SOC at the current time T]Placing the window in the current window and occupying one window width;

step 13: repeating the step 12 until the total number of data in the current window is equal to the width of the window;

step 14: establishing a nuclear parameter sigma_kThe equation of state of (c): sigma_k＝Fσ_k-1+v_kWherein the transition matrix F is a one-dimensional identity matrix,v_kis noise information;

will be provided withModel as a measurement equation for CKF, nuclear parameter σ_kThe state equation and the measurement equation form a filtering model of the CKF;

step 15: setting initial parameters of CKF, and searching for optimal kernel parameter sigma through CKF filtering algorithm_kThe parameters are used as kernel parameters of a dynamic weighted least square vector machine model in a (k + 1) th window;

step 16: for all SOC in the k window_tCalculating the variation amount [ Delta ] SOC^kThen calculate Δ SOC_tVariance of (2)

And step 17: by varianceSum varianceDetermining the width P of the next sliding window_k+1；

Step 18: using data in the kth window, and the Gaussian kernel parameter σ_kTraining a new WLSSVM model, and recording the model as

Step 19: and (6) repeating the step 9-the step 18, and stopping when no new data is input.

2. The method of claim 1, wherein the method comprises the steps of: width P of next sliding window in step 17_k+1The calculation is as follows:

3. the method of claim 1, wherein the method comprises the steps of: in the process of executing step 11, when solving the model, the key is to require the model parameter b_k-1And alpha_k-1；

Wherein the content of the first and second substances,is a process parameter when the model is solved, and N represents a matrix dimension;

when new data is added to the window, the new data information is used for updatingSuppose m new sets of data are added, Q at this time_NIs shown as

Wherein P is_N+mAndall are information obtained from new data, and then obtained according to the block matrix inversion method

Removing Q_NThe old data information about the last window is assumed to be L in total number; here will Q_N+mIs re-expressed as

Wherein P is_LAndare all old data information; obtained by solving the formula (4) and the formula (5)

Then, the parameter b of the model is obtained according to the formulas (2) and (3)_k-1And alpha_k-1Model of lawI.e. can be determined by means of such a dynamic matrix calculation.

Background

The power lithium battery is widely applied to the electric automobile because of the advantages of high density, low memory, environmental friendliness and the like. The battery residual capacity (SOC) is an important index for representing the charging and discharging State Of the battery, and the accurate SOC estimation can avoid the overcharge and the over-discharge Of the battery, thereby avoiding the battery fire event and prolonging the service life Of the battery. However, SOC cannot be directly measured by a sensor, and can only be indirectly estimated by a measurement quantity (such as current, voltage, temperature, and the like) strongly related to SOC, so that the estimation method becomes a key for solving the SOC estimation problem.

In recent years, due to rapid development of machine learning methods, data-driven SOC estimation methods attract a lot of attention, and the main idea of the methods is to perform deep mining on data information to build a mathematical model meeting engineering requirements, so that extracting accurate features from massive dynamic data to build a proper model becomes a key factor for improving SOC estimation. The service life of the power battery is limited, and at the same time, the internal characteristics of the battery are changed along with the advance of the service time of the battery, so that the establishment of a dynamic adaptive data-driven SOC estimation model becomes a key problem for improving SOC estimation.

The patent with publication number CN108872866A, "a lithium ion battery state of charge dynamic evaluation and long-term prediction fusion method", establishes two independent SOC estimation models by using extended kalman and neural network, and performs weighted fusion on the two independent SOC estimation models to obtain a final SOC value, however, since the scheme adopted in the patent is a fixed and unchangeable SOC estimation model, once the latest input data characteristics change, the established model will become inconsistent with the data characteristics. In the patent publication CN107132490A, "a method for estimating the state of charge of a lithium battery pack", a Support Vector Machine (SVM) model is built using small samples, and an optimal parameter of the SVM is found by using a particle swarm algorithm, and then used for SOC estimation. The same problem exists in that the support vector machine or particle swarm optimization process is staged, and an adaptive SOC estimation mechanism is not established.

Therefore, aiming at the problems that the SOC estimation is lack of adaptivity and the optimization process is staged, the SOC estimation model based on the dynamic support vector machine is provided, and the Capacity Kalman Filter (CKF) is used for assisting the adaptive update model parameters of the support vector machine, so that the long-term and effective SOC estimation model of the power lithium battery can be ensured to be established.

Establishing a dynamic adaptive SOC estimation model faces two special challenges: 1) when an SOC estimation model based on dynamic WLSSVM is established, when training data change, as the WLSSVM needs to spend a certain time in training, and the dynamic SVM faces frequent change of the training data, the re-training of the model undoubtedly increases the operation burden of the model; 2) the dynamic WLSSVM realizes the update iteration of the model by means of a sliding window, and under the sliding window mechanism, two problems are faced to how to select the length of the sliding window and how to find the optimal parameters for the WLSSVM model in the next window.

Disclosure of Invention

The invention aims to provide a long-term effective self-adaptive SOC estimation method, which aims to perform online estimation on internal SOC parameters of a lithium battery according to external collected data of the power lithium battery, provide accurate SOC estimation results for a battery management system and a user, increase the reliability of battery management and provide accurate information for the user to reasonably use the battery.

In order to achieve the purpose, the technical scheme of the invention is as follows:

the invention comprises an offline stage and an online stage, wherein the offline stage in the invention comprises:

step 1: in the complete charging and discharging process of the power lithium battery, a plurality of groups of working data are selected as training data, and the data information contained in each group of training data comprises: 1) a working voltage; 2) operating current; 3) the average temperature of the battery; 4) the remaining capacity of the battery.

Step 2: and (3) normalizing the training sample data in the step (1) to enable the training sample data to be in an interval of [ -1,1 ].

And step 3: WLSSVM model under set-up lineNuclear parameter of₀，

And 4, step 4: completing WLSSVM training by using training data to obtain trained WLSSVM

The on-line stage of the invention:

initial stage of on-line stage:

and 5: setting an initial sliding window length to P₀。

Step 6: collecting the working voltage (U), current (I) and average temperature (T) of the battery on line, and inputting the collected voltage, current and average temperature into the trained batteryAnd outputting a corresponding SOC value in the model. Simultaneously will [ U, I, T, SOC]Placed into the initial window and occupies one window width.

And 7: step 6 is repeated until the width in the initial sliding window is filled.

And 8: calculating variance of SOC variations in an initial windowΔ SOC represents the variation of SOC at two adjacent time instants, where var represents the variance of Δ SOC.

Initial post-on-line stage:

and step 9: suppose that the width of the current (i.e. kth window) sliding window is P_k(P_kIs a positive integer). While normalizing the data in the k-1 th window.

Assume that the variance of the SOC variance in the previous window isk denotes the window number.

Step 10: the core parameter of the current latest WLSSVM model is assumed to be sigma_k-1。

Step 11: on the basis of step 10, the data in the k-1 th window after normalization in step 9 is used to complete the training of the current WLSSVM model, and the model is recorded as

Step 12: collecting the working voltage, current and average battery temperature at the current moment t, and inputting the working voltage, current and average battery temperature into the WLSSVM model which is trainedAnd outputs the corresponding SOC value. Simultaneously will [ U, I, T, SOC]Is placed into the current window and occupies one window width.

Step 13: step 12 is repeated until the total number of data sets in the current window equals the width of the window.

Step 14: establishing a nuclear parameter sigma_kThe equation of state of (c): sigma_k＝Fσ_k-1+v_kWhere the transition matrix F is a one-dimensional identity matrix, k is 0,1,2 … indicating the current window number, σ_kAnd the WLSSVM model kernel parameters corresponding to the training data representing the kth window. Will be provided withAs a measurement equation for CKF, the kernel parameter σ_kThe state equation and the measurement equation form a filtering model for CKF.

Step 15: setting initial parameters of CKF, and searching optimal kernel parameter sigma through CKF filtering algorithm_k。

Step 16: for all SOC in the current window_tCalculating the variation amount [ Delta ] SOC^kThen calculate Δ SOC_tVariance of (2)

And step 17: according to

Determining the width P of the next sliding window_k+1。

Step 18: using the new nuclear parameter σ obtained in step 11_kTraining a new WLSSVM model with data from the kth window

Step 19: and (6) repeating the step 9-the step 18, and stopping when no new data is input. In the process of executing step 11, the specific steps are as follows:

retraining WLSSVM model on new data basis when data in a sliding window changesIn order to accelerate the training process, the invention provides a dynamic WLSSVM model solving method to accelerate the model solving process.

In solving forWhen modeling, the key is to require modeling parameters (b)_k-1，α_k-1)

Wherein the content of the first and second substances,is the process parameter when solving the WLSSVM model, N represents the matrix dimension. Obviously, the key to solving the pair of parameters is the requirement solutionIs Q_NThe inverse of (c).

When new data is added to the window, the new data information is used for updatingSuppose m new sets of data are added, Q at this time_NCan be expressed as

Wherein P is_N+mAndall are information obtained from new data, and then obtained according to the block matrix inversion method

Then, removing Q_NRegarding the old data information of the previous window, assume that the total number of the old data is L. Here will Q_N+mIs re-expressed as

Wherein P is_LAndare all old data information. Obtained by solving the formula (4) and the formula (5)Then, the parameters (b) of the model can be obtained from the equations (2) and (3)_k-1，α_k-1)

The new WLSSVM model can be determined by this dynamic matrix calculation.

Compared with the prior art, the invention has the following beneficial effects:

1. the modeling and the online estimation of the SOC estimation model of the power lithium battery are realized by adopting a weighted least square support vector machine with strong learning capacity, and a plurality of researches prove that the weighted least square support vector machine has more robustness compared with the original support vector machine.

2. The selection process of the kernel parameter model of the weighted least square support vector machine can be completed through online self-adaption by CKF.

3. And establishing a dynamic WLSSVM process, and establishing an SOC estimation model by using the latest working condition data so as to better reflect data characteristics.

4. The rapid solving method of the dynamic WLSSVM is provided, model parameters can be rapidly solved under the condition of facing brand new input data, and a new SOC estimation model is established.

Drawings

FIG. 1 is a flow chart of a method of the present invention;

FIG. 2 is a graph of validation results for an example;

fig. 3 is a graph of the verification results of the example.

Detailed Description

The technical scheme of the invention is specifically explained below with reference to the accompanying drawings.

The embodiment provides a method for estimating remaining capacity of a power lithium battery, as shown in fig. 1, including the following steps:

an offline stage:

step 1: in the complete charging and discharging process of the power lithium battery, a time interval between adjacent samples is 5s as a sampling criterion, a plurality of groups of battery working data are selected as training data, and each group of training data comprises data information as follows: 1) a working voltage; 2) operating current; 3) the average temperature of the battery; 4) the remaining capacity of the battery.

Step 2: and (3) carrying out normalization processing on the training data in the step (1) to enable the training data to be in an interval of [ -1,1 ].

And step 3: setting an initial nuclear parameter to σ₀When the kernel function of WLSSVM is

Wherein x is_iAnd x_jAre the ith and jth data in the training data set.

And 4, step 4: completing the training of the current WLSSVM model by using the training data after normalization in the step 2, and recording the trained model as

When the power lithium battery of the electric automobile is in the initial stage of charging or discharging working condition, the off-line training is utilized to completeThe model carries out SOC estimation on the line, and input data are working voltage on the line; operating current; average temperature of power lithium battery.

Initial phase of on-line phase:

the WLSSVM model is dynamically updated using a sliding window, which will be described in detail below, the sliding window mechanism, and how the WLSSVM model is quickly solved after the window is moved.

And 5: setting an initial sliding window length to P₀。

Step 6: collecting the working voltage (U), current (I) and average temperature (T) of the battery every 5 seconds on line, and inputting the working voltage (U), current (I) and average temperature (T) to the trained batteryAnd outputting a corresponding SOC value in the model. Simultaneously will [ U, I, T, SOC]Placed into the initial window and occupies one window width.

And 7: step 6 is repeated until the width in the initial sliding window is filled.

And 8: calculating variance of SOC variations in an initial window

ΔSOC_t＝SOC_t+1-SOC_t,t＝0…P₀ (8)

Where t represents the time of acquisition.

Initial post-on-line stage:

and step 9: assuming that the current is in the kth window, normalizing the data (working voltage, working current, average temperature of the power battery and residual battery capacity) in the kth-1 th window to the interval [ -1,1 ].

Assume that the variance of the SOC variance in the previous window isAnd the width of the current (i.e. kth window) sliding window is p_k(p_kIs a positive integer).

Step 10: setting the core parameter of the current latest WLSSVM model as sigma_k-1In the present embodiment, a gaussian kernel function is adopted:

wherein x is_iAnd x_jAre the ith and jth data in the kth-1 window.

Step 11: on the basis of step 10, the data in the k-1 th window after normalization in step 9 is used to complete the training of the current WLSSVM model, and the model is recorded as

Step 12: collecting the working voltage (U), current (I) and average temperature (T) of the battery at the current moment, and inputting the working voltage, current and average temperature to the batteryAnd outputs the corresponding SOC value in the model, and then,the [ U, I, T, OOC of the current time]Is placed into the current window and occupies one window width.

Step 13: step 12 is repeated until the width in the current window is full.

Step 14: establishing a nuclear parameter sigma_kThe equation of state of (c): sigma_k＝Fσ_k-1+v_kWhere the transition matrix F is a one-dimensional identity matrix, k is 0,1,2 … indicating the current window number, will beThe mathematical model, which is a measurement equation for CKF (equation 10), then combines the state equation and the measurement equation into a filtered model for CKF.

Wherein the content of the first and second substances,

wherein alpha is_k-1And b_k-1Is thatSolution of the model, x_i ^(k)Andis the data in the k window, v_kAnd ω_kIs noise information, SOC_kIs a column vector, u, composed of all SOC values in the k-th window in the order of generation_kIs the input to the measurement equation, in this example the matrix of U, I, T in the kth window.

Step 15: starting CKF algorithm to obtain sigma_kAs a result of the estimation of (2), as a model of WLSSVM in the k +1 th windowThe nuclear parameters of (1).

Step 16: for all SOC in the k window_tCalculating the variation amount [ Delta ] SOC^kThen calculate Δ SOC_tVariance of (2)

And step 17: according to

Determining the width P of the next sliding window_k+1。

Step 18: using data in the kth window, and the Gaussian kernel parameter σ_kTraining a new WLSSVM model, and recording it as

Two key model parameters need to be solved during the execution of step 11, b_k-1And alpha_k-1. The process of solving these two parameters is:

wherein omega_i,j＝K(x_i,x_j) N denotes that the matrix is an N-dimensional square matrix; y ═ y₁ … y_N]；α＝[α₁ … α_N]；V is the weight matrix of the WLSSVM,are elements in the weight matrix.

Definition of Q_N＝Ω_N+ V, as can be seen from the above equation, solving for α_k-1And b_k-1Can be determined by the following formula:

where N is the total number of data in the training set.

Generating new information for updating Q by using data in current window_N. Assume that the number of new data is m, and m is p_k. Then there is

Wherein, P_N+m＝[P_N+1 P_N+2 … P_N+m]

Wherein, P_N+i＝K(x_N+i，x_j)，i＝1，…，m，j＝1，…，N+i-1

Wherein v is_N+1，…，v_N+mIs a new weight matrix value brought by the new data information.

According to the block matrix inversion method, the following results can be obtained:

wherein the content of the first and second substances,

then the information of the old data in the last window is transferred from Q_N+mThe number of old data is assumed to be L, and L is equal to p_k-1. Then there is

Wherein the content of the first and second substances,P_t＝K(x_i，x_j) I-1, …, L, j-i, …, N wherein,

according to equation (5), the inverse knowledge of the blocking matrix can be obtained:

wherein the content of the first and second substances,

thereby, it is possible to deduce:

then, according to the formula, it can be calculatedThe model solution of (2):

step 19: and repeating the steps 9-18 until no new data appears.

The simulation results are shown in fig. 2 and 3. Data is provided by Hangzhou Jie code energy science and technology limited, real SOC (real-SOC) is an SOC curve generated by a vehicle battery management system in the driving process of a power lithium battery which is just put into use on an urban road, and Estimated SOC is a result Estimated by an algorithm.

Fig. 2 is the estimated effect of WLSSVM, and fig. 3 is the estimated effect of dynamic WLSSVM-CKF in the present invention. It can be seen that the Estimated curve in fig. 3 differs less from the true SOC curve.