CN110346734B - Machine learning-based lithium ion power battery health state estimation method - Google Patents

Abstract

The invention discloses a machine learning-based lithium ion power battery state of health estimation method, which is used for estimating the state of charge and the state of health of a power battery in real time. The method comprises the steps of establishing an equivalent circuit model of the lithium ion battery, carrying out parameter identification on the equivalent circuit model, then establishing a Uoc-SOC model, and estimating the SOC. And training by using a large amount of off-line data to obtain a neural network model with the parameters of the Uoc-SOC model as input and the maximum available capacity as output. And performing curve fitting on the Uoc and the SOC at the same moment to obtain parameters to be identified in the model, inputting the parameters to be identified into the trained neural network model to obtain the maximum available capacity, returning the obtained parameters of the Uoc-SOC model and the maximum available capacity to the SOC estimation step, and updating the parameters of the state equation and the observation equation. The invention provides a lithium ion battery state of health estimation method, which carries out online estimation on the state of health of a battery, updates parameters of SOC estimation and improves the estimation precision of the SOC estimation.

Description

Machine learning-based lithium ion power battery health state estimation method

Technical Field

The invention relates to the technical field of state estimation of a power battery management system of an electric automobile, in particular to joint estimation of the state of charge and the state of health of a power battery.

Background

With the gradual depletion of global petroleum resources and the increasing severity of environmental pollution, electric automobiles are receiving attention as an energy-saving, environment-friendly and sustainable development vehicle. The performance of the power battery pack as a power source of the electric automobile is always the focus of research. When the battery of the electric automobile is used, the battery needs to work in reasonable voltage, current and temperature ranges. Therefore, there is a need for effective management of the use of electric batteries on electric vehicles. A specific device for managing the Battery in the electric vehicle is a Battery Management System (BMS). The method not only ensures the safe and reliable use of the battery, but also ensures the full exertion of the capacity of the battery and prolongs the service life of the battery, is used as a bridge for communication between a battery vehicle controller and a driver, controls the charging and discharging of the battery pack, and reports basic parameters and fault information of the power battery system to the vehicle controller. The level of the battery management system determines to a large extent the performance of the power battery pack. Therefore, a real-time, efficient battery management system is very important.

A Battery Management System (BMS) is an important link for the development of a new energy automobile power system assembly and a large-scale energy storage system. The state of charge (SOC) of a battery is used to characterize the remaining capacity of the battery, i.e., the percentage of remaining capacity to rated capacity. The state of charge (SOC) of the battery cannot be obtained directly from the battery itself, but can only be indirectly estimated by measuring external characteristic parameters (e.g., voltage, current, etc.) of the battery pack. When the power battery of the electric automobile is used, the battery characteristics show high nonlinearity due to internal complex electrochemical reaction, so that the accurate estimation of the state of charge (SOC) of the battery has great difficulty. Nonlinear kalman filtering (e.g., extended kalman filtering, unscented kalman filtering, etc.) for dealing with the nonlinear problem is considered for estimating SOC. When using these algorithms, the SOC and Uoc relationships are involved, as well as the maximum available capacity of the battery, and current methods generally do not consider the variation, and default to a constant value. However, in practice, as the battery ages, the relationship between SOC and Uoc changes, and the maximum available capacity of the battery also changes. The estimation error of the SOC becomes larger and larger if these changes are not updated in time.

SOH refers to the state of health of the battery, i.e., the ratio of the current maximum available capacity to the initial maximum available capacity. As the battery is used for a longer time, the battery is gradually aged, and phenomena such as increase in internal resistance and decrease in battery capacity occur. The reason for the decline of the battery capacity is complicated, involves many factors, and changes slowly. Currently there is no accurate regression physical model. When a model is established by using machine learning, the selection of the health factor has a great influence on the final precision of the model, so that the selection of the proper health factor is very important.

Disclosure of Invention

In order to solve the problems in the prior art, the SOC and the SOH of the battery are jointly estimated, and the SOC estimation precision is improved and the estimation error is reduced by updating the coefficient and the maximum possible capacity of the relational expression of the Uoc and the SOC in real time; and taking the coefficient of the relationship between the Uoc and the SOC as a health factor as the input of the BP neural network model. The method has great significance for estimating and controlling the state of the whole battery, prolonging the service life of the battery and giving full play to the capacity of the battery.

The invention adopts the following technical scheme to realize the technical purpose.

A lithium ion power battery state of health estimation method based on machine learning includes the following steps:

step (1), establishing an equivalent circuit model of the lithium ion power battery, and identifying unknown parameters in the model;

step (2), establishing a Uoc-SOC model:

wherein a is_i、b_i、c_iIs the parameter to be identified in the model, U_OCRepresenting the open circuit voltage of the battery, and SOC is the state of charge of the battery;

step (3), estimating the SOC;

step (4), curve fitting is carried out on the off-line data to obtain a_i、b_i、c_iA value of (d);

step (5), for a_i、b_i、c_iAnd the maximum available capacity C is subjected to normalization treatment to obtain a'_i、b'_i、c'_iAs input, C 'as output, adopting machine learning algorithm pair a'_i、b'_i、c'_iAnd C 'corresponding to the alpha-'_i、b'_i、c'_iA machine learning model for input and C' for output;

step (6), performing curve fitting on the Uoc and the SOC at the same time to obtain an input value a_i、b_i、c_iAfter normalization, inputting the normalized data into a machine model to obtain the maximum available capacity C;

step (7) of subjecting a obtained in step (6) to_i、b_i、c_iAnd C value returns to the step (3), and corresponding parameters in the state equation and the observation equation are updated.

Furthermore, the equivalent circuit model is a Thevenin equivalent circuit model or a second-order RC equivalent circuit model.

Further, the Uoc-SOC model uses U_OC＝K₀+K₁z+K₂z²+K₃z³+K₄z⁴+K₅z⁵+K₆z⁶Or

Or

Or

Alternatively, where z is the state of charge SOC, K₀、K₁、K₂K₃、K₄、K₅、K₆、α₁、α₂Is the parameter to be identified in the model.

Further, the SOC is estimated by using an extended Kalman filter algorithm, and the polarization voltage and the state equation of the SOC of the battery are

Wherein U is_p、R_p、C_pRepresenting the polarization voltage, resistance and capacitance of the battery, eta is coulombic efficiency, delta T is sampling time interval, I_LFor charging and discharging current of battery, omega_kIs the system noise; an observation equation of terminal voltage is

Wherein upsilon is_kTo measure noise.

The invention has the beneficial effects that: the lithium ion power battery state of health estimation method based on machine learning provided by the invention can update parameters in a Uoc-SOC model and the maximum available capacity of the current battery in real time when the SOC is estimated by adopting a model method, and improves the estimation precision of the SOC. In addition, the parameters in the Uoc-SOC model are used as the health factors of the data-driven health state estimation method, so that the estimation precision is high; and when in combined estimation, the parameters in the Uoc-SOC model in the estimated SOC are updated, and the updated parameters are used as the input of SOH estimation, so that the SOC estimation and the SOH estimation share the same set of parameters, and the calculation amount is reduced.

Drawings

FIG. 1 is a schematic diagram of an equivalent circuit of a lithium ion power battery model;

fig. 2 is a flow chart of the lithium ion power battery state of health estimation method based on machine learning according to the present invention.

Detailed Description

The following will further describe the embodiments of the present invention with reference to the drawings, but the scope of the present invention is not limited thereto.

A lithium ion power battery state of health estimation method based on machine learning includes the following steps:

step (1), an equivalent circuit model of the lithium ion power battery is established, the Thevenin equivalent circuit model or the second-order RC equivalent circuit model can be selected, and the embodiment takes the Thevenin equivalent circuit as an example (as shown in fig. 1), wherein U is_OCRepresents the open circuit voltage, U, of the battery_tRepresents the terminal voltage of the battery, R₀Is the ohmic internal resistance, U, of the battery_p、R_p、C_pRepresenting the polarization voltage, resistance and capacitance of the battery; i is_LThe battery is charged and discharged. According to the circuit schematic diagram in fig. 1, the equivalent circuit model is analyzed by using an electrical engineering theory, and a continuous time equation of the battery model is established:

U_t＝U_oc-I_LR₀-U_p (1)

the transfer function is obtained by laplace transform:

bilinear transformation of

To obtain

Order to

Equation (4) can be simplified as:

U_L，k＝a₁U_L，k-1+U_OC，k-a₁U_OC，k-1+a₂I_L，k+a₃I_L，k-1 (5)

and because the sampling time T is very short, then:

ΔU_OC，k＝U_OC，k-U_OC，k-1≈0 (6)

equation (6) can be simplified to:

U_t，k＝(1-a₁)U_OC，k+a₁U_t，k-1+a₂I_L，k+a₃I_L，k-1 (7)

then, using a recursive least squares method for parameter identification, equation (7) can be written as:

wherein phi_Ls，kIs the data matrix of the system, theta_Ls，kIs a parameter matrix.

Step (2), establishing a Uoc-SOC model:

adopting a battery test cabinet to perform a cycle life test on a certain 18650 type lithium ion battery to obtain corresponding SOC and Uoc values under different cycle times, and performing curve fitting on the obtained data to obtain a relationship of Uoc-SOC: wherein a is_i、b_i、c_iThe parameters to be identified in the model are changed along with the change of the battery state of health, and can be used as health factors for representing the state of health.

The Uoc-SOC model may also employ the following model:

U_OC＝K₀+K₁z+K₂z²+K₃z³+K₄z⁴+K₅z⁵+K₆z⁶ (11)

wherein z is the state of charge SOC, K₀、K₁、K₂K₃、K₄、K₅、K₆、α₁、α₂The parameters to be identified in the model, which change with the change of the battery state of health, can also be used as the health factors for representing the state of health.

Taking formula (9) as an example, taking n as 3, fitting the n to obtain the following parameters at different cycle numbers:

number of cycles	a1	b1	c1	a2	b2	c2	a3	b3	c3
										1	3.936	1.242	0.8202	0.095	0.5856	0.1909	3.004	-0.2176	0.7787
25	2.985	1.196	0.3515	2.863	0.7151	0.4342	3.12	0.04232	0.5155
										50	3.51	1.196	0.4159	2.654	0.6447	0.4176	3.061	0.02453	0.482
75	3.471	1.219	0.6137	0.1646	0.6018	0.2187	3.353	0.1179	0.8297

Therefore, under different cycle times, the parameters to be identified are transformed and can be used as health factors.

And (3) estimating the SOC, wherein the SOC is estimated by taking an extended Kalman filter algorithm as an example, and the polarization voltage and the SOC of the lithium ion battery are expressed by a state equation (10) and the terminal voltage observation equation is expressed by a formula (11).

Wherein: omega_kIs systematic noise, which is assumed to conform to a Gaussian distribution, upsilon_kFor noise measurement, it is also assumed that it follows a gaussian distribution, η is the coulombic efficiency and Δ T is the sampling time interval.

Step (4), acquiring a large amount of off-line data: the Uoc and SOC data of the test battery under different temperatures and different discharge rates (such as 0.1C, 0.5C, 1C and 2C) under different health states and under various dynamic working conditions (such as new European cycle test-NEDC and urban road cycle-UDDS); selecting a UOC-SOC model of formula (9), taking the model precision and the calculation complexity into consideration, taking n as 3, and then obtaining a by a curve fitting method_i、b_i、c_iThe value of (c).

Step (5), a is carried out under different health states (namely the current maximum available capacity C is different)_i、b_i、c_iBy correlation analysis, a having a large correlation with SOH is selected_i、b_i、c_iWith C as output; to improve the training precision, the maximum and minimum normalization method is used for the input a_i、b_i、c_iIs normalized with the output C to obtain a'_i、b'_i、c'_iAnd C'; adopting a 'of neural network (taking BP neural network as an example) algorithm'_i、b'_i、c'_iAnd C ' (i ═ 1, 2, 3;) were subjected to learning training, and finally obtained was ' a '_i、b'_i、c'_iIs the BP neural network model with input and C' as output.

Step (6), real-time Uoc and SOC data at the same time can be obtained through the steps (1) and (3), SOC and corresponding Uoc within the range of 10% -90% are selected, the SOC is used as an abscissa, the Uoc is used as an ordinate, and an input value a is obtained through curve fitting_i、b_i、c_iTo a, a_i、b_i、c_iAnd (4) after evolution and normalization, inputting the BP neural network model obtained in the step (5), and performing reverse normalization on the output C' to obtain the current maximum available capacity C of the lithium battery.

Step (7) of subjecting a obtained in step (6) to_i、b_i、c_iAnd C value returns to the step (3), and corresponding parameters in the state equation and the observation equation are updated.

And (8) after the parameters are updated in the step (7), executing the steps (1), (3), (5), (6) and (7) and estimating the health state of the battery in the next cycle.

It should be noted that, although the present disclosure has been illustrated by the above-mentioned embodiments, the illustration should not be construed as limiting the present disclosure. Modifications and equivalents of the present invention as described above will occur to those skilled in the art and are intended to be included within the scope of the appended claims.

Claims (4)

1. A lithium ion power battery state of health estimation method based on machine learning is characterized by comprising the following steps:

step (1), establishing an equivalent circuit model of the lithium ion power battery, and identifying unknown parameters in the model;

step (2), establishing a Uoc-SOC model:

wherein a is_i、b_i、c_iIs the parameter to be identified in the model, U_OCRepresenting the open-circuit voltage of the battery, wherein SOC is the state of charge of the battery, and the positive integer n is 2, 3 and 4;

step (3), estimating the SOC;

step (4), obtaining the offline data through curve fittinga_i、b_i、c_iA value of (d);

step (5), for a_i、b_i、c_iAnd the maximum available capacity C is subjected to normalization treatment to obtain a'_i、b'_i、c'_iAs input, C 'as output, adopting machine learning algorithm pair a'_i、b'_i、c'_iAnd C 'corresponding to the alpha-'_i、b'_i、c'_iA machine learning model for input and C' for output;

step (6), performing curve fitting on the Uoc and the SOC at the same time to obtain an input value a_i、b_i、c_iAfter normalization, inputting the normalized data into a machine model to obtain the maximum available capacity C;

step (7) of subjecting a obtained in step (6) to_i、b_i、c_iAnd C value returns to the step (3), and corresponding parameters in the state equation and the observation equation are updated.

2. The machine learning-based lithium ion power battery state of health estimation method according to claim 1, wherein the equivalent circuit model is selected from Thevenin equivalent circuit model or second-order RC equivalent circuit model.

3. The machine learning-based lithium ion power battery state of health estimation method of claim 1, wherein the Uoc-SOC model uses U_OC＝K₀+K₁z+K₂z²+K₃z³+K₄z⁴+K₅z⁵+K₆z⁶Or

Or

Or

Alternatively, where z is the state of charge SOC, K₀、K₁、K₂ K₃、K₄、K₅、K₆、α₁、α₂Is the parameter to be identified in the model.

4. The machine learning-based lithium ion power battery state of health estimation method according to claim 1, wherein the state of charge SOC is estimated by using an extended Kalman filter algorithm, and the polarization voltage and the state of charge equation of the battery are

Wherein U is_p、R_p、C_pRepresenting the polarization voltage, resistance and capacitance of the battery, eta is coulombic efficiency, delta T is sampling time interval, I_L,kThe charge-discharge current of the battery at the time k, omega_kIs the system noise; an observation equation of terminal voltage is

Wherein upsilon is_kTo measure noise.

Images