Battery residual charge estimation method based on Gaussian process regression and Kalman filtering

Abstract

The invention discloses a battery residual charge estimation method based on Gaussian process regression and Kalman filtering. The method specifically comprises the following steps: step 1, extracting an initial feature set according to an existing battery data set; 2, predicting the health degree of the battery by taking the initial feature set as the input of Gaussian process regression; 3, substituting the predicted health degree of the battery obtained in the step 2 into a Kalman filtering method to estimate the state of charge of the battery; according to the method, the SOC estimation of the battery of the electric vehicle is not influenced by sampling noise and other interference, the SOH of the battery is predicted, and the method has the advantages of being high in battery residual charge estimation precision, accurate in battery state prediction, easy to implement, high in robustness for noise measurement and the like, and can be conveniently applied to application scenes where the battery residual charge needs to be accurately estimated.

Description

Technical Field

[0001] The present invention relates to the field of new energy technology, and in particular to a battery remaining charge estimation method based on Gaussian process regression and Kalman filtering. Background Art

[0003] Electric vehicles use electricity, which has a wide range of sources, such as wind power, solar power, etc. Electric vehicles do not produce greenhouse gases, have a wide range of energy sources, and have zero carbon emissions. This provides a possibility for solving the problem of environmental pollution caused by traditional vehicles, and has become one of the industries that countries around the world are vigorously developing.

[0004] As a means of transportation, electric vehicles also need an indicator to indicate the remaining power of the vehicle, just like the fuel gauge of traditional internal combustion engine vehicles. This indicator is the state of charge. Based on the indication of the fuel gauge, the driver can well infer the remaining mileage of the car. Similarly, based on the indication of the state of charge (SOC), the remaining mileage of the electric vehicle can also be inferred. Accurate estimation of the state of charge is also conducive to protecting the battery, preventing overcharging and over-discharging from causing battery performance degradation or explosion, ensuring the safe use of the battery, extending the service life, and giving full play to the battery potential and making full use of the battery capacity. In order to ensure the driving safety of electric vehicles, it is very necessary to accurately estimate the remaining charge of the battery. Summary of the invention

[0005] The purpose of the present invention is to provide a battery remaining charge estimation method based on Gaussian process regression and Kalman filtering. The method combines the battery SOC estimation state space model with neural network prediction, predicts the battery SOH through an existing data set to obtain the charge capacity of the battery after aging, and substitutes it into the state space model to realize accurate estimation of the battery remaining charge.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The battery remaining charge estimation method based on Gaussian process regression and Kalman filtering specifically includes the following steps:

[0008] Step 1, extracting an initial feature set based on an existing battery data set, where the initial feature set includes features during the charging process and features during the discharging process;

[0009] Step 2, using the initial feature set obtained in step 1 as Gaussian process regression input to predict battery health;

[0010] Step 3: Estimate the battery state of charge using the Kalman filter method based on the battery health estimated in step 2.

[0011] Preferably, the specific method of step 1 is:

[0012] Based on the existing curve of battery capacity changing with the number of charge and discharge cycles, a series of feature sets that can characterize battery aging are extracted from the battery voltage curve and current curve. The features extracted during the charging process are: the time and value of the highest voltage when charging, the time and value of the highest temperature when charging, the time when the battery output current starts to decrease during charging, and the time when the battery terminal voltage drops from 3.8V to 3.5V. The features extracted during the discharging process are: the time and value of the lowest voltage when discharging, the time and value of the highest temperature when discharging, and the time when the output current starts to rise during discharging.

[0013] Since the units of each feature are different and the magnitudes vary greatly, the data is normalized for ease of calculation and to improve model accuracy. The minimum-maximum normalization calculation formula is:

[0014]

[0015] In formula (1), x is the sequence before standardization, x' is the sequence after standardization, max(x) and min(x) are the maximum and minimum values ​​of the sequence respectively.

[0016] Preferably, the specific method of step 2 is:

[0017] According to the initial feature set extracted in step 1, it is used as the input variable of GPR, the evolution generation is 20 generations, the first 50% of the data is selected as the training set, and the last 50% of the data is selected as the test set to predict the battery SOH.

[0018] Preferably, the specific method of step 3 is:

[0019] The Kalman filter state space model of electric vehicle battery is established as:

[0020]

[0021] In formula (2), x k+1 =[U1(k+1)SOC(k+1)] T is the state variable of the Kalman filter state space model, where U1(k+1) and SOC(k+1) are the voltage across the polarization resistor and SOC at time k+1, respectively, and T represents the matrix transposition; k =[U1(k)SOC(k)] T are the voltage across the polarization resistor and SOC at time k; e represents the natural base; T is the sampling period; the time constant τ=R1C1, R1 and C1 represent the polarization resistor and capacitor respectively; I(k) is the battery operating current at time k; Q N is the nominal charge capacity of the battery; ω k=[ω1(k)ω2(k)] T is noise; k+1 =U o (k+1) is the system observation value at time k+1, where U o (k+1) is the battery terminal voltage at time k+1; OCV is the voltage source; d represents the derivative operation; R0 is the ohmic internal resistance; ν(k+1) is the noise; is the system parameter matrix; is the observation matrix; D = [R0] is the output matrix; input variable u k =I(k).

[0022] According to the SOH obtained in step 2, calculate the current battery charge capacity Q after multiple charge and discharge processes C , and its calculation formula is

[0023] Q C =Q N ×SOH (3)

[0024] Use the current battery charge capacity Q C Substituting the nominal charge capacity Q of the battery in equation (2) N , the current remaining battery charge is calculated according to formula (2).

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) Compared with the traditional SOC calculation method, the Kalman filter method can suppress the influence of noise and other interference in the estimation process, and further improve the estimation accuracy;

[0027] (2) Gaussian process regression is used to predict battery SOH and obtain a more accurate battery rated capacity;

[0028] (3) The Kalman filter method and Gaussian regression process are combined to achieve accurate estimation of battery SOC, which is in line with the actual application status of electric vehicle batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A circuit diagram of a battery equivalent model according to an embodiment of the present invention;

[0030] Figure 2 A graph showing the variation of battery capacity with the number of charge and discharge cycles according to an embodiment of the present invention;

[0031] Figure 3 This is a graph showing the battery health prediction result of Gaussian process regression according to an embodiment of the present invention;

[0032] Figure 4A Gaussian process regression battery health prediction error diagram of an embodiment of the present invention;

[0033] Figure 5 This is a Kalman filter flow chart of an embodiment of the present invention;

[0034] Figure 6 4 is a diagram showing the battery remaining charge estimation result according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] The present invention provides a technical solution: a battery remaining charge estimation method based on Gaussian process regression and Kalman filtering, a battery equivalent model circuit diagram as shown in Figure 1 As shown, the ohmic internal resistance R0 = 0.011Ω, the polarization resistance R1 = 0.01962Ω, the polarization capacitance C1 = 508.8F, and the battery terminal voltage Voltage source OCV=3.6V.

[0037] Step 1: Extract the initial feature set based on the existing battery data set;

[0038] The present invention uses the B0005 battery data of NASAPCoE to perform battery aging analysis. The battery is subjected to a cyclic charge and discharge experiment at room temperature of 24°C. The nominal capacity of the battery is 2Ah. The charging process adopts a constant current charging (CC) mode followed by a constant voltage charging (CV) mode. The constant current charging of 1.5A is continued until the battery voltage reaches 4.2V and then the constant voltage charging is switched to. During the constant voltage charging process, the current gradually decreases. When the current decreases to 20mA, the charging is stopped. The discharge process adopts a constant current discharge of 1C (2A current) until the battery voltage drops to 2.7V, 2.5V, 2.2V, and 2.5V, respectively, and the discharge process ends. The capacity of the battery changes with the number of charge and discharge cycles as shown in Figure 2. Figure 2 shown.

[0039] It is observed that different charge and discharge cycles correspond to different numbers of sampling points. At the same time, a large amount of data must contain redundant information. It is necessary to reduce the dimension of the data to extract data points that can characterize the degree of battery degradation and actual performance. The features extracted during the charging process are: the time and value of the highest voltage when charging reaches the highest voltage, the time and value of the highest temperature when charging reaches the highest temperature, the time when the battery output current starts to decrease during charging, and the time when the battery terminal voltage drops from 3.8V to 3.5V. The features extracted during the discharge process are: the time and value of the lowest voltage when discharging reaches the lowest voltage, the time and value of the highest temperature when discharging reaches the highest temperature, and the time when the output current starts to rise during discharging.

[0040] Since the units of each feature are different and the magnitudes vary greatly, the data is normalized for ease of calculation and to improve model accuracy. The minimum-maximum normalization calculation formula is:

[0041]

[0042] In formula (1), x is the sequence before standardization, x' is the sequence after standardization, max(x) and min(x) are the maximum and minimum values ​​of the sequence respectively.

[0043] Step 2: Use the initial feature set obtained in step 1 as the input of Gaussian Process Regression (GPR) to predict the battery state of health (SOH);

[0044] A Gaussian process is a set of any finite random variables that have a joint Gaussian distribution. Its properties are determined by m(x) and k(x,x'), where x and x' are arbitrary D-dimensional random variables.

[0045] The Gaussian process can be defined as:

[0046]

[0047] Among them, m(x) is the mean function, k(x,x') is the covariance function, satisfying

[0048]

[0049] For regression problems, consider the model:

[0050] Y=f(X)+ε (4)

[0051] f(X)~N(μ(X),K(X,X)) (5)

[0052] Where X is N D-dimensional input vectors, X = {x1, x2, ..., x N} T,x∈R D , f is the mapping function, ε is the noise, ε~N(0,σ 2 ), Y is the observed value, Y={y1,y2,...,y N} T For the test input X*, there is a predicted value f(X*). From equations (4) and (5), we can get the prior distribution of Y and the joint prior distribution of the observed value Y and the predicted value f(X*) as (6) and (7), respectively:

[0053] Y~N(μ(X),K(X,X)+σ 2 I) (6)

[0054]

[0055] In equations (6) and (7), I is the N-order unit matrix, σ 2 I is the covariance matrix of the noise, and K(X,X) is a symmetric positive definite matrix that satisfies:

[0056]

[0057] From this, the posterior distribution of the predicted value f(X*) can be calculated as:

[0058] f(X*)|X,Y,X*~N(μ*,Σ*) (9)

[0059] in:

[0060] μ*=K(X*,X)[K(X,X)+σ 2 I] -1 (Y-μ(X))+μ(X*)

[0061] ∑*=K(X*,X*)-K(X*,X)[K(X,X)+σ 2 I] -1 K(X,X*)

[0062] For convenience, the input data can be preprocessed to make the mean zero, and μ* can be simplified to:

[0063] μ*=K(X*,X)[K(X,X)+σ 2 I] -1 Y (10)

[0064] The covariance function commonly used in training is the square exponential function:

[0065]

[0066] Where: M = diag (l 2 ).

[0067] Consider that each input vector is D-dimensional, then:

[0068]

[0069] Among them, l d is the variance scale, is the signal variance, l d and The hyperparameters Hyperparameters are generally obtained by the maximum likelihood method. First, the negative log-likelihood function of Y is obtained as:

[0070]

[0071] Then let L(θ) take the partial derivative of θ and get:

[0072]

[0073] At this point, the gradient descent method, Newton's method, etc. can be used to find θ that minimizes L(θ).

[0074] According to the initial feature set extracted in step 1, it is used as the input variable of GPR, the evolution generation is 20 generations, the first 50% of the data is selected as the training set, and the last 50% of the data is selected as the test set to predict the battery SOH.

[0075] Based on the filtered feature set and normalized data set, the current Gaussian process regression prediction results and errors for battery health are as follows: Figure 3 and Figure 4 shown.

[0076] Step 3: Substitute the SOH estimated in step 2 into the Kalman filter (KF) method to estimate the battery state of charge (SOC);

[0077] The Kalman filter state space model of electric vehicle battery is established as:

[0078]

[0079] In formula (15), x k+1 =[U1(k+1)SOC(k+1)] T is the state variable of the Kalman filter state space model, where U1(k+1) and SOC(k+1) are the voltage across the polarization resistor and SOC at time k+1, respectively, and T represents the matrix transposition; k =[U1(k)SOC(k)] Tare the voltage across the polarization resistor and SOC at time k; e represents the natural base; T is the sampling period; the time constant τ=R1C1, R1 and C1 represent the polarization resistor and capacitor respectively; I(k) is the battery operating current at time k; Q N is the nominal charge capacity of the battery; ω k =[ω1(k)ω2(k)] T is noise; k+1 =U o (k+1) is the system observation value at time k+1, where U o (k+1) is the battery terminal voltage at time k+1; OCV is the voltage source; d represents the derivative operation; R0 is the ohmic internal resistance; ν(k+1) is the noise; is the system parameter matrix; is the observation matrix; D = [R0] is the output matrix; input variable u k =I(k).

[0080] In practical applications, the Kalman filter calculation process can be divided into a prediction part and an update part.

[0081] Prediction part:

[0082]

[0083] In formula (16) is the optimal value of the state estimate at time k-1, u k is the control input at time k, To use and u k The calculated k-time forecast estimate (prior value). P is the variance of the state estimate, and Correspondingly, P k-1 and Corresponding to. Q is the covariance of the process noise.

[0084] Updated part:

[0085]

[0086] In formula (17), H T is the transpose of H, and the Kalman gain K at time k can be obtained k ; In formula (18), the measured value y at time k is known k and the forecast estimates The optimal state estimate at time k can be updated Then, the covariance P at time k is updated according to formula (19): kThe prediction update is iterated according to (16)-(19) until the estimation process is completed. The above Kalman filter calculation process can be summarized as follows: Figure 5 Flowchart shown.

[0087] According to the SOH obtained in step 2, calculate the current battery charge capacity Q after multiple charge and discharge processes C , and its calculation formula is

[0088] Q C =Q N ×SOH (20)

[0089] Use the current battery charge capacity Q C Substituting the nominal charge capacity Q of the battery in equation (15) N , the current remaining battery charge is calculated according to formula (15).

[0090] Figure 6 is the battery remaining charge estimation result using the method of the present invention. Figure 6 It can be seen that under the conditions that the battery has aged, its health has decreased, and its capacity has been lost, the state of charge estimated by the Kalman filter method taking into account its health will be closer to the true value.

[0091] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A battery remaining charge estimation method based on Gaussian process regression and Kalman filtering, characterized in that: The specific steps include: Step 1, extracting an initial feature set based on an existing battery data set, where the initial feature set includes features during the charging process and features during the discharging process; Step 2, using the initial feature set obtained in step 1 as Gaussian process regression input to predict battery health; Step 3: Estimate the battery state of charge using the Kalman filter method based on the battery health estimated in step 2.

2. The method for estimating the remaining charge of a battery based on Gaussian process regression and Kalman filtering according to claim 1, characterized in that: The specific steps of step 1 are: Based on the existing curve of battery capacity changing with the number of charge and discharge cycles, a series of feature sets that can characterize battery aging are extracted from the battery voltage curve and current curve. The features extracted during the charging process are: the time and value of the maximum voltage when charging, the time and value of the maximum temperature when charging, the time when the battery output current starts to decrease during charging, and the time when the battery terminal voltage drops from 3.8V to 3.5V. The features extracted during the discharge process are: the time and value of the lowest voltage when the discharge reaches the lowest voltage, the time and value of the highest temperature when the discharge reaches the highest temperature, and the time when the output current starts to rise during discharge; Since the units of each feature are different and the magnitudes vary greatly, the data is normalized for ease of calculation and to improve model accuracy. The minimum-maximum normalization calculation formula is: In formula (1), x is the sequence before standardization, x' is the sequence after standardization, max(x) and min(x) are the maximum and minimum values ​​of the sequence respectively.

3. The method for estimating the remaining charge of a battery based on Gaussian process regression and Kalman filtering according to claim 1, characterized in that: The specific steps of step 2 are: According to the initial feature set extracted in step 1, it is used as the input variable of GPR, the evolution generation is 20 generations, the first 50% of the data is selected as the training set, and the last 50% of the data is selected as the test set to predict the battery SOH.

4. The method for estimating the remaining charge of a battery based on Gaussian process regression and Kalman filtering according to claim 1, characterized in that: The specific steps of step 3 are: The Kalman filter state space model of electric vehicle battery is established as: In formula (2), x k+1 =[U1(k+1)SOC(k+1)] T is the state variable of the Kalman filter state space model, where U1(k+1) and SOC(k+1) are the voltage across the polarization resistor and SOC at time k+1, respectively, and T represents the matrix transposition; k =[U1(k)SOC(k)] T are the voltage across the polarization resistor and SOC at time k; e represents the natural base; T is the sampling period; the time constant τ=R1C1, R1 and C1 represent the polarization resistor and capacitor respectively; I(k) is the battery operating current at time k; Q N is the nominal charge capacity of the battery; ω k =[ω1(k)ω2(k)] T is noise; k+1 =U o (k+1) is the system observation value at time k+1, where U o (k+1) is the battery terminal voltage at time k+1; OCV is the voltage source; d represents the derivative operation; R0 is the ohmic internal resistance; ν(k+1) is the noise; is the system parameter matrix; is the observation matrix; D = [R0] is the output matrix; input variable u k =I(k); According to the SOH obtained in step 2, calculate the current battery charge capacity Q after multiple charge and discharge processes C , and its calculation formula is Q C =Q N ×SOH (3) Use the current battery charge capacity Q C Substituting the nominal charge capacity Q of the battery in equation (2) N , the current remaining battery charge is calculated according to formula (2).

Images